THREE ESSAYS ON THE ECONOMICS OF EDUCATION

By

Riley Acton

A DISSERTATION

Submitted to

Michigan State University

in partial fulfillment of the requirements

for the degree of

Economics – Doctor of Philosophy

2020

ABSTRACT

THREE ESSAYS ON THE ECONOMICS OF EDUCATION

By

Riley Acton

Chapter 1: Effects of Reduced Community College Tuition on College Choices and Degree

Completion

Recent efforts to increase college access concentrate on reducing tuition rates at community

colleges, but researchers and policymakers alike have expressed concern that such reductions may

not lead to long-run college completion gains. In this chapter, I use detailed data on students’

college enrollment and completion outcomes to study how community college tuition rates affect

students’ outcomes across both public and private colleges. By exploiting spatial variation in tu-

ition rates, I find that reducing tuition at a student’s local community college by $1,000 increases

enrollment at the college by 3.5 percentage points (18%) and reduces enrollment at non-local

community colleges, for-profit institutions, and other private, vocationally-focused colleges, by

1.9 percentage points (15%). This shift in enrollment choices increases students’ persistence in

college, the number of credits they complete, and the probability that they transfer to and earn

bachelor’s degrees from four-year colleges.

Chapter 2: Community College Program Choices in the Wake of Local Job Losses

Deciding which field to study is one of the most consequential decisions college students make,

but most research on the topic focuses on students attending four-year colleges. In this chapter,

I study the extent to which community college students’ program choices respond to changes in

local labor market conditions in related occupations. To do so, I exploit the prevalence of mass

layoffs and plant closings across counties, industries, and time, and create occupation-specific lay-

off measures that align closely with community college programs. I find that declines in local

employment deter students from entering closely related community college programs and instead

induce them to enroll in other vocationally-oriented programs. Using data on occupational skill

composition, I document that students predominantly shift enrollment between programs that re-

quire similar skills. These effects are strongest when layoffs occur in business, health, and law

enforcement occupations, as well as when they take place in rural counties.

Chapter 3: Do Health Insurance Mandates Spillover to Education? Evidence from Michi-

gan’s Autism Insurance Mandate (with Scott Imberman and Michael Lovenheim)

Social programs and mandates are usually studied in isolation, but interaction effects could

create spillovers to other public goods. In this paper, we examine how health insurance cover-

age affects the education of students with Autism Spectrum Disorder (ASD) in the context of

state-mandated private therapy coverage. Since Medicaid benefits under the mandate were far

weaker than under private insurance, we proxy for Medicaid ineligibility and estimate effects via

triple-differences. We find little evidence of an overall shift in ASD identification, but we do find

substantial crowd-out of special education services for students with ASD from the mandate. The

mandate led to increased mainstreaming of students in general education classrooms and a re-

duction in special education support services like teacher consultants. There is little evidence of

changes in achievement, which supports our interpretation of the service reductions as crowd-out.

To Lily and Grace Acton. You give me hope for the future.

iv

ACKNOWLEDGMENTS

Many people deserve acknowledgment for the role they have played in helping me complete

this dissertation. First, I would like to express my sincere gratitude to my advisor, Scott Imberman,

for his guidance throughout graduate school. I know that I would not be the economist I am today

without his feedback, advice, and never-ending support. The other members of my committee

played equally integral roles in helping me to develop my research and navigate this Ph.D. Steven

Haider continuously pushed me to think bigger and go further than I previously thought possible,

Stacy Dickert-Conlin painstakingly read my drafts and provided sound advice through every stage

of the process, and Amanda Chuan stepped in to provide a fresh perspective on labor economics,

the job market, and the transition from graduate student to faculty member. I also thank Mike

Conlin for many valuable conversations, Mike Lovenheim for his coauthorship and mentorship

from afar, and the economics department staff (Lori Jean Nichols, Margaret Lynch, Belen Feight,

Jay Feight) for keeping everything running smoothly behind the scenes.

While research can often be a lonely pursuit, I am very grateful that I did not have to travel

this path alone. Both my work and well-being have benefited greatly from conversations with the

brilliant, funny economists in the basement of Berkey Hall. I am especially thankful to Cody Orr,

Hannah Gabriel, Luke Watson, Gabrielle Pepin, Nick Rowe, and Chris Fowler for their friendship

and thoughtful insights, as well as our coffee breaks and their well-timed “TFP shocks.” Thanks

also to the MSU students who came before me and passed along institutional knowledge and ad-

vice. In particular, I thank Dylan Brewer, Alyssa Carlson, Katie Harris-Lagoudakis, and Sarajane

Parsons, all of whom have shared invaluable wisdom with me over the past five years.

The broader economics and education policy fields have also provided me with much support

on this journey. Thank you to the University of Michigan’s Education Policy Institute for sev-

eral opportunities to share my work with leaders in the field and forge friendships with the next

generation of education economists. Thanks also to the Association for Public Policy Analysis

& Management (APPAM) and Association for Education Finance & Policy (AEFP) for providing

v

venues to present my research and opportunities to become part of a supportive community. I am

also grateful to the many seminar participants, at MSU and elsewhere, who have provided feedback

on these dissertation chapters.

Finally, I thank those in my “inner circle” who have seen me through the completion of this

Ph.D. I would not be here without the support of my mentors at Ursinus College. Thank you

to Jennifer VanGilder for enthusiastically steering me towards graduate school in economics and

continuing to check-in on me throughout the process of becoming an economist. Equal thanks go

to Jeff Schepers for always believing in me, calling me when I most needed it, and reminding me

that the only way to eat an elephant is “one bite at a time.”

My largest source of encouragement and inspiration has always been, and continues to be, my

family. Mom and Dad: thank you for everything —from the frequent texts and phone calls, to the

wine deliveries, and the plane tickets home. For the past 27 years, you have given me all of the tools

and resources I needed to be successful and I could not be more grateful for your love and support.

Lily and Grace: I am very proud to be your big sister. Thank you for keeping me grounded,

making me laugh, and giving me so much to be hopeful about. Opa: I am honored to follow in

your footsteps as I enter the next phrase of my academic career, and am so happy that Oma has

been able to watch it unfold. I hope to have a fraction of an impact on this world that you did. Last,

but certainly not least, I express my utmost appreciation for my partner, Samuel Christensen. Your

love, humor, and wisdom have gotten me through the many trials and tribulations of these past few

years. I can’t wait to see what our next chapter together has in store.

vi

PREFACE

This dissertation used data collected and maintained by the Michigan Department of Education

(MDE) and Michigan’s Center for Educational Performance and Information (CEPI). The results,

information and opinions presented here solely represent the analysis, information and opinions of

the author and are not endorsed by, or reflect the views or positions of, grantors, MDE and CEPI

or any employee thereof.

vii

TABLE OF CONTENTS

LIST OF TABLES .

LIST OF FIGURES .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

x

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. xiii

CHAPTER 1 EFFECTS OF REDUCED COMMUNITY COLLEGE TUITION

.

.

.

.

.

. .

Introduction .

1.3 Data and Sample

.
1.3.1 Data Sources
1.3.2

ON COLLEGE CHOICES AND DEGREE COMPLETION . . . . . . . .

1.1
1.2 Michigan’s Postsecondary Education Market . . . . . . . . . . . . . . . . . . . .

.
1
1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
.
6
1.2.1 Michigan’s Community Colleges . . . . . . . . . . . . . . . . . . . . . . .
1.2.2
9
Private Competitors to Community Colleges . . . . . . . . . . . . . . . . .
1.2.3 Other Postsecondary Options . . . . . . . . . . . . . . . . . . . . . . . . . 11
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Sample Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.1 College Enrollment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.5.2 College Completion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5.3 Heterogeneity .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.5.4 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

1.4 Empirical Strategy .
. .
1.5 Results .

1.6 Conclusion .

. .

. .

. .

.
.

.
.

.

.

.

.

.
.

.

.

.

.

.
.

.
.

CHAPTER 2 COMMUNITY COLLEGE PROGRAM CHOICES

.

.

.

. .

Introduction .

2.4 Measuring Local Job Losses

.
2.1
2.2 Conceptual Framework .
2.3

Institutional Setting & Enrollment Data
2.3.1
2.3.2

IN THE WAKE OF LOCAL JOB LOSSES . . . . . . . . . . . . . . . . . . 32
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
. . . . . . . . . . . . . . . . . . . . . . . 37
Programs Offered by Michigan’s Community Colleges . . . . . . . . . . . 38
Students Enrolled in Michigan’s Vocational Programs . . . . . . . . . . . . 40
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.4.1 Using WARN Data to Generate Occupation-Specific Layoff Exposure . . . 43
2.4.2 Distribution of Layoffs Across Occupations . . . . . . . . . . . . . . . . . 46
Potential Measurement Error . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.4.3
2.5 Effect of Job Losses on Enrollment in Related Programs . . . . . . . . . . . . . . . 49
2.5.1 Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.5.2 Main Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.5.3 Robustness
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Substitution out of Vocational Programs . . . . . . . . . . . . . . . . . . . 54
. . . . . . . . . . . . . . . . . 56
Substitution Between Vocational Programs
. . . . . . . . . . 58

2.6.1
2.6.2
2.6.3 Explaining Substitution with Occupation Characteristics

.
2.6 Substitution Effects .

.
.
.

.
.
.

viii

2.6.4 Heterogeneity & Robustness . . . . . . . . . . . . . . . . . . . . . . . .

. 60
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

2.7 Conclusion . .

. .

.

.

.

CHAPTER 3 DO HEALTH INSURANCE MANDATES SPILLOVER TO EDUCA-

3.1
3.2 Background . .

.
.
.

.
.
.

.
.
.

.
.
. .

.
.
. .

TION? EVIDENCE FROM MICHIGAN’S AUTISM INSURANCE MAN-
DATE .
Introduction .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.1 Austism Spectrum Disorder and Therapy Options . . . . . . . . . . . . . . 72
3.2.2 The Michigan Autism Insurance Mandate . . . . . . . . . . . . . . . . . . 74
3.2.3
Special Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
.
3.3.1 Michigan Administrative K-12 Schooling Data . . . . . . . . . . . . . . . 78
3.3.2 Measuring Insurance Status . . . . . . . . . . . . . . . . . . . . . . . . . . 81

. .

. .

. .

.

.

.

3.3 Data .

3.4 The Effect of the Autism Insurance Mandate on ASD and Special Education

.

.

.

.

Incidence
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.4.1 Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.4.2 Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

. .

. .

.

.

.

.

.

3.5 The Effect of the Autism Insurance Mandate on Educational Services and Test

.

.

.

.

. .

. .

. .

Scores .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.5.1 Empirical Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.5.2 Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3.5.3 Heterogeneous Treatment Effects and Robustness Checks . . . . . . . . . . 95
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

. .

. .

. .

.

.

.

.

3.6 Conclusion .

.

.

.

.

.

.

.

APPENDICES .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
APPENDIX A CHAPTER 1 APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . 103
APPENDIX B CHAPTER 2 APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . 131
APPENDIX C CHAPTER 3 APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . 169

.

.

.

.

BIBLIOGRAPHY .

.

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

ix

LIST OF TABLES

Table A.1.1 Mean Tuition Rates at Michigan Community Colleges, 2008-2016 . . . . 109

Table A.1.2 Associate Degree Programs Offered by Community & Vocational Colleges 110

Table A.1.3 Baker College vs. Private Two-Year Colleges

. . . . . . . . . . . . . . . 111

Table A.1.4 Michigan’s Traditional Four-Year Colleges . . . . . . . . . . . . . . . . . 112

Table A.1.5 Michigan’s Community College Districts

. . . . . . . . . . . . . . . . . 113

Table A.1.6 Descriptive Statistics, 2009-2016 High School Graduates . . . . . . . .

. 114

Table A.1.7

First Stage Estimate of In-District Status on Tuition . . . . . . . . . . . . 115

Table A.1.8 Balance Tests of Student Characteristics . . . . . . . . . . . . . . . . .

. 116

Table A.1.9 Balance Tests of Census Tract Characteristics . . . . . . . . . . . . . .

. 117

Table A.1.10 Balance Tests of Distance to Postsecondary Institutions . . . . . . . . . . 118

Table A.1.11 Effect of In-District Status and Reduced Tuition on College Enrollment

. 119

Table A.1.12 Heterogeneous Effects by Graduation Year . . . . . . . . . . . . . . . . . 120

Table A.1.13 Characteristics of Community and Vocational Colleges . . . . . . . . . . 121

Table A.1.14 Effect of In-District Status and Reduced Tuition on College Completion . 122

Table A.1.15 Distribution of Degree Completion Increases Across Majors

. . . . . .

. 123

Table A.1.16 Academic Program Categories . . . . . . . . . . . . . . . . . . . . . . . 124

Table A.1.17 Distribution of Bachelor’s Degree Increases Across Professional Majors . 125

Table A.1.18 Heterogeneity by Student Characteristics . . . . . . . . . . . . . . . . .

. 126

Table A.1.19 Balance Tests of Student Characteristics, Varying Bandwidths

. . . . .

. 127

Table A.1.20 Local Community College Enrollment Results, Within Same School

District

.

.

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

x

Table A.1.21 Full Enrollment Results for Within Same School District Sample . . . . . 129

Table A.1.22 Placebo Tests

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

Table B.1.1

Programs Offered by Michigan’s Community Colleges

. . . . . . . . .

. 142

Table B.1.2

Program Groups and Associated Occupation Codes . . . . . . . . . . .

. 143

Table B.1.3

Summary Statistics of Michigan’s High School Graduates . . . . . . . . . 144

Table B.1.4

Summary Statistics of Vocational Students by Program . . . . . . . . .

. 145

Table B.1.5

Industries with Highest Concentration of Occupation Groups . . . . . . . 146

Table B.1.6

Correlation Between Occupation Composition Across Industries . . . . . 147

Table B.1.7

Summary Statistics of Layoffs in Michigan, 2001-2017 . . . . . . . . .

. 148

Table B.1.8

Largest Layoffs by Occupation Group, 2001-2017 . . . . . . . . . . . .

. 149

Table B.1.9

Effect of Job Losses on Enrollment in Related Community College
Programs

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

.

. .

Table B.1.10 Effect of Job Losses in Alternative Geographic Areas . . . . . . . . . . . 151

Table B.1.11 Effect of Community College Layoffs on Overall Vocational Program

Enrollment

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

Table B.1.12 Effect of Layoffs on College Enrollment Outcomes . . . . . . . . . . . . 153

Table B.1.13 Effect of Layoffs on Composition of Vocational Students . . . . . . . . . 154

Table B.1.14 Effect of Layoffs on First-Year Course-Taking . . . . . . . . . . . . . .

. 155

Table B.1.15 Substitution Between Community College Program Groups . . . . . . .

. 156

Table B.1.16 Substitution Between Narrower Community College Programs . . . . . . 157

Table B.2.1

Relationship Between Estimated Layoffs & Employment Change . . . . . 160

Table B.3.1

Effect of Layoffs on Retention in Related Programs . . . . . . . . . . . . 165

Table B.3.2 Own-Layoff Effects on Program Retention Rates

. . . . . . . . . . . .

. 165

Table C.1.1 Descriptive Tabulations of Analysis Variables . . . . . . . . . . . . . . . 177

xi

Table C.1.2 Overlap Between Free/Reduced Price Lunch and Medicaid in Michi-

gan, by Family Income . . . . . . . . . . . . . . . . . . . . . . . . . . . 178

Table C.1.3

The Effect of the ASD Insurance Mandate on Disability Incidence . . .

. 179

Table C.1.4

The Effect of the ASD Insurance Mandate on Special Education Services 180

Table C.1.5

The Effect of the ASD Insurance Mandate on Other Special Education
Services .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

.

. .

Table C.1.6

The Effect of the ASD Insurance Mandate on Test Scores . . . . . . . .

. 182

Table C.1.7

The Effect of the ASD Insurance Mandate on Taking Regular Exams . . . 183

Table C.1.8 Heterogeneous Effects of the ASD Insurance Mandate on ASD Incidence 184

Table C.1.9

The Effect of the ASD Insurance Mandate, by Gender and Race

. . . . . 185

Table C.1.10 The Effect of the ASD Insurance Mandate, by Grade

. . . . . . . . . .

. 186

Table C.1.11 Heterogeneous Effects of the ASD Insurance Mandate on Test Scores,

using Non-Special Education Control Group . . . . . . . . . . . . . . . . 187

Table C.1.12 The Effect of the ASD Insurance Mandate – Robustness Checks

. . . . . 188

Table C.1.13 The Effect of the ASD Insurance Mandate on Test Scores - Robustness

Checks .

.

.

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

Table C.1.14 The Effect of the ASD Insurance Mandate – No Sample Exclusion

Based on Disadvantaged Status . . . . . . . . . . . . . . . . . . . . . . . 190

Table C.1.15 The Effect of the ASD Insurance Mandate on Disability Incidence - No

Sample Exclusion Based on Disadvantaged Status . . . . . . . . . . . .

. 191

Table C.1.16 The Effect of the ASD Insurance Mandate on Test Scores - No Sample

Exclusion Based on Disadvantaged Status . . . . . . . . . . . . . . . .

. 192

xii

LIST OF FIGURES

Figure A.1.1 Identified Community College District Boundaries

. . . . . . . . . . . . 103

Figure A.1.2 Washtenaw Community College District Analysis Sample . . . . . . . . . 104

Figure A.1.3 Distribution of Border Pair Tuition Differentials . . . . . . . . . . . . . . 105

Figure A.1.4 Correlation Between Tuition Differentials and Area Characteristics . . . . 106

Figure A.1.5 Reduced Form Estimates with Alternative Bandwidths

. . . . . . . . . . 107

Figure A.1.6 School Districts Overlapping Community College Districts . . . . . . . . 108

Figure B.1.1 Differences in Course-Taking and Credit Completion by CC Program

Group .

.

.

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

Figure B.1.2 Labor Market Shocks in Michigan, 2001-2017 . . . . . . . . . . . . . . . 132

Figure B.1.3 Average Layoffs in Michigan Counties, 2001-2017 . . . . . . . . . . . . 133

Figure B.1.4 Correlation Between National and State-Specific Industry Employment

Shares, 2016 .

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

Figure B.1.5 Distribution of Layoffs by County, 2001-2017 . . . . . . . . . . . . . . . 135

Figure B.1.6 Robustness Checks for Pooled Specification . . . . . . . . . . . . . . .

. 136

Figure B.1.7 Substitution into Program Groups Requiring Similar Skills . . . . . . . . 137

Figure B.1.8 Relationship Between Substitution Effects and Skill Distance . . . . . . . 138

Figure B.1.9 Alternative Measures of Skill Distance . . . . . . . . . . . . . . . . . .

. 139

Figure B.1.10 Heterogeneous Own-Layoff Effects

. . . . . . . . . . . . . . . . . . . . 140

Figure B.1.11 Robustness Checks for Own-Layoff Effects

. . . . . . . . . . . . . . . . 141

Figure B.2.1 Comparison of Employment Counts in QCEW & CBP . . . . . . . . . . 159

Figure B.2.2 Relationship between Layoffs and Employment Changes, by Sector

. .

. 161

Figure B.3.1 Effect of Layoffs on Program Choice for Later Enrollees

. . . . . . . . . 162

xiii

Figure B.4.1 Effect of Layoffs on Enrollment in Narrower Program Groups

. . . . .

. 166

Figure B.4.2 Substitution into Narrower Program Groups Requiring Similar Skills . .

. 167

Figure B.4.3 Relationship Between Substitution Effects & Skill Distance, Narrower

Programs

.

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

Figure C.1.1 ASD and Non-ASD Special Education Incidence Event Studies . . . . . . 169

Figure C.1.2 Test Scores Event Studies, using Non-Disabled Control Group . . . . . . 170

Figure C.1.3 Event Study Estimates of Main Outcomes . . . . . . . . . . . . . . . .

. 171

Figure C.1.4 Test Scores Event Studies, using All Non-ASD Control Group . . . . .

. 175

Figure C.1.5 Test Score Event Studies, using Non-ASD Special Ed Control Group . . . 176

xiv

CHAPTER 1

EFFECTS OF REDUCED COMMUNITY COLLEGE TUITION

ON COLLEGE CHOICES AND DEGREE COMPLETION

1.1 Introduction

Community colleges enroll nearly 40% of U.S. undergraduate students and are increasingly the
focus of college access initiatives (National Center for Education Statistics, 2018).1 These insti-

tutions offer a variety of educational programs, including vocationally focused certificates, two-

year associate degrees, and pathways to transfer to four-year colleges and universities. Moreover,

community colleges offer these opportunities at a lower price than nearly all other postsecondary

options, making them accessible to a large and diverse group of students, many of whom come

from low-income backgrounds or are the first in their families to attend college (Ma and Buam,

2016). In recent years, policymakers have capitalized on community colleges’ commitment to ac-

cess in their local communities by implementing programs that make community college low-cost

or completely tuition-free (Smith, 2017).

As these types of programs grow in popularity, so too do questions about their potential con-

sequences for students’ educational attainment and labor market outcomes. Policymakers and

researchers alike have expressed concern that reducing the price of community college may de-

ter students from enrolling in four-year colleges, potentially decreasing the probability that they

earn bachelor’s degrees and receive wage premiums in the labor market. Notably absent from this

discussion, however, is the possibility that reducing the price of community college could deter

students from enrolling in private colleges that offer certificates and associate degrees —here-

after referred to as vocational colleges. These colleges primarily operate as for-profit entities,

which have grown rapidly in the past two decades and now produce over 40% of less-than-two-

1In this paper, I use the term “community college” to refer to any publicly funded college that
primarily offers sub-baccalaureate credentials. These institutions are also sometimes referred to as
junior colleges, technical colleges, or city colleges.

1

year certificates and nearly 20% of associate degrees in the U.S., despite having higher average

tuition rates, and lower average completion rates and wage premiums than their public, not-for-

profit counterparts (Deming et al., 2012; Cellini and Turner, 2018; Armona et al., 2018). While

there is some evidence that community colleges and for-profit colleges compete for students in the

two-year college market, particularly in the presence of declines in state funding for public higher

education (Cellini, 2009; Goodman and Volz, 2019) or local labor demand shocks (Armona et al.,

2018), there is currently no direct evidence on how tuition rates at public institutions alter students’

enrollment decisions in private institutions in the two-year sector, or how such a substitution effect

may impact students’ longer-run educational outcomes.

In this paper, I empirically estimate the effects of community college tuition on students’ col-

lege enrollment decisions and outcomes across different sectors of the postsecondary education

market. To isolate exogenous variation in community college tuition rates, I exploit an institu-

tional feature of Michigan’s community college system in which students residing on either side

of a “community college district” boundary face substantially different tuition rates at their lo-

cal community college due to a locally provided tuition subsidy. This feature allows me to use a

boundary fixed effects strategy that compares the college choices and outcomes of students who

live just inside of a community college district and face an average community college tuition rate

of $2,300 per year to their peers who live just outside of a community college district and face

an average tuition rate of $4,100. While this approach is similar to that used by Denning (2017)

and McFarlin et al. (2018) to study community college taxing districts in Texas, I am able to build

upon both studies through the use of detailed, student-level administrative data from the Michigan

Department of Education that contains students’ precise census blocks of residence, as well as

comprehensive college enrollment and completion records across public and private colleges.

Obtaining students’ census blocks of residence enables me to very accurately determine whether

students reside within community college districts and to avoid the potential measurement error in-

duced by inferring in-district status from the schools they attend. McFarlin et al. (2018) show that

precisely measuring community college tuition is important in determining its effects on college

2

enrollment, but are unable to observe in which colleges students enroll due to their use of restricted-

access Census data. Meanwhile, Denning (2017) observes detailed college enrollment and com-

pletion records but must proxy for in-district status with the location of a student’s high school. By

combining data on students’ precise residences with specific college enrollment records, I am bet-

ter able to identify the direct effect of a community college’s tuition rate on a student’s decision to

enroll in the college. In addition, the detailed college records in my dataset come from the National

Student Clearinghouse (NSC), which covers 97% of all postsecondary institutions in the U.S., and

now covers several of the largest national for-profit colleges (National Student Clearinghouse Re-

search Center, 2017). This coverage allows me to determine the underlying substitution effects that

drive an increase in community college attendance, including whether reduced community college

tuition crowds out enrollment in similar private colleges, which previous work has not been able

to consider.

Among students graduating from Michigan public high schools between 2009 and 2016, I

find that reducing the tuition rate that a student faces at her local community college by $1,000

increases the probability of enrollment at the college within a year of high school graduation by

3.5 percentage points, about 18% of the mean enrollment rate. A portion of this increase can

be attributed to students enrolling in their local community college who would not have initially

enrolled in any postsecondary education program in the absence of the tuition reduction, as a

$1,000 decrease in local community college tuition increases overall college enrollment by 0.7

percentage points (1% of the mean). At the same time, this tuition decrease causes students to

reduce enrollment in non-local community colleges by 1.6 percentage points (8% of the mean)

and in for-profit and other private, vocationally-focused colleges that offer two-year degrees by

0.4 percentage points (11% of the mean). The remainder of the increase in local community

college attendance can be attributed to a 1.0pp decline in four-year college attendance; however,

this estimate is statistically insignificant and is quite small compared to its mean.

Using longer-run data from cohorts who graduated high school between 2009 and 2011, I find

further evidence that reduced community college tuition increases persistence in college and degree

3

completion. A $1,000 decrease in local community college tuition induces students to complete

2.5% more semesters of college, 2.7% more college credits, and to transfer to four-year colleges

at a rate 6.5% higher than their peers who do not receive discounted tuition. This $1,000 tuition

decrease also increases bachelor’s degree completion by 1.1pp (3.5%), particularly in business and

professional fields such as teacher education and exercise science. These improved outcomes are

driven in part by students switching from higher-cost and lower-resourced vocational colleges that

focus on labor market preparation to higher-resourced community colleges that promote transfer

to four-year colleges. Consistent with this mechanism, I also find that reduced community college

tuition induces students to earn general liberal arts associate degrees, which are designed to prepare

students to transfer, rather than associate degrees in vocational subjects.

These results contribute to several strands of literature on college choice and the consequences

of public subsidization of postsecondary education. First, the results add to a large body of empir-

ical work on the effect of college costs on students’ college enrollment decisions. Most previous

analyses find approximately a 3-5 percentage point increase in the probability of enrollment for

each $1,000 decrease in the cost of a college option (Deming and Dynarski, 2010; Page and Scott-

Clayton, 2016), with potentially even larger effects at the community college level. However,

recent estimates of students’ sensitivity to community college costs come from large-scale policy

changes, such as the introduction of free tuition policies (Carruthers and Fox, 2016) or the ex-

pansion of community college districts (Denning, 2017), which may affect students’ choices and

outcomes through multiple channels, such as informational campaigns, mentoring programs, or

the construction of new college campuses. The results presented here isolate tuition variation by

comparing observationally similar students who likely have similar exposure to college informa-

tion, marketing, and campuses, and are very much in line with that of the broader literature. This

finding suggests that, despite the already low cost of most community colleges in the U.S., stu-

dents are responsive to the sticker prices advertised by community colleges and that policies that

reduce advertised tuition rates by even small amounts may have meaningful impacts on students’

educational and labor market outcomes.

4

Second, this research provides the first direct evidence that students substitute towards commu-

nity colleges and away from similar private colleges, including those in the for-profit sector, when

community college tuition is low. Cellini (2009) and Goodman and Volz (2019) document a simi-

lar phenomenon in the context of changes in state funding for higher education, whereby increases

in funding for public colleges deter students from attending for-profit institutions. In this paper,

I find that this private-to-public enrollment shift also occurs as a direct result of a reduction in

community college tuition and that the shift improves students’ educational attainment. However,

as in Denning (2017), I do not find that, on average, students forgo initially attending four-year

colleges when they have access to a low cost community college or that students forgo opportuni-

ties to earn bachelor’s degrees by attending community colleges. This finding comes in contrast to

Carruthers and Fox (2016) who find that a broad, tuition-free community college program in Ten-

nessee reduces four-year college enrollment, suggesting that the structure of community college

tuition policies may play an important role in determining their effects on students’ college choices

and outcomes.

Finally, this work contributes to an expanding literature on the effects of community college

attendance on educational and labor market outcomes. Because community colleges are uniquely

situated between the labor market and four-year colleges, their impact on students’ longer-term out-

comes is often ambiguous and depends on students’ counterfactual enrollment decisions (Rouse,

1995). Some students who attend community college may be made better off because in the ab-

sence of community colleges they would not have attended any college, while others may be made

worse off because they are diverted from attending four-year colleges. Empirically, students who

are deterred from attending four-year colleges tend to experience an educational attainment and

labor market penalty (Reynolds, 2012; Goodman et al., 2017), while students who are induced

to attend their local community college rather than not attending any college experience positive

educational and labor market gains (Mountjoy, 2019). I find that students who are induced to at-

tend their local community college rather than attending other predominantly two-year colleges are

more likely to transfer to four-year colleges and earn bachelor’s degrees. This result implies that

5

gains from community college attendance can extend to a broader group of students than identified

in prior work and suggests that policies that increase community college access without deter-

ring students from attending four-year colleges could increase educational attainment and improve

labor market outcomes.

1.2 Michigan’s Postsecondary Education Market

The institutional setting for this analysis is the postsecondary education market in the state of

Michigan. There are over 90 accredited colleges and universities in Michigan offering a wide range

of academic programs, and over 90% of the state’s high school graduates who enroll in college

choose to attend one of them. Michigan has a largely decentralized community college system

in which tuition rates are determined independently by each college and are based on a student’s

place of residence relative to specific geographic boundaries. This creates large differences in

the tuition rates faced by otherwise observationally similar students who reside on either side of a

given boundary. In addition, Michigan is home to a large private vocational college, Baker College,

which has multiple locations throughout the state and enrolls over 25,000 students annually. Baker

offers sub-baccalaureate academic programs similar to Michigan’s community colleges but spends

less per student on instruction and has much lower transfer rates than its public counterparts. The

presence of this potential competitor in the two-year college market allows me to examine whether

subsidizing community college tuition crowds out enrollment in similar private colleges.

1.2.1 Michigan’s Community Colleges

Michigan is home to 28 public community colleges which together enroll over 300,000 stu-

dents annually (Michigan Community College Association, 2019). Each college is designed to

serve a distinct geographic area, known as a community college district, and is given substantial

autonomy over its administration. There is no overarching state law nor agency governing the

specific operations of community colleges and state intervention in their practices is rare (Hilliard,

2016). The state government does, however, provide annual appropriations funding to community

6

colleges, which accounts for approximately 20% of the community colleges’ operating revenues.2

To supplement this funding, the colleges rely heavily on both tuition and fees (43% of operating

revenues) and local property taxes (35% of operating revenues). For each college, local property

taxes may only be assessed on properties within its community college district (Michigan House
Fiscal Agency, 2017).3

Community college district boundaries are governed by the trustees of each college under state

guidelines and may be primarily comprised of counties, public school districts, or public interme-

diate school districts (ISDs), which are administrative organizations that support multiple school
districts.4 Community college districts may also include or exclude specific cities, townships, or

other geographic features, although any changes to boundaries must be voted on by residents of

the district. Currently, 15 of the state’s 28 community college districts are comprised primarily
of counties and 13 are comprised primarily of school districts or ISDs.5 Based on conversations

with state employees and community college staff members, it is my understanding that no com-

munity college boundaries changed during the time frame of the data, and that most have remained

unchanged for several decades.

Community colleges offer tuition rates based on a student’s place of residence relative to their

2This funding is allocated based on a weighted performance funding model that takes into
account prior-year funding, enrollment, and performance indicators, and rewards colleges for low
administrative costs and adherence to best practices for community engagement (Michigan House
Fiscal Agency, 2017).

3In 2015-2016, the average millage rate for community colleges was 2.51, i.e. $2.51 per $1,000
of taxable property value (Michigan Center for Educational Performance & Information, 2017).
This millage rate is assessed on all properties in a community college’s district, in addition to any
other local property taxes (e.g., county, school district, township, or municipality taxes). Using
data on aggregate real estate taxes and home values at the census tract level from the American
Community Survey, I estimate that in-district areas in Michigan have an average total millage rate
of 17.4, while out-of-district areas have an average total millage rate of 12.3.

4More information about Michigan’s ISDs is available here:

https://www.gomaisa.org/

value-of-isds/.

5Table A.1.5 lists the geographic areas that comprise each community college district. I gather
this information from individual community college websites, course catalogs, and conversations
with colleges’ institutional research staff.

7

community college district boundaries.6 In exchange for property tax funding, students residing

within the boundaries of a district are offered the lowest tuition rate at their district’s community

college, averaging approximately $90 per credit. Students residing within Michigan, but outside of
the district, are offered the next lowest rate,7 and students residing outside of the state are offered
the highest rate.8 Critically for the analysis at hand, a sizable portion of Michigan high school

students reside outside of any community college district and, therefore, face the higher, out-of-

district tuition rate at any community college they wish to attend. Using data on students’ census

blocks of residence, I estimate that approximately 23% of Michigan’s high school graduates reside

in an area that is not part of any community college district. On average, these students face

tuition rates at their local community college —the college whose district area they reside nearest

—that are 65% higher than those faced by their peers who live within the community college’s
district boundaries.9 This equates to an average annual cost difference of nearly $1,500 for a

student taking a course load of 12 credits per semester. Given that the annual median family

income of Michigan’s community college students is approximately $60,000 (Chetty et al., 2017),

this represents a difference of approximately 2.5% of annual median family income. Table A.1.1

provides summary statistics on the average in-district and out-of-district tuition rates between 2008

6Tuition rates are set based on students’ residences regardless of whether students enroll in
courses in-person or online. However, students who reside within a community college district are
also able to enroll in online courses offered by other community colleges at a discounted rate (https:
//www.micollegesonline.org/courses.html). If anything, this feature should attenuate the estimates
that follow as it reduces the incentive for in-district students to enroll in their local community
college.

7Macomb Community College also offers an “affiliate” tuition rate to students who reside out-
side of their district but in areas near their boundaries, which I incorporate in the empirical analysis.
Results are also robust to treating this area as out-of-district.

8Michigan’s community colleges differ in how long a student must be a resident of the district
to qualify for in-district tuition. However, most require several months of residency, which makes
it unlikely that students who do not reside in a district while attending high school would be able
to claim in-district residency upon initial enrollment.

9The tuition prices used in this paper are the colleges’ advertised tuition prices, also known
as sticker prices. Both in-district and out-of-district students may qualify for federal, state, local,
or institutional financial aid that will reduce their net price of attendance. Across Michigan’s
community colleges, data from IPEDS indicates that the average net price for in-district students
is approximately 80% lower than the average net price for out-of-district students.

8

and 2016, measured in 2016 dollars. Following Denning (2017), I calculate semester tuition as the

tuition rate for 12 credits and annual tuition as the tuition rate for 24 credits.

In addition to the tuition variation induced by community college district boundaries, students

residing in different areas of the state and graduating in different years may also face substantially

different local community college tuition rates. Without government oversight of tuition-setting

policies, individual community colleges are free to differ in their relative in-district and out-of-

district rates and may update these rates annually. Over the time frame of the data, real mean in-

district tuition (measured in 2016 dollars) ranged from $76.90 per credit at Oakland Community

College to $114.89 per credit at Mott Community College. Real mean out-of-district tuition ranged

from $114.05 per credit at Wayne Community College to $221.22 per credit at Grand Rapids Com-

munity College. This range means that, on average between 2008 and 2016, it was less costly to

be an out-of-district student at Wayne Community College than to be an in-district student at Mott

Community College. Community college tuition rates, particularly for out-of-district students,

have also steadily increased over the past decade. For the graduating high school class of 2008, the

real average in-district tuition rate per credit was $82.47 and the average out-of-district rate was

$134.46. By 2016, these average rates had increased to $106.10 and $176.58, respectively.

1.2.2 Private Competitors to Community Colleges

Michigan’s other postsecondary institutions may be grouped into two mutually exclusive cat-

egories: vocational colleges, which predominantly offer sub-baccalaureate degree programs, and

traditional four-year colleges, which predominantly offer bachelor’s and graduate degrees. I define

a vocational college as a private institution that is either (1) a for-profit institution or (2) a not-for-

profit institution that offers more than 25% of its degrees at the associate degree level and accepts

90% or more of applicants. These colleges are similar to the state’s community colleges in that they

provide access to a vast majority of interested students and offer academic programs that can be

completed in two years or less: namely, associate degrees and short-term certificates. Community

and vocational colleges also tend to offer degrees in similar fields and both have an emphasis on

9

health and business subjects. Table A.1.2 highlights this point by comparing the types of associate

degrees offered by the community and vocational colleges attended by Michigan’s high school

graduates. Given the overlap in program offerings, it is reasonable to believe that these vocational

institutions compete with community colleges in the market for sub-baccalaureate education.

In Michigan, the colleges identified under this vocational college criteria and available in the

NSC data are: Baker College (not-for-profit), Davenport University (not-for-profit), Everest Insti-

tute (for-profit), ITT Technical Institute (for-profit), and The International Academy of Design &
Technology (for-profit).10 I also observe enrollment in other large national for-profit chains, such

as the University of Phoenix, DeVry University, and Kaplan University, although these institutions

do not report in which campus a student is enrolled so I am unable to observe whether students
enroll in Michigan, online, or elsewhere in the country.11 However, I do not observe enrollment in
any smaller for-profit institutions located within Michigan, such as cosmetology schools.12 This

lack of coverage includes institutions that do not participate in federal financial aid programs,

which Cellini and Goldin (2014) show account for over half of for-profit enrollment in Michigan.

It is not obvious that these types of non-degree granting institutions would be popular among recent

high school graduates, but to the extent that they are, I will overestimate the share of students not

enrolling in college and will underestimate the share enrolling in vocational colleges. As such, my

results should be interpreted as an upper bound of the effect of reduced community college tuition

on overall college enrollment and a lower bound of the effect of reduced tuition on substitution

away from vocational colleges.

The most popular private vocational institution among Michigan’s high school graduates is

Baker College, which has thirteen campuses throughout the state and enrolls over 70% of Michi-

10The three for-profit colleges in this list (Everest, ITT, and The International Academy of De-
sign & Technology) shut down operations within Michigan during the time frame of the data. To
my knowledge, no new colleges opened.

11Students who enroll in exclusively online programs are included in the NSC data, but I am

unable to distinguish between on-campus and online enrollment within an institution.

12In 2017, the NSC reported coverage of 78% of multi-state for-profit institutions but 0% cov-
erage of for-profits operating only in Michigan (National Student Clearinghouse Research Center,
2017).

10

gan’s vocational students.13 Baker is a private, not-for-profit institution, that primarily offers de-

gree programs designed to take two years or less. Such institutions are not common in the U.S. For

example, according to the 2016 College Scorecard, there are 369 private, predominantly associate-

or certificate-degree granting institutions in the U.S. but 2,587 for-profit private institutions offer-

ing the same types of degrees. However, in many ways, Baker College operates similarly to the

more popular model of a private, for-profit two-year college. Table A.1.3 compares Baker to the

universe of private colleges that predominantly grant associate degrees and certificates. Across

several measures of institutional quality and outcomes, Baker appears more similar to its for-profit

counterparts rather than its not-for-profit peers. Given these similarities, the results from this pa-

per should provide suggestive evidence on how reductions in local community college tuition may

affect enrollment at for-profit vocational colleges.

1.2.3 Other Postsecondary Options

The remainder of undergraduate, degree-granting postsecondary institutions in Michigan are

either public or private traditional four-year colleges. In recent years, public universities have pri-

marily relied on students’ tuition payments for operating expenses as state appropriations have

declined and now account for only 21% of the universities’ operating budgets (Michigan House

Fiscal Agency, 2017). Similar to the state’s community colleges, there is little government over-

sight of the universities’ practices and, as a result, there is a substantial amount of heterogeneity in

tuition rates, expenditures, and program offerings among them. However, in contrast to commu-

nity colleges, all public universities offer the same tuition rate to all in-state students regardless of

their location of residence. Michigan also has several private four-year institutions, which finance

their operating expenditures with students’ tuition payments, private donations, and endowments

as they receive minimal support from the state. They tend to be much smaller and somewhat more

expensive than the state’s public universities and, overall, make up a small share of the postsec-

13Because of this large market share, my results are robust to any definition of vocational col-

leges that includes Baker College.

11

ondary education market. Table A.1.4 provides summary statistics on these institutional attributes

across the public and private sectors.

Students who choose not to enroll in community, vocational, or traditional four-year colleges

generally enter the state’s low-skill labor market.

In the years following the Great Recession,

young adults who have chosen this option in Michigan have faced high rates of unemployment

and underemployment. Those who are employed are most likely to work in service and retail

occupations, which have low median wages and minimal opportunities for advancement (Bureau

of Labor Market Information and Strategic Initiatives, 2014).

1.3 Data and Sample

1.3.1 Data Sources

The data used in this paper primarily come from a student-level, administrative dataset pro-

vided by the Michigan Department of Education (MDE) and the state’s Center for Education Per-

formance and Information (CEPI). This dataset contains academic records for all students enrolled

in grades 9-12 in Michigan’s public schools between 2007 and 2017 and further links these stu-

dents to college enrollment and completion records from the NSC and a state-run data repository

(STARR). The high school academic records provide rich information on students’ demographic

characteristics, including race/ethnicity, gender, free and/or reduced price lunch (FRPL) eligibility,

English language learner (ELL) status, and special education enrollment; academic performance,

including math and reading tests scores on a state standardized test administered in eleventh grade;

and place of residence measured at the census block level. The final component is a key advantage

of the MDE/CEPI dataset as it allows me to very accurately determine whether a student resides
within a community college district.14 The college link provided through the NSC and STARR

14This feature of the data is a particular advantage in Michigan because the state has generous
school choice policies and nearly 6% of K-12 students to attend a school other than that to which
they are assigned (either within or outside their school district of residence). An additional 7% of
students attend a charter school (Cowen et al., 2015). Thus, using the location of a student’s high
school to proxy for her place of residence —as is common in other settings with spatial variation
(e.g., Denning, 2017) —would likely introduce measurement error to the estimation procedure.

12

contains all dates and records of students’ enrollments in colleges covered by either database. The

data also include information on the academic programs in which they enroll, using six-digit Clas-

sification of Instructional Program (CIP) codes, the credits they complete, and the awards they

receive. I match these data to postsecondary institutional information, including campus latitudes

and longitudes, from the NCES’ Integrated Postsecondary Education Data System (IPEDS). I also

gather annual in-district and out-of-district tuition rates at each of Michigan’s community colleges

from Michigan’s Workforce Development Agency.

1.3.2 Sample Construction

The goal of this paper is to estimate the causal effect of the tuition rate a student faces at her

local community college on her postsecondary enrollment decisions and outcomes. To do so, I

exploit the fact that students who live inside one of Michigan’s community college districts face

a substantially discounted tuition rate at their local community college. The challenge of this

approach is that community college district areas may be spatially correlated with unobservable

determinants of college choice. For example, community colleges may form their districts in geo-

graphic areas that have strong preferences for community college education, which would then bias

any estimates of the effect of in-district status on college enrollment or outcomes. To mitigate this

type of bias, I limit the sample to students who reside near a community college district boundary

and use fixed effects to compare the outcomes of students who reside in close geographic prox-

imity to one another and graduate from high school in the same year but differ in their in-district
status at the local community college.15

15The two mile bandwidth is chosen to maximize sample size and minimize observed differ-
ences between adjacent in-district and out-of-district students. Results using alternative band-
widths are included in the appendix and discussed in Section 1.5.5. Note that this approach is
similar in spirit to regression discontinuity (RD) designs that exploit geographically-discontinuous
treatments. However, because I do not observe students’ exact addresses and must aggregate to the
census block level, there is a mass point in the running variable at the geographic discontinuity and
I cannot use standard RD inference techniques that rely on a smooth distribution of individuals at
the discontinuity (Keele et al., 2017).

13

To implement this empirical strategy, I first identify the census blocks that are located within

each community college district. For community college districts consisting solely of counties, this

is straightforward: I assign a census block to the community college district if the census block is

contained within the county of interest. For community college districts that include public K-12

school districts, I first calculate the amount of geographic overlap between each census block and

all overlapping school districts. I then match a census block to the school district with which it

shares the most overlap and assign it to the community college district of that school district. Once

I have mapped all census blocks to their corresponding community college districts, I identify

community college district boundaries that divide a collection of census blocks that are contained

within a given community college district from a collection of census blocks that are not contained

within any community college districts. Figure A.1.1 displays all 28 community college districts
and bolds the district boundaries used in the analysis.16

To limit the analysis to students who differ in their in-district status, but reside within a small

distance of one another, I divide each identified boundary into equal segments, each of which

is no more than 5 miles long. Throughout the remainder of the text, I refer to these segments

as “boundary segments.” I next calculate the distance from the centroid of each student’s census

block to the nearest boundary segment and, in my main empirical specification, restrict the sample
to students residing within two miles of their nearest boundary segment.17 An example of this

sample restriction for the Washtenaw Community College district area is provided in Figure A.1.2.

Each dot represents a single census block centroid that is no more than two miles from the nearest

boundary segment, and dots displayed in the same shade are located closest to the same boundary

16Both Bay de Noc Community College and Glen Oaks Community College have “service dis-
tricts” in which students face tuition rates that are greater than the in-district but lower than the
out-of-district rate. I do not include boundaries that divide these areas from areas not in any com-
munity college district as they are less salient than the true community college district boundaries.
17In order to only include students who are likely to be affected by the local community
college’s listed tuition rate, I further exclude 6,687 students who are eligible for place-based
promise scholarships upon high school graduation, or whose area of residence becomes eligi-
ble for a promise scholarship during the time frame of the data. I identify areas that are eligible
for promise scholarships from the Upjohn Institute’s Promise Database: https://www.upjohn.org/
promise/promiseSearch.html.

14

segment. Intuitively, the empirical strategy compares the outcomes of students who live in census
blocks shown in the same shade, but reside inside or outside of the community college district.18

Figure A.1.3 presents visual evidence on the differences in local community college tuition

rates among students residing on either side of the identified boundary segments by plotting the

distribution of in-district vs. out-of-district tuition differentials across all border-year pairs. The

average difference in tuition between in-district and out-of-district students is $1,617, which is only

slightly higher than the average college-level difference of $1,463 (see Table A.1.1). However,

there is some variation in this differential, with the interquartile range stretching from $1,315 to

$2,036. To further explore this variation, Figure A.1.4 plots the tuition differentials against various

demographic characteristics. There is no identifiable relationship between a border-year pair’s

tuition differential and the share of economically disadvantaged students, the median household

income of the area, or students’ average test scores. This finding suggests that the variation in

tuition differentials likely come from different tuition-setting policies and practices at colleges

throughout the state rather than differences in local economies or preferences for education.

Table A.1.6 then provides descriptive statistics on the entire sample of students who graduate

from Michigan public high schools between 2009 and 2016, and on the analysis sample who live
within two miles of their nearest boundary segment.19 I also present separate means for the in-

district and out-of-district students in each sample. All variables are measured when a student

takes the Michigan Merit Exam (MME), a required standardized test that is typically administered

during a student’s junior year of high school. Panel A shows that there are some differences

in demographic characteristics between in-district and out-of-district students. For example, in-

district students are less likely to be white and are more likely to be English language learners

18I do not consider boundaries that divide two distinct community college districts, so students
residing outside of a community college district of interest do not reside within any community
college district.

19Students who graduate before 2009 or after 2016 are dropped from the sample due to incom-
plete college enrollment and completion data collection. Students enrolled in juvenile detention
centers, adult education, or alternative education programs, as well as those missing academic or
demographic variables, are also dropped from the sample.

15

(ELL). This is not surprising since community college districts tend to be located in more urban

and diverse areas of the state. Panel B then shows that in-district students score slightly lower on

their state standardized tests than their out-of-district peers.

Panel C reports college enrollment outcomes for the first year following a student’s graduation

from high school. I maintain all college enrollment spells that occur within this time frame, which

may include enrollment at multiple institutions. As a result, the sum of enrollment in different

college types is slightly larger than the total number of students who enroll in some form of post-

secondary education. In both samples, about 30% of high school graduates enroll in a community

college, with more in-district students doing so than out-of-district students, especially in the all

students sample. About 3% enroll in vocational colleges, with less in-district students doing so

than out-of-district students. In the sample of all students, about 41% of graduates enroll in a four-

year college, while about 38% do so in the analysis sample. There are little differences in this rate

between in-district and out-of-district students. In total, about 70% of all Michigan public high

school graduates enroll in college within one year, while about 67% of the analysis sample does.

1.4 Empirical Strategy

The boundary fixed effects approach, as outlined in Figure A.1.2, naturally lends itself to the

following reduced form estimating equation:

Yibt = γ + δ Districti + XiΨ + µbt + νibt

(1.1)

where Yibt is an outcome of interest for student i who resides along boundary segment b and grad-
uates from high school in year t. Districti is a dummy variable equal to 1 if student i resides in
a community college district and equal to 0 otherwise. Xi is a vector of individual control vari-
ables that may affect college enrollment choices, such as a student’s socioeconomic background
and academic aptitude. µbt is a full set of boundary segment by year fixed effects, which will
hold constant any factors affecting graduates who live in the same area along a community college

boundary segment, such as local economic conditions or changing preferences for higher educa-
tion. νibt is an idiosyncratic error term. The coefficient of interest is δ , which represents the effect

16

of residing in a community college district on Yibt.

To estimate how community college tuition itself affects students’ choices and outcomes, I also

use a two-stage least squares approach similar to Denning (2017). I choose to use this approach

because it is a straightforward way to scale the results by the mean difference between in-district

and out-of-district tuition rates. The first stage equation is:

Tuitionibt = ζ + λ Districtib + XiΦ + µbt + υibt

and the second stage equation is:

Yibt = α + β (cid:92)Tuitionibt + XiΓ + µbt + εibt

(1.2)

(1.3)

where (cid:92)Tuitionibt is predicted from the first stage, and the remainder of the variables are defined as
in previous equations.20

In order for β to represent the causal effect of local community college tuition in the 2SLS ap-
proach outlined above, it must be the case that (1) Cov(Districtib,Tuitionibt|Xi, µbt) (cid:54)= 0 and that
(2) Cov(Districtib,εibt|Xi, µbt) = 0. The first assumption states that, within a narrowly defined ge-
ographic area and graduation year, and after controlling for observable characteristics, a student’s

in-district status is related to the tuition rate he or she faces at the local community college. Given

that all community colleges in Michigan set different tuition rates for in-district and out-of-district

students, this assumption should hold. However, it is also directly testable in the data. Table A.1.7

presents the estimated first stage value of λ in three specifications of equation (1.2): including

no control variables, including only distance-related control variables, and including a full set of
distance and student control variables.21 The estimated values are quite stable across the different

20One could also estimate this relationship via ordinary least squares, but this would impose that
the relationship between the tuition differential (in dollars) and the effect of residing in-district
is linear, i.e.
that the largest in-district effects occur when there are the largest raw tuition dif-
ferentials. This may be reasonable, considering I do not see much correlation between tuition
differentials and observable characteristics, but it is also possible that boundaries with high tuition
differentials are unobservably different than boundaries with low differentials. Thus, to remain
agnostic about this relationship, I prefer the 2SLS approach.

21The distance-related variables are the distance between a student’s census block of residence

17

specifications and indicate that in-district students face a local community college tuition rate that

is approximately $1,800 lower than that of their out-of-district peers. All three estimates also have

partial F-statistics greater than 40, limiting the probability that the 2SLS estimates suffer from

weak instrument bias.

The second assumption states that, within a narrowly defined geographic area and graduation

year, and after controlling for observable characteristics, a student’s in-district status is uncorre-

lated with unobservable determinants of college choices or outcomes. This is also the assumption

needed for the identification of δ in the reduced form equation. This assumption rules out the pos-

sibility that, for example, families choose to live in community college districts due to unobserved

preferences for community college attendance. This is inherently untestable. However, there are

several reasons to believe that this assumption is likely to hold. First, community college district

boundaries are not well-publicized by the state of Michigan. The state does not maintain any

publicly-available record of community college district boundaries and each community college

has discretion over whether and how they make this information available to potential students.

Thus, it is possible that a family could select a place of residence without knowing whether or not
it is contained within a community college district.22

Second, very few students move into community college districts between 9th and 12th grade.

This suggests that families do not anticipate community college attendance and move to take ad-

vantage of the subsidized tuition rates offered to in-district students. While nearly 14% of all

students move census blocks during high school, less than 1% move from an out-of-district cen-
sus block to an in-district census block.23 Moreover, conditional on beginning high school in a

and the nearest campus of the local community college, the nearest vocational college, the nearest
public university, and the nearest private four-year college. The student control variables are: a
student’s race (white, black, or Hispanic), gender, FRPL status, special education participation,
ELL status, math and reading test scores, school of choice participation, on-time graduation status,
and dual enrollment experience.

22Property taxes for the local community college are displayed on the tax bills of property own-
ers who reside within community college districts, but there is no indication of in-district status,
nor tuition rates, on these bills.

23Author’s own calculation based on a sample of students who have records for all grades 9-12

and non-missing census block information in at least two of those grades.

18

community college district, a student has a probability of finishing high school in a community

college district of 99%. In contrast, conditional on beginning high school outside of a community

college district, a student has a probability of finishing high school in a community college district

of 4%. While I do not observe students’ residences after they graduate from high school, I restrict

outcomes to students’ enrollment choices within one year of high school graduation to avoid the

possibility that students move into community college districts as adults.

Third, students residing on either side of a community college district boundary appear quite

similar across observable characteristics. Table A.1.8 reports balance tests of observable student
characteristics and predicted community college enrollment along the boundary segments.24 The

results indicate that students residing near one another, but on opposite sides of a community

college district boundary, are quite similar. These students are similarly likely to be white, to be

eligible for free or reduced price lunch, and to be English language learners. They also score

similarly on standardized tests, graduate on-time from high school at similar rates, and have equal

predicted community college attendance rates. The only attributes across which the two groups

differ are special education status and dual enrollment participation: in-district students are both

less likely to be classified as special education students and slightly less likely to dual enroll in

a college course while in high school, although the latter result is only marginally statistically
significant.25 Tables A.1.9 and A.1.10 provide additional evidence of balance across neighborhood

characteristics and distance to local colleges.

Despite these mitigating factors, the largest threat to identification is the fact that community

college district boundaries are often congruent with either county or school district boundaries,

24I predict enrollment on the full sample of high school graduates using a probit equation that
includes the observable characteristics of the other balance tests. Specifically, I estimate enroll-
ment as a function of a student’s race, gender, FRPL eligibility, special education status, ELL
status, math test score, reading test score, on-time graduation status, and dual enrollment status.
This approach explicitly tests for differences in observable characteristics that are correlated with
community college attendance.

25Additional analyses suggest that in-district students are also less likely to be school of choice
students, but this is unsurprising given that in-district school districts tend to be larger and more
suburban and students residing in rural areas are more likely to choice in to suburban school dis-
tricts than suburban students are to choice in to rural districts.

19

inducing compound treatments at the cutoff points (Keele et al., 2017).26 To my knowledge,

there are no other specific community college policies that are discontinuously applied along com-

munity college district boundaries. Nevertheless, school districts may provide different college

information and guidance to students and families often select where to live based on school dis-

trict attributes (Caetano and Macartney, 2014) including the quality of the school district’s college

advising. A related concern is that families choose where to live based on preferences for other

types of taxes or public goods, which may be correlated with their preferences for education more

generally. However, I find that, along the boundaries, in-district residents face an average millage

rate of 15.4, while out-of-district residents face an average rate of 12.3. Given that the average

community college millage rate is about 2.5, this suggests that there is only about a 0.6 millage

difference (i.e., 0.60per1,000 of taxable value) attributable to other types of taxes, which is rather

small and unlikely to explain residential choices.

To address potential sorting into school districts, in Section 1.5.5, I repeat the analysis using a

subset of students who live in school districts which are bisected by a community college district.

Students in this sample come from families who chose to live within the school district’s bound-

aries, and therefore likely have similar preferences for education, and overwhelmingly attend the

same high school, and therefore likely receive similar college counseling. However, only a frac-

tion of the students live within the local community college’s district. I find very similar effects

of in-district status using this subsample of students, suggesting that neither residential sorting nor

school-level policies are likely driving my main results.

26The overlap of counties and community college districts is less concerning as the vast major-
ity of college advising and implementation of college access policies occurs at the school, school
district, or intermediate school district level, rather than the county level. Moreover, specifications
that include county fixed-effects produce qualitatively similar results, indicating that, among stu-
dents residing along community college district boundaries, there are not unobserved differences
in preference for higher education institutions along county lines. These results are available from
the author upon request.

20

1.5 Results

1.5.1 College Enrollment

Table A.1.11 presents the reduced form and 2SLS estimates for student’s college enrollment

choices within one year of high school graduation. The first four columns present estimates for

four mutually exclusive college categories —the local community college (that at which in-district

students receive reduced tuition), non-local community colleges (both in Michigan and in other

states), private vocational colleges, and four-year colleges. However, students may enroll in more

than one type of college within their first year following high school graduation, such that the sum

of these estimates need not exactly equal the overall college enrollment effect presented in column

(5). Panel A presents estimates for all cohorts of students, while Panel B presents estimates only

using the 2009-2011 cohorts who will be used for analyses of college completion.

The first row of each panel presents the reduced form effects of residing in a community college

district. For the “all cohorts” sample, residing in a community college district increases enrollment

in the local community college within one year of high school graduation by 6.4pp (31%), while

decreasing enrollment in non-local community colleges by 2.8pp (31%) and in private vocational

colleges by 0.7pp (20%). All three of these estimates are statistically significant at the 99% confi-

dence level and imply that students shift enrollment away from other two-year colleges and towards

their local community college when they reside in a community college district. In contrast, there

is no statistically significant effect of in-district status on enrollment in four-year colleges, and the

point estimate is quite small: -1pp, or 2.7% of the mean enrollment rate. On net, these enrollment

effects increase overall college enrollment within one year of high school graduation by 1.3pp,

or approximately 1.9% of the mean enrollment rate of 67.3%. The community college and voca-

tional college enrollment effects are qualitatively similar for the 2009-2011 cohorts, but the overall

college enrollment effect for this subsample is much smaller (0.6pp) and not statistically different
from zero.27 All changes in enrollment behavior for these cohorts come from switching out of

27In Table A.1.12 I estimate the main specification including an interaction term between the

21

non-local community colleges and vocational colleges.

The second row of each panel presents the 2SLS estimates of the effect of reducing the tuition

rate at a student’s local community college by $1,000. Across all students, this reduction in tuition

increases enrollment at the local community college by 3.5pp (18%) and is primarily driven by

a 1.5pp (17%) decrease in enrollment in non-local community colleges and a 0.4pp decrease in

enrollment in private vocational colleges (11%). Taken together, these enrollment effects increase

overall college enrollment in the year following high school graduation by a statistically significant

amount of 0.7pp, or approximately 1% of the mean enrollment rate of 67.3%. Again, the commu-

nity college and vocational college enrollment effects are qualitatively similar using the 2009-2011

subsample, but the overall enrollment effect is smaller (0.4pp) and statistically insignificant.

1.5.2 College Completion

Table A.1.14 estimates how residing in a community college district affects longer-run educa-

tional outcomes for the 2009-2011 cohorts. The first row of the table presents reduced form effects.

In-district status significantly increases the total number of college semesters students complete by

0.344 (4.2%) and the total number of credits students complete by 3.46 (4.5%), indicating that

students increase their educational attainment when they have access to low-cost local community

college. Residing in a community college district also increases the probability that a student will

transfer to a four-year college by 1.1pp (9.6%), where transfer is defined as a student beginning

college at a community or vocational college but later enrolling in a four-year college. The 2SLS

results in the second row indicate that reducing a student’s local community college tuition rate by

$1,000 increases the number of semesters of college she completes by 0.206 (2.5%), the number

of credits she completes by 2.07 (2.7%), and her probability of transferring from to a four-year

college by 0.7pp (6.5%).

in-district dummy variable and a dummy variable for being in the 2009-2011 cohorts. I find that
the effects for local community college, vocational college, and four-year college are statistically
no different for the 2009-2011 cohorts, compared to the 2012-2016 cohorts. However, the effects
for non-local community college and overall college enrollment are statistically different between
the two groups.

22

Columns (4) and (5) show that residing in a community college district does not significantly

affect students’ completion of certificates nor associate degrees, although the coefficient for asso-

ciate degree completion is positive. This lack of a degree completion effect could be driven by

the fact that community colleges have lower completion rates than their vocational counterparts.

On average, only 13.5% of students at Michigan’s community colleges complete programs within

150% of their intended length, whereas 19.6% of students at vocational colleges do so. However,

column (6) indicates that in-district status increases bachelor’s degree completion by a statistically

significant amount of 1.8pp (5.7%). The 2SLS estimate shows that reducing a student’s local com-

munity college tuition rate by $1,000 increases her probability of completing a bachelor’s degree

by 1.1pp (3.5%).

To better understand these degree completion outcomes, Table A.1.15 reports the distribution

of associate and bachelor’s degree increases across seven categories of majors: (1) general studies,

which primarily consists of pre-transfer programs at community colleges; (2) liberal arts and sci-

ences; (3) health; (4) business; (5) technical fields, such as engineering and technology programs;

(6) professional fields, such as education, criminal justice, and journalism; and (7) other or unspec-
ified fields, which primarily consists of degrees awarded without a major recorded in the data.28

For each estimate, the outcome of interest is whether a student completes a given degree in a given

field. These outcomes are mutually exclusive such that the sum of their coefficients must equal the

overall degree completion increases presented in Table A.1.14.

Panel A reports the reduced form and 2SLS results for associate degree completion by field,

indicating that a $1,000 decrease in a student’s local community college tuition rate increases her

probability of earning a general studies associate degree by 0.6pp (17.1%) and an associate degree

28For all students who enroll in a postsecondary institution covered by the NSC, the MDE/CEPI
dataset records the six-digit federal Classification of Instructional Program (CIP) code of the pro-
grams in which students enroll. I define a student as earning a degree in a given field of study if
the student is enrolled in the field of study when she earns her degree. Table A.1.16 lists the set of
two-digit CIP codes included in each category. If a student earns more than one degree of the same
type (e.g., multiple associate degrees), only the field of study for her first degree is considered in
this analysis.

23

in other or unspecified fields by 0.2pp (12.5%). These estimates indicate that, while reduced lo-

cal community college tuition does not statistically significantly increase overall associate degree

completion, it shifts the fields in which students earn associate degrees. Specifically, students are

more likely to earn degrees that enable transfer to four-year colleges than degrees which lead to

labor market entry. Panel B reports the effects of in-district status and reduced tuition on bache-

lor’s degree completion by field, and shows that the increase in bachelor’s degree completion is

primarily driven by increases in bachelor’s degree completion in business and professional fields

of study. Given that business majors experience substantial earnings gains in the labor market
(Andrews et al., 2017), this increase is likely to have longer-term payoffs for students.29

Taken together, these completion results indicate that having access to in-district tuition induces

students to complete associate degrees that enable transfer to four-year colleges, and to ultimately

complete bachelor’s degrees. These improved outcomes are likely the result of differences in

institutional resources and objectives between Michigan’s community and vocational colleges. For

example, community colleges spend about $1,166 more per student on instruction than vocational

colleges and also award a large share of their degrees in general liberal arts fields (two-digit CIP

code 24).

In contrast, vocational colleges rarely award degrees in this area. Given that these

degrees are generally intended for students transferring to four-year colleges, it is not surprising

that community colleges also have substantially higher rates of transfer than vocational colleges:

36% compared to 11%. Table A.1.13 provides additional summary statistics on the differences

between these institutions that further explain why attending a community college, rather than a

vocational college, could improve students’ educational attainment.

29To further explore the increase in professional fields, Table A.1.17 presents separate estimates
for disaggregated majors contained within this category. The results indicate that the increase is
driven by more students completing degrees in education majors and parks, recreation, leisure, and
fitness studies majors. The largest majors in the latter category are exercise science (CIP 31.0505)
and sports administration (CIP 31.0504). It is not obvious why the degree increases are largest
in these fields as community colleges in Michigan have transfer programs for a wide variety of
majors; future work could explore further reasons why students primarily choose these pathways.

24

1.5.3 Heterogeneity

Table A.1.18 reports heterogeneous treatment effects by a student’s FRPL eligibility, gender,
and academic achievement for select college enrollment and completion outcomes.30 Panel A

shows that FRPL eligible and ineligible students respond similarly to residing in a community

college district with regards to local community college enrollment, but their substitution patterns

are different. FRPL ineligible students, who come from higher-income families, respond to living

in a community college district by changing which community college they attend: they are 3.3pp

less likely to enroll in a non-local community college and 6.7pp more likely to enroll in their local

community college. In contrast, FRPL eligible students respond to in-district status by reducing

non-local community college enrollment by only 1.5pp. These students also decrease enrollment

in vocational colleges by 0.8pp and increase overall college enrollment by 1.8pp, indicating that

having access to a low-tuition local community college option is particularly important for overall

college enrollment for lower income students. However, FRPL eligible and ineligible students earn

associate and bachelor’s degrees at comparable rates.

Panel B shows that male students are more responsive to in-district status than female students:

they are 7.2pp more likely to attend the local community college than their out-of-district peers,

whereas female students are 5.6pp more likely to do so. The underlying substitution effects are also

different by gender. Female students respond to in-district status by significantly reducing enroll-

ment in non-local community colleges and vocational colleges, while male students only somewhat

reduce enrollment in non-local community colleges and also reduce enrollment in four-year col-

leges. This difference in substitution patterns may stem from the fact that vocational colleges tend

to offer degrees in female-dominated fields, such as healthcare. Nevertheless, as in the case of

FRPL eligible and ineligible students, these differences do not persist when looking at completion

30For the binary FRPL status and gender variables, I extend equation (1.1) to include an inter-
action term between Districti and the demographic variable of interest. For the test score variable,
students are assigned to score quartiles among all students who took the MME exam in the same
year based on their combined scores on the math and reading exams. I then modify equation (1.1)
to include a dummy variable for the middle two quartiles, a dummy variable for the top quartile,
and interaction terms with these dummy variables and Districti.

25

outcomes. That is, even male students who forgo initially attending four-year colleges to attend

their local community college do not forgo ultimately earning bachelor’s degrees.

Lastly, Panel C reports the estimated effects by students’ test scores. Students from the bot-

tom three test score quartiles are very responsive to residing in a community college district: it

increases their probability of enrolling in the local community college by 7.4-7.5pp. In contrast,

students from the top quartile respond to in-district status by increasing their enrollment in the local

community college by only 2.9pp. There are also differences among these groups when consider-

ing substitution effects. Students from bottom quartile forgo enrollment in non-local community

colleges, vocational colleges, and four-year colleges, whereas students from the middle quartiles

primarily forgo enrollment in other community colleges. However, there are no decreases in bach-

elor’s degree attainment among any group of students, which again suggests that the students who

are deterred from attending four-year colleges do not forgo opportunities to earn bachelor’s de-

grees.

1.5.4 Robustness Checks

The reduced form and 2SLS results both rely on the assumption that there are no unobservable

differences between students residing on either side of a community college district boundary that

affect their college choices and outcomes. One threat to this assumption is that the two mile band-

width does not create appropriate treatment and control groups because individual students may

live several miles from one another, and therefore, may have different preferences over postsec-

ondary education options or may be exposed to different social networks and information about
college.31 To test whether the results hold across comparisons of students who reside farther from

or closer to one another, I re-run the reduced form analysis for local community college enrollment

using varying bandwidths from 0.1 miles to 4 miles. Figure A.1.5 presents the estimates from these

specifications, which range from 2.5pp to 8.0pp and are all statistically significant at the 90% level

31Observed differences in student characteristics do not necessarily decrease as the bandwidth
is narrowed, and in some cases, actually increase. Table A.1.19 documents this fact by providing
the balance tests from Table A.1.8 for varying bandwidths.

26

or greater. Moreover, the 90% confidence intervals of all of the point estimates contain the 6.4pp

estimate from the main specification, indicating that the two mile bandwidth selection is not the

main driver of the results.

A greater threat to the identifying assumption is the fact that community college district bound-

aries are often congruent with school district boundaries, and a non-trivial share of families choose

where to live based on school district characteristics. To test whether differences in school dis-

tricts drive the college enrollment and completion results, I provide an alternative specification

that compares the college choices and outcomes of students who reside in the same school district

but live on opposite sides of a community college district boundary. This situation occurs when a

community college district is congruent with a county (or multiple counties), but school districts

in the area span more than one county. Figure A.1.6 identifies the 25 school districts in the state

in which at least 10% of the high school residents reside within the community college district

and at least 10% reside outside. Using these school districts as the analysis sample eliminates the

concern that families sort into more desirable school districts that are located in community college

districts. In addition, this approach holds constant college counseling information provided by the

school district as the majority of students residing within one of these school districts attend the

same high school: twenty-four of these twenty-five school districts contain only one high school,

and 92% of students attend a high school that is located within their district of residence.

I repeat the reduced form and 2SLS analyses on this selected sample, but replace the boundary

segment by year fixed effect with a school district of residence by year fixed effect. Table A.1.20

presents results from this analysis for enrollment in the local community college for the 2009-

2016 cohorts and bachelor’s degree completion for the 2009-2011 cohorts. The first column of

the table presents the local community college enrollment results from the main specification. The

second column presents results from the within school district specification. Using this sample and

specification, residing in-district increases enrollment at the local community college by 5.0pp,

and reducing the tuition rate by $1,000 increases enrollment by 3.2pp. Neither of these estimates

27

is statistically different from the analogous estimates produced by the main specification.32 The

third and fourth columns show that the estimated degree completion effect is also similar when

using the within school district specification, indicating that the results are unlikely to be driven by

selection into particular school districts.

Another way to check the robustness of the main results is to examine whether college en-

rollment choices and completion outcomes discontinuously change along geographic boundaries

other than community college districts. If the differences in college outcomes between in-district

and out-of-district students residing along a community college district border are truly driven by

differences in tuition rates, then there should be no differences in college choices nor outcomes

along borders where tuition rates do not differ and there are not other related policies in place. To

test whether this is true, I conduct two different placebo tests. First, I contract all community col-

lege district perimeters by two miles and compare the college choices of students residing within

two miles of the new placebo boundary. This approach compares the choices and outcomes of

students who all live within the same community college district, and face the same low tuition

rate, but differ in how close they live to the center of the community college district. Second, I

expand all community college district perimeters by two miles and compare the college choices

of students residing within two miles of the new placebo boundary. In this approach, I compare

students who live outside of a community college district but differ in how close they live to the

nearest community college district boundary.

Table A.1.22 presents the results from these approaches. The first column indicates that stu-

dents residing within a community college district, but on either side of the contracted placebo

boundary, do not differ in their likelihood of attending the local community college. The second

column shows that students residing outside of a community college district, but on the side of

the expanded placebo boundary that is closer to the true community college district, are slightly

more likely to attend the local community college. However, this estimate (0.7pp) is quite small

32Table A.1.21 contains estimates for all one-year enrollment outcomes using this alternative
specification. Given the reduced sample size, these estimates lack precision but are qualitatively
similar to those produced by the main specification.

28

compared to the estimate of 6.4pp along the true community college district boundaries and is only

marginally significant. The third column indicates that students who reside within a community

college district, but inside the contracted placebo boundary, are slightly less likely to obtain bach-

elor’s degrees, while the fourth column indicates that out-of-district students living on either side

of the expanded placebo boundary are equally likely to obtain a bachelor’s degree. Both sets of

results indicate that enrollment and completion outcomes do not change in meaningful ways along

non-community college district boundaries, providing additional validation that the main results

capture the true effect of reduced community college tuition.

1.6 Conclusion

Community colleges serve millions of undergraduate students each year and are increasingly

the focus of college access policies, making it critical to understand how students respond to their

costs. In this paper, I provide new evidence on the effect of community college tuition rates on

students’ college enrollment decisions, persistence in college, and degree completion. To do so, I

exploit the fact that Michigan’s community colleges offer students different tuition rates depend-

ing on whether they live within our outside a college’s district boundaries, as well as the fact that

nearly one-quarter of Michigan’s high school graduates do not live within the boundaries of any

community college district. This geographic variation allows me to use a boundary fixed effects

design that compares the outcomes of students who reside on either side of a community college

district but who are otherwise observationally similar. I combine this approach with detailed ad-

ministrative records from the Michigan Department of Education to track students’ residences,

college enrollment choices, and college completion outcomes over time.

Among students graduating from Michigan public high schools between 2009 and 2016, I find

that a $1,000 decrease in the advertised tuition rate at a student’s local community college upon

graduating high school increases the probability of enrollment at the college by 3.5pp, or about

18%. This increase in local community college enrollment can be partially attributed to an increase

in overall college enrollment but is also due to a decrease in enrollment at non-local community

29

colleges and at private vocationally-focused colleges who offer similar degree programs to com-

munity colleges. However, I find little evidence that students forgo attending four-year colleges or

decrease their overall educational attainment in response to a low community college tuition rate.

Instead, for students who graduate from high school between 2009 and 2011, I find an increase

in persistence in college, credit completion, transfer to four-year colleges, and bachelor’s degree

completion. These improved outcomes may be attributed to the substitution towards local com-

munity colleges and away from non-local community colleges and vocational colleges, as overall

college enrollment is not affected by reduced community college tuition for this subset of students.

This finding suggests that gains from community college attendance can extend to more students

than identified in prior work (Rouse, 1995; Reynolds, 2012; Mountjoy, 2019): namely, students

who would have attended a private vocational college in the absence of a community college.

These results have meaningful policy implications, both for Michigan and for community col-

lege policies throughout the country. Approximately 100,000 students graduate from Michigan

public high schools in a given year; of these, about 23,000 do not live within a community col-

lege district. Based on this paper’s estimates, reducing local community college tuition by $1,000
for these students would induce 253 more students to earn bachelor’s degrees.33 Given that the

average discounted lifetime premium to earning a bachelor’s degree is about $300,000-$600,000

(?), the total discounted earnings benefits to students under such a policy would be between $76

million and $152 million. These figures far exceed the $5-$6 million cost of reducing tuition by
$1,000 for all out-of-district students who attend community colleges.34 In fact, the income tax

gains alone (assuming students continue to reside in Michigan) would total $3-6 million under

Michigan’s current state income tax rate of 4.25%. Other policies that induce students to attend

33Currently, about 6,828 (29.7%) out-of-district students in each cohort earn a bachelor’s degree.
Increasing this percentage by 1.1pp (estimated increase in overall community college enrollment)
to 30.8% would mean 7,081 students would complete a bachelor’s degree —a difference of 253
students.

34About 5,267 (22.9%) out-of-district students attend community colleges each year. Increasing
this percent to 24.9% would bring the total to about 5,727. At $1,000 per student, the cost of
implementing the proposed policy would be $5,727,000 plus administrative costs.

30

community colleges rather than not pursuing postsecondary education or attending lower quality

private colleges, including the regulation of the for-profit industry and funding for new community

college campuses, are likely to be similarly cost-effective and should continue to be a focus of

education policy research.

However, the findings of this paper are not without limitations. One limitation of this study is

that the results are estimated from an empirical design that compares students living very near one

another, and thus, does not address the role of distance in college choices. Given the documented

relationship between college proximity and college attendance (Card, 1995; Currie and Moretti,

2003; Lapid, 2017), it is likely that rural students who live far from colleges face additional chal-

lenges in accessing higher education and may not respond to reduced tuition as strongly as their

non-rural peers. Future work should seek to identify how reduced tuition policies differentially af-

fect rural students and should investigate alternative policy interventions to increase college-going

behavior among this population. Second, the tuition reduction studied in this paper does not in-

clude changes in marketing, mentoring, or college campuses. Policies that include such factors,

such as broad free-tuition programs or the expansion of community college districts, may influence

students in different ways and should continue to be rigorously evaluated as they are implemented.

31

CHAPTER 2

COMMUNITY COLLEGE PROGRAM CHOICES

IN THE WAKE OF LOCAL JOB LOSSES

2.1 Introduction

The educational decisions that young people make can substantially affect their long-run labor

market outcomes and overall economic well-being. The typical college graduate will earn more

than double the typical high school graduate over her lifetime (Hershbein and Kearney, 2014),

while also experiencing improved health, less reliance on social safety net programs, and fewer

interactions with the criminal justice system (Oreopoulos and Salvanes, 2011). Equally large earn-

ings gaps exist among students with the same level of education who pursue different fields of study

(Altonji et al., 2012), and a growing body of literature shows that students take these earnings gaps

into account when selecting college majors (Montmarquette et al., 2002; Beffy et al., 2012; Long

et al., 2015), particularly when provided with reliable information about the labor market (Wiswall

and Zafar, 2015; Hastings et al., 2015; Baker et al., 2018).

However, the vast majority of college major choice research focuses on the four-year college

sector. The nearly ten million students who attend two-year community colleges (National Center

for Education Statistics, 2018) also must decide which fields to study, and their decisions have

similarly large implications for their labor market outcomes. For example, students who enroll in

healthcare programs can expect to experience large earnings gains in the labor market, while stu-

dents who select other programs may not earn more than their peers who do not enroll in postsec-

ondary education (Bahr et al., 2015; Belfield and Bailey, 2017; Stevens et al., 2018; Grosz, 2018).

In response to these earnings differences, policymakers have begun to introduce programs that

aim to steer students into programs that align with local economies. Several states tie community

colleges’ appropriation funding to their ability to produce degrees in high-demand areas (Snyder

and Boelscher, 2018), and some recent financial aid programs incentivize students to choose in-

32

demand fields of study (Allen, 2019; Natanson, 2019). Yet, there is little evidence on the extent to

which labor market opportunities affect community college students’ program choices.

In this paper, I use administrative data on the education decisions of recent high school gradu-

ates in Michigan to analyze how labor market conditions influence students’ choices of community

college programs. Specifically, I consider how students’ choices respond to local, occupation-

specific job losses that alter the relative benefit of pursuing different programs. These types of

job losses are likely to be particularly influential to community college students for several rea-

sons. First, community college students tend to remain close to home when attending college and

after graduating, making it likely that local labor demand shapes students’ expected labor market
prospects more than state or national demand.1 Second, community college programs are generally

designed to take two years or less to complete. Thus, while four-year college students may con-

sider longer-run labor market trends when choosing college majors, community college students

may be more likely to consider short-term fluctuations in labor demand. Finally, many programs

at the community college level are closely tied to specific occupations, such as nursing or weld-

ing, rather than the broad subjects that typically define majors at four-year colleges. As a result,

the expected labor market opportunities associated with programs align closely with labor market

opportunities in specific occupations.

My empirical approach exploits plausibly exogenous variation in students’ exposure to local

job losses resulting from mass layoffs and plant closings.

I further rely on the distribution of

occupations across industries to create estimated measures of occupation-specific labor demand

shocks that align closely with six broad groups of community college programs. Intuitively, these

measures isolate job losses that affect the types of occupations community college graduates would

expect to enter after completing their educational programs. For example, hospitals employ a

large number of healthcare workers with community college credentials, such as nurses and health

1The median distance a community college student travels to campus is only eight miles (Hill-
man and Weichman, 2016), and over 60% of community college graduates live within 50 miles of
the college they attended (Sentz et al., 2018). In Michigan, I estimate that 66% of students who
attend community colleges within six months of high school graduation attend one located in their
county. This number is 86% for students who live in a county with a community college.

33

assistants. Therefore, hospital closures should change the benefit to local students of enrolling in

community college healthcare programs. In contrast, mass layoffs at prisons will mostly affect law

enforcement professionals and, in turn, should alter the benefits of entering community college law

enforcement programs.

By comparing cohorts in the same county that were exposed to different local job losses as they

exited high school, I show that students’ program choices are sensitive to occupations’ local labor

market conditions. On average, an additional layoff per 10,000 working-age residents in a county

reduces the share of the county’s high school graduates enrolling in related community college

programs by 0.8%. Correspondingly, a one standard deviation increase in layoff exposure reduces

enrollment by 3.8%. This effect is most pronounced when layoffs occur in a student’s county

during her senior year of high school, and is driven by students substituting enrollment between

community college programs, rather than forgoing higher education opportunities.

To explain these substitution patterns, I leverage data on the skills required in different occupa-

tions from the U.S. Department of Labor’s Occupational Information Network (O*NET) to create

measures of skill similarity between community college programs. I then document that students

primarily shift their enrollment into programs that require similar skills to the field affected by

layoffs. Moreover, when occupations that do not have close substitutes experience negative em-

ployment shocks, students exhibit a lower degree of responsiveness. This finding suggests that

students’ abilities to adapt to labor market changes depends on the set of available educational

choices and further indicates that supply-side responses by colleges could alter the effects of local

labor market downturns.

These results contribute to two related lines of literature on how individuals make human capital

investment decisions. First, the results add to a large body of empirical work on factors affecting

what students study in college, particularly how expected wages affect students’ college majors.

Most prior work at the four-year college level finds that, to some extent, expected wages influence

students’ choices (Altonji et al., 2016). Consistent with this finding, a recent line of work shows

that the composition of college majors changed following the Great Recession, with more students

34

pursuing “recession-proof” majors (Shu, 2016; Liu et al., 2018; Ersoy, 2019). Choi et al. (2018)

also show that the occurrence of “superstar” firms with abnormally high stock returns increases the

number of four-year college students majoring in related fields.

Related research at the community college level is limited, but two recent studies indicate that

students attending these institutions are sensitive to expected labor market prospects. Baker et al.

(2018) perform an information experiment and find that students’ program choices respond to new

information about labor market outcomes, particularly the salaries earned by previous graduates.

Meanwhile, Grosz (2018) uses a shift-share approach to show that, in California, the distribution of

community college program completions has kept pace with statewide employment composition

changes. He further shows that these trends are primarily due to changes in student demand rather

than supply-side responses by colleges. I build on these findings by showing that exposure to job

losses also affect students’ choices across community college programs. In line with prior work,

these effects are rather small in magnitude, suggesting that factors outside of the labor market play

a substantial role in determining students’ choices.

Second, this research provides new evidence that local labor market shocks can affect education

choices across a variety of margins. Several recent papers exploit mass layoffs and similar events

to study how labor market conditions affect college enrollment (Charles et al., 2018; Hubbard,

2018; Foote and Grosz, 2019). They generally find that poor labor market conditions lead to

an increase in college enrollment, and conversely, that economic booms decrease postsecondary

enrollment and completion. A line of literature on the sensitivity of community college enrollment

to the business cycle confirms this finding (Betts and McFarland, 1995; Hillman and Orians, 2013).

However, few papers consider the occupation- or industry-specific nature of local labor market

shocks. Two recent exceptions are Weinstein (2019), who finds that various industry-level shocks

affect the composition of college majors at nearby four-year universities, and Huttunen and Riukula

(2019), who find that Finnish children are less likely to enter the same field of study as their parent

when their parent has been laid off. I find similar responses to local shocks among a previously

unstudied population of students and also show that students shift enrollment towards programs

35

that require similar skills, which has not been documented in prior work.

2.2 Conceptual Framework

This paper estimates how community-level job losses affect students’ postsecondary choices,

particularly at the community college level. The basic economic intuition of this analysis is that

job losses occurring through labor market shocks (e.g., mass layoffs and plant closings) represent

changes in local labor demand, which in turn can affect students’ expected benefits of pursuing

different postsecondary education programs. To see the potential changes in students’ decisions

arising from a change in expected benefits, consider a simplified setting where student i decides

between four different postsecondary options: (1) a community college vocational program that

leads to a career in occupation group A (e.g., health), (2) a community college vocational program

that leads to a career in occupation group B (e.g., business), (3) a four-year college program (lead-
ing to a bachelor’s degree), or (4) directly entering the labor market.2 Each alternative is associated

with an expected lifetime benefit, Bi j, where j denotes one of the choices. This expected benefit
term is a function of student i’s expected earnings in related occupations and the student’s taste

for the occupations and/or coursework. That is, Bi j = Yi j + µi j, where Yi j is an expected earnings
term and µi j is a taste parameter. For example, the expected benefit to student i of pursuing a com-
munity college health program is a combination of the expected earnings in community college

health occupations and how much a student expects to enjoy the nature of healthcare work and

coursework. Each alternative is also associated with an expected cost, Ci j.

Students choose the alternative that maximizes Ui j = Ui(Bi j −Ci j), where Ui is some increas-
ing, concave function. That is, a student will choose alternative j if Ui j > Uik for all j (cid:54)= k and the
probability that student i chooses alternative j can be expressed as Pi j = P(Ui j > Uik). Suppose
that student i observes a plant closing or mass layoff while she is deciding which postsecondary

option to pursue. Her response to the shock will depend on how it affects the occupations associ-

2Students may also choose to enroll in a non-vocational program at a community college. Be-
cause these programs are typically designed to assist students in transferring to four-year colleges,
I implicitly consider them as part of option (3), a four-year college program.

36

ated with each alternative. Consider two extreme examples. In one, the labor market shock only

affects community college health occupations and reduces the expected earnings of pursuing health
programs by ε1, while holding all other components of the model constant. In another, the labor
market shock affects all occupations in the economy and reduces Yi j by ε2 for all alternatives. In
the first example, the utility student i receives from entering a community college health program

will decrease and, if the decline is large enough, she will choose a different postsecondary option.

If the student has a strong taste for vocational education —that is, a high µi j term for the vocational
program options —she will likely shift her enrollment into the other vocational program. If not,

may no longer enroll in college or may enroll in a four-year college program instead. In contrast,

in the second example, the utility student i receives from each alternative will decrease and the

student’s choice should be less affected.

These examples highlight that the anticipated effects of layoffs depend on the distribution of

layoffs across different segments of the economy. Moreover, they show that labor market shocks

can have large effects without inducing students to change whether or where they enroll in college.

Namely, students can choose to enter other programs within the vocational community college

sector. Previous studies that only consider the effects of layoffs on college entry do not capture this

response and potentially miss important labor market implications since the returns to a community

college education vary significantly across programs.

2.3 Institutional Setting & Enrollment Data

The institutional setting for this analysis is the community college market in the state of Michi-

gan. Michigan is home to 28 public community colleges, which together enroll more than 300,000

students annually (Michigan Community College Association, 2019). Local boards of trustees

control and govern the colleges, but all institutions share two key features. First, all colleges are
open enrollment institutions, meaning students can enroll regardless of academic preparation.3

3Colleges may set admissions standards for select programs, but most programs do not have
such requirements. For example, at Lansing Community College, one of the largest in the state,
only 7 out of 209 programs use selective admissions (https://www.lcc.edu/academics/documents/

37

Second, the colleges primarily confer certificates and associate degrees, which may either be vo-
cational or non-vocational in nature.4 Vocational programs are designed to prepare students for

immediate entry into the labor market and have direct links to specific occupations, whereas non-

vocational programs typically consist of general education courses and prepare students to transfer

to four-year colleges and universities.

2.3.1 Programs Offered by Michigan’s Community Colleges

Due to the deregulated nature of Michigan’s community college system, the state does not

systematically track the programs offered by each college over time. However, in 2011 and 2013,

the Department of Treasury published the “Michigan Postsecondary Handbook,” which provides

a listing of all programs offered by each of Michigan’s community colleges and includes their

degree level, number of credits, and six-digit Classification of Instructional Program (CIP) codes.

The Workforce Development Agency also maintains an online database of all current programs

offered by the state’s community colleges. I use data from the handbooks and online database to

classify programs into vocational and non-vocational categories, as well as to create the program

groups that I use to analyze students’ responses to job losses in related occupations.

To begin, I match each CIP code listed in one of the program listings to its associated occu-

pation code in the Standard Occupation Classification System (SOC) using a crosswalk developed
by the Bureau of Labor Statistics (BLS) and National Center for Education Statistics (NCES).5 In

the crosswalk, a CIP code is only matched to an occupation if “programs in the CIP category are

preparation directly for entry into and performance in jobs in the SOC category” (National Center

for Education Statistics, 2011). For example, physical therapy assistant programs (CIP 51.0806)

are matched to physical therapy assistants (SOC 31-2021) and welding technology programs (CIP

48.0508) are matched to welders (SOC 51-4121). One limitation of the crosswalk is that CIP codes
pdf-policies/selective-admission-programs-criteria.pdf).

4Since 2012, Michigan’s community colleges have been able to confer bachelor’s degrees in a
small number of fields. However, as of 2016, community colleges had only awarded 116 bachelor’s
(House Fiscal Agency, 2017).

5The crosswalk can be accessed at: https://nces.ed.gov/ipeds/cipcode/resources.aspx?y=55.

38

are constant across levels of education. As a result, some programs may be matched to occupa-

tions that are unlikely to be obtained by recent community college graduates. For example, the

CIP code for registered nursing (51.3801) is matched to the SOC codes for both registered nurses

(29-1141), which is a career attainable by graduates of community college nursing programs, and

postsecondary nursing instructors (25-1072), which requires an advanced degree. To ensure all

programs are only mapped to attainable occupations, I further match the SOC occupation codes

to data on job preparation requirements from O*NET and limit the occupation matches to those

that require at least a high school diploma but not necessarily a bachelor’s degree. I then define a

program as a vocational program if it is matched to an occupation within this subset of attainable

occupations. All other programs are considered non-vocational. These programs include general

studies programs in which students take core classes that transfer to four-year colleges, pre-transfer

programs in specific areas (such as pre-engineering), or academic programs that do not have close
occupation links (such as foreign languages).6

Table B.1.1 provides summary statistics on the programs offered by Michigan’s community

colleges in 2011. On average, a college offers 117 unique academic programs, with 81% being

vocational. The five most commonly offered vocational programs, according to broader four-digit

CIP codes, are those in vehicle maintenance and repair technologies (CIP 47.06), industrial pro-

duction technologies (CIP 15.06), allied health (CIP 51.09), criminal justice and corrections (CIP

43.01), and business administration (CIP 52.02). To analyze students’ choices across this large

set of programs, I create six broad groups of programs based on programs’ matched occupations:

business, health, skilled trades, STEM, law enforcement, and other. I create these groupings by

combining programs that are matched to similar two digit SOC occupation codes and, throughout
the remainder of the text, refer to the occupations they contain as community college occupations.7

6Any programs that explicitly state in their name that they are “pre-transfer” programs are

considered non-vocational, regardless of whether an occupational match exists.

7Programs can be matched to more than one detailed SOC occupation code, but 95% of pro-
grams are matched to only one two-digit SOC occupation code. For the 5% (22 programs) that
are matched to more than one two-digit SOC code, I merge in data on occupational employment
from the BLS Occupational Employment Series and assign programs to the occupation group of

39

Table B.1.2 provides a list of the two-digit SOC codes contained within each group.

2.3.2 Students Enrolled in Michigan’s Vocational Programs

To analyze how enrollment in community college programs responds to job losses in related

occupations, I rely on a student-level administrative dataset provided by the Michigan Department

of Education (MDE) and Center for Educational Performance and Information (CEPI). The dataset

contains high school academic records for all students who attended public high schools from 2009

to 2016 and further links students to college enrollment and completion records from the National

Student Clearinghouse (NSC) and a state-run data repository (STARR). The high school academic

records provide rich information on students’ demographic characteristics and academic perfor-

mance, including race/ethnicity, gender, standardized test scores, and census block of residence.

All variables are measured during students’ eleventh grade year, when they complete state stan-

dardized tests. The college link provided through the NSC and STARR contains all records of

students’ enrollments in colleges covered by either database, as well as information on the aca-

demic programs in which they enroll, the credits they complete, and the awards they receive. Like

the information on colleges’ program offerings, program enrollment is recorded using six-digit

CIP codes each semester a student is enrolled in a postsecondary institution.

I focus my analysis on high school graduates’ first college enrollment and program choices
within six months (180 days) of graduating from high school.8 This restriction ensures that the

county in which a student resides during high school is a valid measure of her local labor market

when she is deciding her postsecondary choice. Once students graduate from high school, I no

longer observe where they reside, and therefore, cannot reasonably assume that the labor market

shocks occurring in their high school county are the labor market shocks they actually observe.

Moreover, by limiting enrollment choices to those occurring soon after high school graduation,

the matched occupation that had higher statewide employment in 2009.

8In order to focus on students who are likely to consider postsecondary education, I drop stu-
dents enrolled in juvenile detention centers, adult education, or alternative education programs
from the analysis. Results are robust to including these students.

40

I limit the possibility that supply-side responses by colleges drive my results. For example, it is

unlikely that colleges can respond to labor market shocks by altering the programs or courses they
offer, as these decisions are typically made months or years in advance.9

Table B.1.3 provides summary statistics on Michigan’s high school graduates disaggregated by

their first postsecondary education choices. A non-trivial share of students enroll in vocational and
non-vocational community college programs each year: 9% and 14% of graduates, respectively.10

Students who enroll in vocational programs are more likely to be economically disadvantaged than
students in non-vocational programs and also score lower on state standardized tests.11 They are

also more likely to be male and a racial minority. Compared to their peers who do not enroll in

college, they are less disadvantaged and more academically prepared.

Table B.1.4 disaggregates the summary statistics by students’ vocational program choices.12

Across the eight cohorts in the sample, about 24% of vocational students enroll in business pro-

grams, while 23% enroll in health programs, 8% enroll in the skilled trades, 13% enroll in STEM,

13% enroll in law enforcement, and 20% enroll in other programs, such as culinary arts or graphic

design. There are some demographic differences across the program groups. For example, stu-

9Because Michigan does not provide annual information on the programs offered by each com-
munity college, I am unable to directly analyze whether colleges alter course or program offerings
in response to local job losses. However, Grosz (2018) provides evidence that student demand is
much more responsive to labor market trends than college supply.

107.9% of community college students simultaneously enroll in a vocational and non-vocational
program. I classify these students as enrolling in vocational programs. 6.3% of vocational students
enroll in more than one six-digit CIP code. If a student enrolls in two programs and one of the
programs is in the “other” category, I assign the student to the alternative program. Otherwise, I
randomly assign the student to enroll in one of the programs they have selected. In Section 2.6, I
show that the results are robust to dropping students who enroll in multiple program groups.

11Students are classified as economically disadvantaged if they qualify for free or reduced-price
meals under the National School Lunch Program, are in a household that receives food (SNAP) or
cash (TANF) assistance, are homeless, are a migrant, or are in foster care.

12 To verify that program choices accurately capture students’ educational experiences, I cat-
egorize community college courses into the same six occupation groups and tabulate the share
of courses taken in different subject areas among students enrolled in different programs. Figure
B.1.1 presents these results. The figures show that students who indicate enrollment in a given
program group take disproportionately more courses, and earn disproportionately more credits, in
the subject area of their program than students in other program groups.

41

dents who enroll in skilled trades programs are overwhelmingly white (84%) and male (94%). In

contrast, students who enroll in health programs tend to be non-white (29%) and female (78%).

There is less sorting across academic abilities: average math and reading test scores are similarly

low across the programs, but nearly all students in each group graduate from high school on time.

2.4 Measuring Local Job Losses

In my empirical approach, I build on work by Hubbard (2018) and Foote and Grosz (2019)

that uses the prevalence of mass layoffs and plant closings to proxy for changes in local labor

demand. A key advantage of this type of data is that events are reported at the establishment level.

Therefore, I can generate counts of reported job losses in small industries and small counties that

are typically suppressed or imputed in county-level databases. For example, of 8,217 possible

county-industry pairs in Michigan (83 counties, 99 NAICS 3-digit subsectors), only 2,633 (32%)

have a complete panel of employment data available in the BLS’ Quarterly Census of Employment

and Wages (QCEW) series. Other data series, such as County Business Patterns, have similar

limitations, which I detail in Appendix B. Layoff data are also advantageous because they represent

sharp declines in local employment that are plausibly exogenous to students’ educational choices,

and are likely representative of the employment changes students observe through newspapers and

other media outlets.

My primary source of layoff data is a listing of all mass layoffs and plant closings reported

to the Michigan Workforce Development Agency (WDA) under the federal Worker Adjustment

and Retraining Notification (WARN) Act of 1989. The WARN Act requires employers with 100

or more employees to provide at least 60 days notice to employees ahead of large, permanent

reductions in employment. Two types of events may trigger a WARN notice: (1) a plant closing

affecting 50 or more employees at a single employment site, or (2) a mass layoff affecting either

500 or more employees or between 50 and 499 employees that account for at least one-third of the

employer’s workforce (U.S. Department of Labor, 2019). Employers must give written notice of

the anticipated layoff to the employees’ representative (e.g., a labor union), the chief local elected

42

official (e.g., the mayor), and the state dislocated worker unit. If employers do not provide such

notice, they are liable to provide each aggrieved employee with back pay and benefits for up to

60 days. Krolikowski and Lunsford (2020) offer additional information on the WARN act and

document its value as a labor market indicator.

All WARN notices filed in Michigan are publicly available on the WDA’s website. However,

the WARN Act does not apply to government entities, which limits my ability to observe layoffs

in law enforcement professions —one of Michigan’s most popular community college program

groups. To overcome this limitation, I supplement the WARN data with a listing of correctional

facility closures and corresponding staff reductions from Michigan’s Senate Fiscal Agency (SFA).

These events are analogous to plant closures in the private sector but particularly affect public law

enforcement occupations, such as corrections officers.

2.4.1 Using WARN Data to Generate Occupation-Specific Layoff Exposure

The layoff data available from the WDA include a record of the date that each mass layoff or

plant closing event was reported to the state, along with the name of the company, the city where
the affected operation is located, and the number of affected workers.13 The correctional facility

closure data available from the SFA include a record of the name of the correctional facility that

closed, along with the year and number of affected full-time equivalent (FTE) workers. For each

correctional facility closure, I find related local news articles to approximate the date the closure

was announced and the county in which the correctional facility was located.

Panel A of Figure B.1.2 plots the number of mass layoffs, plant closures, and correctional

facility closings reported during each academic year from 2001 to 2017, where I define academic

years as July 1 of year t to June 30 of year t + 1. For example, the 2005 academic year runs from

July 1, 2005 to June 30, 2006. On average, there are about 75 layoff events each year, with 24

being mass layoffs, 50 being plant closures, and 1.4 being correctional facility closures. The total

13I drop 19 layoff events (1.35% of the sample) that do not provide sufficient geographic infor-

mation to assign to a county.

43

number of layoff events spiked to 193 during the 2008 academic year when the Great Recession

and automotive industry collapse hit Michigan especially hard. Panel B shows that the total number

of job losses also spiked during 2008. These layoffs occur throughout the state, in both rural and

urban areas, which I highlight in Figure B.1.3 by plotting the average amount of per capita layoffs

that occur in each county from 2001 to 2017.

The layoff data do not contain information on the occupations of laid-off workers. Therefore, I

estimate students’ exposure to job losses in each community college occupation group by exploit-

ing the fact that different occupations are concentrated in different industries. I first match all 1,024

entities that experience a layoff to their respective three-digit NAICS industry code using informa-

tion from company websites and online business databases. There are 99 unique three-digit codes

in the NAICS system, each of which represents a subsector of economic activity. I observe 72 of

the 99 subsectors in the layoff data, with the three most common subsectors being transportation

equipment manufacturers (21% of observations); general merchandise stores (6% of observations);

and professional, scientific, and technical services (5% of observations).

I then calculate the distribution of community college occupations across industries. Explic-

itly, let g denote one of the six program/occupation groups outlined in Appendix Table B.1.2 (for

example, health or business) and k denote a three-digit NAICS industry (for example, hospitals or

general merchandise stores). The share of industry k’s employment that belongs to occupations in

group g in year t can be calculated as:

αgkt =

Employmentgkt
Employmentkt

(2.1)

where Employmentgkt is the total employment in occupations in group g in industry k in year t and
Employmentkt is total employment in industry k in year t. For example, if g is the health occupa-
tion group and k is the hospital subsector, then α will capture the share of employment in hospitals

that belongs to health-related occupations that community college graduates can reasonably en-
ter. I calculate αgkt for each year, occupation group, and industry using nationally-representative
data from the BLS’ Occupational Employment Series (OES) for non-government sectors and the

44

American Community Survey (ACS) for government sectors.14 Continuing with the example from

above, I find that, on average, community college health occupations account for 54.4% of employ-

ment in the hospital subsector. In contrast, community college health occupations only account for
only 1% of employment at general merchandise stores.15 As a result, layoffs that occur at hospitals

should affect these occupations, and therefore the benefit of enrolling in community college health
programs, much more than layoffs that occur at general merchandise stores.16

I operationalize this intuition by using the occupation-by-industry employment shares to esti-

mate layoff exposure within a given occupation group, county, and academic year. Specifically, I

estimate the number of layoffs in occupation group g in county c in academic year t as:

Layoffsgct = ∑
k

αgktLayoffskct

(2.2)

where Layoffskct is the number of layoffs in industry k in county c in academic year t, which is
identified in the mass layoff data. These measures take into account both the occupations which

likely experience layoffs and the size of the layoff events occurring in a given county and year.

For example, consider Kalamazoo County during the 2012 academic year. During this year, three

firms reported mass layoffs: Hostess Brands, a food manufacturer (15 layoffs); International Paper,
a paper manufacturer (77 layoffs); and OneWest Bank, a credit intermediary (168 layoffs).17 In
14The BLS only began publishing state-specific estimates in 2012 and cautions that they are
subject to more error than the national-level estimates. Nevertheless, I also construct the α
values using Michigan-specific data and find a strong correlation with my preferred nationally-
representative estimates. Appendix Figure B.1.4 plots the α values for each community college
occupation group using each 2016 national and Michigan data. The figure shows a strong correla-
tion between the two measures, with a Pearson coefficient of 0.95.

15Appendix Table B.1.5 presents the three largest average values of α for each occupation group.
16In Appendix Table B.1.6 I compute the correlation between the α values across the six com-
munity college occupation groups. Most correlations are negative, indicating that different com-
munity college occupations are concentrated in different industries and, therefore, will be affected
by different layoff events. Only two correlations are positive: business and STEM occupations,
and health and other occupations.

17Note that the Hostess Brands layoff is below the 50 job loss threshold for required WARN
reporting. Firms sometimes voluntarily report smaller layoffs, particularly when they are reporting
simultaneous layoffs at facilities across the state. In Section 2.6, I repeat the empirical specifica-
tions only using layoffs that meet the 50 job loss threshold and obtain very similar results to the
main specification.

45

this same year, community college business occupations, i.e., business occupations which com-

munity college graduates can enter, accounted for 6.7% of employment in food manufacturing,

10.9% of employment in paper manufacturing, and 44.5% of employment in credit intermedi-

aries nationally. As such, a reasonable estimate of the number of business occupation layoffs

reported under the WARN system in Kalamazoo County during the 2012-2013 academic year is
0.067(15) + 0.109(77) + 0.445(168) ≈ 84.18

2.4.2 Distribution of Layoffs Across Occupations

Table B.1.7 provides summary statistics on the layoffs occurring in Michigan counties between

the 2001 and 2017 academic years.

In addition to estimating the number of layoffs occurring

in community college occupations, I use equations (1) and (2) to generate the number of layoffs

occurring in low-skilled occupations that require less than an associate’s degree and the number of

layoffs occurring in high-skilled occupations that require more than an associate’s degree. These

layoff measures correspond to the types of occupations students would expect to enter if they did

not pursue any postsecondary education or if they obtained four-year college degrees.

Panel A presents summary statistics on the number of layoffs occurring per 10,000 working-age
residents in a given county, year, and occupation group.19 On average, a county-year observation

with 10,000 working-age residents experiences 5.3 layoffs in low-skilled occupations, 4.1 lay-

offs in middle-skill community college occupations, and 1.3 layoffs in high-skilled occupations.

Among the community college occupations, 2.1 layoffs occur in the skilled trades, 1.0 occurs in

business, 0.5 occur in law enforcement, 0.3 occur in STEM, 0.2 occur in health, and 0.1 occur

in other community college occupations. There is substantial variation in the number of layoffs

18To illustrate more examples of county layoffs, Appendix Table B.1.8 provides information on
the three county-year pairs with the largest amount of per capita layoffs in each occupation group
from 2001 to 2017.

19I define working-age residents as those aged 20 to 64 and obtain annual county-level estimates
of this population from the Census Bureau’s Population Estimates Program (https://www.census.
gov/programs-surveys/popest.html). The average county-year observation in the data has 71,131
working-age residents.

46

occurring in different occupations, with the standard deviations for each category far exceeding the

means. For example, the number of skilled trade layoffs occurring in a county ranges from 0 to

nearly 96 per 10,000 working-age residents. Panel B then calculates the share of layoffs occurring

in each category for county-year observations that experience non-zero layoffs. Across the time

frame, 369 county-year observations (26%) experience layoffs. On average, 51% layoffs are in

low-skilled occupations, while about 37% occur in middle-skill occupations, and 11% occur in

high-skilled occupations.

Figure B.1.5 further highlights the variation in layoffs across counties by plotting the layoffs

that occur in each occupation group in each county between 2001 and 2017. I do not include

counties that do not experience layoffs over this time frame and order all other counties by their

average working-age population over this time frame. The left-hand panel plots the total number

of layoffs per 10,000 working-age residents in each occupation group while the right-hand panel

shows the share of layoffs occurring in each occupation group. The total number of layoffs varies

substantially across counties, with both small and large counties experiencing a high number of

local labor market shocks over the time frame. For example, the two counties that experience

the most per capita layoffs are Ingham County, which is home to the state capitol of Lansing and

has about 200,000 residents, and Ontonagon County, which only has 4,000 residents. The share

of layoffs occurring in each occupation group also varies considerably across counties, further

emphasizing the importance of separating layoffs by the types of jobs they affect.

2.4.3 Potential Measurement Error

Because the layoff data does not contain information on the occupations of laid off workers,

the layoff measures I construct rely on the distribution of occupations across industries. Implicitly,

these measures assume that layoffs in an occupation are proportional to its national employment

shares in industries that experience layoffs. Any deviation of layoffs from these proportions could

lead to measurement error in the layoff terms, whereby I inaccurately classify layoffs as affecting

one occupation group when, in reality, they affect another. For example, suppose that a hospital

47

reports a mass layoff of 100 workers. Based on industry-by-occupation shares, I estimate that

about 55 layoffs should affect community college health occupations, while only about 8 should

affect community college business occupations. However, suppose that a hospital was to layoff

only their billing or financial services department. This type of layoff would disproportionately

affect business occupations rather than health occupations, causing me to overstate the effect of

the event on health occupations and understate the effect on business occupations.

More formally, suppose that a single layoff in occupation X occurs. Further, suppose that with

probability ε, I will incorrectly classify this layoff as affecting occupation Y . Then, the estimated

effect of the layoff on the probability that a student chooses program X will be:

(cid:98)δXX = (1− ε)δXX + εδXY

where δXX is the true effect of layoffs in occupation group X on enrollment in group X programs
and δY X is the true effect of layoffs in occupation group Y on enrollment in group X programs.
Because δXX ≤ 0 (layoffs deter students from entering related programs) and δXY ≥ 0 (students
substitute into other programs), the estimated response will be of a smaller magnitude than the true

response and could even be positive if either ε or δXY is sufficiently large. Correspondingly, the
estimated effect of the layoff on the probability that a student chooses program Y will then be:

(cid:98)δY X = (1− ε)δY X + εδYY

where δYY is the true effect of layoffs in occupation group Y on enrollment in group Y programs
and δY X is the true effect of layoffs in occupation group Y on enrollment in group X programs.
Because δY X ≥ 0 and δYY ≤ 0, the estimated term will be biased downward toward zero and could
be negative if either ε or δYY are sufficiently large.

Given the non-classical nature of this measurement error and the fact that ε is unknown, there

is no straightforward way to empirically correct for it. However, there are circumstances where

measurement error is less likely to occur. Specifically, plant and prison closures are likely to

affect all jobs contained within a given facility and, therefore, should align more closely with the

industry-by-occupation employment shares than layoffs that only affect a subset of jobs. In Section

48

2.5.3, I conduct the empirical analysis using only layoffs that are a result of facility closures and

find quite similar results to my main specification.

2.5 Effect of Job Losses on Enrollment in Related Programs

The conceptual framework in Section 2.2 outlines two key outcomes of interest for the empir-

ical analysis: (1) the effect of local job losses on enrollment in related community college pro-

grams, and (2) the corresponding substitution into other postsecondary options (including direct
labor market entry) if students are indeed deterred from entering related programs.20 I begin by

estimating the average effect of job losses on enrollment in related community college programs.

Then, in Section 2.6, I consider heterogeneous effects across occupation groups and document how

students substitute between postsecondary programs in response to job losses.

2.5.1 Empirical Approach

I create measures of program enrollment at the county-year-program level and estimate speci-

fications of the following form:

Enrollgct = α + Layoffsgctβ + XctΓ + θgc + δgt + εgct

(2.3)

where Enrollgct is the number of students from county c and cohort t who enroll in community
college programs in group g, per 100 high school graduates, and Layoffsgct is a vector of the
number of layoffs per 10,000 working-age residents in occupation group g that may affect cohort t

in county c. I consider two sources of variation in layoffs: timing and location. That is, I consider

how students respond to layoffs that occur in different points during their pre-college years and
that occur in different geographic areas. The vector Xct contains time-varying county control
variables that may affect students’ enrollment choices, such as the average test scores of the cohort

or the unemployment rate. θgc is a program-by-county fixed effect that accounts for unobserved
differences in program preferences across counties. δgt is a program-by-cohort fixed effect that
20In Appendix B.3, I further consider how related educational outcomes, such as delayed enroll-

ment or program retention, respond to layoffs.

49

captures unobserved changes in program preferences over time. Finally, εgct is an idiosyncratic

error term. Throughout the analysis, I cluster all standard errors at the county level.

The fixed effects capture two important sources of unobserved heterogeneity: differences in

preferences for community college programs across counties and across time. The vector of con-

trols further accounts for changes in economic conditions across counties and time. As such, the

identifying assumption for β to represent the causal effect of job losses on students’ choices is

that there are no unobserved changes in preferences at the county-program level that are correlated

with job losses. This assumption rules out the possibility that, for example, firms lay off workers

because they know the next cohort of high school graduates has different preferences for college

education than the last cohort. While such a phenomenon seems unlikely, the assumption could

be threatened if there are county-specific trends in occupation-specific job prospects and program

preferences. Thus, I also estimate specifications that include county-by-program linear time trends.

Similarly, layoffs may not represent true changes in occupation-specific employment conditions if

job losses are absorbed by increased employment in nearby counties. For this reason, I estimate

specifications that interact the cohort-by-program fixed effects with commuting zone (CZ) fixed

effects to account for any unobservable changes in an occupation group’s employment in a broader
geographic region.21

2.5.2 Main Results

Table B.1.9 presents estimates of equation (2.3), measuring layoffs at different times during

a cohort’s academic career. Column (1) includes only layoffs occurring during a cohort’s senior

year of high school: the time period during which students must decide what educational program,

if any, they will enter following graduation. The point estimate is negative and statistically sig-

nificant, indicating that an additional layoff per 10,000 county residents during this year reduces

enrollment in related programs by 0.012 students per 100 graduates, or about 0.012pp. There are

21Commuting zones are groups of counties that reflect a local labor market (see: https://www.
ers.usda.gov/data-products/commuting-zones-and-labor-market-areas/). Throughout the analysis,
I use the 1990 CZ delineations.

50

several ways to interpret this estimate. At the mean enrollment rate of 1.5%, this estimate repre-

sents a 0.8% decrease in enrollment in related programs. Correspondingly, a one standard deviation

increase in layoff exposure reduces enrollment in related programs by 3.83% of the mean. Alterna-

tive, doubling the amount of per capita layoff exposure the average county-cohort pair experiences

reduces enrollment by about 0.6%. These estimates imply that, for the average county, 52 workers
being laid off in a given occupation induces one less student to enroll in a related program.22

Columns (2) and (3) then add measures of layoffs occurring in earlier years. The estimate

on layoffs occurring in a cohort’s senior year of high school remains negative and statistically

significant, but there are little effects of layoffs occurring prior to this year. The largest point

estimate comes from layoffs occurring in students’ sophomore year of high school, but this effect

is about half the size of the effect of layoffs occurring in the senior year of high school and is

not consistently statistically significant. These results indicate that students primarily respond to

layoffs occurring in the year leading up to their postsecondary decision point. Such evidence is

consistent with a growing literature highlighting the importance of salience in decision-making

(Mullainathan, 2002; Genniaoli and Shleifer, 2010), and particularly, the sensitivity of college

major choice to recent events (Xia, 2016; Patterson et al., 2019)

Finally, Column (4) adds a measure of layoffs occurring in the year following a cohort’s high

school graduation. Because I restrict the analysis to students’ first program choices within six

months of high school graduation, including this measure serves as a natural placebo test: these

layoffs have not occurred when students make their postsecondary choices, and thus, should not

affect enrollment in related vocational programs. Indeed, I find that they do not. The point estimate

on this variable is positive, but close to zero statistically insignificant. Meanwhile, the estimate

on layoffs occurring during a cohort’s senior year of high school remains negative, statistically

significant, and close to the -0.012.

22I obtain this estimate by re-estimating equation (2.3) with the dependent variable scaled by
the average number of graduates in a county and the independent variable scaled by the average
number of working-age residents in a county. The β parameter then represents the effect of an
additional layoff on enrollment in the average county. Thus, 1/β provides the number of layoffs
needed to induce one student not to enroll in the average county.

51

Next, I consider how layoffs in other areas of the state affect students’ program enrollment

decisions. To do so, I estimate equation (2.3) without including the occupation group by cohort

fixed effects (δgt), as this term absorbs any statewide changes in student preferences for a program,

including the effects of statewide layoffs. Table B.1.10 presents these results. Column (1) includes

only layoffs occurring during a cohort’s senior year of high school within their own county. This

specification produces a very similar estimate to the main specification in Table B.1.9, despite the

lack of a program-by-year fixed effect. Column (2) then adds a measure of layoffs occurring in

the rest of the state. The coefficient on this measure is close to zero and statistically insignifi-

cant, indicating that, on average, layoffs occurring elsewhere in the state do not affect students’

program choices. Column (3) then separates this measure into layoffs occurring elsewhere in the

county’s commuting zone and layoffs occurring outside of the commuting zone. The coefficient

on layoffs occurring elsewhere in the commuting zone is negative, indicating that students also

respond to layoffs occurring outside of their county but in their general area of the state. However,

the coefficient is smaller than the coefficient on county layoffs and is not statistically significant,

again indicating that saliency plays a role in students’ decision-making process and that students

primarily respond to layoffs that occur in their immediate local area.

2.5.3 Robustness

Figure B.1.6 presents several robustness checks of the main specification from column (1)

in Table B.1.9: the effect of layoffs in a student’s county during her senior year of high school

on enrollment in related programs. First, Panel A shows how the results change when including
different control variables in the Xct vector. Including the number of layoffs occurring in low-
skill and high-skill occupations, either together or separately, does not meaningfully change the

estimated coefficient. Similarly, replacing the vector of covariates with a county-by-cohort fixed

effect produces a nearly identical estimate. Next, Panel B estimates specifications with county-by-

52

program linear time trends and program-by-year-by-CZ fixed effects.23 These specifications also

similar estimates to the main specification, indicating that unobserved changes in local economic

conditions are not driving the results.

Panel C then shows how the estimates change when dropping events that are the result of

mass layoffs rather than plant closings, or events that report less than the required 50 job losses.

The estimates are similar when using all layoffs and when using only layoffs that are a result of

closings. Moreover, the point estimate using only closings is slightly larger in magnitude, which

is consistent with the expected effects of measurement error outlined in Section 2.4.4. I also find

quite similar estimates when only including layoffs that reach the 50 job loss threshold, indicating

that the voluntary reporting of smaller layoff events does not contaminate the main results. Finally,

Panel D estimates non-linear specifications that can better handle fractional dependent variables.

First, I estimate equation (2.3) using the inverse hyperbolic sine of a county’s program enrollment
as the dependent variable.24 I then estimate Poisson and fractional logit (Papke and Wooldridge,
1996) specifications.25,26 All specifications produce similar semi-elasticities to the main linear

specification, providing evidence that functional form selection is not driving the results.

23In all specifications that include year-by-CZ fixed effects, Monroe County is dropped from the

analysis because all other counties in its commuting zone are in Ohio.

24The inverse hyperbolic sine (IHS) function approximates the log function but allows values
of zero (Burbidge et al., 1988). I use the transformations proposed by Bellemare and Wichman
(2019) to estimate elasticities at the mean values of the dependent and independent variables.

25In the Poisson specification, the dependent variable remains the share of students from a given
county and cohort who enroll in a given program (rather than a raw count variable). This speci-
fication may be interpreted the same as estimating a linear model with the dependent variable as
log program enrollment and controlling for log total vocational enrollment and restricting the co-
efficient to be equal to 1. However, like the IHS specification, the Poisson approach allows for the
inclusion of dependent variables equal to zero. See Lindo et al. (2018) for more details.

26The fractional logit specification is analogous to estimating a standard logit demand specifica-
tion where the dependent variable is the log of the enrollment share, but allows for the inclusion of
county-program-years where no students enroll in a given program.

53

2.6 Substitution Effects

The results in Section 2.5 indicate that fewer students enroll in community college programs

when exposed to related job losses. This response primarily occurs when the job losses take place

in a student’s own county during her senior year of high school. In order to better understand how

this response may affect students’ longer-run outcomes, I now estimate how these job losses affect

students’ decisions to enroll in other postsecondary options.

2.6.1 Substitution out of Vocational Programs

I begin by estimating how layoffs in community college occupations affect students’ decisions

to enroll in vocational community college programs overall. To do so, I estimate the following

equation:

VocationalEnrollct = α +

6
∑
g=1

βgLayoffsgc,t−1 + XctΓ + θc + δt + εct

(2.4)

where VocatonalEnrollct is the number of students from county c and cohort t, per 100 graduates,

who enroll in vocational community college programs at community colleges. The vector of layoff
variables, Layoffsgc,t−1, captures the number of layoffs, per 10,000 working-age residents, that
occur in different community college occupation group g in county c during cohort t’s senior
year of high school. As in equation (2.3), the vector Xct contains time-varying county control
variables that may affect students’ choices, including the number of layoffs that occur in non

community college occupations. θc is a county fixed effect that absorbs county-specific preferences
for different types of postsecondary education (as θgc does in the previous estimating equation)
and δt is a cohort fixed effect that accounts for changing preferences over time (as δgt does in the
previous estimating equation). εct is the error term. I continue to cluster all standard errors at the

county level.

The β vector identifies how layoffs in different types of occupations affect students’ decisions

to enroll in related types of college programs. The identifying assumption is that, after controlling

for secular trends through the cohort fixed effects, any within-county variation in layoffs is uncor-

54

related with within-county variation in unobserved college preferences. As in Section 2.5, this as-

sumption seems reasonable, but could be threatened if there are unobserved changes in preferences

or economic opportunities over time. Therefore, I also estimate specifications with county-specific

linear time trends or cohort dummies interacted with commuting zone fixed effects.

Table B.1.11 presents the estimates of equation (2.4). Column (1) is the baseline specification,

column (2) includes county-specific linear time trends, and column (3) includes cohort-by-CZ

fixed effects. Across the three columns, the effects of layoffs are small and none are statistically
significant at the 5% level.27 Moreover, in all specifications, I fail to reject the joint hypothesis

that all six coefficients are equal to zero, indicating that layoffs in community college occupations

do not affect enrollment in vocational programs.

In Table B.1.13, I further consider whether layoffs in community college occupations affect

the composition of students enrolling in vocational programs by regressing mean demographic

values of vocational students against the vector of layoff measures.

I find little evidence that

layoffs affect who enrolls in vocational programs, and, in all specifications, I fail to reject the joint

hypothesis that the coefficients on all community college layoff terms are equal to zero. Similarly,

in Table B.1.14, I estimate how layoffs in community college occupations affect credit completion

within vocational students’ first year of community college enrollment. I find no evidence that
layoffs affect total credit completion, nor completion of vocational vs. non-vocational courses.28

Taken together, these findings show that layoffs in community college occupations do not dissuade

27In Table B.1.12, I show that, overall, layoffs increase college enrollment. This finding is con-
sistent with prior work that shows college enrollment increases when local economic conditions
worsen. I further show that this increase in college enrollment is concentrated in programs that
should lead to four-year college degrees, including non-vocational programs at community col-
leges, while layoffs slightly decrease enrollment in community college vocational programs. This
finding is slightly different from Hubbard (2018), who also uses Michigan data and finds that
layoffs predominantly increase enrollment in community colleges. However, he uses an earlier
sample (2002-2011 academic years) and measures layoffs within a 30-mile radius of a student’s
high school rather than at the county level, which could explain the differences in our results.

28I use course codes and information from community college catalogs to divide all courses into
vocational and non-vocational groups. I define vocational courses as those in the same fields that
are included in the six vocational program groups of interest, while all other courses are considered
non-vocational.

55

students from enrolling in community colleges and pursuing vocational education programs, nor

do they change students’ intensity of enrollment. Thus, the response documented in Section 2.5

must come from students changing which types of vocational programs they pursue.

2.6.2 Substitution Between Vocational Programs

Because job losses do not deter students from entering vocational community college programs

overall, I now consider how students substitute between vocational programs in response to layoffs.

I restrict the sample to students who enroll in vocational programs and estimate the following

system of six equations:

ProgramEnroll jct = α +

6
∑
g=1

βgLayoffsgc,t−1 + XctΓ + θc + δt + εct

(2.5)

where ProgramEnroll jct is enrollment in occupation group j among students from county c and
cohort t, per 100 students enrolling in vocational programs, and Layoffsgc,t−1 is the number of
layoffs in occupation group j in county c occuring in school year t − 1, per 10,000 working-age
residents in the county.29 The vector Xct contains the same variables as in equation (2.4), θc is a
county fixed effect, δt is a cohort fixed effect, and εct is the error term. I again cluster all standard

errors at the county level.

The coefficient βg will represent the “own-layoff” effect when j = g and will represent a “cross-
layoff” effect when j (cid:54)= g. As predicted in Section 2.2, the own-layoff terms should be negative
because layoffs should deter students from enrolling in related programs. The cross-layoff terms

should be positive since students would then substitute between programs, but could be negative

if there is some measurement error. Moreover, because the dependent variable shares must sum

to 100, the sum of a βg term across the six enrollment outcomes must equal 0. This restriction

implies that any decrease in enrollment in a given program group due to related layoffs must be

offset by students enrolling in other vocational community college programs.

29Because the same regressors appear in every equation and there are no cross-equation re-
strictions, estimating each equation separately is algebraically equivalent to jointly estimating the
system using feasible generalized least squares (Wooldridge, 2010).

56

The identifying assumption for the β j terms to represent causal effects of layoffs on students’
choices is that, conditional on all other variables, layoffs in occupation group j must be uncor-

related with unobservable determinants of enrollment in program group g. When j = g, this as-

sumption imposes that occupation-specific layoffs are not correlated with changing preferences for
corresponding programs within a county. When j (cid:54)= g, the assumption is that occupation-specific
layoffs are not correlated with changing preferences for other programs within a county. As in

the previous sections, unobserved changes in preferences or economic opportunities could violate

this assumption, so I again estimate specifications with county-specific linear time trends or cohort

dummies interacted with commuting zone fixed effects.

Table B.1.15 presents the substitution matrix created from estimating equation (2.5) for each
of the six occupation groups.30 The bold diagonal terms represent the effect of an additional

layoff per 10,000 county residents in occupation group g on enrollment in related programs. For

example, an additional layoff per 10,000 county residents in business programs reduces enrollment

in business programs by 1.02 students per 100 enrollees, or by 1.02pp. An analogous increase

in layoffs reduce enrollment in health programs by 0.61pp and in law enforcement programs by

0.15pp, in other programs by 0.81pp, and by smaller but negative amounts in the skilled trades and

STEM. In the bottom panel of the table, I present the own-layoff elasticities at the mean values

of both the dependent and independent variables. An additional layoff per 10,000 working-age

county residents reduces enrollment in related programs by between 0.6% and 4.7%, with the

largest statistically significant effects coming from the business and health fields.

Moving horizontally across the columns shows how layoffs induce students to substitute into

other types of vocational programs. For example, an additional business layoff per 10,000 county

residents increases enrollment in law enforcement programs by about 1.7pp. This coefficient shows

that business layoffs induce students to substitute away from business programs and towards law

enforcement programs. Similarly, students primarily substitute from health programs into other

30The sample consists of 657 (98.9%) county-cohort pairs where at least one student enrolls in
vocational programs. Restricting the sample to counties that have non-zero vocational enrollment
in every year of the data produces nearly identical results.

57

programs when there are health layoffs. In Table B.1.16, I further disaggregate the “other” category

and find that most of the substitution occurs in social service programs, such as childcare, although

there is also statistically significant substitution into arts and media programs and personal care

and culinary programs. Although not statistically significant, the estimates further suggest that

students substitute from law enforcement programs towards business, STEM, and health programs

when there are law enforcement layoffs.

2.6.3 Explaining Substitution with Occupation Characteristics

While it is interesting to document that health layoffs induce students to substitute towards pro-

grams in the “other” category, this finding raises yet another question: why do students substitute

towards these fields? Based on the conceptual framework presented in Section 2.2, students should

substitute into their “next best” alternative program. Given that programs are closely tied to oc-

cupations, the next best programs are likely to share similar occupation characteristics. For exam-

ple, health programs and several programs in the other category —such as childcare professionals

—focus on serving one’s community and require a high level of person-to-person interaction, so it

seems reasonable that students would substitute between these programs.

To empirically assess the extent to which students substitute into similar programs, I use data

on occupation characteristics from the U.S. Department of Labor’s Occupational Information Net-

work (O*NET), which contains a wealth of information on worker and job characteristics, includ-

ing the skills required in different occupations. I characterize community college program groups

using measures of three dimensions of skill requirements for related occupations: cognitive skills,

social skills, and technical skills. The cognitive skill category contains ten measures of skills “that

facilitate learning or the more rapid acquisition of knowledge,” such as mathematics, reading com-

prehension, and writing. The social skills category contains six measures of skills that are “used

to work with people to achieve goals,” such as negotiation and service orientation. The techni-

cal skills category contains eleven measures of skills “used to design, set-up, operate, and correct

malfunctions involving application of machines or technological systems,” such as repairing and

58

programming. For each occupation and skill measure, O*NET reports a standardized importance

score and standardized level score. Both measures range from 0 to 100, but each provides different

information. The importance score describes how important a particular skill is to an occupation,

with higher values indicating more importance. The level score characterizes the degree to which

the skill is required to perform the occupation, with higher values indicating a higher requirement.

I use these data elements to create a Euclidean distance measure that identifies program groups

that require similar skills. The measure is similar to that used by O*NET to identify similar careers

but, to my knowledge, has not previously been used to identify similar college programs. I define

the distance between program group p and program group s, which experiences the labor market

shock, as:

Distanceps =

(cid:114)
∑27

j=1 Importance js(Level jp − Level js)2

(2.6)

where Importance js is the importance of skill j for program group s, Level jp is the required level
of skill j for program p, and Level js is the required level of skill j for program group s. As a result,
the programs that are most similar to program group s in terms of the skills that are most important
for careers in group s will have the lowest distance measures.31 I standardize the measures such

that the least similar pair of program groups has a distance measure of 1.

Figure B.1.7 plots the coefficients in Table B.1.15 against this skill distance measure. Each

panel shows the effect of a different type of layoff on enrollment in each program group. For

example, the upper left panel shows that business layoffs decrease enrollment in business programs

but increase enrollment in law enforcement programs, which is the most similar program group to

business. A similar pattern emerges in the second panel, where health layoffs decrease enrollment

in health programs but increase enrollment in law enforcement and other programs, both of which

are fairly similar to health. Layoffs in law enforcement and other community college occupations

also induce students to enroll in similar programs. However, when there are layoffs in STEM

31To create level and importance measures for program groups, I create a weighted average of
all occupations that belong to the group where weights are proportional to the total enrollment of
Michigan students over the time frame of the data. For example, nursing receives a high weight in
the health program group because it is one of the most popular programs.

59

and skilled trades, students are not substantially deterred from enrolling in related programs. This

lack of a response may be due to the lack of nearby substitutes in which students could enroll. For

example, all of the non-STEM program groups have a distance measure of 0.5 or greater, indicating

that they require quite different skills than STEM occupations do. This difference is not surprising

as STEM occupations tend to require much more mathematical skills than non-STEM occupations.

Figure B.1.8 provides further evidence that students substitute into similar programs by pooling

all of the substitution effects and plotting them against their respective skill distance measures.

The largest substitution effects appear at the left end of the x-axis, indicating that students mostly

substitute into programs that are similar to those affected by layoffs. Moving across the x-axis,

there is a downward slope showing that students are less likely to enroll in programs that require

substantially different skills. A simple linear fit of the data indicates that moving from the most

similar to the most different program group reduces the substitution effect by 0.55, where I measure

effect sizes as the impact of an additional layoff per 10,000 county residents on enrollment per 100
vocational students.32 In Appendix B.4, I consider substitution patterns between more narrowly-

defined program groups and find that the general pattern of students substituting towards similar

programs still holds.

2.6.4 Heterogeneity & Robustness

Figure B.1.10 considers heterogeneous responses to layoffs by re-estimating the system of

equations in equation (2.5) using different subgroups of students. First, in Panel A, I consider how

the effects vary across genders. Because there is substantial sorting across genders in community

college programs, it is reasonable to think that male and female students may respond differently

to layoffs in various fields. Indeed, I find that the responses to health layoffs are predominantly

driven by female students, who account for nearly 80% of enrollment in health programs. The

32In Figure B.1.9, I re-create the figure using alternate measures of skill distance. The results
are quite similar, with an additional layoff per 10,000 county residents reducing the effect size by
0.73 when using only differences in skill levels and by 0.62 when using only differences in skill
importance.

60

responses to business, skilled trades, STEM, and law enforcement layoffs tend to come from male

students, who make up the majority of enrollment in these programs. However, the estimates for

these fields are noisier and are not significantly different between male and female students.

In Panel B, I show how the effects vary across urban and rural counties.33 This type of het-

erogeneity is particularly relevant in Michigan because a majority of the state’s residents reside in

urban areas, but those areas comprise little of the state’s land area. Moreover, there are documented

differences in racial composition, political leanings, and educational attainment across rural and

urban areas in the state (Citizens Research Council of Michigan, 2018). I find that the responses

to layoffs are predominantly driven by rural counties, except for law enforcement layoffs, which

mostly affect urban counties. This strong response could be the result of geographic preferences of

students’ in rural areas to remain in their local communities or differences in information networks

in these areas. For example, rural news outlets may have fewer events to cover and, therefore,

may devote more attention to a local layoff or business closure. Layoffs in rural areas may also

be better indicators of future labor market prospects than layoffs in urban areas, particularly if an

occupation’s employment is heavily concentrated in one firm that then closes or downsizes.

I next perform a series of robustness checks that test the sensitivity of the results to alternative

specifications. First, because scaling the dependent variable by the number of vocational students

in a given county and cohort may introduce heteroskedasticity, I estimate the substitution matrix

using the refined weighting schemed proposed by Solon et al. (2015). Panel A of Figure B.1.11

presents the own-layoff effects using this approach. The point estimates and corresponding stan-

dard errors are quite similar with or without weights. Second, because layoffs may be more likely

to occur when a county is on a downward economic trajectory, Panel B of Figure B.1.11 shows how

the estimates change when including county-specific linear time trends. The results are also quite

similar with and without trends. I also estimate specifications that include cohort-by-commuting

zone fixed effects to capture changing economic conditions or program preferences that are unique

33I define urban counties as those that the U.S. Census Bureau classifies as “mostly urban” and
define all other counties as rural. A list of Michigan’s urban and rural counties is available here:
https://www.mlive.com/news/2016/12/michigans urban rural divide o.html.

61

to geographic regions within the state. Panel C shows how the results change when including this

additional set of fixed effects. Again, the estimates are quite similar to the main specification.

Panel D then shows how the results change when dropping the 2009 cohort, who graduated

during the height of the Great Recession in Michigan and may have faced additional challenges in

both accessing higher education and entering the labor market. The estimates are somewhat noisier

when I do not include this cohort, but the effect sizes remain similar. Panel E further shows how

the estimates change when I drop any student who enrolls in more than one program group from

the analysis. The results are nearly identical when restricting the sample in this way.

Finally, because the dependent variable represents county-level enrollment shares, I estimate

several alternative specifications that are designed to handle fractional data. As in Section 2.5.3, I

first estimate inverse hyperbolic sine, Poisson, and fractional logit specifications. I then implement

a fractional multinomial logit specification that jointly estimates all coefficients in Table B.1.15,

while imposing that each enrollment outcome must fall between 0 and 100, and the shares must

sum to 100 (Buis, 2017).

In Panel F of Figure B.1.11, I compare the results from these three

specifications to the estimated elasticities obtained from the main linear specification. The semi-

elasticities are quite similar across the specifications, with an additional layoff per 10,000 working-

age residents reducing enrollment in related programs by up to 5% and effects varying across fields

of study.

2.7 Conclusion

More than 8 million students enroll in public community colleges in the United States each

year, with many entering vocational programs that prepare them for a continually evolving labor

market. The returns to these programs vary substantially by field of study, but there is little ev-

idence on how students choose which programs to pursue.

In this paper, I study the extent to

which students’ program choices respond to changes in local labor market conditions in related

occupations. To do so, I match detailed administrative data on students’ educational decisions

with establishment-level data on plant closings and mass layoffs in the state of Michigan. While

62

previous researchers have used similar data to study how local economic conditions affect col-

lege enrollment, I provide the first analysis in the literature that matches layoffs to corresponding

academic programs and considers how they affect what students study once they enroll in college.

I find that local labor market shocks deter students from entering related programs at commu-

nity colleges. Instead, students shift their enrollment into other types of vocationally-oriented com-

munity college programs. Using rich data on occupation characteristics, I document that students

primarily substitute into programs that lead to occupations that require similar skills. However,

when layoffs occur in fields that do not have clear substitutes, such as STEM occupations and the

skilled trades, students are less likely to shift their enrollment to alternative programs.

These results have several policy implications for Michigan’s community colleges and national

education policy efforts. For example, colleges should prepare for students to enter different pro-

grams when there are local labor market shocks. Providing community colleges with the resources

to expand the supply of alternative programs, particularly those with high labor market returns,

could be beneficial to students. High schools and colleges should also carefully consider the type

of labor market information they provide students. I find that students are particularly sensitive to

local labor market conditions. However, it is not clear whether this responsiveness is a result of

the salience of local events or geographic preferences. Ideally, educators would urge students to

consider both local and non-local labor market opportunities to make informed choices that best

align with their geographic preferences and constraints.

Nevertheless, these results also have limitations. First, the majority of local labor market shocks

I observe come during the aftermath of the Great Recession in a state that was particularly affected

by the collapse of the automotive industry. While this setting produces substantial variation in

local labor market conditions, the results may not generalize to future cohorts or other areas of the

country. Additional work analyzing how students respond to local labor market shocks in other

contexts would be a valuable contribution to the literature. Second, my results are limited in that

they apply only to the decisions of recent high school graduates. Adults enrolling in commu-

nity college programs, especially those who lose their jobs during local labor market downturns,

63

may have different preferences for program characteristics and may respond quite differently to

local labor market shocks than younger students who are enrolling in college for the first time.

Understanding the choices of this population and evaluating interventions meant to promote their

employment and earnings are important areas of both future research and public policy.

64

CHAPTER 3

DO HEALTH INSURANCE MANDATES SPILLOVER TO EDUCATION? EVIDENCE

FROM MICHIGAN’S AUTISM INSURANCE MANDATE

3.1 Introduction

How policy decisions in one area spill over to other areas in which there are no direct policy

connections is a core question in economics. These spillovers often are unintended by policy-

makers, but they can have large impacts on how individuals respond to policy changes and the

resulting social welfare effects of those policies. The opportunity for these unintended spillovers

is particularly large in the United States, where a large array of different government organizations

at the federal, state and local levels enact separate policies that interact with one another in com-

plex ways. These interactions mostly have been studied with respect to the social safety net in the
US.1 Little research has addressed spillovers into education, particularly with respect to the health
care system.2 This lack of research is surprising, since education and health together accounted

for 25.2% of GDP in 2017. Health and education are strongly linked through their central role

in the development of human capital, and there also are direct policy linkages through the special

education system that services students with disabilities.

In this paper, we provide one of the first analyses of how health care policies spill over to

the education sector by examining the effect of Michigan’s autism insurance mandate on the ed-

ucational services received by, and achievement outcomes of, students diagnosed with Autism

1See for example Elwell (2018), Ham and Shore-Sheppard (2005), Yelowitz (1995), Moffitt
and Wolfe (1992), and Blank (1989). These studies all find evidence that changing one program
affects participation in other programs.

2Recent work by Benson (2018) estimates the effect of special education participation on Sup-
plemental Security Income receipt and shows strong evidence of interactions among these pro-
grams. There also is some research that shows how direct health interventions in public schools
affects student health and educational achievement, but these studies do not identify spillover ef-
fects across programs or policy areas (e.g., Lovenheim et al., 2016; Reback and Cox, 2018; Buckles
and Hungerman, 2018).

65

Spectrum Disorder (ASD). While our analysis contains broad lessons for how health care policies

affect educational services and outcomes, the specific focus on students with ASD also is of high

importance.

ASD is one of the fastest-growing developmental disabilities in the United States. The ASD

diagnosis rate among eight-year-olds increased from 6.7 per 1,000 students in 2000 to 16.8 per
1,000 in 2014.3 Among students 3-21 years old, the rate of special education primary disability

identifications with ASD rose from 0.2% in 2000 to 1.2% in 2015 (a 500% increase). The overall
student disability rate declined slightly over this time period, from 13.3% to 13.2%.4 A recent

study using self-reports from 2016 found that among children aged 3 to 17, 2.8% were diagnosed

with ASD at some point (Xu et al., 2019). Students with an ASD diagnosis are growing in absolute

terms (617,000 children in 2015) and are an increasing proportion of all students with disabilities

(9.2% in 2015, up from 1.2% in 2000). Further, students with ASD are some of the most ex-

pensive students to teach. Children with ASD typically have substantial learning disabilities that

require intensive therapy services throughout childhood as well as coordination between the health

care system and the education system. These students cost schools $8,610 more than the average

non-ASD student (Lavelle et al., 2014), while the cost to families varies dramatically by health
insurance coverage but has been estimated to be as high as $47,000 per year in the US.5

Applied Behavioral Analysis (ABA) is the main therapy used to treat students with ASD. It is

not a “cure” but has been shown to substantially improve symptoms through behavioral modifica-

tion therapy (Peters-Scheffer et al., 2011; Dawson et al., 2010; Viru´es-Ortega, 2010; Howlin et al.,

2009; Eldevik et al., 2009; Foxx, 2008). It is most effective when implemented early in life and

when done intensively, often for at least 20 hours per week. Because of the high cost of these ther-

apies, schools and families often lack the resources to provide sufficient services to ASD students.

Until recently ABA and other autism therapies were often excluded from health insurance because

3Source: https://www.cdc.gov/ncbddd/autism/data.html.
4These tabulations are taken from the 2017 Digest of Education Statistics, Table 204.30. https:
5Source: https://www.special-learning.com/article/funding$\ $overview.

//nces.ed.gov/programs/digest/

66

they were considered “experimental” and/or “educational.” Even when a therapy is not excluded,

the coverage is uneven across health insurance plans both in terms of what is covered and the ages

for which therapies are covered.

The lack of coverage for autism services in many private health insurance plans highlights that

what treatments students receive relies on the interaction between school-based services and the

health insurance plan to which a family has access. The interaction of health insurance and school

special education services is not unique to ASD, as these issues are present for all student disabili-
ties.6 Currently, there is very little understanding of how the health insurance and special education

systems interact in the production of education services for students with disabilities. The need for

expensive extra-curricular treatments that are unevenly covered by health insurance plans makes

students diagnosed with ASD an informative group through which to study this interaction.

We provide the first analysis in the literature of how mandating private insurance coverage of

ASD treatments such as ABA affects special education diagnoses (including ASD), the educational

supports students with ASD receive, and the educational outcomes of students diagnosed with

ASD. Beginning in October 2012, Michigan required that all private state-regulated insurance

plans cover ASD treatment services through age 18. Self-insured plans, while not mandated to

cover ASD therapy, were provided generous financial incentives to do so. Medicaid also provided
coverage but only for children under age 6 due to a lack of funding.7 The difference between the

Medicaid and private insurance coverage forms the basis of our empirical strategy.

We use administrative data on all public K-12 students in the State of Michigan from the 2009-

2010 to the 2014-2015 school years. The data are extremely rich and include not only traditional

test scores, demographics, and schools attended but also specific disability diagnoses and the ser-

vices students receive in school through their Individual Education Plans (IEPs). The data do not

6Given the strong positive correlation between health and education and the central role both
play in the development of human capital, health policy is likely to have an impact on education for
non-disabled students as well. For example, Cohodes et al. (2016) show that Medicaid expansions
led to higher educational attainment among affected cohorts.

7As of January, 2016, Michigan began covering all youth in Medicaid up to age 18. Our analysis

thus ends prior to the Medicaid expansion to focus on the private insurance mandate.

67

include information on the health insurance plan under which each student is covered, so we rely

on the close overlap of Medicaid and free/reduced price lunch (FRPL) eligibility. Between 2008

and 2016, tabulations from the American Community Survey show that only 31% of FRPL-eligible

students had private insurance coverage while 89% of those not eligible for FRPL had private in-
surance.8 We use this overlap to estimate intent-to-treat models that examine how outcomes among
non-economically disadvantaged students9 (who are mostly covered by private insurance) change

when the mandate is enacted in 2012 relative to economically disadvantaged students (who are less

likely to be covered by private insurance). To further increase the accuracy of our proxies for insur-

ance coverage, we restrict our sample to students who are economically disadvantaged in all years

that we observe them in grades 2 through 8 and those never observed during those grades as eco-

nomically disadvantaged, conditional on being observed for at least two years. We use this sample

because students who are persistently eligible for FRPL are the most disadvantaged (Michelmore

and Dynarski 2017).

In a difference-in-differences setting, we first show that the mandate has little effect on the
likelihood of receiving an ASD special education identification in grades 2 through 8.10 This is

interesting in its own right, as private insurance can cover diagnostic services. However, most

medical diagnoses for ASD occur before the age of six (Fountain et al., 2011), which is likely why
we find no effect on ASD identification.11 The lack of any meaningful effect on overall ASD inci-

dence supports a triple difference strategy when we examine education services and achievement.

We estimate how outcomes among ASD students who are not economically disadvantaged change

relative to ASD students who are and how this change relates to the change in outcomes among
8Note that insurance counts may exceed 100% as some people remain eligible for Medicaid

while enrolled in private plans.

9“Economic disadvantage” status refers to students in poverty and the vast majority of students

qualify based off eligibility for free or reduced-price lunch.

10Generally, free/reduced price lunch eligibility is more accurately measured in primary grades

than secondary grades, which is one benefit of focusing on younger students.

11While medical diagnosis and identification of ASD for education purposes are similar, they
are not the same. Some students may be diagnosed but not have an IEP or have a different primary
identification. Alternatively, while it is extremely likely a child with an ASD identification also
has a medical diagnosis, the latter is not a necessary condition for the former.

68

non-ASD students who differ in disadvantage status.

We find that the insurance mandate reduces the set of special education resources ASD students

receive and induces students to be placed in less restrictive environments. Among ASD students

who are never disadvantaged, the mandate causes a 6.4 percentage point (9.5%) reduction in place-

ment in resource room (pull-out) or cognitive impairment programs (self-contained classes for stu-

dents with cognitive impairments) and a 3.4 percentage point (26.4%) increase in placement in no

special education program. Furthermore, these ASD students are 2.3 percentage points (17.7%)

less likely to be assigned an ASD-specialized teacher consultant (who provides oversight and sup-

port to the students’ teachers and help develops instructional plans), though they are slightly more

likely to receive any special education support services. Taken together, these measures indicate

that the private insurance mandate led to lower special education resource provision for affected

students in schools.

Our data do not allow us to observe the use of ASD therapy services outside of school. Such

data would be useful in assessing the costs and benefits of this policy and whether our results

indeed reflect crowd out of special education services rather than students not requiring as many

services in school. To provide some evidence on whether the mandate generates crowd out versus

reducing the need for in-school services, we examine student test scores that yield insight into

the extent to which the mandate supports or detracts from student learning. There could be a

negative effect if service crowd-out is more than 100% or if the privately provided services are

of lower quality. Conversely, student learning may increase if overall services increase and/or if

the quality of services provided increases. Additionally, providing ABA outside of school may

facilitate more inclusion of ASD students in general education classrooms, which some research

suggests is productive for learning among students with disabilities (Ruijs and Peetsma, 2009)

and more time to focus on direct instruction. The inability to observe privately provided services

precludes a direct analysis of these mechanisms, but we are able to identify the net policy effect

that shows how the policy impacts achievement. The achievement results thus provide suggestive

evidence of the mechanisms at work.

69

We find little evidence of a net change in reading or math test scores. In our preferred model

that uses non-disabled students as a comparison group, the 95% confidence interval for math in

standard deviation units is [-0.046,0.056] and in reading it is [-0.039,0.075]. We thus can rule

out anything but modest-to-small changes in math and reading scores due to the mandate. This

finding suggests either that crowd-out was complete (and thus total services did not change) or that

any reduction in services is balanced by the effects of being in a more inclusive general education

environment. Nonetheless, these results suggest there was no adverse effect on students’ academic

performance from the crowd-out.

We also examine heterogeneous effects by gender, race, and grade level. Our results do not vary

much across groups, but we do find that girls are more likely to be removed from resource/cognitive

programs and more likely to be placed in no special education programs than are boys. Effects are

similar for White and Asian versus Black and Hispanic students, but we lack power to estimate

precise effects for the latter group. We also find that the effects only begin to appear in grade 2.

This grade heterogeneity is sensible, as nearly all students in Kindergarten and many in grade 1

are 6 years old or younger and therefore receive increased private ASD services under the man-

date regardless of their health insurance status. In terms of test scores, we find little evidence of

heterogeneous treatment effects.

This paper contributes to several different strands of research. The first is the small literature

that examines the effects of health care policies on student achievement. Most of the prior liter-

ature focuses on Medicaid (Cohodes et al., 2016; Levine and Schanzenbach, 2009) or examines

direct health interventions in schools (Lovenheim et al., 2016; Reback and Cox, 2018; Buckles and

Hungerman, 2018). To our knowledge this is the first paper to examine the interaction between the

health insurance and special education systems. More specifically, our study is the first to identify

causal effects of Autism insurance mandates on the educational services disabled students receive

and their subsequent educational achievement.

The second literature to which we contribute is the crowd-out of public goods by private pro-

vision (Bergstrom et al., 1986). Crowd-out of public services from private provision has been doc-

70

umented in several contexts like Medicaid (Cutler and Gruber, 1996; Gruber and Simon, 2008),

charitable donations (Payne, 1998; Gruber and Hungerman, 2007; Anderoni and Payne, 2011), re-

ligion (Hungerman, 2005), and school funding (Gordon, 2004). Ours is the first analysis to show

that private health insurance mandates crowd out special education services in public schools. This

is an important contribution because special education is by design at the intersection of publicly

provided education and often privately provided health care. That changes to private insurance can

affect the services that disabled students receive in public schools is a novel finding that has impli-

cations for health insurance policies and the funding and provision of special education services.

We also contribute to a growing body of work on policies surrounding ASD students. ASD is

a very expensive disability to treat, with current estimates in the US indicating that it costs about

$17,000 per year to treat a student with ASD through health care and special education services

(Lavelle et al. 2014). There also is suggestive evidence that ASD leads to lower labor force

attachment and earnings among parents (Cidav et al., 2012), although identifying causal estimates

is difficult in this context. The lifetime cost of supporting a child with ASD, including potential

labor force effects among parents, is between $1.4 and $2.4 million in the US (Buescher et al.,

2014). A large literature has arisen that examines the causes of the rise in ASD (see e.g., Hansen

et al., 2015; Matson and Kozlowski, 2011), but to date very little work has been done on what

school or health policies can support the academic development of ASD students and how best to

deliver services to them in a cost-effective manner.

Finally, we present direct estimates of the effect of the Michigan insurance mandate on special

education services and academic outcomes. These mandates are growing in prevalence: 46 states

(plus D.C.) currently have some form of regulation that requires ASD services to be covered by

health insurance plans. However, the scope of what is covered and the ages of children included
in the regulations vary considerably across states.12 As small literature has arisen that examines

these mandates. Mandell et al. (2016), Barry et al. (2017), and Candon et al. (2019) find that

these mandates increase treatment prevalence and spending for those with private insurance: ASD

12Regulations

for

each state

can be

found at http://www.ncsl.org/research/health/

autism-and-insurance-coverage-state-laws.aspx.

71

mandates lead to a $77 increase in monthly spending on ASD services. Chatterji et al. (2015)

use a similar triple difference design as in this study and find that Autism Insurance Mandates

have no effect on financial burden, access to care, and unmet need for services. These results are

suggestive of full crowd-out of services due to the mandate. The Michigan mandate is among the

most expansive in terms of what must be covered and in terms of the ages of children included.

Thus, our analysis is informative with respect to the potential for these types of policies to impact

educational services and outcomes among students with ASD.

3.2 Background

3.2.1 Austism Spectrum Disorder and Therapy Options

Autism spectrum disorder (ASD) is a developmental disability that generates problems with

social, emotional, and communication skills (Centers for Disease Control, 2018). The categoriza-

tion combines disorders that were previously viewed as distinct - Autism, Asperger’s Syndrome,

Pervasive Developmental Disorder (Not Otherwise Specified) - and are now considered by psychi-

atrists to be variations of the same spectrum of conditions. Children with ASD show many different

symptoms and often exhibit some but not others. Common symptoms include difficulty with social

interactions, delayed speech and inability to communicate verbally, repetitive behaviors, and stim-

ming. These start to appear as early as 18 months of age, and diagnoses can be obtained as early as

24 months (Centers for Disease Control, 2018). Even so, diagnosis this early is uncommon. The

median age of first diagnosis in the US as of 2012 was 4.2 years, and only 46% of children with

ASD had a full evaluation prior to 3 years of age (Baio et al., 2018).

ASD is considered a lifelong disorder. While there is no cure, there are treatments that can

help alleviate symptoms and improve the ability of individuals with ASD to perform well behav-

iorally, both in school and in society more broadly. Children with ASD usually receive a variety

of therapeutic interventions. These often include occupational therapy, speech therapy, sensory
integration therapy, and Applied Behavior Analysis (ABA) therapy.13 Students may receive these

13While there is some research on how nutritional changes can help, these studies are largely

72

services from private practitioners, through the special education system, which we discuss further

in Section 3.2.3, or through a combination of private and public providers.

Applied Behavior Analysis (or Early Intensive Behavioral Intervention), which involves using

positive reinforcement and repetitive application of behavioral situations where cause and effect
are outlined, has become one of the most widely used strategies for addressing autism.14 As noted

above, there is substantial experimental and observational research that shows ABA to be effective

at improving educational and behavioral outcomes for children. Further, providing intervention

early on when the child is very young has been shown to be more effective than starting later

(Zwaigenbaum et al. 2015; Dawson et al. 2010; Granpeesheh et al. 2009; Corsello (2005)). ABA

therapy is typically provided by licensed Board Certified Behavior Analysts (BCBAs), many of
whom work outside of the public school system.15 However, many students also receive some

form of ABA therapy in school. For example, in 2011 - prior to the insurance reform - 59% of

public school educators in Michigan reported using ABA therapies with students with ASD (Ferreri

and Bolt, 2011).

Data on the costs of these interventions are sparse, but the therapies are generally considered

to be quite expensive. Total costs of treatment combined with opportunity costs (e.g.

for lost

work by a parent or caretaker) have ranged from $17,000 per year in the US to $44,000 in the UK

and $68,000 in Sweden (J¨abrink (2007); Knapp et al. (2009); Lavelle et al., 2014). Additionally,

estimates of medical expenditures for individuals with ASD indicate they exceed those without

ASD by $4,110 to $6,200 per year, 4 to 6 times larger than average (Shimabukuro, Grosse, and

Rice 2008). Given these large costs, insurance coverage is potentially a very important factor

in whether children receive treatment. Cost-benefit analyses have shown ABA interventions to

be highly cost effective over the long run. Jacobson et al. (1998) find lifetime benefits for the

observational or small sample experiments and show limited evidence of impacts on symptoms
(Mari-Bauset et al., 2014)

14https://www.autismspeaks.org/applied-behavior-analysis-aba-0.
15Less than 30% of job postings for BCBAs come from the education sector, while nearly 60%
come from the health care and social assistance sectors. See: https://www.bacb.com/wp-content/
uploads/2017/09/151009-burning-glass-report.pdf

73

individual of up to $1 million. Ganz (2007) estimates the lifetime social costs of untreated autism

at $3.2 million as of 2003, though it is unclear how much this can be mitigated by therapeutic

treatment.

3.2.2 The Michigan Autism Insurance Mandate

Until recently, treatments for Autism Spectrum Disorder beyond therapies for co-morbidities

like speech and occupational therapy were not commonly covered by health insurance. As a result,

states started mandating coverage for ABA and related therapies. Today, 46 states and the District

of Columbia have some coverage requirements for autism services. Even in these states, however,

coverage can be limited. Affected children often have to go through a time consuming evaluation

process where access may only be available in a few locations with long wait lists. Furthermore,

coverage mandates do not extend to all health insurance plans - since self-insured firms are covered

under Federal law, states have little ability to mandate coverage in these cases.

In 2012, the state of Michigan passed a law that expanded access to insurance coverage for

children with ASD, with implementation starting in October, 2012. The law had three main pillars.

First, for people covered under state regulated insurance plans, mainly employer sponsored plans

for small or medium sized employers and individually purchased plans, the law mandated coverage

for “evidence-based behavioral health treatment” - typically ABA, pharmaceuticals, psychiatric

care, psychological care, and other therapies - from birth through 18 years of age for children
with diagnosed ASD.16 Coverage requirements are generous: the maximum annual benefit starts

at $50,000 for children under six and decreases with age to a floor of $30,000 at age 18. Co-pays,

deductibles, and co-insurance rates cannot exceed those required by the individual’s insurance plan
for physical illness.17

A difficulty often faced by states in ensuring widespread coverage of Autism insurance man-

dates is that only a subset of insurance plans are subject to state regulation. Self-insured plans,

16Diagnosis by a physician is required and insurance companies are permitted to require the

evaluations be done through designated evaluation centers (Peters et al., 2014).

17Michigan Public Acts 99 and 100 of 2012.

74

mostly used by large employers, are covered under Federal law and so are not typically subject to

state mandates. As of 2011, 61% of Michiganders with employer-provided coverage were in self-

insured plans (Fronstin, 2012). Michigan addressed this gap via Public Act 101 of 2012, which set

up a reimbursement fund for self-insured plans that provided benefits in line with those required

for regulated insurance plans. Plans were permitted to request up to 100% of the claims from
their beneficiaries for reimbursement.18 While there are no data on how many self-insured firms

provided coverage under this law, the very generous reimbursement likely led to high take-up.

It is worth noting that the Autism Alliance of Michigan maintains a list of self-funded firms in

Michigan that offer the insurance benefit, including many of largest employers in the state such as

General Motors, Ford, Meijer (a supermarket chain), and Beaumont Health System, along with the
state government and most major universities.19

Children not covered under employer-provided or individually-purchased insurance plans in

Michigan are almost all covered under Medicaid, including those covered through the Children’s

Health Insurance Program (CHIP). In 2014, 58% of children aged 0 to 18 in Michigan were covered
by private insurance, while 39% were covered under Medicaid and only 3% were uninsured.20 The

Michigan reform provided insurance coverage for ASD to Medicaid beneficiaries as of April 2013,

but the benefits were considerably less generous than the private insurance mandate, a key aspect of

the reform for our identification strategy. In particular, while pharmacy, psychiatric, psychological,

and co-morbid therapies like speech therapy were already covered prior to the reform, the only

ASD-specific therapy added to coverage from the law is ABA. Other evidence based therapies are

not covered. This itself is only a minor difference as most therapy for ASD is based on ABA.

18In FY2016, the fund ran out of money and hence claim processing has been suspended since
then. While it is possible some firms have since removed their benefits due to the lack of reim-
bursement, since our data cover only through the 2015-16 school year and firms typically make
insurance coverage decisions towards the end of the calendar year, this is unlikely to affect our
results.

19A full

list of self-insured employers with ASD benefit can be found at https://
autismallianceofmichigan.org/insurance-facts/. Data on the largest employers in Michigan are
from https://www.zippia.com/advice/largest-companies-in-michigan/.

20Kaiser Family Foundation estimates based on the Census Bureau’s March Current Population

Survey, 2014-2017.

75

More importantly, underfunding of the Medicaid benefit led to coverage expiring once the child
reaches an age of six.21 Generally, ABA therapy continues beyond this age and many years of

therapy are needed for benefits to emerge and be maintained. Further reducing the value of this

benefit is that often children are diagnosed relatively close to the age cutoff. According to the most

recent report available, the average age at first ASD diagnosis for Medicaid recipients across the

US was 5.4 years in 2002 - 2004 (Mandell et al., 2010). This leaves virtually no time for therapy

to have an effect before access is cut off. Even if age of diagnosis has improved, nationwide data

regardless of insurance coverage showed that the median age of diagnosis in 2012 was 4.2 years,

again leaving little time to garner substantial benefits from therapy prior to reaching six years of

age (Christensen et al., 2016).

3.2.3 Special Education

Students deemed eligible for special education services are covered under the Individuals with

Disabilities Education Act (IDEA). To receive such services, students first must be evaluated. An

evaluation can be initiated by parents or the school and involves a review of the child’s educational

progress as well as factors potentially related to the suspected disability like health, vision and hear-

ing, social and emotional development, academic performance, communication, and motor skills.

Hence, eligibility is not simply based on standardized test scores or pure academic performance.

Autism is classified under IDEA as a specific disability category. The evaluation process for

students who may be on the Autism spectrum involves examining the students existing academic

record, behavioral outcomes, interviews with teachers and parents, and an assessment by special-

ists trained in ASD diagnosis and treatment. These specialists can be provided by the school or

outside of the school. Critically, under IDEA these assessments are to be provided at no cost to

families, though parents can and often do use external assessments to inform the process. Once a

child has been evaluated, parents meet with some combination of teachers, school administrators,

21As of January 2016, due to requirements of the Affordable Care Act, the age limit was in-

creased to 21. For this reason we focus our analysis on school years prior to 2015-2016.

76

school counselors, and specialists to determine whether the child qualifies for special education

services. If so, they develop an individualized education plan (IEP) that specifies what educational

environment and educational services the student will receive and the benchmarks that will be used

to determine whether the student needs to continue receiving these services in the future. Typically,

IEPs are updated every year with a full reassessment every three years.

The special education services agreed to in the IEP need to be provided by the school at no

cost to the family. However, the services being received outside of school can influence the in-

school services on the IEP because the IEP is developed with direct input from the parents. It

is reasonable to assume that parents and schools consider the sum total of therapies and services

available to students when crafting an IEP, and parents of students with ASD often report needing

to ask schools to provide more services than initially offered (Ferreri and Bolt, 2011). However, we

are aware of no research that examines the extent to which outside therapies influence IEP services.

IDEA also includes a Least Restrictive Environment (LRE) provision that requires students be

placed in the most general education setting possible. This provision is designed to avoid special

education students being segregated from the rest of the school population, which could have

negative consequences for educational and social/emotional development.

As this discussion highlights, the special education process is complex and involves many

participants and constituencies. Together with the LRE provision of IDEA there is significant

scope for non-school resources and factors to play a role in the specific education services student

receive and the educational environment to which they are exposed. Research on factors that

influence how IEP plans are developed is thin; our paper is the first to empirically examine how

external factors such as health insurance affect special education services, which is an important

advancement in our understanding of how the special education system operates.

77

3.3 Data

3.3.1 Michigan Administrative K-12 Schooling Data

Our analysis relies on a student-level dataset provided by the Michigan Department of Educa-

tion (MDE) and Center for Education Performance and Information (CEPI). The dataset contains

administrative educational records on all students enrolled in pre-K through grade 8 in Michigan’s

public schools from the 2009-10 to 2014-15 school years. These records provide rich informa-

tion on students’ demographic characteristics, disabilities, educational settings, special education

programs and services, and achievement levels. Student demographic characteristics are reported

by schools to MDE and include a student’s race, gender, and eligibility for Limited English Pro-

ficiency (LEP) services. Our key demographic variable of interest is a student’s “economically

disadvantaged” status, which we use as a proxy for measuring a student’s insurance status. A

student is defined as economically disadvantaged if she qualifies for free or reduced-price meals

under the National School Lunch Program, is in a household that receives food (SNAP) or cash

(TANF) assistance, is homeless, is a migrant, or is in foster care. Students’ economic disadvantage
statuses are updated annually to reflect changes in families’ economic circumstances.22 Typically

a student who qualifies for any of the latter also qualifies for free/reduced price lunch, and so we

interchangeably refer to this status as “economically disadvantaged” or FRPL.

For students with IEPs, which is synonymous with qualifying for special education, we obtain

additional information on a student’s primary disability, as defined on her IEP, and special educa-
tion resources provided to the student.23 The special education resources variables are classified

22One concern with this measure is that some schools qualify for the Community Eligibility
Provision for free-lunch that allows all students in a school to qualify regardless of individual
circumstances. However, Michigan still requires schools to collect family income information to
determine individual FRPL eligibility for record keeping purposes. Only 0.3% of observations in
our data are in school-years with 100% FRPL eligibility, indicating that CEP does not affect our
classifications.

23Students may also receive services through the use of 504 plans which typically provide access
for students who are not classified under conditions recognized via the Individuals with Disabilities
Education Act. Since autism is a category in IDEA and IEPs provide more legal protections and
guarantees of educational (as opposed to simply disability related) services, most children with

78

into three distinct categories: (1) a student’s special education program, (2) a student’s educational

setting, and (3) the special education support services received by a student.

The program category contains the IEP-designated programs in which a student is enrolled.

Programs are state-defined special education settings that must adhere to specific regulations. To be

considered an ASD program, a classroom must not have more than 5 students and must be served

by a state-endorsed teacher of students with ASD who has completed ASD-specific education and

training. Therefore, not all schools or school districts offer all special education programs, and a

student’s program need not exactly correspond with her disability. For example, students with ASD

are commonly enrolled in a “cognitive impairment” program, which has classrooms with up to 10

students and is designed to provide instruction to students with an array of learning disabilities.

There are 14 specific types of special education programs in Michigan, and students can be enrolled

in up to 3 of them. We focus on four categories that are the most relevant for ASD students: ASD-

specific special education programs, cognitive impairment programs, resource programs - which

usually involve pull-out time in a “resource room” with a special education specialist, and no

program (e.g. in a general education classroom 100% of the time). It is also important to note that

even if a student is attached to a program, he/she could spend any percentage of time below 100%

in a general education classroom. So in some cases students are technically placed in a program

but spend very little time in the actual program classroom.

The educational setting category contains information on the primary educational setting where

a student receives his education. Our data include eight different measures of the education setting:

enrollment in a special education school, whether students are in a general education classroom

more than 80% of the time, 40-79% of the time, or less than 40% of the time, whether students

spend any time in a general education classroom, and the proportion of the student’s full-time

equivalent (FTE) enrollment that is in a special education setting. Since the special education FTE

rate is measured in the fall and spring, we use the average rate across the two semesters. For

ASD are covered under IEPs. While we do not observe 504 plans directly, these students would
have an ASD identification but no data on services, which accounts for only 0.5% of students with
ASD in our analysis sample.

79

brevity, we focus on whether a student is enrolled in a separate special education school, whether

she is enrolled in a general education classroom more than 80% of the time, and the percentage of

FTE enrollment that is in a special education setting averaged across the year. We hypothesize ex-

ante that these education settings are the most likely to be associated with changes in ASD-related

school services, however we also show estimates for our other measures.

The services category records any special education support services a student receives within

an academic year. These services include therapies, such as speech, occupational, and physical

therapy; work with school social workers and/or psychologists; special transportation to and from

school; and assignment to teacher consultants who provide assistance to general education teach-

ers. Teacher consultants in Michigan must have a master’s degree in education or in a field related

to special education as well as teaching experience in a special education classroom. Moreover,

teacher consultants may be approved to work with special education students generally or may be

approved to work with students with particular disabilities by meeting additional education and

experience standards. There are 30 specific special education support service categories listed in

the data, and up to 10 are recorded for a student. Our primary estimates examine ASD teaching

consultants, language support, occupational therapy, social workers, and whether a student receives

any support services. We focus on these support services because they are the most relevant for

ASD student. We also examine non-ASD teaching consultants, physical therapy, and transporta-

tion services to demonstrate that services that are less important for ASD students are unaffected

by the insurance mandate.

Finally, we obtain information on students’ test-taking behavior and achievement on standard-

ized math and reading exams. MDE reports the type of exams (standard or special education)

students took in a given year, as well as any special education accommodations used. Test scores

are recorded for students who take standard or particular types of special education exams. How-

ever, the special education exams assess different material and are scored on a different scale than

the standard exams. Thus, we only consider test scores for students who take standard exams. As

we discuss below, we do not find that the mandate led to changes in regular exam taking by stu-

80

dents with ASD relative to non-disabled students, allowing us to analyze the regular exams while

avoiding sample selection concerns. We further standardized the scaled scores for these exams

across all students in the state within a grade level and school year.

Our main sample consists of students in grades 2-8 in school years 2009-2010 through 2014-

2015. While as noted above we have data for more grades and years, students in kindergarten and

grade 1 are excluded because it is likely that publicly insured students in these grades also received

increased access to private services through the Medicaid benefit, though we also look at grade

specific effects for the excluded grades. Further, the 2015-2016 school year is dropped due to the

age six cutoff of Medicaid ASD benefits being removed in 2016. Table C.1.1 presents means of all

analysis variables for several subsamples of these students: all students, ASD students, non-ASD
special education students, and non-special education students.24 About 1% of the sample has an

ASD identification; these students are more likely to be male and white than the sample overall

and are less likely to be disadvantaged. ASD students also are less likely to be disadvantaged and

more likely to be white and male than non-ASD special education students. Further, there are

substantive differences in the programs, educational environments and support services received

by ASD and non-ASD disabled students.

3.3.2 Measuring Insurance Status

One of the central data challenges in this analysis is the inability to measure student health

insurance status. The insurance mandate was only binding for students covered under a private

health insurance plan. In some cases children may be uninsured but this is very rare. According to

the American Community Survey, 96% of individuals in Michigan under age 18 were covered by
insurance in 2012, the year Michigan passed the Autism mandate.25 Thus, those students who do

24The sample sizes of ASD students in the top and bottom panel do not match because a small
number of students receive an ASD identification through a 504 plan but receive no special educa-
tion services.

25This is likely a lower bound of the child health insurance coverage rate because those who are
eligible for Medicaid but who are not signed up would be signed up and receive treatment upon
arrival at a hospital. They therefore are functionally covered by Medicaid even if they are not

81

not have private insurance are almost certainly covered by Medicaid/CHIP.

To proxy for private insurance coverage, we use a student’s status as being economically disad-

vantaged, which is primarily based on free/reduced-price lunch (FRPL) eligibilty. The motivation

for using this proxy is that the eligibility criteria for FRPL status overlaps closely with eligibility

for public health insurance. In order to qualify for public health insurance, children must be in
families that earn less than 200% of the Federal Poverty Line.26 Eligibility for free or reduced

price lunches in schools is set at 185% of the poverty line.

Table C.1.2 shows health insurance status by free/reduced price lunch eligibility (top panel)

and by family income as a percent of the federal poverty line (bottom panel) from the 2008-2016
American Community Survey among K-8th grade students in Michigan. The top panel shows

health insurance coverage is near-universal and varies little by whether students are eligible for

free/reduced price lunch. What does vary across these groups is what type of insurance students

have. Almost 89% of ineligible students have private insurance, while 73% of eligible students

receive Medicaid. While there is some overlap, FRPL eligibility is strongly correlated with whether

students receive Medicaid.

FRPL status is a somewhat noisy proxy for family income. Research using data from education

records linked to tax data indicates that there is a wide range of family incomes among students

in the same free/reduced price lunch category (Domina et al., 2010). The bottom panel of Table

C.1.2 shows that the poorest students, those whose families earn under 135% of the poverty line,

are the most likely to be on Medicaid (81%). The percent on Medicaid declines with income,

even in the eligible range. In order to strengthen the proxy we use for Medicaid eligibility, we use

the fact that those who persistently receive free/reduced price lunch are the most disadvantaged

students (Michelmore and Dynarski, 2017). These students are likely to come from the bottom of

the income distribution, and Table C.1.2 indicates that they are unlikely to have private insurance.

formally enrolled in the program.

26Federal Medicaid eligibility is stricter, at 133% of the Federal Poverty Line. The Michigan
Child Health Insurance Program (MCHIP) extends public insurance eligibility up to 200% of the
poverty line for children of Michigan residents.

82

Conversely, higher income students who are above 250% of the poverty line are very likely to have

private insurance. Hence, our main analysis sample is comprised of those who receive free/reduced

price lunch in every observed year of school enrollment from grade 2 to 8 and those who do not

receive free/reduced price lunch in any observed year of school enrollment between grades 2 and

8. Students who receive free/reduced price lunch in some years are excluded, as are students

who are only observed for one year. For completeness, we show robustness checks that include

the “sometimes free/reduced price lunch” students. The estimates that include these students are

qualitatively similar but attenuated as expected because of the use of a weaker treatment proxy.

3.4 The Effect of the Autism Insurance Mandate on ASD and Special Edu-

cation Incidence

3.4.1 Empirical Approach

Using data on students in grades 2-8 in Michigan from school years 2009-2014 as described in

Section 3.3, we estimate difference-in-differences models that identify how the insurance mandate

affects the likelihood that students are diagnosed with ASD or have any special education diagno-

sis. Our measure of special education is whether students have an individual education plan, and

we designate a student as being diagnosed with ASD if the IEP lists ASD as the primary disability.

Note that we only observe specific services, programs, and educational settings for students with

IEPs. Further, in our main model we exclude any student who is identified as economically dis-

advantaged in some but not all years and conditions on students being observed in grades 2-8 for

at least two years. Our difference-in-differences model for ASD identification is of the following

form:

ASDig jt = β0 + β1NonDisadvi + β2PostMandatet × NonDisadvi

(3.1)

+ ΩXit + γgt + δ j + εig jt,

where ASDig jt is an indicator equal to 1 if student i in grade g and school j is identified as hav-
ing ASD (or an alternative disability that generates an IEP in companion estimates) in year t,

83

NonDisadvi is an indicator that equals 1 if we never observe the student as economically disad-
vantaged in grades 2 through 8 (i.e., treated students), and PostMandatet is an indicator that is

equal to 1 in the 2012-2013 school year and beyond. The model includes school fixed effects

(δ j) as well as grade-by-year fixed effects (γgt). The inclusion of school fixed-effects ensures that
we are only comparing students in the same school, since special education implementation often

varies considerably across schools. Furthermore, special education policies are usually defined

at the district level, and the school fixed effects subsume school district fixed effects. We show

below that our estimates are similar when we only use school district fixed effects. Finally, we
include controls for whether a student is white, male, or limited English proficient (Xit). Standard
errors are clustered at the school district level since students in the same district experience sim-

ilar education environments and face the same special education evaluation and service provision

practices.

The coefficient of interest in equation (3.1) is β2, which is the difference-in-differences es-
timate of how the ASD (or other disability) rate of non-disadvantaged children changes in 2012

relative to disadvantaged students. The core identifying assumption is that trends in special edu-

cation diagnoses among disadvantaged students are a valid counterfactual for trends among non-

disadvantaged children, conditional on the controls. It is important to stress that for identification

it need not be the case that disadvantaged and non-disadvantaged students with a disability are

treated similarly, only that their treatment would have changed in similar ways in the absence of

the insurance mandates. This assumption can functionally be broken down into two pieces: 1)

outcome trends between disadvantaged and non-disadvantaged children must be similar prior to

2012, and 2) there must be no shocks that occur in 2012 that disproportionately affect students by

disadvantage status.

Using data prior to 2012, we generate direct evidence on whether there are pre-treatment rela-

tive trends. These figures are presented below in Section 3.4.2 and provide strong support for the

assumption that ASD rates are trending similarly across disadvantaged and non-disadvantaged stu-

dents prior to 2012. The second assumption of no correlated shocks is more difficult to examine in

84

the data, as such shocks are by definition unobserved. Nonetheless, we are aware of no other state

policy that was enacted during 2012 that would have disproportionately affected students across

the SES distribution. The economy was recovering from the Great Recession, but this should be

reflected in trends rather than a 2012-specific shock. The Affordable Care Act individual mandate

came into effect in 2012. However, it was effective January 1 2012, while the ASD mandate went

into effect in October 2012. Thus, any effects of the ACA mandate should be evident in the prior

school year. We also do not believe it is plausible that the ACA affected these students: the ACA

was focused on uninsured adults rather than children. Health insurance coverage among children

was nearly universal prior to 2012 due to Medicaid and CHIP, and Medicaid rules did not change
in Michigan during this period.27 To the extent the mandate caused some parents of children who

are disadvantaged to switch from Medicaid to private insurance to use the ASD service benefits,

our estimates will understate the effect of the mandate.

3.4.2 Results

Table C.1.3 presents estimates of β2 from equation (3.1). Panel A shows the baseline estimates.
In column (1), we show estimates using ASD diagnosis as the dependent variable; while the point

estimate is statistically different from zero at the 10% level, it is small in magnitude. The point

estimate indicates that non-disadvantaged students experience a relative decline in ASD diagnoses

of 0.045 percentage points after 2012. This is 4.5% relative to the mean, and the 95% confidence

interval rules out an ASD increase among this group of more than 0.03 percentage points. Thus,

our estimates indicate that the mandate led at most to a very small change in ASD special education

identification.

As discussed above, a core assumption underlying our approach is that there are no trends

in disability diagnosis that differ across students who do and do not receive free/reduced price

lunch. The top panel of Figure C.1.1 presents event study estimates of equation (3.1), where

27Michigan expanded Medicaid for adults under the ACA in April 2014, but this expansion did

not affect eligibility among children.

85

PostMandatet × NonDisadvi is replaced with a set of interactions between NonDisadvi and year
indicators. The figure demonstrates that there is no systematic change in the likelihood of being

diagnosed with ASD across the two groups prior to 2012. Furthermore, the year-to-year changes

are extremely small, even relative to the low baseline mean of 1%. The figure also shows that there

is little post-2012 change in the ASD diagnosis likelihood across the groups, which is consistent

with estimates in Table C.1.3.

Column (2) of Table C.1.3 presents estimates for the incidence of all other non-ASD disabili-

ties. The estimate is positive, not statistically significant, and is small in magnitude. Taken at face

value, it suggests that non-ASD diagnoses decreased by 0.04 percentage points (0.03% relative

to the mean) among economically disadvantaged students relative to non-disadvantaged students

post-2012. The bottom panel of Figure C.1.1 presents event study estimates for non-ASD disability

incidence. The estimates in this specification are noisy, but similar to the ASD event study, there

is little evidence of pre-2012 or post-2012 relative changes. Taken together, the panels of Figure
C.1.1 support the use of the free/reduced price lunch students as a control group in this analysis.28

Although there is no aggregate change in non-ASD disability incidence, the remaining columns

of Table C.1.3, Panel A present evidence of a shift in the composition of disabilities in this broad

group. The mandate is associated with an increase in the prevalence of emotional and other health

disabilities and a decline in the prevalence of speech disability. These relative changes in the

composition of the special education groups complicates our preferred triple difference analysis in

which we compare changes in outcomes by poverty status among ASD versus non-ASD special

education students when the mandate comes into effect. It is possible that outcomes in the non-

ASD population are affected by this change in the composition of disabilities.

In Panel B of Table C.1.3, we show estimates that include a linear time trend interacted with

non-poverty status. Including this control, the ASD effect shrinks substantially and is no longer

statistically significantly different from zero at even the 10% level. Furthermore, the non-ASD

disability effects are attenuated such that only emotional disability is significantly different from
28Note that we are unable to use non-disabled students as a control group, since they by defini-

tion are not diagnosed with ASD.

86

zero (at the 10% level). A comparison of Panel A and Panel B of Table C.1.3 shows that the relative

shifts in diagnoses are due to linear secular trends by poverty status. Event studies that exclude

this control also demonstrate this point: the changes in disability incidence are mostly driven by
secular trends that appear prior to 2012.29 The linear non-disadvantaged trend accounts for these

secular trends. Critically, we show below that the rest of our results and conclusions are robust to

including linear non-poverty time trend controls. Hence, our results are not being driven by the

small compositional changes that are evident in Panel A of Table C.1.3.

Panel C of Table C.1.3 demonstrates the robustness of our estimates to school district fixed

effects. Since special education policies often are district-specific, it is not clear that school fixed

effects are necessary or desirable. The results in Panel C are almost identical to those in Panel A,

which further supports the validity of our empirical approach.

3.5 The Effect of the Autism Insurance Mandate on Educational Services

and Test Scores

3.5.1 Empirical Approach

Motivated by the finding that the insurance mandate has a negligible effect on ASD inci-

dence, we employ triple difference models that compare changes in outcomes among students

with ASD by free/reduced price lunch status to changes in outcomes among non-ASD students by

free/reduced price lunch status. Specifically, we estimate models of the following form:

Yig jt = β0 + β1NonDisadvi + β2ASDit + β3PostMandatet × NonDisadvi

+β4PostMandatet × ASDit + β5NonDisadvi × ASDit
+β6PostMandatet × NonDisadvi × ASDit + ΩXit + γgt + δ j + εig jt,

(3.2)

where Yig jt is an outcome for student i in grade g, school j, and year t. All other variables are as
previously defined. In all the models, the vector (Xit) includes controls for whether a student is
white, male, or limited English proficient. In models with test scores as the dependent variable,

29These event study estimates are available from the authors upon request.

87

the vector also includes the lagged test score for the same subject. As with equation (3.1), standard

errors are clustered at the school district level throughout.

For special education outcomes, we are restricted to using other special education students as

one of the comparison groups in the triple-difference analysis. This is because each of the out-

comes we analyze are only available for students with an IEP; students without an IEP do not

receive special education services. Hence, all outcomes for students who are not special education

will have values of zero. One potential limitation is that since special education funding is largely

fixed, districts may respond to higher (lower) needs in one group by reducing (increasing) services

in other groups. There are several reasons to believe that this is not a first-order issue in our anal-

ysis. First, districts are not restricted to using only special education funding for special education

services. In fact it is very common for special education funding to be insufficient and for districts

to use general funds to supplement the costs (Degrow, 2017). Since our results show a decline in

ASD service provision due to the mandate, it is likely that most, if not all, of the money saved will

go towards general education. Second, the fixed-effects in our model ensure that we are comparing

students in the same school. This ensures that differences in how schools (or districts) respond to

the funding changes do not affect the estimates.

The main variable of interest in equation (3.2) is β6, which yields the triple difference esti-
mate of the effect of the Michigan insurance mandate on student outcomes. The identification

assumptions underlying this model are similar to those discussed above for equation (3.1). How-

ever, this model relaxes the common trends assumption somewhat: any differences in trends in

outcomes across students who do and do not receive free/reduced price lunch must be similar for

ASD and non-ASD students. Put differently, the non-ASD relative trends by free/reduced price

lunch status need to be an accurate counterfactual for these trends among ASD students. Similar

to the difference-in-differences model, there are two main sources of bias. The first is differential

relative trends across treatment and control groups. In this case, ASD students who are/are not

disadvantaged would have to exhibit different relative trends to non-ASD students prior to 2012.

We present graphical evidence that such relative trends are not present. The second is contem-

88

poraneous and persistent shocks that differentially impact ASD students who are not eligible for

free/reduced price lunch. We know of no reason to suspect that these shocks exist, especially since

the ASD incidence rate does not change substantially when the mandate comes into effect.

The plausibility of the identification assumptions rests heavily on the composition of the control

group. When estimating impacts on programs, educational settings, and support services, the

sample consists of all students with a disability (i.e., with an individual education plan). The

control group thus is students with a non-ASD disability. Since only students with an IEP receive

special education services, non-ASD disabled students are a natural control group. When we

estimate effects on test scores, we are able to consider both non-ASD disabled and non-disabled

students as potential comparison groups. One important issue is that Michigan changed the format

and structure of achievement exams, particularly those taken by students with disabilities, in 2014-

15. Given this change, we restrict our achievement analysis to 2013-14 and earlier.

3.5.2 Results

One of the strengths of our data that is unique to our administrative education data and is not

commonly available in other states is the detailed information on special education services. As

discussed in Section 3.3.1, we consider three types of services: education programs, the educa-

tional setting, and special education support services. These categories are correlated with one

another but imperfectly so. For example, the program in which one is enrolled can affect spe-

cial education support services and the extent to which students are in a general education setting.

However, students can receive support services even if they are not enrolled in special education

programs and if they are in a general education setting. Examining these three categories of ed-

ucational inputs thus paints a rich picture of how the ASD insurance mandate affects the type of

education students with ASD receive.

Table C.1.4 presents our baseline estimates for the main set of special education service out-

comes that are most associated with ASD. The special education program outcomes we examine

are whether the student is in an ASD program (column 1), whether the student is in a resource or

89

cognitive impairment program (column 2) and whether the student is in no special program (col-
umn 3).30 The program does not have to match the disability listed on the IEP, so students with

ASD diagnoses can be in non-ASD focused programs, and attachment to cognitive impairment

programs in particular is common for students with ASD. Each column of the table shows results

from a separate regression, and the first row presents the triple difference coefficient of interest.

Column (1) shows that the ASD mandate increases the likelihood that students are placed in an

ASD program by 3.1 percentage points, which is 15.7% of the ASD-specific mean (shown at the

bottom of the table). However, the estimate is not statistically significantly different from zero. En-

rollment in resource and cognitive programs decline substantially, by 6.4 percentage points (9.6%),

and this estimate is statistically significant at the 5% level. The likelihood a student is enrolled in

no special education program increases by a statistically significant 3.4 percentage points, which

is a 4% drop in enrollment in any program - 87% of students with ASD are enrolled in at least

one special education program. That the percentage of ASD students not in any special education

program rises substantially suggests that the insurance mandate leads to lower intensity of special

education interventions and less placement in self-contained special education classrooms. These

results thus reflect crowding out of special education services offered by public schools from the

private insurance mandate.

As discussed in Section 3.3.1, one of the main assumptions under which our estimates are

identified is that relative trends among free/reduced price lunch and non-free-reduced price lunch

students are similar for ASD and non-ASD disabled students prior to 2012. Figure C.1.3 presents

evidence on the plausibility of this assumption by showing event study estimates of the mandate
on program placement.31 For no outcome do we see any evidence of differential pre-2012 trends,

which supports our identification strategy. Furthermore, there is a clear decline in the likelihood

of being in a cognitive or resource program after 2012 and an increase in the likelihood of being

30The estimates for other programs are provided in Table C.1.5 and show no effect.
31Specifically, we replace PostMandatet × NonDisadvi × ASDit with NonDisadvi × ASDit in-
teracted with a set of year dummies in equation (3.2). Year 2011 is excluded, so all estimates are
relative to that year.

90

enrolled in no program. These figures match the results in Table C.1.4 closely.

Columns (4)-(6) of Table C.1.4 present estimates of the effect of the mandate on students’ ed-

ucational settings. We focus on three outcomes: whether a student is placed in a special education

school (column 4), whether the student is in a general education classroom more than 80% of the

time (column 5), and the percentage of time students are in a special education classroom (column

6). The point estimates are small in magnitude and are not statistically significant at conventional

levels. Thus it appears that while the special education environment students are attached to during

their time outside the general education classroom is changing to less intensive environments, they

are not spending substantially more or less time in a general education classroom. While our data

do not allow us to dig deeper there are some explanations that are consistent with this pattern.

First, it may be the case that students who were attached to a cognitive impairment classroom were

already spending more than 80% of their day in the general education classroom in the absence

of the mandate, but with insurance coverage they leave the cognitive impairment program but still

spend some time outside the general education classroom for therapeutic services. Second, we are

limited to analyzing rather broad categories of general education classroom time, so it possible

these are masking some changes. For example, a student could have been attached to a resource

room and spent 80% to 90% of time in the general education classroom but with insurance the
student now no longer uses the resource room and is in general education 100% of the day.32

Figure C.1.3 presents event study estimates for these outcomes. The event study models are

less informative for these measures because education setting variables are only available begin-

ning in 2010 and the FTE measures are only available beginning in 2011. Hence, we have fewer

pre-treatment years for these outcomes with which to diagnose any selection on relative trends.

Nonetheless, given the data available, the event study estimates support the validity of our empiri-

cal approach and match the findings in Table 4 closely.

Our final set of educational input measures – special education support services – are shown

in columns (7)-(11) of Table C.1.4. We focus on ASD teaching consultants (column 7), language
32In Table C.1.5 we also do not see shifts in likelihood of being in a general education classroom

< 40% or between 40% - 79% of the time.

91

support (column 8), occupational therapy (column 9), access to social workers (column 10), and

whether students receive any support services (column 11). Column (7) shows that the insurance

mandate reduced the likelihood that a student received an ASD teacher consultant by 2.3 per-

centage points (17.7% relative to the mean), and this estimate is statistically significant at the 5%

level. Column (10) further shows that students are 2.1 percentage points (3% relative to the mean)

less likely to be assigned a social worker. Columns (8) and (9) indicate that there is little change

in students’ access to language and occupational therapy services following the mandate. While

these results are consistent with a crowd-out of special education services, we also find that the

likelihood of receiving any special education support services increases by 2.0 percentage points,

or about 2% of the mean. One possibility is that very marginal students are moved to having no

program but given access to services they did not have before to compensate. Nonetheless, given

the consistent patterns for the other measures, this also may simply be a spurious result that is

a function of having multiple outcomes. Overall, the pattern is again consistent with crowd-out.

Figure C.1.3 shows event studies for these outcomes and supports the identification assumption

of common relative trends in special education support services prior to 2012 between ASD and

non-ASD disabled children.

The results presented thus far suggest that ASD students who are likely to have private insur-

ance receive fewer special education supports in more inclusive environments after the insurance

mandate is passed. The effect of this change on student achievement is unclear. If the crowd-out

we find is incomplete, overall support levels increase and student achievement is likely to increase.

This is particularly the case if the quality of services provided by the private market are higher than

those provided through the schools. However, if crowd-out is full (or more than full) or if support

quality declines, student achievement should decline. Alternatively, if the services are focused on

one aspect of instruction for a student (e.g. math instead of ELA), that subject could worsen while

the other subject, which may have been crowded out by extra time spent on the focus subject, could

improve. Unfortunately, since we do not observe the instructional foci of the special education in-

terventions, it is likely that in aggregate we would see those effects offset. Hence, a finding of no

92

impact on achievement (which we show below to be the case) could be indicative of these shifts in

achievement across subjects.

In Table C.1.6, we focus on two achievement measures: standardized math test scores and
standardized reading test scores for students in grades 4-8.33 Given Michigan changed their exam

structures in 2014-2015, we exclude that year from the estimates, leaving us with two pre-mandate

and two post-mandate years. An added complication when we analyze these outcomes is that

it is unclear which control group is most appropriate. This closely relates to the issue that we

can only measure test scores for students who take Michigan’s traditional standardized exams.

Special education specific exams do not assess the same material, are scored on a different scale,

and experienced substantial changes over the time period. In Table C.1.7, we show that, when

comparing ASD students to non-disabled students, there is no evidence that test-taking behavior

was altered by the insurance mandate. However, when comparing ASD students to other disabled

students, there is some evidence that the mandate induced fewer students to take traditional exams.

Given these two complications, we show triple difference estimates using all three comparison

groups - all non-ASD, non-Sped, and non-ASD sped - but note that we prefer the estimates that

compare ASD students to their non-disabled peers. Even so, the results are relatively similar across

comparison groups.

The first three columns of Table C.1.6 show results for math scores and the second three show

results for reading test scores, both of which are in standardized units. Across all columns, we find

little evidence that academic achievement is affected by the private insurance mandate. For math

all of the estimates are very close to zero. Our preferred estimate using all non-Sped students as

the comparison group is in column 2 and has a coefficient of 0.005 with a 95% confidence interval

of [-0.046,0.056]. Thus, we can rule out at the 95% level that the mandate changes math scores by

more than 5-6% of a standard deviation.

33Students begin standardized testing in grade 3 but cannot be analyzed until grade 4 given our
inclusion of lagged achievement. In these specifications, our sample consists of all students who we
observe taking a regular exam for at least two years and are either never or always disadvantaged
in the years in which we observe them.

93

The point estimates for reading (columns 4-6) are somewhat larger than for math but are also

not statistically different from zero. In this case, the largest estimate is in column (4) using all non-

ASD as a comparison group. The 95% confidence interval for reading with our preferred control

group (column 5) is [-0.039,0.075]. We thus can rule out anything larger than modest effects on

test scores at the 95% level.

Figure C.1.2 shows event study estimates for math and reading using non-ASD special edu-
cation students as the comparison group.34 The figures do not show any evidence of differential

pre-2012 trends that would bias our triple difference estimates, though we acknowledge that we

are limited in this assessment by the need to control for prior achievement, which restricts us to

only having two years of pre-mandate testing data.

Taken together, the results from Table C.1.6 and Figure C.1.2 indicate that academic achieve-

ment is likely unaffected and certainly does not increase substantially due to the private insurance

mandate. While, as noted above, improvements in one subject may be offset by lower scores in

the other, it remains the case that in total we see little evidence of achievement impacts. Fur-

thermore, it is important to note that this is an intention-to-treat; for those students who do have

service changes (the treatment-on-the-treated effect) we cannot rule out sizable achievement im-

pacts. Nonetheless, the analysis indicates that the mandate policy itself has at most very small

impacts on achievement. This null result is an important finding given the changes in education

services we document. Because we cannot observe services provided outside of school, we can-

not determine whether our results indeed reflect crowd-out or just a reduced demand for services

among students because their conditions improve due to increased access to private services. That

student achievement does not substantially increase suggests our findings are most consistent with

a crowd-out story, which is our preferred interpretation of the results. This interpretation also

is consistent with the findings in Chatterji et al. (2015), who find no change in unmet need for

services due to ASD mandates. That achievement does not substantially decrease in our setting

further suggests that the shifting of responsibility to the private sector does not academically harm
34Event studies for the other control groups are shown in Figures C.1.4 and C.1.5 and are very

similar.

94

students with ASD. We caution, however, that these achievement analyses are short term, and so it

is possible that cognitive improvements do not show up until more exposure time has elapsed.

3.5.3 Heterogeneous Treatment Effects and Robustness Checks

We examine several sources of heterogeneity: gender, race, and grade.35 Table C.1.8 presents

estimates of the effect of the mandate on ASD incidence for each of these different groups. We

see no evidence of a change in diagnoses for girls or for Whites & Asians vs. Black & Hispanic

students. While there is evidence of reduced ASD incidence of 0.07 percentage points among boys

due to the mandate, as with the overall estimates, this is very small relative to the mean rate of

1.7%. When we examine heterogeneity by grade in Panel B, there is a decline in ASD diagnosis in

second grade that is significant at the 5% level. However, there is little evidence of any effects in

higher grades, which suggests the mandate may shift the timing of diagnosis slightly to these later

grades. It is worth noting that testing begins in grade 3, which may provide an impetus to identify

students who would have been identified earlier in the absence of the mandate. Overall, there is

little evidence that different groups experience a meaningful increase in ASD diagnosis, and there

is no evidence of a positive shift along any dimension we examine.

We now turn to examining outcomes for these different groups. As shown in Table C.1.1, 85%

of students with an ASD diagnosis are boys and in general, the condition is far more common

among males: according to the Centers for Disease Control, boys are 3 times as likely to be
diagnosed as girls nationwide.36 It thus is informative to examine effects separately by gender.

Table C.1.9 shows triple difference estimates of educational service outcomes for boys (Panel A)

and girls (Panel B). The direction of the estimates are similar for boys and girls, but the effects

tend to be larger both in absolute value and relative to the ASD identification rate for girls. Effects

on resource/cognitive programs, no special education programs, and ASD teacher consultant are

35When assessing heterogeneity by grade, our sample consists of all students in grades K-8 who
we observe for at least two years and are either disadvantaged in all years in which we observe
them or are never disadvantaged in the years in which we observe them.

36https://www.cdc.gov/nchs/products/databriefs/db291.htm

95

all larger for girls than for boys, but they are qualitatively similar across genders. Combined, the

results show substantial crowd out effects on both genders that are somewhat larger for girls. We

also note, however, that among girls there is a statistically significant increase in the likelihood

of receiving any special education services. This appears to operate through language services.

Given that this is the only outcome across both genders consistent with an increase in school-

based provision, we are unable to say whether this is a real effect or simply an artifact of multiple

hypothesis testing.

Panels C and D show effects of the mandate for Black and Hispanic as well as White and Asian

students, respectively. The crowd-out effects are most evident for White and Asian students, who

make up nearly 3/4 of the ASD population in Michigan. The effects on Black and Hispanic students

are quite noisy as a result. The results for Whites and Asians mirror the overall estimates quite

closely while the findings for Black and Hispanic students are qualitatively similar but imprecise,

which limits our ability to draw strong conclusions for this group.

Table C.1.10 shows effects by grade, including kindergarten and 1st grade. Kindergarten can

be interpreted as a specification check, as ASD services for the vast majority of these students are

covered by both private and Medicaid insurance plans after 2012. Furthermore, many first grade

students are under age 6 and would also be able to receive services through Medicaid. Hence, if our

identification strategy is valid we would not expect to see significant impacts in these grades and

indeed that is what the table shows. The estimates in both of these grades are small, and none are

statistically significantly different from zero at even the 10% level. These estimates suggest that

we are not picking up unobserved shocks or trends that differentially influence outcomes among

non-disadvantaged ASD students.

The remaining estimates in the table test for heterogeneous treatment effects by grade for

higher grades. Autism therapies like applied behavioral analysis can have differential effects by

age, and the ability of schools to alter special education services also can differ for older versus

younger students. Resource and cognitive program reductions are largest for students in grades 2
through 6 and start to fade after 4th grade, though estimates never turn positive. Consistent with

96

resource/cognitive program effects, the effects on non-participation in special education programs

are positive in grades 2 through 6. Further, the reduction in the use of ASD teacher consultants is

concentrated in early elementary grades. For the other outcomes, the estimates are generally small

and are not statistically significant regardless of grade level. The crowd-out effects are mostly con-

centrated in elementary rather than middle school, which one would expect given ASD therapy is

more effective, and hence more commonly used, when the child is younger.

We also examine test score effects along these dimensions of heterogeneity. Table C.1.11

shows reading and math score estimates by gender, race, and grade using non-disabled students as

the comparison group. There are no strong statistically significant patterns across groups, though

for girls and Black/Hispanic students the estimates are very imprecise due to there being relatively

few ASD students in each of these groups. When we look by grade level, the estimates are again

very noisy, but only one estimate is significant at the 10% level and they generally have a mix of

positive an negative coefficients.

We next estimate a series of robustness checks that assess the validity of several data limitations

and identifying assumptions. First, we examine the importance of including the linear time trend

interacted with non-disadvantaged status. As discussed in Section 3.4, this control accounts for

secular linear trends in special education incidence, particularly in the comparison group of non-

ASD disabled, and removes the composition changes we see within the comparison group in Table

C.1.3. In Panel A of Table C.1.12, we present triple difference results that include this control.

The estimates are virtually identical to baseline, which suggests that the composition changes have

little effect on our program, educational setting, and special education service provision results.

Panel A of Table C.1.13 shows a similar robustness check for math and reading test scores; the test

score results are also robust to including this control.

In Panel B of Table C.1.12, we present estimates that include school district (rather than school)

fixed effects. This robustness test assesses the stability of the estimates to not accounting for

unobserved heterogeneity across schools within each district. The estimates are very similar to

baseline, suggesting that this heterogeneity is not correlated with the treatment.

97

Panel C shows results that address the concern that ASD designation is endogenous to the

mandate. While we find little evidence to suggest this is the case, the point estimate for ASD in

Table C.1.3 is significant at the 10% level. To check if this is a concern, we use the pre-2012

ASD assignment of students in place of their contemporaneous identification for post-treatment
years. Hence, we assign everyone their pre-treatment ASD identification status.37 The estimates

again are very similar to those in Table C.1.4 except the reduction in use of social workers and the

increase in enrollment in no special education program become larger. Hence if anything we may

be slightly underestimating the crowd-out effects. Nonetheless, since students for this analysis

have to be observed before 2012, the age profile in the post-treatment period skews older and so

we believe the baseline estimates are more accurate.

Throughout this analysis we have compared students who are always disadvantaged to students

who never are to increase the strength of the proxy for Medicaid eligibility. In Table C.1.14, we
show results using the full sample of students.38 Relative to our baseline sample, this sample

adds those who sometimes receive free/reduced price lunch or are observed only once in the data

and identifies treatment as being non-disadvantaged in a given year. The sample sizes increase,

and the point estimates are attenuated relative to the main results as expected, since free/reduced

price lunch receipt is a worse proxy for Medicaid eligibility among the sometimes-eligible students

(Michelmore and Dynarski, 2017; Domina et al., 2010). Nonetheless, the qualitative patterns do

not change; the conclusion that the Autism insurance mandate led to crowd-out of education ser-

vices is robust to including these students. Table C.1.16 shows that we obtain similar results for

math and reading test scores when we include these students as well. In Panel B of these tables, we

show that the estimates using the full sample of students are robust to including linear year trend

interacted with non-poverty status. Together, the results in Table C.1.12 and Tables C.1.6 through

C.1.16 demonstrate that our results and conclusions are robust to the way in which we construct

37Data restricted to students who are observed at least once before 2012.
38Table C.1.15 shows disability incidence estimates using the full sample of students both with
and without time trends by non-poverty status. The estimates are very similar to those in Table
C.1.3.

98

our analysis sample and to the use of linear time trends by non-poverty status.

The results thus far have focused on a set of special education services and outcomes that are

most closely associated with ASD. As noted previously, in Table C.1.5, we present estimates for

other services in our data that are less likely to be affected by the Autism insurance mandate. If we

find effects on many of these outcomes, it is suggestive of bias in our main results. Specifically,

we examine enrollment in another special education program, two categories of general education

participation, any general education participation, non-ASD teacher consultant use, physical ther-

apy services, and transportation services. None of the point estimates in Table C.1.7 is statistically

significant at even the 10% level, and each estimate is close to zero. There is no evidence that these

other service measures are affected by the insurance mandate, which supports the validity of our

main findings; the services that change are those that are most closely aligned with the needs of

ASD students.

Finally, in Panel D of Table C.1.11, we show that our test score estimates are robust to in-

cluding the 2014-15 testing year when Michigan changed to a new exam format that could have

affected ASD versus non-ASD students differently. In general the estimates are similar with the

exception of math relative to non-ASD sped students which becomes negative and marginally sig-

nificant. More importantly, however, these results are consistent with our overall conclusion that

any achievement gains were at most modest.

3.6 Conclusion

We present the first estimates in the literature of how a mandate that requires private insurance

to cover therapeutic services for children diagnosed with Autism Spectrum Disorder affects the

special education services students receive in public schools, as well as their educational achieve-

ment as measured by test scores. While we study Michigan’s mandate, passed in 2012, 46 states

and D.C. currently have some form of coverage mandate for ASD students. The prevalence of

these mandates makes them important to study, but our results also provide more general insight

into how the effect of health policy spills over to education services and outcomes. The close con-

99

nection between health and education in the production of human capital underscores the relevance

of studying such policy spillovers more broadly.

Using administrative K-12 data on all 2nd through 8th grade students in the state of Michigan

from 2009-2010 to 2014-2015 school year, we estimate how the insurance mandate affected a wide

range of special education services as well as student test scores. The data do not contain informa-

tion on private insurance coverage, so we use the strong overlap between economic disadvantage

and Medicaid (the near-universal alternative to private insurance among children in Michigan) to

proxy for exposure to this mandate. Because we find little evidence that the incidence of ASD di-

agnoses is altered by the mandate, we estimate triple difference models that compare how services

and outcomes change in 2012 among non-disadvantaged ASD students relative to disadvantaged

ASD student and disadvantaged vs. non-disadvantaged differences among non-ASD students.

Our main findings indicate that the ASD coverage mandate led to sizable declines in the special

education services students receive. ASD students who are not economically disadvantaged expe-

rienced declines in the likelihood of being placed in a resource or cognitive impairment special

education program, the likelihood of being placed in any special education program, the likelihood

of being given an ASD teaching consultant, and the likelihood of being provided in-school access

to a social worker. However, test scores did not change on average. Taken together, we argue

the evidence is most consistent with a crowd-out story, where the private provision of ASD thera-

pies reduces special education services in schools. This would generate the service reductions we

document and would lead to no change in academic achievement, as we find.

Our results are important in showing that supply-side health policies focused on health insur-

ance have spillover effects to the education system that likely were unintended by policymakers.

The findings from this paper suggest that the crowd-out of special education services largely un-

does the intent of policymakers to help provide more therapy services to autistic children. Nonethe-

less, we see little evidence that policy change was harmful to students and there is a potential

for welfare enhancement if provision through the health insurance system is more efficient than

through the education system or if it frees up instruction time for students. Nonetheless, that these

100

spillovers occur in this setting is suggestive that other health care policies, such as recommenda-

tions against teens taking anti-depressants or medical practices surrounding ADHD disabilities,

also may have effects on the services students receive in schools and their academic achievement.

Further understanding these interactions between health policies and schools and how they affect

students is a ripe area for future research.

101

APPENDICES

102

APPENDIX A

CHAPTER 1 APPENDIX

A.1 Tables & Figures

Figure A.1.1: Identified Community College District Boundaries

103

Figure A.1.2: Washtenaw Community College District Analysis Sample

104

Figure A.1.3: Distribution of Border Pair Tuition Differentials

105

051015Percent0100020003000Tuition DifferentialFigure A.1.4: Correlation Between Tuition Differentials and Area Characteristics

(a) % Economically Disadvantaged

(b) Math Test Scores

(c) Local CC Size

(d) Local CC Quality

106

Correlation: -0.0620100020003000Tuition Differential0.2.4.6.81% Economically DisadvantagedCorrelation: 0.0650100020003000Tuition Differential-2-1012Average 11th Grade Math ScoreCorrelation: -0.0940100020003000Tuition Differential05000100001500020000Local CC Undergraduate EnrollmentCorrelation: 0.1390100020003000Tuition Differential20004000600080001000012000Local CC Instructional Spending per FTEFigure A.1.5: Reduced Form Estimates with Alternative Bandwidths

107

0.05.1.15Effect of In-District Status on Local CC Enrollment0.1   0.5    1.0    1.5    2.0    2.5    3.0    3.5    4.0Miles to BorderFigure A.1.6: School Districts Overlapping Community College Districts

108

Table A.1.1: Mean Tuition Rates at Michigan Community Colleges, 2008-2016

Per Credit Per Semester Per Year Annual/Income

$2266.56
$3729.36
$1462.80

3.78%
6.22%
2.44%

$94.44
$155.39
$60.95

$1133.28
$1864.68
$731.40

In-District
Out-of-District
Difference
Notes: Tuition rates are provided by Michigan’s Workforce Development Agency and con-
verted into real 2016 dollars. All amounts are averaged across academic years 2008-2009 to
2015-2016. “Per semester” rates are calculated as the cost of 12 credits and “per year” rates are
calculated as the cost per 24 credits. The final column “Annual/Income” presents the “per year”
estimates divided by 60,000, the approximate median household income of students attending
Michigan’s community colleges (Chetty et al., 2017).

109

Table A.1.2: Associate Degree Programs Offered by Community & Vocational Colleges

CIP
Code CIP Title
1
3
4
5
9
10
11
12
13
14
15
16
19
22
23
24
25
26
27
29
30
31
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
54

Agriculture, Agriculture Operations, and Related
Natural Resources and Conservation
Architecture and Related Services
Area, Ethnic, Cultural, and Gender Studies
Communication, Journalism, and Related
Communications Technologies/Technicians and Support
Computer and Information Sciences and Support
Personal and Culinary Services
Education
Engineering
Engineering Technologies/Technicians
Foreign Languages, Literatures, and Linguistics
Family and Consumer Sciences/Human Sciences
Legal Professions and Studies
English Language and Literature/Letters
Liberal Arts and Sciences, General Studies, Humanities
Library Science
Biological and Biomedical Sciences
Mathematics and Statistics
Military Technologies
Multi/Interdisciplinary Studies
Parks, Recreation, Leisure, and Fitness Studies
Philosophy and Religious Studies
Theology and Religious Vocations
Physical Sciences
Science Technologies/Technicians
Psychology
Security and Protective Services
Public Administration and Social Service Professions
Social Sciences
Construction Trades
Mechanic and Repair Technologies/Technicians
Precision Production
Transportation and Materials Moving
Visual and Performing Arts
Health Professions and Related Clinical Sciences
Business, Management, Marketing, and Related
History

MI Comm.
Colleges

Vocational
Colleges

0.342
0.258
0.249
0.097
0.536
0.615
0.996
0.723
0.864
0.750
0.999
0.371
0.838
0.782
0.346
1.000
0.191
0.585
0.377
0.000
0.429
0.381
0.185
0.018
0.416
0.439
0.397
0.996
0.483
0.284
0.567
0.931
0.842
0.264
0.930
1.000
1.000
0.231

0.198
0.006
0.001
0.000
0.726
0.724
0.957
0.130
0.735
0.427
0.778
0.302
0.725
0.343
0.006
0.084
0.000
0.056
0.006
0.001
0.007
0.115
0.000
0.000
0.006
0.000
0.007
0.939
0.010
0.008
0.016
0.802
0.421
0.745
0.765
0.945
0.962
0.011

Diff.

0.144
0.252
0.248
0.097
-0.191
-0.109
0.039
0.593
0.129
0.323
0.221
0.069
0.112
0.440
0.339
0.916
0.190
0.529
0.372
-0.001
0.422
0.266
0.185
0.017
0.411
0.438
0.389
0.056
0.473
0.276
0.551
0.129
0.421
-0.482
0.165
0.055
0.038
0.220

Notes: All data comes from the U.S. Department of Education’s College Scorecard. All variables are averaged
across all 2009-2016 high school graduates who enroll in college within one year of high school graduation to
reflect the characteristics of the colleges that students attend.

110

Table A.1.3: Baker College vs. Private Two-Year Colleges

Variable:
Avg. Net Price
Instruction $ per FTE
% Full-Time Faculty
200% Graduation Rate
% Liberal Arts Degrees
Median Earnings
Median Debt

Baker
College

(1)

$12,333
$4,010
0.103
0.168
0.003
$26,880
$8,447

All Private
Two-Years For-Profit Not-For-
Profit
(4)

(2)

(3)

$16,320
$4,361
0.481
0.682
0.007
$26,129
$8,927

$16,200
$3,908
0.457
0.688
0.000
$24,512
$8,673

$17,172
$7,525
0.580
0.635
0.058
$35,725
$10,637

Institutions
Notes: Data comes from the College Scorecard, averaged over all years 2009-2016.

1,838

2,092

1

336

111

Table A.1.4: Michigan’s Traditional Four-Year Colleges

Variable:
Undergraduates
Avg. Net Price
Instruction $ per FTE
% Full-Time Faculty
200% Graduation Rate
Median Earnings
Median Debt

Flagships Other
Public
(2)

(1)

32,475
$15,477
$17,943
0.840
0.850
$55,220
$18,771

12,522
$13,245
$8,448
0.640
0.512
$42,229
$14,170

Private

(3)
1,669
$18,995
$8,430
0.685
0.521
$41,938
$16,106

Institutions
Notes: Data comes from the College Scorecard, averaged over all years
2009-2016.

26

2

13

112

Table A.1.5: Michigan’s Community College Districts

Community College
Alpena

Bay de Noc

Delta

Gogebic

Counties

Delta
Dickinson*
Bay
Midland
Saginaw
Gogebic

Grand Rapids

Henry Ford
Jackson

Kalamazoo Valley

Kellogg

Kirtland

Lake Michigan

Berrien

Lansing

Macomb

Macomb

Mid Michigan

Monroe County

Monroe

Montcalm

Mott

School Districts
Alpena

Cities/Townships

Kent ISD: Byron Center, Caledonia, Cedar Springs,
Comstock Park, East Grand Rapids, Forest Hills,
Godfrey Lee, Godwin Heights, Grand Rapids,
Grandville, Kelloggsville, Kenowa Hills, Kent City,
Kentwood, Lowell, Northview, Rockford, Sparta,
Thornapple Kellogg, Wyoming
Dearborn
Jackson
Climax-Scotts, Comstock, Galesburg-Augusta, Gull
Lake, Kalamazoo, Mattawan, Parchment, Portage,
Schoolcraft, Vicksburg
Albion, Athens, Battle Creek, Harper Creek, Homer,
Lakeview, Mar-Lee, Marshall, Pennfield, Tekonsha,
Union City
C.O.O.R. ISD: Crawford-AuSable, Fairview Area,
Houghton Lake, Mio-AuSable, Roscommon Area,
West Branch-Rose City
South Haven
Bath, Dansville, Dewitt, East Lansing, Grand Ledge,
Haslett, Holt/Diamondale, Lansing, Leslie, Mason,
Okemos, Stockbridge, Waverly, Webberville,
Williamston

Clare-Gladwin RESA: Beaverton, Clare, Farwell,
Gladwin, Harrison

Montcalm Area ISD: Carson City-Crystal, Central
Montcalm, Greenville, Lakeview, Montabella,
Tri County, Vestaburg
Genesee ISD: Atherton, Beecher, Bendle, Bentley,
Carman-Ainsworth, Clio, Davison, Fenton, Flint,
Flushing, Genesee, Goodrich, Grand Blanc, Kearsley,
Lake Fenton, Lakeville, Linden, Montrose, Mt. Morris,
Swartz Creek, Westwood Heights

Covert

Muskegon

Muskegon
North Central Michigan Emmet
Northwestern Michigan Grand Traverse
Oakland

Oakland

Schoolcraft

Southwestern Michigan Cass

St. Clair County

Washtenaw

Washtenaw

Wayne County

Wayne

West Shore

Notes: * denotes service district areas.

Clarenceville, Garden City, Livonia, Northville, Novi
(part), Plymouth-Canton

Algonac, Capac, East China, Marysville, Memphis,
Port Huron, Yale

Keeler, Hamilton

NOT INCLUDED: Dearborn, Garden
City, Highland Park, Livonia, Northville,
Plymouth, Canton (part)

Bear Lake, Free Soil, Kaleva-Norman-Dickson,
Ludington, Manistee, Mason County Central, Mason
County Eastern, Onekama, Walkerville

Crystal, Elbridge,
Weare

113

Table A.1.6: Descriptive Statistics, 2009-2016 High School Graduates

All Students

In

All

Variable

Panel A. Demographics
White
Black
Hispanic
Male
FRPL eligible
Special education
English language learner
Resides in CC district

0.760
0.150
0.041
0.490
0.333
0.082
0.025
0.779

Panel B. High School Academics
Math standardized score
Reading standardized score
School of choice
On-time graduation
Dual enrollment in HS

0.095
0.087
0.096
0.966
0.095

Panel C. One-Year College Enrollment
Community college
Vocational college
Four-year college
Any college

0.294
0.031
0.407
0.697

Out

0.906
0.015
0.043
0.498
0.320
0.085
0.010
0.000

0.169
0.141
0.104
0.972
0.121

0.226
0.046
0.393
0.642

Analysis Sample

All

In

Out

0.851
0.081
0.029
0.499
0.300
0.081
0.021
0.616

0.120
0.104
0.124
0.970
0.108

0.295
0.035
0.375
0.674

0.814
0.110
0.029
0.497
0.315
0.078
0.029
1.000

0.090
0.078
0.120
0.968
0.102

0.314
0.031
0.373
0.684

0.911
0.034
0.030
0.503
0.278
0.086
0.007
0.000

0.168
0.144
0.130
0.974
0.117

0.265
0.043
0.378
0.658

0.719
0.189
0.041
0.488
0.337
0.082
0.030
1.000

0.075
0.071
0.094
0.965
0.088

0.314
0.027
0.411
0.712

734,928

24,853
Observations
Notes: The “All Students” sample include all students who graduate from a traditional public high school in
Michigan between 2009 and 2016, take the Michigan Merit Exam (MME), and have non-missing geographic
and test score information. The “Analysis Sample” further restricts the sample to students who reside within
two miles of a community college district boundary. Students who attend alternative education high schools
or juvenile detention centers are not included in either sample.

572,581 162,347

64,667

39,814

114

Table A.1.7: First Stage Estimate of In-District Status on Tuition

Variable

In-District Status

No Controls Distance Controls All Controls
-1,797***
-1,814***
(269.0)
(240.0)

-1,795***
(240.6)

64,667
45.46
0.901

Observations
Partial F-Statistic
Adjusted R2
Notes: Each coefficient is estimated from a single regression and corresponds to λ
in equation (1.2), representing the difference in local community college tuition faced
by students residing inside of a community college district, as compared to students
residing outside of a community college district. All standard errors are clustered at the
boundary segment level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

64,667
55.67
0.905

64,667
56.05
0.905

115

Table A.1.8: Balance Tests of Student Characteristics

Outcome

White Male
(2)

(1)

FRPL

(3)

SPED

(4)

In-District Effect

0.001
(0.010)

-0.004
(0.005)

-0.015
(0.012)

-0.009***
(0.003)

Observations
Mean

64,667
0.851

64,667
0.499

64,667
0.300

Outcome

Math
Score
(6)

Reading
Score
(7)

On-Time

Grad
(8)

64,667
0.081

Dual
Enroll

(9)

ELL
(5)

0.006
(0.006)

64,667
0.021

Pred. CC
Enrollment

(10)

In-District Effect

0.012
(0.013)

0.015
(0.012)

-0.001
(0.003)

-0.008*
(0.004)

0.002
(0.001)

64,667
0.120

64,667
0.104

64,667
0.970

64,667
0.108

Observations
Mean
Notes: The sample consists of all students who reside within two miles of the nearest community
college district boundary segment and graduated from high school between 2009 and 2016. Each
coefficient is estimated from a single regression that regresses the student characteristic of interest
on a dummy variable for in-district status and the full set of boundary segment by year fixed
effects. The coefficients represent the average difference in characteristics among students who
reside within two miles of the same community college district boundary and graduate from high
school in the same year. All standard errors are clustered at the boundary segment level. ∗ p <
0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

64,667
0.295

116

Table A.1.9: Balance Tests of Census Tract Characteristics

Outcome

In-District Effect

Observations
Mean

Outcome

Median

HH Income

(1)

1,427
(1,065)

64,645
59,505

Poverty
Share

(2)

-0.003
(0.005)

64,667
0.118

Mean 3rd Grade

Math Score

(3)

-0.077
(0.116)

64,653
2.976

2 Bedroom
Rental Price

(4)

13.08
(12.69)

46,927
748.58

Single Parent

Non-White

Share

(5)

Share

(6)

High-Paying
Job Share

(7)

Job Growth
2004-2013

(8)

In-District Effect

0.001
(0.011)

0.001
(0.001)

0.008
(0.006)

0.004
(0.009)

64,667
0.142

64,667
0.245

64,667
0.387

Observations
Mean
Notes: The sample consists of all students who reside within two miles of the nearest community college
district boundary segment and graduated from high school between 2009 and 2016. Each coefficient
is estimated from a single regression that regresses the student characteristic of interest on a dummy
variable for in-district status and the full set of boundary segment by year fixed effects. The coefficients
represent the average difference in characteristics among students who reside within two miles of the same
community college district boundary and graduate from high school in the same year. All data comes from
the Equality of Opportunity Project and is publicly available at: https://opportunityinsights.org/data/. All
standard errors are clustered at the boundary segment level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

64,667
0.006

117

Table A.1.10: Balance Tests of Distance to Postsecondary Institutions

Outcome

Local
CC
(1)

Public

Four-Year

(2)

Private
Four-Year

(3)

Private

Vocational

(4)

In-District Effect

-1.462***
(0.178)

0.023
(0.432)

-0.929**
(0.442)

-0.717
(0.468)

64,667
10.30

64,667
19.93

64,667
23.73

64,667
19.50

Observations
Mean
Notes: The sample consists of all students who reside within two miles of the nearest
community college district boundary segment and graduated from high school between
2009 and 2016. Each coefficient is estimated from a single regression that regresses the
student characteristic of interest on a dummy variable for in-district status and the full
set of boundary segment by year fixed effects. The coefficients represent the average
difference in characteristics among students who reside within two miles of the same
community college district boundary and graduate from high school in the same year.
All standard errors are clustered at the boundary segment level. ∗ p < 0.10, ∗∗ p < 0.05,
∗∗∗ p < 0.01.

118

Table A.1.11: Effect of In-District Status and Reduced Tuition on College Enrollment

Outcome

Local
CC
(1)

Non-Local

CC
(2)

Vocational

College

(3)

Four-Year
College

(4)

Any

College

(5)

Panel A. All Cohorts
In-District Effect

0.064***
(0.007)

-0.028***
(0.006)

-0.007***
(0.002)

Tuition Effect

0.035***
(0.004)

-0.015***
(0.004)

-0.004***
(0.001)

Observations
Mean

64,667
0.209

64,667
0.089

64,667
0.035

Panel B. 2009-2011 Cohorts
0.060***
In-District Effect
(0.010)

-0.035***
(0.007)

-0.007**
(0.003)

Tuition Effect

0.036***
(0.006)

-0.021***
(0.006)

-0.004***
(0.001)

-0.010
(0.007)

-0.005
(0.003)

64,667
0.375

-0.005
(0.008)

-0.003
(0.005)

0.013**
(0.005)

0.007**
(0.003)

64,667
0.674

0.006
(0.008)

0.004
(0.004)

23,734
0.096

23,734
0.368

23,734
0.691

23,734
0.040

23,734
0.225

Observations
Mean
Notes: The sample in Panel A consists of all students who reside within two miles of the nearest
community college district boundary segment and graduated from high school between 2009 and
2016. Panel B further restricts the sample to students who graduated from high school between 2009
and 2011. In both panels, each coefficient is estimated from a single regression. The coefficients
in the “in-district effect” rows correspond to δ in equation (1.1), representing the estimated change
in the probability of an outcome due to a student residing in a community college district. The
coefficients in the “tuition effect” rows correspond to −β ∗ 1000, where β is defined as in equation
(1.3). These coefficients represent the estimate change in the probability of an outcome due to a
$1,000 decrease in the annual tuition rate at a student’s local community college. All regressions
include controls for a student’s race/ethnicity, gender, FRPL status, special education participation,
ELL status, math and reading test scores, school of choice participation, on-time graduation, and
dual enrollment experience, as well as the distance between the centroid of a student’s census block
of residence and the nearest campus of the local community college, the nearest vocational college,
the nearest in-state public university, and the nearest in-state private four-year college. All standard
errors are clustered at the boundary segment level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

119

Table A.1.12: Heterogeneous Effects by Graduation Year

Local
CC
(1)

Non-Local

CC
(2)

Vocational

College

(3)

Four-Year
College

(4)

Any

College

(5)

0.066***
(0.008)

-0.024***
(0.006)

-0.007***
(0.002)

Variable:
In-District Effect

-0.009
(0.007)

-0.001
(0.007)

0.019***
(0.006)

-0.017**
(0.007)

In-District x 2009-2011 Grad

-0.005
(0.008)

-0.010**
(0.004)

-0.001
(0.003)

64,667
0.209

64,667
0.089

64,667
0.035

64,667
0.375

Observations
Mean
Notes: The sample consists of all students who reside within two miles of the nearest community college district
boundary segment and graduated from high school between 2009 and 2016. The coefficients in the “in-district
effect” rows correspond to δ in equation (1.1) for the 2012-2016 cohorts, representing the estimated change in the
probability of an outcome due to a student residing in a community college district. The coefficients in the second
row represent the difference in the in-district effect between the 2009-2011 cohorts and the 2012-2016 cohorts.
All regressions include controls for a student’s race/ethnicity, gender, FRPL status, special education participation,
ELL status, math and reading test scores, school of choice participation, on-time graduation, and dual enrollment
experience, as well as the distance between the centroid of a student’s census block of residence and the nearest
campus of the local community college, the nearest vocational college, the nearest in-state public university, and the
nearest in-state private four-year college. All standard errors are clustered at the boundary segment level. ∗ p < 0.10,
∗∗ p < 0.05, ∗∗∗ p < 0.01.

64,667
0.674

120

Table A.1.13: Characteristics of Community and Vocational Colleges

Variable
Avg. Net Price
Instruction $ per FTE
% Full-Time Faculty
Transfer Rate
150% Graduation Rate
% Liberal Arts Degrees
Median Earnings
Median Debt

MI Community

Colleges

(1)

$5,325.38
$4,993.05

0.400
0.360
0.135
0.349

Vocational
Colleges

(2)

$14,004.62
$3,897.80

0.213
0.111
0.196
0.006

$29,326.50
$4,211.58

$29,018.95
$8,867.49

Difference

(3)

-$8,679.24
$1,095.25

0.188
0.249
-0.061
0.343
$307.55
-$4,655.91

21,720

144

28

191,394

Students
Institutions
Notes: All data comes from the U.S. Department of Education’s College Scorecard,
except for the transfer rate variable which is calculated on the full sample of Michigan’s
2009-2011 high school graduates who enroll in community or vocational colleges within
one year of high school graduation. All variables are averaged across all 2009-2016
high school graduates who enroll in college within one year of high school graduation
to reflect the characteristics of the colleges that students attend.

-
-

121

Table A.1.14: Effect of In-District Status and Reduced Tuition on College Completion

Semesters
Bachelor’s
of College Completed Four-Year Completion Completion Completion

Transfer to Certificate

Associate

Credits

(5)
0.005
(0.005)

0.003
(0.002)

(6)

0.018**
(0.008)

0.011**
(0.005)

Outcome

In-District Effect

Tuition Effect

(1)

(2)

(3)

(4)

0.344***
(0.097)

3.463***
(1.302)

0.206***
(0.062)

2.069***
(0.656)

0.011**
(0.005)

0.007**
(0.003)

-0.003
(0.004)

-0.002
(0.003)

23,734
8.133

23,734
76.46

23,734
0.055

23,734
0.115

23,734
0.126

Observations
Mean
Notes: The sample consists of all students who reside within two miles of the nearest community college district boundary
segment and graduated from high school between 2009 and 2011. Each coefficient is estimated from a single regression. The
coefficients in the “in-district effect” rows correspond to δ in equation (1.1), representing the estimated change in the probability
of an outcome due to a student residing in a community college district. The coefficients in the “tuition effect” rows correspond
to −β ∗ 1000, where β is defined as in equation (1.3). These coefficients represent the estimate change in the probability of
an outcome due to a $1,000 decrease in the annual tuition rate at a student’s local community college. All regressions include
controls for a student’s race/ethnicity, gender, FRPL status, special education participation, ELL status, math and reading test
scores, school of choice participation, on-time graduation, and dual enrollment experience, as well as the distance between the
centroid of a student’s census block of residence and the nearest campus of the local community college, the nearest vocational
college, the nearest in-state public university, and the nearest in-state private four-year college. All standard errors are clustered
at the boundary segment level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

23,734
0.316

122

Table A.1.15: Distribution of Degree Completion Increases Across Majors

General
Studies

Liberal
Arts
(2)

Health Business Technical

(3)

(4)

(5)

Prof.
(6)

Other

(7)

Outcome

(1)
Panel A. Associate Degree
In-District Effect

0.010***
(0.002)

-0.000
(0.002)

-0.002
(0.002)

-0.001
(0.002)

Tuition Effect

0.006***
(0.002)

-0.000
(0.001)

-0.001
(0.002)

-0.001
(0.001)

Observations
Mean

23,734
0.033

23,734
0.010

23,734
0.023

23,734
0.013

Panel B. Bachelor’s Degree
0.002**
In-District Effect
(0.001)

0.001
(0.005)

0.001
(0.003)

0.007*
(0.004)

Tuition Effect

0.001**
(0.001)

0.000
(0.003)

0.001
(0.002)

0.004*
(0.002)

-0.002
(0.002)

-0.001
(0.001)

23,734
0.016

-0.002
(0.004)

-0.001
(0.002)

-0.002
(0.002)

0.003*
(0.001)

-0.001
(0.001)

0.002*
(0.001)

23,734
0.016

23,734
0.016

0.009**
(0.004)

0.001
(0.003)

0.005***
(0.002)

0.001
(0.001)

23,734
0.003

23,734
0.100

23,734
0.034

23,734
0.067

23,734
0.051

23,734
0.038

23,734
0.022

Observations
Mean
Notes: In both panels, the sample consists of all students who reside within two miles of the nearest community col-
lege district boundary segment and graduated from high school between 2009 and 2011. Each coefficient is estimated
from a single regression. The coefficients in the “in-district effect” rows correspond to δ in equation (1.1), represent-
ing the estimated change in the probability of an outcome due to a student residing in a community college district.
The coefficients in the “tuition effect” rows correspond to −β ∗ 1000, where β is defined as in equation (1.3). These
coefficients represent the estimate change in the probability of an outcome due to a $1,000 decrease in the annual tu-
ition rate at a student’s local community college. All regressions include controls for a student’s race/ethnicity, gender,
FRPL status, special education participation, ELL status, math and reading test scores, school of choice participation,
on-time graduation, and dual enrollment experience, as well as the distance between the centroid of a student’s census
block of residence and the nearest campus of the local community college, the nearest vocational college, the nearest
in-state public university, and the nearest in-state private four-year college. All standard errors are clustered at the
boundary segment level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

123

Major
Group
General Studies

Liberal Arts

Health

Business

Technical

Professional

Table A.1.16: Academic Program Categories

Two-Digit
CIP Code CIP Title

24

1
3
5
16
23
26
27
30
38
40
42
45
50
54

51

52

4
10
11
14
15
41
46
47
48
49

9
12
13
19
22
25
31
43
44

Liberal Arts and Sciences, General Studies and Humanities

Agriculture, Agriculture Operations, and Related Sciences
Natural Resources and Conservation
Area, Ethnic, Cultural, and Gender Studies
Foreign Languages, Literatures, and Linguistics
English Language and Literatures
Biological and Biomedical Sciences
Mathematics and Statistics
Multi/Interdisciplinary Studies
Philosophy and Religious Studies
Physical Sciences
Psychology
Social Sciences
Visual and Performing Arts
History

Health Professions and Related Clinical Sciences

Business, Management, Marketing, and Related Support Services

Architecture and Related Services
Communications Technologies/Technicians and Support Services
Computer and Information Sciences and Support Services
Engineering
Engineering Technologies/Technicians
Science Technologies/Technicians
Construction Trades
Mechanic and Repair Technologies/Technicians
Precision Production
Transportation and Materials Moving

Communication, Journalism, and Related Programs
Personal and Culinary Services
Education
Family and Consumer Sciences/Human Sciences
Legal Professions and Studies
Library Science
Parks, Recreation, Leisue, and Fitness Studies
Security and Protective Services
Public Administration and Social Service Professions

124

Table A.1.17: Distribution of Bachelor’s Degree Increases Across Professional Majors

Outcome

In-District Effect

Tuition Effect

Protective
Service

(1)
0.001
(0.002)

0.001
(0.001)

Family &
Consumer
Sciences

(2)

-0.001
(0.001)

-0.001
(0.001)

Personal
Care &
Culinary

(3)
0.000
(0.000)

Legal
Studies

(4)
0.000
(0.001)

0.000
(0.000)

0.000
(0.000)

Education
& Library

Science

(5)

0.004**
(0.002)

0.002**
(0.001)

Comm. &
Journalism

Public
Admin.

(6)

-0.000
(0.002)

-0.000
(0.001)

(7)

0.002*
(0.001)

0.001
(0.001)

Parks, Rec.,
Leisure, &

Fitness

(8)

0.003**
(0.001)

0.002*
(0.001)

23,734
0.008

23,734
0.0002

23,734
0.018

23,734
0.001

23,734
0.004

Observations
Mean
Notes: The sample consists of all students who reside within two miles of the nearest community college district boundary segment and graduated
from high school between 2009 and 2011. Each coefficient is estimated from a single regression. The coefficients in the “in-district effect” rows
correspond to δ in equation (1.1), representing the estimated change in the probability of an outcome due to a student residing in a community
college district. The coefficients in the “tuition effect” rows correspond to −β ∗ 1000, where β is defined as in equation (1.3). These coefficients
represent the estimate change in the probability of an outcome due to a $1,000 decrease in the annual tuition rate at a student’s local community
college. All regressions include controls for a student’s race/ethnicity, gender, FRPL status, special education participation, ELL status, math and
reading test scores, school of choice participation, on-time graduation, and dual enrollment experience, as well as the distance between the centroid
of a student’s census block of residence and the nearest campus of the local community college, the nearest vocational college, the nearest in-state
public university, and the nearest in-state private four-year college. All standard errors are clustered at the boundary segment level. ∗ p < 0.10,
∗∗ p < 0.05, ∗∗∗ p < 0.01.

23,734
0.011

23,734
0.017

23,734
0.008

125

Table A.1.18: Heterogeneity by Student Characteristics

Outcome

Overall effect

One year enrollment:

Local
(1)

Non-
Local
(2)

Voc.
(3)

Four-
Year
(4)

Any
(5)

0.064***
(0.007)

-0.028***
(0.006)

-0.007***
(0.002)

-0.010
(0.007)

0.013***
(0.005)

Completion:

Assoc.

(6)
0.005
(0.005)

Bach.
(7)

0.018**
(0.008)

Panel A. FRPL Eligibility

Ineligible

Eligible

0.067***
(0.008)
0.056***
(0.009)

-0.033***
(0.007)
-0.015**
(0.006)

-0.007***
(0.002)
-0.009**
(0.003)

-0.010
(0.008)
-0.008
(0.008)

0.010
(0.006)
0.019***
(0.007)

0.006
(0.005)
0.002
(0.009)

0.015
(0.009)
0.028**
(0.013)

Ineligible = eligible?

0.244

0.026

0.647

0.859

0.312

0.686

0.436

Panel B. Gender

Female

Male

0.056***
(0.008)
0.072***
(0.008)

-0.034***
(0.007)
-0.022***
(0.008)

-0.011***
(0.003)
-0.004*
(0.002)

-0.005
(0.007)
-0.014*
(0.008)

-0.001
(0.009)
0.026***
(0.007)

0.004
(0.008)
0.006
(0.006)

0.019*
(0.011)
0.017*
(0.010)

Female = male?

0.008

0.138

0.033

0.206

0.017

0.832

0.913

Panel C. Test Score

Bottom quartile

Middle two quartiles

Top quartile

0.074***
(0.011)
0.075***
(0.008)
0.029***
(0.010)

-0.018**
(0.009)
-0.036***
(0.007)
-0.19**
(0.008)

-0.011***
(0.004)
-0.009***
(0.003)
0.001
(0.003)

-0.023**
(0.009)
-0.003
(0.007)
-0.008
(0.015)

0.021**
(0.011)
0.017***
(0.007)
-0.004
(0.008)

0.012
(0.010)
-0.001
(0.007)
0.011
(0.0010)

0.008
(0.011)
0.026**
(0.011)
0.014
(0.014)

Bottom = middle?
Top = middle?

0.934
0.000

0.031
0.017

0.677
0.022

0.061
0.702

0.730
0.029

0.363
0.288

0.150
0.488

64,667

64,667

64,667

23,734

64,667

64,667

N
Notes: For outcomes (1)-(5), the sample consists of all students who reside within two miles of the nearest community college
district boundary segment, graduated from high school between 2009 and 2016. For outcomes (6) and (7), the sample is further
restricted to students who graduated from high school between 2009 and 2011, and students who earn postsecondary degrees
in high school are dropped from the sample. Coefficients are estimated from regressions with interaction terms, as described in
section 1.5.3. All regressions include controls for a student’s race/ethnicity, gender, FRPL status, special education participation,
ELL status, math and reading test scores, school of choice participation, on-time graduation, and dual enrollment experience, as
well as the distance between the centroid of a student’s census block of residence and the nearest campus of the local community
college, the nearest vocational college, the nearest in-state public university, and the nearest in-state private four-year college.
All standard errors are clustered at the boundary segment level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

23,734

126

Table A.1.19: Balance Tests of Student Characteristics, Varying Bandwidths

Bandwidth: White Male
(2)
0.001
(0.003)

4 Miles
(N=145,775)

-0.015
(0.013)

(1)

FRPL

(3)

-0.009
(0.016)

SPED

(4)

-0.008*
(0.004)

ELL Math Reading On-Time
(5)
0.004
(0.003)

(6)
0.018
(0.016)

0.026**
(0.012)

-0.001
(0.003)

(8)

(7)

3 Miles
(N=102,791)

-0.008
(0.009)

-0.000
(0.004)

-0.012
(0.012)

-0.009**
(0.004)

0.004
(0.004)

0.022*
(0.013)

0.023**
(0.011)

2 Miles
(N=64,667)

0.001
(0.010)

-0.004
(0.005)

-0.015
(0.012)

-0.009***
(0.003)

0.006
(0.006)

0.012
(0.013)

0.015
(0.012)

1 Mile
(N=31,541)

0.016
(0.013)

0.004
(0.007)

-0.023
(0.019)

-0.010**
(0.005)

0.010
(0.010)

0.020
(0.016)

0.008
(0.015)

0.5 Miles
(N=15,185)

0.020
(0.014)

0.005
(0.009)

-0.032
(0.025)

-0.009
(0.007)

0.008
(0.010)

0.017
(0.020)

0.026
(0.022)

-0.001
(0.002)

-0.001
(0.003)

-0.002
(0.003)

-0.003
(0.006)

Dual
(9)

-0.007
(0.006)

-0.007*
(0.004)

-0.008*
(0.004)

-0.011*
(0.006)

-0.014*
(0.008)

-0.011
(0.018)

0.001
(0.003)

-0.015
(0.052)

-0.014
(0.022)

0.051**
(0.023)

0.054**
(0.025)

-0.023
0.1 Miles
(0.028)
(N=1,136)
Notes: The sample consists of all students who reside within the specified distance of the nearest community college district boundary
segment and graduated from high school between 2009 and 2016. Each coefficient is estimated from a single regression that regresses
the student characteristic of interest on a dummy variable for in-district status and the full set of boundary segment by year fixed effects.
The coefficients represent the average difference in characteristics among students who reside within the specified distance of the same
community college district boundary and graduate from high school in the same year. All standard errors are clustered at the boundary
segment level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

-0.015
(0.024)

0.020
(0.070)

127

Table A.1.20: Local Community College Enrollment Results, Within Same School District

Local CC Enrollment

Bachelor’s Degree

Main Strategy

School District Main Strategy

School District

(1)

(2)

(3)

In-District Effect

Tuition Effect

0.064***
(0.007)

0.035***
(0.004)

0.050***
(0.014)

0.032***
(0.011)

0.018**
(0.008)

0.011**
(0.005)

(4)
0.015
(0.022)

0.011
(0.015)

6,946
0.292

64,667
0.209

23,734
0.316

17,783
0.233

Observations
Mean
Notes: Columns (1) and (3) repeat the estimates for local community college enrollment and bachelor’s degree
completion presented in Tables 5 and 7, respectively. Here, the sample consists of all students who reside within
two miles of the nearest community college district boundary segment and graduated from high school between
2009 and 2016. Standard errors are cluster at the boundary segment level. Columns (2) and (4) present reduced
form and 2SLS estimates on the sample of school districts that overlap community college districts (see Section
1.5.4). The sample consists of all students who reside in one of the overlapping school districts and graduated
from high school between 2009 and 2016. In these columns, standard errors are clustered at the school district
level. All regressions include controls for a student’s race/ethnicity, gender, FRPL status, special education
participation, ELL status, math and reading test scores, school of choice participation, on-time graduation, and
dual enrollment experience, as well as the distance between the centroid of a student’s census block of residence
and the nearest campus of the local community college, the nearest vocational college, the nearest in-state public
university, and the nearest in-state private four-year college. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

128

Table A.1.21: Full Enrollment Results for Within Same School District Sample

Outcome

In-District Effect

Local
CC
(1)

Non-Local

CC
(2)

0.050***
(0.014)

-0.020*
(0.012)

Tuition Effect

0.032***
(0.011)

-0.013
(0.008)

Vocational

College

(3)

-0.001
(0.005)

-0.001
(0.003)

Four-Year
College

(4)

-0.016
(0.018)

-0.010
(0.012)

Any

College

(5)
0.014
(0.016)

0.009
(0.010)

17,783
0.067

17,783
0.035

17,783
0.233

17,783
0.643

17,783
0.336

Observations
Mean
Notes: The sample consists of all students who reside within a school district that intersects a
community college district and graduated from high school between 2009 and 2016. Each coefficient
is estimated from a single regression that includes a full set of school district by graduation year fixed
effects. The coefficients in the “policy effect” represent the estimated change in the probability of
an outcome due to a student residing in a community college district. The coefficients in the “tuition
effect” row represent the estimate change in the probability of an outcome due to a $1,000 decrease
in the annual tuition rate at a student’s local community college. All regressions include controls
for a student’s race/ethnicity, gender, FRPL status, special education participation, ELL status, math
and reading test scores, school of choice participation, on-time graduation, and dual enrollment
experience, as well as the distance between the centroid of a student’s census block of residence
and the nearest campus of the local community college, the nearest vocational college, the nearest
in-state public university, and the nearest in-state private four-year college. All standard errors are
clustered at the school district level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

129

Table A.1.22: Placebo Tests

Local CC Enrollment

Bachelor’s Degree

Further In Further Out Further In Further Out

(3)

(4)

(1)
0.005
(0.007)

94,582
0.242

Estimate

(2)

0.007*
(0.004)

-0.012*
(0.007)

-0.004
(0.008)

50,527
0.159

33,676
0.318

19,390
0.314

Observations
Mean
Notes: Each column reports the estimates of a placebo test that alters the boundaries of the
community college districts. Columns (1) and (3) contract all community college districts
by 2 miles; columns (2) and (4) expand all community college districts by 2 miles. Each
sample consists of all students who reside within two miles of the nearest placebo commu-
nity college district boundary segment and graduated from high school between 2009 and
2016. Each column then estimates δ from equation (1.1) using the constructed placebo
community college district boundaries. All regressions include controls for a student’s
race/ethnicity, gender, FRPL status, special education participation, ELL status, math and
reading test scores, school of choice participation, on-time graduation, and dual enroll-
ment experience, as well as the distance between the centroid of a student’s census block
of residence and the nearest campus of the local community college, the nearest vocational
college, the nearest in-state public university, and the nearest in-state private four-year col-
lege. All standard errors are clustered at the placebo boundary segment level. ∗ p < 0.10,
∗∗ p < 0.05, ∗∗∗ p < 0.01.

130

APPENDIX B

CHAPTER 2 APPENDIX

B.1 Figures & Tables

Figure B.1.1: Differences in Course-Taking and Credit Completion by CC Program Group

(a) Share of Courses Taken

(b) Share of Credits Completed

Notes: Each bar represents the share of courses taken or credits completed in different areas of study among students
pursuing a program in the designated program group (e.g., business, health, etc.). The sample consists of all students
who enroll in Michigan community colleges within six months of high school graduation. Only courses taken and
credits completed within the first academic year following high school graduation are included.

131

020406080100OtherLaw Enf.STEMTradesHealthBusinessNon-Voc.Non-Voc.BusinessHealthTradesSTEMLaw Enf.Other020406080100OtherLaw Enf.STEMTradesHealthBusinessNon-Voc.Non-Voc.BusinessHealthTradesSTEMLaw Enf.OtherFigure B.1.2: Labor Market Shocks in Michigan, 2001-2017

(a) Layoff Events

(b) Total Job Losses

132

050100150200Number of Layoff Events200120032005200720092011201320152017Academic YearTotal Layoff EventsWARN Mass LayoffsWARN Plant ClosuresPrison Closures010,00020,00030,00040,000Number of Jobs Losses200120032005200720092011201320152017Academic YearTotal Layoff EventsWARN Mass LayoffsWARN Plant ClosuresPrison ClosuresFigure B.1.3: Average Layoffs in Michigan Counties, 2001-2017

133

IronDeltaLuceMarquetteChippewaKentAlgerGogebicHuronBaragaSanilacOntonagonLakeIoniaBaySchoolcraftHoughtonMackinacAlleganClareIoscoOaklandBarryTuscolaAlconaEatonCassSaginawLapeerMenomineeNewaygoWayneSt. ClairAlpenaJacksonGratiotCalhounDickinsonAntrimClintonLenaweeOttawaBerrienMasonOscodaIsabellaOtsegoMontcalmInghamCheboyganMonroeBranchGeneseeEmmetWexfordOsceolaHillsdaleOceanaMecostaMidlandOgemawGladwinKalkaskaCrawfordManisteeWashtenawVan BurenLivingstonPresque IsleMacombMissaukeeArenacBenzieMuskegonSt. JosephKalamazooRoscommonShiawasseeMontmorencyLeelanauCharlevoixKeweenawGrand TraverseKeweenawCharlevoixLeelanauAvg. Layoffs per 10,000 Working-Age Residents0.000 - 2.1042.105 - 5.9885.989 - 9.1329.133 - 11.21311.214 - 13.48813.489 - 18.95618.957 - 29.33329.334 - 47.984Figure B.1.4: Correlation Between National and State-Specific Industry Employment Shares, 2016

134

0.2.4.6National Employment Share0.2.4.6Michigan Employment ShareBusinessHealthTradesSTEMLaw Enf.OtherFigure B.1.5: Distribution of Layoffs by County, 2001-2017

Notes: The sample consists of the 66 (79.5%) Michigan counties that experience layoffs between 2001 and 2017. The
left-hand panel shows the total number of layoffs in each type of occupation per 10,000 working-age residents
(averaged over the time frame). The right-hand panel shows the share of total layoffs occurring in each type of
occupation.

135

0200400600800Layoffs per 10,000 ResidentsWayneOaklandMacombKentGeneseeWashtenawInghamKalamazooOttawaSaginawLivingstonMuskegonSt. ClairJacksonBerrienMonroeCalhounAlleganEatonBayLenaweeLapeerGrand TraverseMidlandVan BurenClintonIsabellaShiawasseeMarquetteIoniaMontcalmSt. JosephBarryTuscolaCassNewaygoHillsdaleBranchGratiotChippewaSanilacDeltaHoughtonEmmetWexfordHuronClareMasonDickinsonOceanaGladwinManisteeIoscoOtsegoMenomineeOsceolaAntrimBenzieArenacCrawfordLakeIronAlgerBaragaSchoolcraftOntonagon0.2.4.6.81Share of Total LayoffsWayneOaklandMacombKentGeneseeWashtenawInghamKalamazooOttawaSaginawLivingstonMuskegonSt. ClairJacksonBerrienMonroeCalhounAlleganEatonBayLenaweeLapeerGrand TraverseMidlandVan BurenClintonIsabellaShiawasseeMarquetteIoniaMontcalmSt. JosephBarryTuscolaCassNewaygoHillsdaleBranchGratiotChippewaSanilacDeltaHoughtonEmmetWexfordHuronClareMasonDickinsonOceanaGladwinManisteeIoscoOtsegoMenomineeOsceolaAntrimBenzieArenacCrawfordLakeIronAlgerBaragaSchoolcraftOntonagonLow SkillCC BusinessCC HealthCC TradesCC STEMCC Law Enf.CC OtherHigh SkillFigure B.1.6: Robustness Checks for Pooled Specification

(a) Different Control Variables

(b) Trends and CZ Fixed Effects

(c) Exclusion of Layoff Events

(d) Non-Linear Specifications (Semi-Elasticities)

136

Main SpecificationControlling for Non-CCLayoffsControlling Separatelyfor High- andLow-Skill LayoffsIncludingCounty-by-Year FE-.03-.02-.010.01Main SpecificationCounty-by-ProgramLinear Time TrendsYear-by-Program-by-CZFixed Effects-.03-.02-.010.01Main SpecificationOnly Plant ClosingsOnly Events with 50 orMore Job Losses-.03-.02-.010.01Main SpecificationInverse HyperbolicSine (IHS)Fractional Logit-.03-.02-.010.01Figure B.1.7: Substitution into Program Groups Requiring Similar Skills

137

-2-10123Effect Size0.2.4.6.81Skill DistanceBusiness Layoff-1-.50.51Effect Size0.2.4.6.81Skill DistanceHealth Layoff-.4-.20.2.4Effect Size0.2.4.6.81Skill DistanceSkilled Trades Layoff-2-1012Effect Size0.2.4.6.81Skill DistanceSTEM Layoff-.4-.20.2.4Effect Size0.2.4.6.81Skill DistanceLaw Enforcement Layoff-2-1012Effect Size0.2.4.6.81Skill DistanceOther LayoffBusinessHealthTradesSTEMLaw Enf.OtherFigure B.1.8: Relationship Between Substitution Effects and Skill Distance

138

-1012Effect Size.2.4.6.81Skill Distancen = 30    RMSE = 0.456Effect = 0.384 - 0.548 Similarity    R2 = 7.3%Figure B.1.9: Alternative Measures of Skill Distance

(a) Differences in Skill Levels Only

(b) Differences in Skill Importance Only

139

-1012Effect Size.2.4.6.81Skill Distancen = 30    RMSE = 0.432Effect = 0.558 - 0.728 SkillOnly    R2 = 16.6%-1012Effect Size.2.4.6.81Skill Distancen = 30    RMSE = 0.441Effect = 0.489 - 0.652 ImpOnly    R2 = 13.2%Figure B.1.10: Heterogeneous Own-Layoff Effects

(a) Heterogeneity by Gender

(b) Heterogeneity by County Urbanicity

140

-4-202Own-Layoff Effect SizeBusinessHealthTradesSTEMLaw Enf.OtherMaleFemale-3-2-1012Own-Layoff Effect SizeBusinessHealthTradesSTEMLaw Enf.OtherUrbanRuralFigure B.1.11: Robustness Checks for Own-Layoff Effects

(a) Weighting for Heteroskedasticity

(b) County-Specific Time Trends

(c) Cohort-by-Commuting Zone Fixed Effects

(d) Dropping 2009 Cohort

(e) Dropping Students in Multiple Programs

(f) Non-Linear Specifications

141

-2-1.5-1-.50.5Own-Layoff Effect SizeBusinessHealthTradesSTEMLaw Enf.OtherNo WeightsSHW Weights-3-2-101Own-Layoff Effect SizeBusinessHealthTradesSTEMLaw Enf.OtherNo TrendsCounty Trends-3-2-101Own-Layoff Effect SizeBusinessHealthTradesSTEMLaw Enf.OtherMain Spec.CZ-by-Year FEs-2-1.5-1-.50.5Own-Layoff Effect SizeBusinessHealthTradesSTEMLaw Enf.OtherAll CohortsDropping 2009-2-1.5-1-.50.5Own-Layoff Effect SizeBusinessHealthTradesSTEMLaw Enf.OtherAll StudentsSingle Majors-.15-.1-.050.05.1Own-Layoff Semi-Elasticity (at mean)BusinessHealthTradesSTEMLaw Enf.OtherLinearIHSPoissonFractional LogitMultinomial LogitTable B.1.1: Programs Offered by Michigan’s Community Colleges

Variable:
Panel A. All Programs
Total Programs
Vocational Programs
Non-Vocational Programs
Share Vocational

Mean
(1)

S.D. Min. Max.
(2)
(4)

(3)

116.54
95.29
21.25
0.81

67.18 41.00
59.00 33.00
13.03
5.00
0.56
0.10

319.00
280.00
51.00
0.94

Panel B. Associate Programs
Total Programs
Vocational Programs
Non-Vocational Programs
Share Vocational

59.75
45.07
14.68
0.75

30.11
24.42
9.94
0.12

10.00
5.00
2.00
0.49

142.00
124.00
37.00
0.92

56.79
50.21
6.57
0.88

177.00
158.00
21.00
1.00

Panel C. Certificate Programs
Total Programs
Vocational Programs
Non-Vocational Programs
Share Vocational
Notes: The sample consists of Michigan’s 28 community colleges during
the academic year 2011-2012. Vocational programs are defined as those
which can be matched to an occupation that is attainable by community col-
lege graduates. Non-vocational programs are all other programs offered by
Michigan’s community colleges.

40.52
17.00
36.47 13.00
0.00
5.45
0.08
0.67

142

Table B.1.2: Program Groups and Associated Occupation Codes

Program Group SOC SOC Title
Business

11
13
23
41
43

Management
Business and Financial
Legal
Sales and Related
Office and Administrative Support

Health

Trades

STEM

Law Enf.

Other

29
31

37
45
47
49
51
53

15
17
19

33

21
25
27
35
39

Healthcare Practitioners and Technical
Healthcare Support

Building and Grounds Cleaning and Maintenance
Farming, Fishing, and Forestry
Construction and Extraction
Installation, Maintenance, and Repair
Production*
Transportation and Material Moving**

Computer and Mathematical
Architecture and Engineering
Life, Physical, and Social Science

Protective Service

Community and Social Service
Education, Training, and Library
Arts, Design, Entertainment, Sports, and Media
Food Preparation and Serving Related
Personal Care and Service

* Programs matched to the 3-digit code 51-3 (Food Processing Workers) are included in
the “Other” group because they are generally part of Culinary Arts programs that are mostly
matched to the 2-digit code 35 (Food Preparation and Serving Related). Results are robust to
including these programs in either group.
** Programs matched to the 6-digit code 53-3011 (Ambulance Drivers and Attendants) are in-
cluded in the “Health” group because they are generally part of Emergency Medical Services
programs that are mostly matched to the 2-digit code 29 (Healthcare Practitioners and Techni-
cal). Results are robust to including these programs in either group.

143

Table B.1.3: Summary Statistics of Michigan’s High School Graduates

Variable:
White
Black
Hispanic
Male
Economically Disadvantaged
English Language Learner
Standardized Math Score
Standardized Reading Score
On-Time Graduation

All

Grads

(1)
0.760
0.150
0.041
0.490
0.333
0.025
0.095
0.087
0.971

CC
Voc.
(2)
0.738
0.176
0.046
0.537
0.366
0.039
-0.165
-0.205
0.984

CC

Non-Voc.

(3)
0.789
0.128
0.040
0.465
0.324
0.036
-0.028
-0.048
0.986

Other
College

(4)
0.785
0.128
0.027
0.443
0.222
0.010
0.532
0.524
0.997

No

College

(5)
0.723
0.178
0.057
0.543
0.461
0.035
-0.305
-0.303
0.931

Students
Share of Graduates
Notes: The sample consists of all graduates of Michigan public high schools from 2009 to 2016 who have non-missing
demographic and geographic information. College and program choices are defined as a student’s first enrollment choice
within 6 months of graduating high school. For example, the sample in column (2) consists of all students who first
enroll in vocational programs in Michigan’s community colleges within 6 months of high school graduation.

259,072
0.353

103,032
0.140

306,532
0.417

734,928
1.000

66,292
0.090

144

Table B.1.4: Summary Statistics of Vocational Students by Program

Business Health Trades STEM Law Enf. Other

Students
Share of Vocational Students
Notes: The sample consists of all graduates of Michigan public high schools from 2009 to 2016 who have
non-missing demographic and geographic information and enroll in a vocational program at one of the state’s
community colleges within 6 months of high school graduation.

12,979
0.196

16,082
0.243

15,080
0.227

5,387
0.081

Variable:
White
Black
Hispanic
Male
Economically Disadvantaged
English Language Learner
Standardized Math Score
Standardized Reading Score
On-Time Graduation

(1)
0.747
0.169
0.041
0.588
0.329
0.044
-0.056
-0.162
0.987

(2)
0.705
0.203
0.051
0.216
0.415
0.053
-0.260
-0.231
0.984

(3)
0.837
0.088
0.045
0.943
0.348
0.034
-0.193
-0.398
0.978

(6)
0.704
0.213
0.046
0.396
0.366
0.019
-0.242
-0.162
0.984

(4)
0.759
0.146
0.042
0.855
0.338
0.048
0.069
-0.072
0.984

8,476
0.128

(5)
0.750
0.171
0.049
0.653
0.389
0.031
-0.306
-0.316
0.984

8,288
0.125

145

Table B.1.5: Industries with Highest Concentration of Occupation Groups

NAICS Industry Title
Business
524
522
425

Insurance Carriers and Related Activities
Credit Intermediation and Related Activities
Wholesale Electronic Markets and Agents and Brokers

Health
621
623
622

Trades
212
811
484

STEM
511
516
518

Ambulatory Health Care Services
Nursing and Residential Care Facilities
Hospitals

Mining (except Oil and Gas)
Repair and Maintenance
Truck Transportation

Publishing Industries (except Internet)
Internet Publishing and Broadcasting
Data Processing, Hosting, and Related Services

Law Enforcement
482
921
922

Rail Transportation
Executive, Legislative, and Other General Government Support
Justice, Public Order, and Safety Activities

Other
515
812
624

Broadcasting (except Internet)
Personal and Laundry Services
Social Assistance

α

0.429
0.443
0.470

0.414
0.508
0.544

0.386
0.449
0.623

0.187
0.216
0.300

0.005
0.010
0.411

0.228
0.313
0.369

146

Table B.1.6: Correlation Between Occupation Composition Across Industries

Business Health Trades STEM Law Enf. Other

Business

1.000

Health

Trades

STEM

-0.133

1.000

-0.258

-0.212

1.000

0.328

-0.106

-0.190

1.000

Law Enf.

-0.106

-0.002

-0.098

-0.051

1.000

-0.138

Other
Notes: Each cell displays a pairwise correlation between the industry employment shares
for the occupation groups of interest. See Section 2.4.1 for more information.

-0.026

1.000

0.071

-0.360

-0.011

147

Table B.1.7: Summary Statistics of Layoffs in Michigan, 2001-2017

(3)

Mean
(1)

S.D. Min. Max.
Layoff category:
(2)
(4)
Panel A. Layoffs per 10,000 Working-Age Residents
Non-CC Low Skill
CC Business
CC Health
CC Trades
CC STEM
CC Law Enf.
CC Other
Non-CC High Skill

16.395 0.000
0.000
2.991
2.647
0.000
0.000
7.134
0.000
0.991
6.302
0.000
0.000
0.596
4.483
0.000

5.250
1.024
0.210
2.080
0.307
0.518
0.106
1.263

290.3
45.75
88.23
95.56
14.98
138.9
14.10
69.81

County-Year Obs.

1,411

1,411

1,411

1,411

Panel B. Share of Total Layoffs
(County-Year Pairs with Non-Zero Total Layoffs)
Non-CC Low Skill
CC Business
CC Health
CC Skilled Trades
CC STEM
CC Law Enf.
CC Other
Non-CC High Skill

0.142
0.155
0.028
0.066
0.000
0.070
0.000
0.120
0.037
0.000
0.0844 0.000
0.029
0.000
0.002
0.075

0.512
0.118
0.019
0.173
0.033
0.020
0.015
0.114

0.909
0.451
0.552
0.648
0.234
0.432
0.219
0.510

369

369

County-Year Obs.
Notes: The sample consists of all county-year observations from
2001 to 2017. Layoffs in each category are estimated using local
industry layoffs and national occupation-by-industry shares. See
Section 2.4.1 for more details.

369

369

148

Table B.1.8: Largest Layoffs by Occupation Group, 2001-2017

County
Business
Lake
Iosco
Ontonagon

Health
Midland
Gladwin
Ontonagon

Trades
Antrim
Ontonagon
Wexford

STEM
Antrim
Ingham
Midland

Year

Size Largest Related Layoff (Jobs Lost)

2005
2008
2009

27.88 Michigan Youth Correctional Facility (204)
29.02 Kalitta Air (219)
45.75 SmurfitStone Container Corp. (150)

2015
2015
2009

13.95 MidMichigan Health - Stratford Village (143)
29.72 MidMichigan Health - Gladwin Pines (85)
88.23 Maple Manor Nursing Home (62)

2007
2009
2010

61.18 Dura Automotive Systems (300)
69.30 SmurfitStone Container Corp. (150)
95.56 AAR Mobility Systems (282)

2007
2004
2015

61.18 Dura Automotive Systems (300)
9.987 General Motors (3,975)
14.98 Dow Chemical Company (700)

Law Enforcement
2011
Lake
2009
Arenac
Lake
2005

87.01 Northlake Correctional Facility (146)
131.2 Standish Maximum Facility (281)
138.9 Michigan Youth Correctional Facility (204)

Other
Oceana
Hillsdale
Ontonagon
Notes: Size is measured as the estimated number of layoffs per 10,000 working-age residents in
the county.

6.03 Double JJ Resort (150)
7.45
14.10 SmurfitStone Container Corp. (150)

2008
2012
2009

The Manor Residential Treatment Facility (140)

149

Table B.1.9: Effect of Job Losses on Enrollment in Related Community College Programs

Layoffs per 10,000 in:
Year following graduation

Senior year of H.S.

Junior year of H.S.

Sophomore year of H.S.

Freshman year of H.S.

8th grade

7th grade

6th grade

5th grade

Enrollment in Occupation Group
Programs per 100 H.S. Graduates
(4)
(1)
0.007
(0.005)

(3)

(2)

-0.012**
(0.006)

-0.014**
(0.007)

-0.014**
(0.007)

-0.011*
(0.006)

-0.002
(0.004)

-0.003
(0.005)

-0.001
(0.005)

-0.008**
(0.004)

-0.008*
(0.004)

-0.006
(0.004)

-0.004
(0.004)

-0.005
(0.004)

-0.002
(0.004)

-0.007
(0.005)

-0.004
(0.004)

0.005
(0.005)

0.007
(0.006)

-0.002
(0.004)

-0.000
(0.004)

0.002
(0.005)

0.004
(0.005)

1.57
3,984
0.488

1.57
3,984
0.489

1.57
3,984
0.490

1.57
3,984
0.490

Outcome Mean
County-Program-Year Obs.
R-squared
Notes: The unit of observation is a county-cohort-program triad. Outcomes are measured
as the number students who initially enroll in a given vocational program within 6 months
of high school graduation per 100 graduates in the county. The coefficients in each column
are estimated from a separate regression and represent variants of β in equation (2.3), the
effect of an additional layoff per 10,000 working age residents in a given occupation group
on enrollment in corresponding programs. All regressions include controls for the share
of graduates that are white, male, and categorized as economically disadvantaged; average
11th grade math and reading test scores; and the county unemployment rate and logged
size of the labor force during a cohort’s senior year of high school. All standard errors are
clustered at the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

150

Table B.1.10: Effect of Job Losses in Alternative Geographic Areas

Layoffs per 10,000 in:
Own county, t-1

Rest of state, t-1

Rest of commuting zone, t-1

State less commuting zone, t-1

Enrollment in Occupation Group

Programs per 100 Vocational Students

(1)

(2)

(3)

-0.012**
(0.006)

-0.012**
(0.006)

-0.012**
(0.006)

0.003
(0.012)

-0.008
(0.009)

0.007
(0.013)

1.57
3,984
0.476

1.57
3,936
0.479

1.57
3,984
0.476

Outcome Mean
County-Program-Year Obs.
R-squared
Notes: The unit of observation is a county-cohort-program triad. Outcomes are measured
as the number students who initially enroll in a given vocational program within 6 months
of high school graduation per 100 graduates in the county. The coefficients in each column
are estimated from a separate regression and represent variants of β in equation (2.3), the
effect of an additional layoff per 10,000 working age residents in a given occupation group
on enrollment in corresponding programs. All regressions include controls for the share of
graduates that are white, male, and categorized as economically disadvantaged; average 11th
grade math and reading test scores; and the county unemployment rate and logged size of the
labor force during a cohort’s senior year of high school. All standard errors are clustered at
the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

151

Table B.1.11: Effect of Community College Layoffs on Overall Vocational Program Enrollment

Layoffs per 10,000 in:
Business, t-1

Health, t-1

Skilled Trades, t-1

STEM, t-1

Law Enforcement, t-1

Other, t-1

P-Value for Joint Test

County-Specific Trends
Year-by-CZ Fixed Effects

Vocational Enrollment per 100 Graduates

(1)
0.009
(0.013)

(2)
0.016
(0.017)

(3)
0.003
(0.012)

0.002
(0.005)

0.002
(0.002)

0.018
(0.015)

-0.000
(0.002)

0.012
(0.027)

0.351

-0.006
(0.005)

0.001
(0.004)

0.001
(0.018)

-0.001
(0.002)

0.021
(0.024)

0.607

X

0.011*
(0.006)

0.003
(0.003)

0.002
(0.014)

-0.000
(0.002)

0.015
(0.023)

0.314

X

9.40
664
0.761

9.40
664
0.671

9.40
656
0.809

Outcome Mean
County-Year Obs.
R-squared
Notes: The unit of observation is a county-cohort pair. Outcomes are measured as the
number of students who enroll in vocational community college programs within 6 months
of high school graduation, per 100 high school graduates in the county and cohort. The
coefficients in each column are estimated from a separate regression and represent the β
parameters in equation (2.4), the effect of an additional layoff per 10,000 working age
residents in a given occupation group on the outcome of interest. The numbers in brackets
below the estimates are the estimated elasticities at the mean dependent and independent
variable values. All regressions include controls for the share of graduates that are white,
male, and categorized as economically disadvantaged; average 11th grade math and reading
test scores; and the county unemployment rate, logged size of the labor force, and the
number of layoffs per 10,000 working-age residents in non community college occupations
during a cohort’s senior year of high school. All standard errors are clustered at the county
level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

152

Table B.1.12: Effect of Layoffs on College Enrollment Outcomes

Enrollment per 100 Graduates in:

No Formal

CC Vocational

CC Non-Voc.

Programs

Programs

Layoffs per 10,000 in:
Panel A. Total layoffs
All occupations, t-1

Outcome Mean
County-Year Obs.
R-Squared

College

(1)

-0.013**
(0.006)

39.60
664
0.787

Panel B. Layoffs by skill group
Low-skill
occupations, t-1

-0.004
(0.020)

Community college
occupations, t-1

High-skill
occupations, t-1

-0.041
(0.035)

0.058
(0.077)

(2)

(3)

-0.004*
(0.002)

9.40
664
0.670

-0.012
(0.013)

0.004
(0.017)

-0.002
(0.037)

0.005
(0.005)

12.56
664
0.731

0.019
(0.016)

0.011
(0.021)

-0.069
(0.052)

Four-Year
Colleges

(4)

0.012**
(0.005)

38.44
664
0.865

-0.002
(0.022)

0.026
(0.027)

0.012
(0.053)

9.40
664
0.670

38.44
664
0.865

12.56
664
0.732

39.60
664
0.788

Outcome Mean
County-Year Obs.
R-Squared
Notes: The unit of observation is a county-cohort pair. Outcomes are measured as the number of students
who enroll in vocational community college programs within 6 months of high school graduation, per
100 high school graduates in the county and cohort. The coefficients in each column are estimated from
a separate regression and represent the β parameters in equation (2.4), the effect of an additional layoff
per 10,000 working age residents in a given occupation group on the outcome of interest. The numbers in
brackets below the estimates are the estimated elasticities at the mean dependent and independent variable
values. All regressions include controls for the share of graduates that are white, male, and categorized
as economically disadvantaged; average 11th grade math and reading test scores; and the county unem-
ployment rate and logged size of the labor force during a cohort’s senior year of high school. All standard
errors are clustered at the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

153

Table B.1.13: Effect of Layoffs on Composition of Vocational Students

% White

% Male

Layoffs per 10,000 in:
Business, t-1

Health, t-1

Skilled Trades, t-1

STEM, t-1

Law Enforcement, t-1

Other, t-1

(1)
0.007
(0.004)

0.004
(0.003)

-0.000
(0.001)

0.008
(0.006)

0.000
(0.001)

-0.016
(0.012)

P-Value for Joint Test

0.456

(2)

-0.005
(0.009)

0.005
(0.003)

0.001
(0.002)

-0.003
(0.009)

0.002
(0.002)

-0.001
(0.011)

0.638

% Econ.

Dis.
(3)

-0.005
(0.008)

-0.002
(0.002)

-0.000
(0.001)

-0.007
(0.008)

0.001
(0.001)

-0.009
(0.006)

0.217

Avg. Math

Score
(4)
0.011
(0.008)

0.002
(0.002)

-0.002
(0.002)

-0.009
(0.008)

-0.001
(0.001)

-0.010
(0.014)

0.217

Avg. Read

Score
(5)

-0.003
(0.008)

-0.001
(0.002)

-0.000
(0.002)

-0.005
(0.010)

-0.003
(0.002)

0.010
(0.011)

0.827

0.870
657
0.728

0.531
657
0.220

0.393
657
0.528

Outcome Mean
County-Year Obs.
R-Squared
Notes: The unit of observation is a county-cohort pair. Outcomes are measured as the mean characteristic across all students
who enroll in vocational programs. The coefficients in each column are estimated from a separate regression and represent the
β parameters in equation (2.4), the effect of an additional layoff per 10,000 working age residents in a given occupation group
on the outcome of interest. All regressions include controls for the share of graduates that are white, male, and categorized as
economically disadvantaged; average 11th grade math and reading test scores; and the county unemployment rate, logged size
of the labor force, and the number of layoffs per 10,000 working-age residents in non community college occupations during a
cohort’s senior year of high school. All standard errors are clustered at the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

-0.067
657
0.474

-0.144
657
0.389

154

Table B.1.14: Effect of Layoffs on First-Year Course-Taking

Layoffs per 10,000 in:
Business, t-1

Health, t-1

Skilled Trades, t-1

STEM, t-1

Law Enforcement, t-1

Other, t-1

Total
Credits

(1)
0.007
(0.216)

0.019
(0.086)

0.019
(0.036)

0.044
(0.346)

0.034
(0.034)

0.140
(0.705)

Vocational

Credits

(2)

-0.082
(0.108)

0.029
(0.050)

0.000
(0.018)

0.006
(0.143)

0.009
(0.018)

-0.150
(0.329)

Non-Voc.
Credits

(3)
0.089
(0.152)

-0.010
(0.049)

0.019
(0.025)

0.039
(0.233)

0.025
(0.021)

0.290
(0.397)

P-Value for Joint Test

0.952

0.920

0.669

6.46
657
0.482

10.88
657
0.505

17.34
657
0.471

Outcome Mean
County-Year Obs.
R-Squared
Notes: The unit of observation is a county-cohort pair. Outcomes are mea-
sured as the mean number of credits completed in the first year of com-
munity college enrollment across all students who enroll in vocational pro-
grams. The coefficients in each column are estimated from a separate re-
gression and represent the β parameters in equation (2.4), the effect of an
additional layoff per 10,000 working age residents in a given occupation
group on the outcome of interest. All regressions include controls for the
share of graduates that are white, male, and categorized as economically
disadvantaged; average 11th grade math and reading test scores; and the
county unemployment rate, logged size of the labor force, and the number
of layoffs per 10,000 working-age residents in non community college oc-
cupations during a cohort’s senior year of high school. All standard errors
are clustered at the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

155

Table B.1.15: Substitution Between Community College Program Groups

Layoffs per 10,000 in:
Business, t-1

Health, t-1

Skilled Trades, t-1

STEM, t-1

Law Enf., t-1

Other, t-1

Business

(1)

-1.025**
(0.456)

-0.120
(0.138)

0.067
(0.078)

0.212
(0.676)

0.076
(0.075)

0.753
(0.617)

0.164
(0.109)

0.206
(0.626)

0.078
(0.082)

0.072
(0.945)

Enrollment per 100 Vocational Students in:
STEM Law Enf.

Trades

Health

(2)

(3)

(4)

(5)

-0.702
(0.682)

-0.056
(0.449)

-0.093
(0.280)

1.736***
(0.592)

-0.610**
(0.232)

-0.281**
(0.122)

0.164
(0.123)

0.250
(0.222)

0.030
(0.123)

-0.086
(0.839)

-0.088
(0.097)

-0.253
(0.674)

-0.014
(0.066)

-0.124
(0.347)

-0.048
(0.061)

0.143
(0.094)

-0.153**
(0.075)

-0.344
(0.518)

-0.688
(0.522)

1.014
(0.678)

Other

(6)
0.141
(0.347)

0.597***
(0.132)

-0.159**
(0.063)

0.044
(0.405)

-0.097
(0.061)

-0.807
(0.511)

Own-layoff semi-elasticities (at mean):

-0.047**
(0.021)

-0.029***
(0.011)

-0.006
(0.007)

-0.010
(0.029)

-0.011**
(0.005)

-0.046
(0.029)

20.67
657
0.506

21.66
657
0.190

14.33
657
0.344

11.84
657
0.266

Outcome Mean
County-Year Obs.
R-squared
Notes: The unit of observation is a county-cohort pair. Outcomes are measured as the number of students who
enroll in a given program within 6 months of high school graduation per 100 students who in the county and
cohort enroll in vocational programs. The coefficients in each column are estimated from a separate regression and
represent the β j terms in equation (2.5), the effect of an additional layoff per 10,000 working age residents in a
given occupation group on the outcome of interest. All regressions include controls for the share of graduates that
are white, male, and categorized as economically disadvantaged; average 11th grade math and reading test scores;
and the county unemployment rate, logged size of the labor force, and the number of layoffs per 10,000 working-age
residents in non community college occupations during a cohort’s senior year of high school. All standard errors
are clustered at the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

13.74
657
0.258

17.75
657
0.353

156

Table B.1.16: Substitution Between Narrower Community College Programs

Enrollment per 100 Vocational Students in:
Trades

STEM Law Enf.

Arts

Personal
& Culinary

Business Health

(1)

-1.025**
(0.456)

(2)

-0.702
(0.682)

(3)

(4)

(5)

-0.056
(0.449)

-0.093
(0.280)

1.736***
(0.592)

& Media

(6)

-0.303
(0.227)

0.107
(0.084)

-0.124***
(0.039)

0.383
(0.316)

Social
Services

(8)

0.440**
(0.184)

0.346***
(0.073)

-0.008
(0.031)

0.196
(0.195)

0.068
(0.053)

-0.031
(0.371)

(7)
0.004
(0.201)

0.144*
(0.083)

-0.027
(0.057)

-0.535**
(0.268)

-0.088**
(0.043)

-0.123
(0.302)

-0.610**
(0.232)

-0.281**
(0.122)

0.164
(0.123)

0.250
(0.222)

0.030
(0.123)

-0.086
(0.839)

-0.088
(0.097)

-0.253
(0.674)

-0.014
(0.066)

-0.124
(0.347)

-0.048
(0.061)

0.143
(0.094)

-0.153**
(0.075)

-0.077***
(0.027)

-0.344
(0.518)

-0.688
(0.522)

1.014
(0.678)

-0.652
(0.404)

Layoffs per 10,000 in:
Business, t-1

Health, t-1

Skilled Trades, t-1

STEM, t-1

Law Enforcement, t-1

Other, t-1

-0.120
(0.138)

0.067
(0.078)

0.212
(0.676)

0.076
(0.075)

0.753
(0.617)

21.66
657
0.190

0.164
(0.109)

0.206
(0.626)

0.078
(0.082)

0.072
(0.945)

20.67
657
0.506

9.11
657
0.542

13.74
657
0.258

11.84
657
0.266

14.33
657
0.344

Outcome Mean
Observations
R-squared
Notes: The unit of observation is a county-cohort pair. Outcomes are measured as the number of students who enroll in a given program within 6
months of high school graduation per 100 students who in the county and cohort enroll in vocational programs. I define social service programs
as those with 2-digit occupation codes of 21 (Community and Social Service) and 25 (Education, Training, and Library), plus childcare programs
(SOC 39-9011); arts and media programs as those with the 2-digit occupation code 27 (Arts, Design, Entertainment, Sports, and Media); and
personal care and culinary programs as those with the 2-digit codes 35 (Food Preparation and Serving) and 39 (Personal Care and Service), other
than childcare, plus baking programs (SOC 51-3011). The coefficients in each column are estimated from a separate regression and represent the
β j terms in equation (2.5), the effect of an additional layoff per 10,000 working age residents in a given occupation group on the outcome of
interest. All regressions include controls for the share of graduates that are white, male, and categorized as economically disadvantaged; average
11th grade math and reading test scores; and the county unemployment rate, logged size of the labor force, and the number of layoffs per 10,000
working-age residents in non community college occupations during a cohort’s senior year of high school. All standard errors are clustered at the
county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

5.26
657
0.322

3.39
657
0.313

157

B.2 Comparing Layoffs to Other Employment Data Sources

The estimated layoff measures used throughout the analysis are designed to capture changes

in local labor demand in a given occupation group and county. They should not, however, be

treated as the exact number of job losses in an occupation group and county because not all layoff

events are required to be reported under the WARN Act and, among events that are required to

be reported, there is non-compliance in reporting. For example, in 2001, the federal government

estimated that only about one quarter of events were required to be reported under the WARN Act

and that, of those that were required to be reported, only one-third of were reported to the correct

government agencies (United States General Accounting Office, 2003).

Nevertheless, to verify that these proxy measurements capture true changes in employment

over time and across counties, I compare county-by-industry layoffs to analogous employment

data from two commonly used employment datasets: the Quarterly Census of Employment and

Wages (QCEW) and the County Business Patterns (CBP). The QCEW is published quarterly by

the Bureau of Labor Statistics and captures employment in more than 95% of U.S. jobs. However,

a large share of its data at the county-by-industry level is suppressed due to privacy concerns. The

CBP is released annually by the U.S. Census Bureau and captures the number of establishments

and total employment during the week of March 12. Like the QCEW, many county-by-industry

cells in the CBP are suppressed to prevent users from inferring information about individual firms.

But in contrast to the QCEW, employment counts for some cells in the CBP can be imputed from

establishment counts and higher-level geographic and industrial classifications. In the analyses

that follow, I use the imputed data provided by Eckert et al. (2020) to maximize the coverage of

Michigan’s counties.

I begin by comparing the county-by-industry employment counts provided by both the QCEW

and CBP. Because the CBP data does not contain information on government employment, I re-

strict the sample to all non-government NAICS 3-digit sectors. I further restrict the sample to

county-by-industry pairs that have non-zero employment counts in all years 2001-2016 in at least

one of the datasets. Figure B.2.1, below, provides a simple scatterplot of employment counts in

158

the two datasets for the 73% of observations (3,630 county-industry pairs) that contain employ-

ment information in both datasets. The two measures of employment are highly correlated, with a

Pearson’s coefficient of 0.95.

Figure B.2.1: Comparison of Employment Counts in QCEW & CBP

Then, with each dataset, I estimate regressions of the following form:

∆Employmentkct = α + β Layoffskc,t−1 + εkct

(B.1)

where ∆Employmentkct is the change in employment in industry k in county c between March of
year t − 1 and March of year t, and Layoffskc,t−1 is the number of layoffs in industry k in county
c between March of year t − 1 and March of year t.1 The parameter of interest, β , captures the
relationship between layoffs and year-over-year employment change in a given county and indus-

try. If β is equal to -1, then, on average, an additional layoff is associated with an employment
reduction of exactly one worker. If |β| is less than 1, then an additional layoff reduces employment
1The CBP provides employment counts as of March 12. To track corresponding employment

changes in the QCEW, I use the first quarter, third month employment counts.

159

025,00050,00075,000100,000125,000QCEW Employment025,00050,00075,000100,000125,000CBP Employmentby less than one worker on average, presumably because some laid-off workers find work at other

firms in the same county and industry or other firms are increasing employment at the same time
as the layoff. Alternatively, if |β| is greater than 1, then an additional layoff reduces employment
by more than one worker on average, indicating that there are additional employment reductions,

including changes in labor supply, that are not captured in the WARN data. Table B.2.1 presents

the results of this specification using each dataset.

Table B.2.1: Relationship Between Estimated Layoffs & Employment Change

Layoff measure:
Panel A. Quarterly Census of Employment & Wages (QCEW)
-1.139***
Layoffs in county and industry, t-1
(0.312)

-1.236***
(0.322)

(1)

(2)

(3)

-0.749***
(0.266)

County, industry, and year FEs
Interacted FEs

X

X
X

County-Year-Industry Obs.

47,399

47,398

47,254

Panel B. County Business Patterns (CBP)
Layoffs in county and industry, t-1

-0.942*** -0.914***
(0.196)
(0.196)

-0.803***
(0.202)

County, industry, and year FEs
Interacted FEs

X

X
X

County-Year-Industry Obs.

58,202

58,202

58,186

Notes: The sample consists of all county-by-industry pairs that have non-zero em-
ployment between 2001 and 2016 in either the QCEW or CBP dataset. All standard
errors are clustered at the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

Column (1) shows that an additional layoff is associated with an employment reduction of 1.2

workers in the QCEW and of 0.94 workers in the CBP data. Column (2) then adds county, industry,

and year fixed effects to assess whether the negative relationship continues to hold after controlling

for factors that may induce layoffs (e.g., overall economic downturns or industry-specific turnover

patterns). When using either dataset, the estimated change in employment due to an additional

160

layoff remains negative, statistically significant and close to -1 when including these fixed effects.

Finally, column (3) interacts these fixed effects to mimick the interacted fixed effects in the most

saturated version of equation (2.3) in the main text. When controlling for county-by-year, county-

by-sector, and sector-by-year effects, an additional layoff reduces employment by 0.75 workers

(QCEW) to 0.8 workers (CBP). The estimates remain statistically significant, indicating that the

layoff measures are indeed capturing changes in local employment counts.

Finally, to ensure that the relationship between is not driven by select industries, I estimate

equation (B.1) separately for the ten NAICS 3-digit subsectors with the most layoffs in the WARN

data. Figure B.2.2 presents these results. The estimated coefficients are overwhelmingly negative

and do not vary substantially by dataset, again indicating that the layoff measures used throughout

the paper capture true changes in local employment conditions.

Figure B.2.2: Relationship between Layoffs and Employment Changes, by Sector

161

336: Transportation Equipment Manufacturing326: Plastics and Rubber Products Manufacturing452: General Merchandise Stores332: Fabricated Metal Product Manufacturing522: Credit Intermediation & Related Activities541: Professional, Scientific, and TechnicalServices481: Air Transportation333: Machinery Manufacturing331: Primary Metal Manufacturing561: Administrative and Support Services-10-50510QCEWCBPB.3 Other Responses to Layoffs

To supplement the main analysis, I also analyze how layoffs affect two other educational out-

comes of interest: the enrollment choices of students who delay community college entrance be-

yond the first six months of high school graduation and the retention rates of students once enrolled.

For the first outcome, I restrict the sample to students who graduate from high school between 2009

and 2013 and enroll in vocational community college programs within at some point before 2017
and re-estimate equation 2.3 in the main text for different enrollment timeframes.1 Figure B.3.1

shows the estimated elasticity of program choice with respect to prior-year layoffs in related occu-

pations. For enrollment within either six or twelve months of high school graduation, an additional

layoff per 10,000 county working-age residents during a cohort’s senior year of high school re-

duces enrollment in related programs by about 1%. This effect continues to hold when I control

for layoffs occurring during students freshman, sophomore, and junior years of high school.

Figure B.3.1: Effect of Layoffs on Program Choice for Later Enrollees

1To control for time-varying county characteristics that I may not observe in my data, I include

county-by-cohort fixed effects in these specifications.

162

-.03-.02-.010Semi-Elasticity w.r.t. Prior-Year Layoffs0-6 Months0-12 Months1-2 YearsNo ControlsPrior Layoff Controls  When analyzing longer-run enrollment choices, I cannot observe where students live in the

years following high school graduation and, therefore, implicitly assume that students’ remain

living in the same county that lived in during high school. Nevertheless, for students enrolling

in vocational community college programs in the 1-2 years following graduation, I find similar

effects of layoffs on program choices. Figure B.3.1 shows that an additional layoff per 10,000

students reduces enrollment in the following year by about 2%. The magnitude of this estimate

suggests that older students may be even more responsive to local labor market shocks, which is

an important topic for future work.

I also consider how layoffs affect program retention rates by including all cohorts and estimat-

ing equations of the following form:

Retentiongct = α + Layoffsgctβ + XctΓ + θgc + δgt + εgct

(B.1)

where Retentiongct is a measure of the year-over-year retention of students from county c enrolled
in program group g in year t, Layoffsgct is a measure of analogous layoffs, and all other terms are
defined as in previous equations in the main text. My main measure of retention is the number of
students from county c who were enrolled in program group g in year t − 1 and remain enrolled
in the same program and community college in year t, per 100 students initially enrolled.2 This

measure is equal to the share of students who remain enrolled in the same college and program in

the following year and multiply the share by 100. I also calculate measures of students switching

between programs and between colleges, graduating from programs, and not being observed in the
data the following year. I measure layoffs as those that occur between July 1st of year t − 1 and
June 30th of year t to capture layoffs that students observe throughout the year in which they are

enrolled in a program.

Table B.3.1 presents these results. Column (1) indicates that an additional layoff per 10,000

working-age residents reduces program retention by 0.26pp, or about 0.6%. This estimate is

2In these calculations, I only consider enrollment in the college at which students earn the most
credits during a given year. That is, if a student enrolls in two colleges within one year, she is
assigned to enrollment only at the college in which she earns more credits.

163

smaller than the decrease in initial program enrollment documented in my earlier results, which

is consistent with the fact that students already enrolled in a program likely face a lower marginal

cost to finishing. For example, they have likely already completed some of the coursework needed

to earn a degree in the subject. I also estimate the effects of layoffs on retention separately for each
program group using a modified version of the systems of equations setup.3 Table B.3.2 presents

these results, which indicate that the largest elasticities come from students’ responses to layoffs

in STEM and other programs.

Columns (2) through (5) of Table B.3.1 document what choices students make when layoffs

deter them from continuing in vocational programs. While the estimates are imprecise, the largest

coefficient appears in Column (5), which measures the share of students who were enrolled in a

program in the prior year but are no longer formally enrolled in postsecondary education. In most

cases, this means that a student has dropped out of her community college program without earning
a degree.4 Given the large labor market returns to degree completion, this type of substitution effect

may negatively impact students’ longer-run outcomes and suggests that policies that assist students

in switching between programs after local labor market shocks could improve student outcomes.

3Specifically, I regress a program’s retention rate on the vector of layoffs occurring in each

occupation group, county control variables, county fixed effects, and cohort fixed effects.

4Students could also be enrolled in colleges not covered by the NSC data. However, these
types of colleges make up less than 1% of U.S. postsecondary institutions overall (National Student
Clearinghouse Research Center, 2017).

164

Table B.3.1: Effect of Layoffs on Retention in Related Programs

Number per 100 Prior-Year Vocational Students:
Same
Not

Different
Program

Different
College

Program

(1)

(2)

(3)

Layoff measure:
Layoffs per 10,000 in
occupation group

-0.264**
(0.128)

-0.034
(0.027)

-0.008
(0.043)

Earned
Degree

(4)
0.027
(0.052)

Observed

(5)

0.279**
(0.129)

25.44
3,364
0.276

Outcome Mean
County-Program-Year Obs.
R-Squared

43.48
3,364
0.246

11.92
3,364
0.300

10.62
3,364
0.270

8.54
3,364
0.374

Notes: The unit of observation is a county-year-program triad. Each coefficient is estimated from
a separate regression and represents β in equation B.1, the effect of an additional layoff per 10,000
working age residents in a given occupation group on retention in related programs. All regressions
include controls for the share of graduates that are white, male, and categorized as economically
disadvantaged; average 11th grade math and reading test scores; and the county unemployment
rate, logged size of the labor force, and the number of layoffs per 10,000 working-age residents in
non community college occupations during a cohort’s first year of college. All standard errors are
clustered at the county level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

Table B.3.2: Own-Layoff Effects on Program Retention Rates

Layoff measure:
Layoffs per 10,000 in
own occupation group

Retention per 100 Students in:

Business Health Trades STEM Law Enf.

(1)

(2)

(3)

(4)

(5)

Other

(6)

-0.250
(0.546)

-0.082
(0.275)

-0.364
(0.246)

-1.307
(0.951)

-0.226
(0.204)

-3.600***
(1.358)

Outcome Mean
County-Year Obs.
R-Squared

41.41
566
0.353

43.93
566
0.291

43.98
560
0.253

45.25
554
0.245

41.97
560
0.285

44.37
558
0.233

Notes: The unit of observation is a county-cohort pair. Each coefficient is estimated from a separate
regression and represents the effect of an additional layoff per 10,000 working age residents in a
given occupation group on retention in related programs. All regressions include controls for the
share of graduates that are white, male, and categorized as economically disadvantaged; average
11th grade math and reading test scores; and the county unemployment rate, logged size of the
labor force, and the number of layoffs per 10,000 working-age residents in non community college
occupations during a cohort’s first year of college. All standard errors are clustered at the county
level. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

165

B.4 Substitution Between Narrower Program Groups

One limitation of the main analysis is that it combines multiple, potentially distinct programs

into a single program group. To investigate substitution patterns between narrower program groups,

I re-estimate the system of equations presented in equation (2.5) of the main text using enrollment

in the two-digit occupation codes that comprise each program group as the dependent variables.

For example, rather than estimating how business layoffs affect enrollment in business programs

overall, I separately estimate how business layoffs affect enrollment in management, business and

financial operations, legal, sales, and administrative support programs. I present these own-layoff

effects in Figure B.4.1.

Figure B.4.1: Effect of Layoffs on Enrollment in Narrower Program Groups

The results indicate that the reduction in business program enrollment is driven by students

forgoing enrollment in management-related programs, such as business administration, and the re-

166

11 Management13 Business &Financial Operations23 Legal41 Sales & Related43 Office &AdministrativeSupport-2-1.5-1-.50.5Business29 HealthcarePracticioners &Technical31 HealthcareSupport-1.5-1-.50.5Health37 Building, GroundsCleaning &Maintenance45 Farming, Fishing,& Forestry47 Construction &Extraction49 Installation,Maintenance, &Repair51 Production53 Transportation &Material Moving-.2-.10.1Skilled Trades15 Computer &Mathematical17 Architecture &Engineering19 Life, Physical, &Social Science-1-.50.51STEM33 ProtectiveService-.3-.2-.10Law Enforcement21 Community &Social Service25 Education,Training, & Library27 Arts, Design,Entertainment,Sports, & Media35 Food Preparation& Serving Related39 Personal Care &Service-1.5-1-.50.5Otherduction in healthcare programs is driven by students forgoing enrollment in healthcare practitioner

programs, such as nursing. The reductions in enrollment in skilled trades programs are driven

by programs in the installation, maintenance, and repair and production categories, which includes

auto mechanic and welding degrees. The responses to STEM and other layoffs are not substantially

different across each program group’s occupational categories.

Next, I analyze substitution patterns relative to the two-digit occupation code that experiences

the largest own-layoff effect in each program group. For example, because the largest decrease in

business program enrollment comes from the management group, I compare the skills of all other

two-digit occupation codes to the skills needed for management occupations to see if students are

substituting into similar programs. Figure B.4.2 shows how the substitution patterns for each pro-

gram group relate to the skill distance measures.

Figure B.4.2: Substitution into Narrower Program Groups Requiring Similar Skills

167

-2-10123Effect Size0.2.4.6.81Skill DistanceBusiness Layoff-1-.50.51Effect Size0.2.4.6.81Skill DistanceHealth Layoff-.2-.10.1.2.3Effect Size0.2.4.6.81Skill DistanceSkilled Trades Layoff-2-1012Effect Size0.2.4.6.81Skill DistanceSTEM Layoff-.4-.20.2Effect Size0.2.4.6.81Skill DistanceLaw Enforcement Layoff-2-1012Effect Size0.2.4.6.81Skill DistanceOther LayoffBusinessHealthTradesSTEMLaw Enf.OtherFigure B.4.3 then plots the pooled substitution effects against the skill distance measures for

all six program groups. As in Figure B.1.8 in the main text, the largest substitution effects occur at

the start of the x-axis, and there is a downward slope, indicating that substitution effects are largest

in the most similar programs and diminish as skill distance increases. However, the results are less

precise when considering enrollment in smaller program categories.

Figure B.4.3: Relationship Between Substitution Effects & Skill Distance, Narrower Programs

168

-1-.50.511.5Effect Size.2.4.6.81Skill Distancen = 110    RMSE =  .26380208Effect = 0.099 - 0.157 Similarity    R2 = 1.7%APPENDIX C

CHAPTER 3 APPENDIX

C.1 Figures & Tables

Figure C.1.1: ASD and Non-ASD Special Education Incidence Event Studies

Notes: This figure plots event study estimates in which we replace PostMandatet × NonDisadvi with NonDisadvi
interacted with a set of year dummies in equation (3.1). The NonDisadv×t control is excluded because of
collinearity with the event study variables. Year 2011 is excluded, so all estimates are relative to that year. Each point
represents the point estimate and the bars extending from each point show the 95% confidence interval that is
calculated from standard errors that are clustered at the school district level.

169

-.0015-.001-.00050.0005Estimate200920102011201220132014YearAny ASD Incidence0.002.004.006.008Estimate200920102011201220132014YearAny Non-ASD IncidenceFigure C.1.2: Test Scores Event Studies, using Non-Disabled Control Group

Notes: This figure plots event study estimates in which we replace PostMandatet × NonDisadvi × ASDit with
NonDisadvi × ASDit interacted with a set of year dummies in equation (3.2). Year 2011 is excluded, so all estimates
are relative to that year. Each point represents the point estimate and the bars extending from each point show the
95% confidence interval that is calculated from standard errors that are clustered at the school district level.

170

-.1-.050.05Estimate2010201120122013YearMath Test Score (vs. Non-Disabled Students)-.1-.050.05.1Estimate2010201120122013YearReading Test Score (vs. Non-Disabled Students)Figure C.1.3: Event Study Estimates of Main Outcomes

171

-.050.05Estimate200920102011201220132014YearASD Program-.1-.050.05Estimate200920102011201220132014YearResource/Cog. Program-.020.02.04.06Estimate200920102011201220132014YearNo ProgramFigure C.1.3 (cont’d)

172

-.01-.0050.005.01.015Estimate20102011201220132014YearSpecial Ed School-.020.02.04.06.08Estimate20102011201220132014YearGen Ed 80% or more-.03-.02-.010.01.02Estimate2011201220132014YearAverage Special Ed FTEFigure C.1.3 (cont’d)

173

-.04-.020.02.04Estimate200920102011201220132014YearASD Teacher Consultant-.04-.020.02.04Estimate200920102011201220132014YearSpeech Therapy-.04-.020.02.04.06Estimate200920102011201220132014YearOccupational TherapyFigure C.1.3 (cont’d)

Notes: This figure plots event study estimates in which we replace PostMandatet × NonDisadvi × ASDit with
NonDisadvi × ASDit interacted with a set of year dummies in equation (3.2). Year 2011 is excluded, so all estimates
are relative to that year. Each point represents the point estimate and the bars extending from each point show the
95% confidence interval that is calculated from standard errors that are clustered at the school district level.

174

-.06-.04-.020.02Estimate200920102011201220132014YearSocial Worker0.01.02.03.04.05Estimate200920102011201220132014YearAnyFigure C.1.4: Test Scores Event Studies, using All Non-ASD Control Group

Notes: This figure plots event study estimates in which we replace PostMandatet × NonDisadvi × ASDit with
NonDisadvi × ASDit interacted with a set of year dummies in equation (3.2). Year 2011 is excluded, so all estimates
are relative to that year. Each point represents the point estimate and the bars extending from each point show the
95% confidence interval that is calculated from standard errors that are clustered at the school district level.

175

-.1-.050.05Estimate2010201120122013YearMath Test Score (vs. All Non-ASD Students)-.1-.050.05.1Estimate2010201120122013YearReading Test Score (vs. All Non-ASD Students)Figure C.1.5: Test Score Event Studies, using Non-ASD Special Ed Control Group

Notes: This figure plots event study estimates in which we replace PostMandatet × NonDisadvi × ASDit with
NonDisadvi × ASDit interacted with a set of year dummies in equation (3.2). Year 2011 is excluded, so all estimates
are relative to that year. Each point represents the point estimate and the bars extending from each point show the
95% confidence interval that is calculated from standard errors that are clustered at the school district level.

176

-.1-.050.05Estimate2010201120122013YearMath Test Score (vs. Non-ASD Disabled Students)-.1-.050.05.1Estimate2010201120122013YearReading Test Score (vs. Non-ASD Disabled Students)Table C.1.1: Descriptive Tabulations of Analysis Variables

Variable

Demographics
White
Male
LEP
Poverty

Disability
ASD
Any Non-ASD
Cognitive
Emotional
Speech
Learning Disability
Other Health

All

ASD

Non-ASD
Special Ed. Special Ed.

Non-

0.754
0.858
0.030
0.434

0.651
0.644
0.053
0.691

0.687
0.489
0.055
0.482

0.683
0.513
0.055
0.509

0.010
0.133
0.011
0.009
0.036
0.055
0.016

Observations

3,854,234 38,803

506,432

3,308,999

Special Education Program
ASD
Resource
Cognitive
Other
None

Education Setting
Special Ed. School
Gen. Ed. > 80%
Gen. Ed. 40-79%
Gen Ed. < 40%
Average FTE

Special Education Support Services
ASD Teaching Consultant
Non-ASD Teaching Consultant
Language
Social Worker
Occupational Therapy
Physical Therapy
Transportation
Other Service
Any Service

0.197
0.551
0.119
0.031
0.129

0.056
0.455
0.149
0.207
0.353

0.130
0.091
0.790
0.691
0.401
0.031
0.041
0.042
0.943

0.001
0.611
0.090
0.046
0.268

0.018
0.594
0.143
0.074
0.194

0.003
0.074
0.478
0.220
0.085
0.027
0.009
0.023
0.674

Observations
Notes: Authors’ tabulations from data on students in grades 2-8 from the 2009-2010 to the 2014-2015
school years. The sample sizes for the ASD groups in the top and bottom panels differ slightly because a
small number of students with an ASD diagnosis do not receive any special education services.

506,432

38,621

177

Table C.1.2: Overlap Between Free/Reduced Price Lunch and Medicaid
in Michigan, by Family Income

By Free/Reduced Price Lunch Status:

FRPL
Status
Eligible
Not Eligible

Percent
Medicaid
72.93%
12.86%

Percent Private

Insurance
31.03%
88.53%

Percent
Insured
95.40%
97.54%

Percent
Insured

By Family Income as Percent of Poverty Line:

Percent
Medicaid

Percent Private

Insurance

Family
Income
≤ 135% FPL
135-185% FPL
185-250% FPL
250-350% FPL
≥ 350% FPL
Notes: Authors’ tabulations from the 2008-2016 American Community Survey
among children who were in grades K-8 at Michigan public schools (N = 84,477).
“FPL” stands for Federal Poverty Line. Note that insurance counts may exceed 100%
as some people remain eligible for Medicaid while enrolled in private plans.

22.50%
54.19%
74.53%
86.82%
95.29%

81.05%
50.87%
29.13%
14.60%
5.16%

95.63%
94.75%
95.95%
96.84%
98.57%

178

Table C.1.3: The Effect of the ASD Insurance Mandate on Disability Incidence

Panel A: Main Estimates

Dependent Variable:

ASD
(1)

Non-ASD
Disability

(2)

Cognitive Emotional
Disability Disability

(3)

(4)

Speech
Disability

(5)

Learning
Disability

(6)

Other Health

Disability

(7)

-0.00045*
(0.00024)

-0.00004
(0.00105)

0.00001
(0.00032)

0.00126***
(0.00028)

-0.00283***
(0.00065)

0.00399***
(0.00088)

-0.00266***
(0.00043)

Panel B: Including Linear Time Trend Interacted with Non-Disadvantaged Status

ASD
(1)

Non-ASD
Disability

(2)

Dependent Variable:

Cognitive Emotional
Disability Disability

(3)

(4)

Speech
Disability

(5)

-0.00004
(0.00022)

-0.00004
(0.00097)

-0.00028
(0.00031)

0.00051*
(0.00029)

-0.00013
(0.00064)

Panel C: School District Fixed Effects
Dependent Variable:

Non-ASD
Disability

(2)

Cognitive Emotional
Disability Disability

(3)

(4)

Speech
Disability

(5)

ASD
(1)

Learning
Disability

(6)

-0.00033
(0.00063)

Learning
Disability

(6)

Other Health

Disability

(7)

0.00034
(0.00039)

Other Health

Disability

(7)

Independent
Variable
Non-disadv*
Post-2012

Independent
Variable
Non-disadv*
Post-2012

Independent
Variable
Non-disadv*
Post-2012

-0.00044*
(0.00023)

-0.00013
(0.00108)

-0.00005
(0.00031)

0.00125***
(0.00031)

-0.00303*** 0.00418***
(0.00066)
(0.00098)

-0.00254***
(0.00041)

0.009

0.011

0.133

0.010

Incidence Rate
Notes: Authors’ estimates of equation (3.1) as described in the text using data on students in grades 2-8 from the 2009-2010 to the 2014-2015
school years. The sample includes only students who are always or never eligible for free/reduced price lunch. Each column is a separate
regression; N=3,854,234. “Non-disadv” is an indicator for whether the student is eligible for free/reduced price lunch in all observed years
of schooling. Students who are eligible for free/reduced price lunch in only some years of schooling are excluded from the regression. All
regressions include controls for whether a student is white, male, and limited English proficient as well as grade-by-year fixed effects. Estimates
in Panels A and B include school fixed effects, while those in Panel C include school district fixed effects. Estimates in Panel B also control for a
linear time trend interacted with non-disadvantaged status. Standard errors are clustered at the school district level: *,**,*** indicate significance
at the 10, 5, and 1 percent level, respectively.

0.016

0.036

0.055

179

Table C.1.4: The Effect of the ASD Insurance Mandate on Special Education Services

Resource or
Cognitive
Program Program

ASD

Program School

Independent
Variable
Non-disadv*Post-
2012*ASD
Non-disadv

ASD

Non-disadv*Post-
2012
Non-disadv*ASD

Post-2012*ASD

(1)
0.031
(0.035)
0.002
(0.002)
0.187***
(0.0449)
0.000
(0.002)
-0.062
(0.040)
-0.022
(0.038)
545,053
0.197

(2)

-0.064**
(0.030)
-0.133***
(0.004)
-0.048
(0.0367)
0.007
(0.005)
0.152***
(0.034)
0.032
(0.032)
545,053
0.670

No
Sped

(3)

0.034***
(0.011)
0.143***
(0.004)
-0.104***
(0.0104)
-0.012***
(0.005)
-0.109***
(0.012)
-0.008
(0.009)
545,053
0.129

Special

General

Ed

(4)
0.004
(0.004)
-0.001**
(0.001)
0.009*
(0.005)
0.000
(0.001)
-0.006
(0.004)
-0.007
(0.006)
455,751
0.056

Ed

>80%

(5)
0.000
(0.015)
0.122***
(0.005)
-0.145***
(0.019)
-0.031***
(0.005)
-0.018
(0.019)
-0.017
(0.014)
455,751
0.455

Dependent Variable:

Sped
FTE
Rate
(6)

ASD

Teacher

Consultant

(7)

-0.008
(0.009)
-0.056***
(0.003)
0.115***
(0.024)
0.012***
(0.002)
-0.001
(0.023)
0.005
(0.009)
332,372
0.353

-0.023**
(0.011)
0.000
(0.001)
0.125***
(0.016)
-0.001
(0.001)
0.039**
(0.018)
-0.015
(0.011)
545,053
0.130

Language
Services

(8)
0.005
(0.013)
0.101***
(0.004)
0.300***
(0.015)
-0.020***
(0.004)
-0.068***
(0.015)
0.009
(0.013)
545,053
0.790

Occupational

Therapy
Services

(9)

Social
Worker

(10)

Any
Sped

Services

(11)

-0.004
(0.012)
0.023***
(0.003)
0.257***
(0.014)
0.001
(0.003)
0.014
(0.015)
0.003
(0.012)
545,053
0.401

-0.021*
(0.013)
-0.089***
(0.005)
0.406***
(0.019)
0.004
(0.006)
0.105***
(0.018)
0.028**
(0.013)
545,053
0.691

0.020**
(0.010)
0.055***
(0.005)
0.255***
(0.011)
-0.015***
(0.005)
-0.053***
(0.011)
-0.021**
(0.010)
545,053
0.943

Observations
ASD Mean
Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the 2010-2011 to the 2014-2015 school years. The sample includes only students who are
always or never eligible for free/reduced price lunch. Each column is a separate regression. “Non-disadv” is an indicator for whether the student is eligible for free/reduced price lunch in all observed
years of schooling and “ASD” is an indicator for whether the student has an ASD diagnosis in that year. All regressions include controls for whether a student is white, male, and limited English
proficient as well as school and grade-by-year fixed effects. The final row of the table provides dependent variable means for the ASD sample. Standard errors are clustered at the school district level:
*,**,*** indicate significance at the 10, 5, and 1 percent level, respectively.

180

Table C.1.5: The Effect of the ASD Insurance Mandate on Other Special Education Services

Gen Ed Gen Ed General

Program 40-79% <40%

Consultant Therapy

Independent
Variable
Non-disadv*Post-
2012*ASD

(2)

(3)

-0.001
(0.005)

-0.005
(0.010)

-0.001
(0.014)

-0.006
(0.004)

Other
Sped

(1)

Dependent Variable:

Any

Ed
(4)

Non-ASD
Teacher

(5)
0.003
(0.007)

Trans-
Physical portation
Services

(6)

(7)

-0.002
(0.005)

-0.001
(0.006)

545,053
0.031

455,751
0.149

455,751
0.207

455,571
0.811

Observations
ASD Mean
Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the 2010-2011
to the 2014-2015 school years. The sample includes only students who are always or never eligible for free/reduced price
lunch. Each column is a separate regression. “Non-disadv” is an indicator for whether the student is eligible for free/reduced
price lunch and “ASD” is an indicator for whether the student has an ASD diagnosis in that year. All regressions include
the full set of controls listed in equation (3.2), including controls for whether a student is white, male, and limited English
proficient as well as school and grade-by-year fixed effects. The final row of the table provides dependent variable means for
the ASD sample. Standard errors are clustered at the school district level: *,**,*** indicate significance at the 10, 5, and 1
percent level, respectively.

545,053
0.041

545,053
0.031

545,053
0.091

181

Table C.1.6: The Effect of the ASD Insurance Mandate on Test Scores

Non-ASD Non-Sped

Non-ASD Non-Sped

Control Group:

Non-disadv*Post-
2012*ASD
Non-disadv

ASD

Non-disadv*Post

Non-disadv*ASD

Post*ASD

Lagged Achievement

All

(1)
0.001
(0.026)
0.156***
(0.005)
-0.091***
(0.016)
0.014***
(0.005)
-0.079***
(0.018)
0.023
(0.021)
0.738***
(0.006)

Math

(2)
0.005
(0.026)
0.144***
(0.005)
-0.129***
(0.016)
0.011**
(0.005)
-0.070***
(0.018)
0.022
(0.021)
0.731***
(0.006)

Non-ASD

Sped
(3)

-0.010
(0.027)
0.139***
(0.006)
0.112***
(0.018)
0.033***
(0.007)
-0.016
(0.020)
0.064***
(0.021)
0.598***
(0.014)

All

(4)
0.021
(0.029)
0.197***
(0.004)
-0.097***
(0.017)
-0.002
(0.003)
-0.186***
(0.022)
0.014
(0.023)
0.644***
(0.003)

Reading

(5)
0.018
(0.029)
0.181***
(0.004)
-0.166***
(0.018)
0.001
(0.003)
-0.168***
(0.022)
0.025
(0.023)
0.618***
(0.003)

Non-ASD

Sped
(6)
0.004
(0.030)
0.157***
(0.007)
0.195***
(0.019)
0.014*
(0.007)
-0.143***
(0.024)
0.029
(0.024)
0.564***
(0.010)

1,754,971

1,579,046

185,814

1,749,290

Observations
Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the 2009-2010
to the 2013-2014 school years. 2014-2015 is excluded as Michigan changed from the Michigan Assessment of Educational
Progress to the M-Step exam and restructured alternative examination options. The sample includes only students who are
always or never eligible for free/reduced price lunch. Each column is a separate regression. “Non-disadv” is an indicator for
whether the student is eligible for free/reduced price lunch in all observed years of schooling and “ASD” is an indicator for
whether the student has an ASD diagnosis in that year. All regressions include controls for whether a student is white, male,
and limited English proficient as well as school and grade-by-year fixed effects. Standard errors are clustered at the school
district level: *,**,*** indicate significance at the 10, 5, and 1 percent level, respectively.

1,578,937

180,172

182

Table C.1.7: The Effect of the ASD Insurance Mandate on Taking Regular Exams

Non-ASD Non-Sped

Non-ASD Non-Sped

Control Group:

Non-disadv*Post-
2012*ASD
Non-disadv

ASD

Non-disadv*Post

Non-disadv*ASD

Post*ASD

All

(1)

-0.025
(0.024)
0.028***
(0.001)
-0.351***
(0.026)
0.017***
(0.001)
0.057**
(0.025)
-0.060**
(0.025)

Math

(2)
0.001
(0.024)
0.001*
(0.000)
-0.406***
(0.028)
-0.000
(0.000)
0.089***
(0.026)
-0.098***
(0.024)

Non-ASD

Sped
(3)

-0.047*
(0.025)
0.058***
(0.006)
-0.203***
(0.028)
0.046***
(0.007)
0.017
(0.027)
0.039
(0.025)

All

(4)

Reading

(5)

-0.030
(0.024)
0.031***
(0.001)
-0.351***
(0.026)
0.018***
(0.001)
0.048*
(0.025)
-0.049**
(0.025)

-0.003
(0.024)
0.000
(0.000)
-0.413***
(0.028)
-0.001***
(0.000)
0.083***
(0.025)
-0.089***
(0.025)

Non-ASD

Sped
(6)

-0.050*
(0.026)
0.065***
(0.006)
-0.183***
(0.029)
0.045***
(0.007)
0.005
(0.027)
0.057**
(0.026)

385,960

2,712,322

2,352,871

2,712,322

Observations
Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the 2009-
2010 to the 2013-2014 school years. 2014-2015 is excluded as Michigan changed from the Michigan Assessment of
Educational Progress to the M-Step exam and restructured alternative examination options. The sample includes only
students who are always or never eligible for free/reduced price lunch. Each column is a separate regression. “Non-
disadv” is an indicator for whether the student is eligible for free/reduced price lunch in all observed years of schooling
and “ASD” is an indicator for whether the student has an ASD diagnosis in that year. All regressions include controls
for whether a student is white, male, and limited English proficient as well as school and grade-by-year fixed effects.
Standard errors are clustered at the school district level: *,**,*** indicate significance at the 10, 5, and 1 percent level,
respectively.

2,352,871

385,960

183

Table C.1.8: Heterogeneous Effects of the ASD Insurance Mandate on ASD Incidence

Independent
Variable

Non-disadv*
Post-2012

Girls
(1)

Boys
(2)

-0.00020
(0.00016)

-0.00073*
(0.00044)

Observations
Incidence Rate

1,878,120

1,976,114

0.003

0.017

Panel A: By Gender and Race

White & Asian Black & Hispanic

(3)

(4)

-0.00027
(0.00024)

3,272,013

0.011

0.00034
(0.00082)

582,221
0.007

Independent
Variable

KG
(1)

Grade 1

(2)

Grade 2

(3)

Panel B: By Grade
Grade 3

(4)

Grade 4

(5)

Grade 5

(6)

Grade 6 Grade 7 Grade 8

(7)

(8)

(9)

Non-disadv*Post-
2012

-0.00036
(0.00051)

-0.00082
(0.00052)

-0.00204***
(0.00055)

-0.00068
(0.00057)

Observations
Incidence Rate

594,630
0.007

545,212
0.008

526,440
0.009

520,105
0.010

-0.00024
(0.00058)

524,539
0.010

0.00121**
(0.00060)

0.00037
(0.00063)

-0.00048
(0.00059)

-0.00065
(0.00054)

538,197
0.011

559,531
0.010

588,427
0.010

503,930
0.010

Notes: Authors’ estimates of equation (3.1) as described in the text using data on students in grades 2-8 from the 2009-2010 to the 2014-2015 school years. The
sample includes only students who are always or never eligible for free/reduced price lunch. Each column is a separate regression. “Non-disadv” is an indicator for
whether the student is eligible for free/reduced price lunch. All regressions include the full set of controls listed in equation (3.2), including controls for whether
a student is white, male, and limited English proficient as well as school and grade-by-year fixed effects. Standard errors are clustered at the school district level:
*,**,*** indicate significance at the 10, 5, and 1 percent level, respectively.

184

Table C.1.9: The Effect of the ASD Insurance Mandate, by Gender and Race

Panel A: Boys

Dependent Variable:
ASD

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(2)

(3)

Teacher

Consultant

(7)

(4)
0.006
(0.005)

(5)
0.000
(0.016)

0.029**
(0.012)

-0.008
(0.008)

-0.021*
(0.011)

359,165
0.130

300,256
0.062

300,256
0.529

218,809
0.349

359,165
0.130

Independent
Variable
Non-disadv*Post-
2012*ASD

(1)
0.032
(0.034)

Observations
ASD Mean

359,165
0.193

-0.057*
(0.029)

359,165
0.662

Independent
Variable
Non-disadv*Post-
2012*ASD

(1)
0.012
(0.047)

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(2)

(5)

(3)

(4)

Teacher

Consultant

(7)

-0.092**
(0.046)

0.066***
(0.023)

-0.005
(0.011)

-0.002
(0.031)

-0.011
(0.029)

-0.031
(0.021)

Panel B: Girls

Dependent Variable:
ASD

Observations
ASD Mean

185,888
0.222

185,888
0.644

185,888
0.123

155,495
0.076

155,495
0.489

113,563
0.380

185,888
0.131

Panel C: Black & Hispanic

Dependent Variable:
ASD

Independent
Variable
Non-disadv*Post-
2012*ASD

ASD

No
Sped

Special General

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)
0.019
(0.028)

(5)
0.005
(0.049)

(1)
0.083
(0.091)

(4)
0.012
(0.012)

-0.016
(0.034)

-0.058
(0.082)

Ed

Ed

(2)

(3)

Observations
ASD Mean

82,393
0.320

82,393
0.579

82,393
0.093

58,337
0.073

58,337
0.428

29,426
0.396

Teacher

Consultant

(7)

-0.012
(0.025)

82,393
0.088

Panel D: White & Asian

Dependent Variable:
ASD

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(1)

(2)

(3)

Teacher

Consultant

(7)

(4)
0.001
(0.003)

(5)
0.009
(0.013)

-0.010
(0.009)

-0.016*
(0.009)

-0.008
(0.009)

-0.033**
(0.014)

0.039***
(0.010)

Independent
Variable
Non-disadv*Post-
2012*ASD

Language
Services

(8)

-0.003
(0.014)

359,165
0.788

Language
Services

(8)

0.050*
(0.026)

185,888
0.805

Language
Services

(8)
0.012
(0.035)

82,393
0.849

Language
Services

(8)

-0.010
(0.012)

Occupational

Therapy
Services

(9)

-0.006
(0.013)

359,165
0.400

Occupational

Therapy
Services

(9)
0.009
(0.030)

Social
Worker

(10)

-0.024**
(0.012)

Any
Sped

Services

(11)
0.012
(0.010)

359,165
0.694

359,165
0.943

Social
Worker

(10)
-0.022
(0.030)

Any
Sped

Services

(11)

0.050**
(0.020)

185,888
0.410

185,888
0.676

185,888
0.944

Occupational

Therapy
Services

(9)

-0.001
(0.043)

82,393
0.387

Occupational

Therapy
Services

(9)

-0.005
(0.012)

Social
Worker

(10)
-0.007
(0.040)

Any
Sped

Services

(11)
-0.011
(0.032)

82,393
0.662

82,393
0.946

Social
Worker

(10)
-0.020
(0.013)

Any
Sped

Services

(11)
0.014*
(0.008)

462,660
0.133

462,660
0.183

462,660
0.668

Observations
ASD Mean
Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the 2009-2010 to the 2014-2015 school years. The
sample includes only students who are always or never eligible for free/reduced price lunch. Each column is a separate regression; N=707,376. “Non-disadv” is
an indicator for whether the student is eligible for free/reduced price lunch and “ASD” is an indicator for whether the student has an ASD diagnosis in that year.
All regressions include the full set of controls listed in equation (3.2), including controls for whether a student is white, male, and limited English proficient
as well as school and grade-by-year fixed effects. The final row of each panel provides dependent variable means for the ASD sample. Standard errors are
clustered at the school district level: *,**,*** indicate significance at the 10, 5, and 1 percent level, respectively.

462,660
0.942

462,660
0.784

462,660
0.403

462,660
0.695

397,414
0.063

397,414
0.533

302,946
0.350

462,660
0.135

185

Table C.1.10: The Effect of the ASD Insurance Mandate, by Grade

ASD

Resource or
Cognitive

Sped
FTE
Independent Program Program Program School >80% Rate
Variable
(6)
KG

Special General

No
Sped

Ed

(5)

Ed

(3)

(1)

-0.006
(0.038)
0.045
(0.041)
0.022
(0.046)
0.053
(0.047)
0.055
(0.044)
0.011
(0.037)
0.004
(0.030)
0.020
(0.040)
0.022
(0.025)

(2)
0.027
(0.038)
-0.037
(0.041)
-0.072*
(0.042)
-0.10**
(0.047)
-0.080*
(0.042)
-0.040
(0.039)
-0.033
(0.031)
-0.020
(0.039)
-0.010
(0.032)

1

2

3

4

5

6

7

8

(4)
0.006
(0.013)
-0.007
(0.012)
-0.011
(0.010)
0.013
(0.012)
-0.010*
(0.006)
-0.004
(0.008)
0.002
(0.008)
0.004
(0.011)
0.010
(0.009)

-0.005
(0.028)
-0.017
(0.028)
0.049**
(0.025)
0.041*
(0.024)
0.029
(0.021)
0.018
(0.020)
0.037**
(0.017)
0.000
(0.019)
-0.004
(0.021)

-0.001
(0.033)
-0.032
(0.032)
-0.021
(0.031)
-0.014
(0.031)
-0.031
(0.033)
0.007
(0.030)
0.013
(0.030)
-0.015
(0.031)
0.020
(0.032)

-0.002
(0.033)
0.004
(0.025)
0.012
(0.031)
0.006
(0.025)
-0.001
(0.021)
0.031
(0.025)
-0.022
(0.023)
-0.022
(0.024)
-0.039*
(0.022)

ASD

Teacher

Consultant

(7)

-0.006
(0.023)
-0.035
(0.025)
-0.047**
(0.022)
-0.040*
(0.021)
-0.003
(0.018)
-0.008
(0.018)
-0.011
(0.018)
-0.015
(0.020)
-0.021
(0.022)

Language
Services

(8)

-0.014
(0.018)
0.022
(0.021)
0.025
(0.024)
0.008
(0.022)
0.007
(0.026)
0.002
(0.026)
-0.018
(0.025)
-0.035
(0.026)
-0.041
(0.028)

Occupational

Therapy
Services

(9)

-0.012
(0.042)
-0.018
(0.032)
-0.037
(0.030)
-0.058**
(0.027)
0.066**
(0.029)
0.040
(0.026)
0.013
(0.024)
-0.029
(0.023)
0.015
(0.023)

Any
Sped

Social
Worker Services

(10)
-0.049
(0.033)
-0.002
(0.033)
-0.022
(0.029)
-0.010
(0.027)
-0.015
(0.026)
-0.041*
(0.024)
-0.020
(0.023)
0.001
(0.026)
0.002
(0.025)

(11)
-0.017
(0.013)
-0.001
(0.015)
0.025
(0.017)
0.019
(0.016)
0.029
(0.021)
-0.002
(0.018)
0.007
(0.017)
0.013
(0.018)
-0.013
(0.021)

Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades Kindergarten-8 from the 2009-2010 to the 2014-2015 school
years. The sample includes only students who are always or never eligible for free/reduced price lunch. Each cell is a separate regression and shows the estimate of
the coefficient on the triple interaction term of Non-disadv*Post-2012*ASD. All regressions include the full set of controls listed in equation (3.2), including controls
for whether a student is white, male, and limited English proficient as well as school and grade-by-year fixed effects. Standard errors are clustered at the school district
level: *,**,*** indicate significance at the 10, 5, and 1 percent level, respectively.

186

Table C.1.11: Heterogeneous Effects of the ASD Insurance Mandate on Test Scores, using Non-
Special Education Control Group

Independent
Variable

Non-disadv*
Post-2012

Girls
(1)

0.079
(0.070)

Panel A: Math, by Gender and Race

Boys
(2)

White/Asian Black/Hisp.

(3)

(4)

-0.005
(0.028)

0.006
(0.028)

-0.008
(0.091)

Observations

803,320

775,726

1,365,614

213,432

Independent
Variable

Non-disadv*
Post-2012

Panel B: Reading, by Gender and Race

Girls
(1)

Boys
(2)

White/Asian Black/Hisp.

(3)

(4)

0.061
(0.087)

0.008
(0.030)

0.035
(0.031)

-0.115
(0.108)

Observations

803,315

775,622

1,365,950

212,987

Independent
Variable
Non-disadv*Post-
2012

Grade 4

(1)

0.136*
(0.073)

Panel C: Math, by Grade

Grade 5

(2)

Grade 6

(3)

-0.079
(0.066)

-0.013
(0.058)

Grade 7

(4)
0.006
(0.059)

Grade 8

(5)
0.019
(0.061)

Observations

298,060

300,174

309,489

326,170

343,020

Independent
Variable
Non-disadv*Post-
2012

Grade 4

(1)
0.024
(0.067)

Panel D: Reading, by Grade
Grade 6

Grade 5

(2)
0.072
(0.063)

(3)

-0.069
(0.066)

Grade 7

(4)
0.039
(0.064)

Grade 8

(5)
0.069
(0.068)

Observations

298,012

299,973

309,512

326,192

343,089

Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from
the 2009-2010 to the 2013-2014 school years. 2014-2015 is excluded as Michigan changed from the Michigan
Assessment of Educational Progress to the M-Step exam and restructured alternative examination options. The
sample includes only students who are always or never eligible for free/reduced price lunch. Each column is a
separate regression. “Non-disadv” is an indicator for whether the student is eligible for free/reduced price lunch
and “ASD” is an indicator for whether the student has an ASD diagnosis in that year. All regressions include the
full set of controls listed in equation (3.2), including controls for whether a student is white, male, limited English
proficient, as well as school and grade-by-year fixed effects. Estimates also include controls for lagged test score.
Standard errors are clustered at the school district level: *,**,*** indicate significance at the 10, 5, and 1 percent
level, respectively.

187

Table C.1.12: The Effect of the ASD Insurance Mandate – Robustness Checks

Panel A: Including Linear Time Interacted with Non-Disadvantaged Status

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(2)

(3)

Teacher

Consultant

(7)

Dependent Variable:
ASD

-0.064**
(0.030)

0.034***
(0.011)

(4)
0.004
(0.004)

(5)
0.000
(0.015)

-0.009
(0.009)

-0.023**
(0.011)

Independent
Variable
Non-disadv*Post-
2012*ASD

(1)
0.031
(0.035)

Observations
ASD Mean

545,053
0.197

545,053
0.670

545,053
0.129

455,751
0.056

455,751
0.455

332,372
0.353

545,053
0.130

545,053
0.790

Language
Services

(8)
0.005
(0.013)

Occupational

Therapy
Services

(9)

-0.004
(0.012)

545,053
0.401

Panel B: School District FEs

Dependent Variable:
ASD

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(2)

(3)

(4)

Teacher

Consultant

(7)

-0.068**
(0.030)

0.041***
(0.011)

-0.006
(0.007)

-0.013
(0.011)

-0.023**
(0.011)

(5)
0.014
(0.016)

Independent
Variable
Non-disadv*Post-
2012*ASD

(1)
0.028
(0.036)

Observations
ASD Mean

545,053
0.197

545,053
0.670

545,053
0.129

455,751
0.056

455,751
0.455

332,372
0.353

545,053
0.130

Language
Services

(8)

-0.000
(0.014)

545,053
0.790

Occupational

Therapy
Services

(9)

-0.017
(0.013)

545,053
0.401

Panel C: Pre-Treatment Assignment of ASD Variable

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(2)

(3)

(5)

Teacher

Consultant

(7)

Dependent Variable:
ASD

-0.072***
(0.028)

0.049***
(0.011)

-0.011
(0.016)

-0.002
(0.008)

-0.011
(0.011)

(4)
0.004
(0.004)

Independent
Variable
Non-disadv*Post-
2012*ASD

(1)
0.020
(0.032)

Language
Services

(8)
0.001
(0.016)

Occupational

Therapy
Services

(9)

-0.012
(0.014)

Social
Worker

(10)

-0.021*
(0.012)

Any
Sped

Services

(11)
0.019*
(0.010)

545,053
0.691

545,053
0.943

Social
Worker

(10)
-0.015
(0.014)

Any
Sped

Services

(11)
0.016
(0.010)

545,053
0.691

545,053
0.943

Social
Worker

(10)

-0.041***
(0.013)

Any
Sped

Services

(11)
0.019*
(0.011)

352,054
0.534

352,054
0.049

422,974
0.133

422,974
0.661

422,974
0.190

Observations
ASD Mean
Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the 2010-2011 to the 2014-2015 school years. The
sample in Panel A includes only students who are always or never eligible for free/reduced price lunch, while the sample in Panels B and C include all students.
Each column is a separate regression. “Non-disadv” is an indicator for whether the student is eligible for free/reduced price lunch and “ASD” is an indicator
for whether the student has an ASD diagnosis in that year. All regressions include the full set of controls listed in equation (3.2), including whether a student is
white, male, and limited English proficient as well as grade-by-year fixed effects. Estimates in Panels A and C include school fixed effects, while those in Panel
B include school district fixed effects. Estimates in Panel A also control for a linear time trend interacted with non-disadv status. The final row of each panel
provides dependent variable means for the ASD sample. Standard errors are clustered at the school district level: *,**,*** indicate significance at the 10, 5, and
1 percent level, respectively.

422,974
0.779

422,974
0.373

422,974
0.700

422,974
0.939

332,338
0.338

252,537
0.132

188

Table C.1.13: The Effect of the ASD Insurance Mandate on Test Scores - Robustness Checks

Math

(2)
0.005
(0.026)

Math

(2)
0.014
(0.026)

Math

(2)
0.002
(0.027)

Math

(2)

All

(1)
0.010
(0.026)

All

(1)
0.000
(0.027)

All

(1)

Panel A: Including Linear Time Trend Interacted with Non-Disadvantaged Status

Control Group:

Non-disadv*Post-
2012*ASD

All

(1)
0.001
(0.026)

Non-ASD Non-Sped

Non-ASD Non-Sped

Non-ASD

Sped
(3)

-0.010
(0.027)

All

(4)
0.021
(0.029)

Reading

(5)
0.018
(0.029)

Non-ASD

Sped
(6)
0.004
(0.030)

Observations

1,754,971

1,579,046

185,814

1,749,290

1,578,937

180,172

Panel B: School District FEs

Non-ASD Non-Sped

Non-ASD Non-Sped

Non-ASD

Sped
(3)

-0.001
(0.027)

All

(4)
0.029
(0.029)

Reading

(5)
0.027
(0.029)

Non-ASD

Sped
(6)
0.014
(0.029)

Control Group:

Non-disadv*Post-
2012*ASD

Observations

1,754,971

1,579,046

185,814

1,749,290

1,578,937

180,172

Panel C: Pre-Treatment Assignment of ASD Variable

Non-ASD Non-Sped

Non-ASD Non-Sped

Non-ASD

Sped
(3)

-0.017
(0.028)

All

(4)
0.031
(0.029)

Reading

(5)
0.028
(0.029)

Non-ASD

Sped
(6)
0.007
(0.032)

Control Group:

Non-disadv*Post-
2012*ASD

Observations

1,542,200

1,387,847

162,767

1,537,785

1,388,220

157,909

Control Group:

Non-ASD Non-Sped

Non-ASD Non-Sped

Panel D: Including 2014-15

Non-ASD

Sped
(3)

-0.045*
(0.025)

All

(4)
0.028
(0.026)

Reading

(5)
0.034
(0.026)

Non-ASD

Sped
(6)
0.008
(0.028)

Non-disadv*Post-
2012*ASD

-0.019
(0.024)

-0.005
(0.024)

Observations

2,184,874

1,961,605

236,273

2,179,192

1,961,404

230,726

Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from
the 2009-2010 to the 2013-2014 school years, except where specified. The sample includes only students
who are always or never eligible for free/reduced price lunch. Each column is a separate regression. “Non-
disadv” is an indicator for whether the student is eligible for free/reduced price lunch and “ASD” is an
indicator for whether the student has an ASD diagnosis in that year. All regressions include the full set of
controls listed in equation (3.2), including whether a student is white, male, and limited English proficient
as well as grade-by-year fixed effects. Estimates in Panels A, C, and D include school fixed effects, while
those in Panel B include school district fixed effects. The estimates in Panel A also include a linear time
trend interacted with non-disadv status. Estimates also include controls for lagged test score. Standard
errors are clustered at the school district level: *,**,*** indicate significance at the 10, 5, and 1 percent
level, respectively.

189

Table C.1.14: The Effect of the ASD Insurance Mandate – No Sample Exclusion Based on Disadvantaged Status

Panel A: Including All Students Who are Ever Observed Disadvantaged

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(2)

(5)

(3)

Teacher

Consultant

(7)

Dependent Variable:
ASD

(4)
0.002
(0.003)

0.020**
(0.010)

-0.006
(0.014)

-0.010
(0.007)

-0.017*
(0.009)

Language
Services

(8)
0.005
(0.011)

Occupational

Therapy
Services

(9)
0.003
(0.012)

Any
Sped

Social
Worker Services

(10)
-0.011
(0.011)

(11)
0.012
(0.008)

705,616
0.125

577,221
0.056

577,221
0.444

421,418
0.359

705,616
0.129

705,616
0.784

705,616
0.395

705,616
0.688

705,616
0.940

Independent
Variable
Non-disadv*Post-
2012*ASD

(1)
0.021
(0.030)

Observations
ASD Mean

705,616
0.200

Independent
Variable
Non-disadv*Post-
2012*ASD

(1)
0.021
(0.030)

Observations
ASD Mean

705,616
0.200

-0.041*
(0.025)

705,616
0.672

-0.041*
(0.025)

705,616
0.672

Panel B: Including All Students Who are Ever Observed Disadvantaged and Linear Time Trend Interacted with Non-Disadvantaged Status

ASD

Resource or
Cognitive

Sped
FTE
Program Program Program School >80% Rate
(6)

Special General

No
Sped

Ed

Ed

(2)

(5)

(3)

Teacher

Consultant

(7)

Dependent Variable:
ASD

(4)
0.002
(0.003)

0.020**
(0.010)

-0.006
(0.014)

-0.010
(0.007)

-0.017*
(0.009)

Language
Services

(8)
0.005
(0.011)

Occupational

Therapy
Services

(9)
0.003
(0.012)

Any
Sped

Social
Worker Services

(10)
-0.011
(0.011)

(11)
0.012
(0.008)

705,616
0.125

577,221
0.056

577,221
0.444

421,418
0.359

705,616
0.129

705,616
0.784

705,616
0.395

705,616
0.688

705,616
0.940

Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the 2010-2011 to the 2014-2015 school years. The sample
includes all students. Each column is a separate regression. “Non-disadv” is an indicator for whether the student is eligible for free/reduced price lunch and “ASD” is an
indicator for whether the student has an ASD diagnosis in that year. All regressions include the full set of controls listed in equation (3.2), including whether a student is white,
male, and limited English proficient as well as school and grade-by-year fixed effects. The estimates in Panel B also include a linear time trend interacted with non-disadv
status. The final row in each panel provides dependent variable means for the ASD sample. Standard errors are clustered at the school district level: *,**,*** indicate
significance at the 10, 5, and 1 percent level, respectively.

190

Table C.1.15: The Effect of the ASD Insurance Mandate on Disability Incidence - No Sample Exclusion Based on Disadvantaged Status

Panel A: Excluding Linear Time Trend Interacted with Non-Disadvantaged Status

Dependent Variable:

ASD
(1)

Non-ASD
Disability

(2)

Cognitive Emotional
Disability Disability

(3)

(4)

Speech
Disability

(5)

Learning
Disability

(6)

Other Health

Disability

(7)

-0.00032
(0.00024)

0.00123
(0.00136)

0.00004
(0.00028)

0.00113***
(0.00027)

-0.00215*** 0.00420***
(0.00053)
(0.00093)

-0.00214***
(0.00040)

Panel B: Including Linear Time Trend Interacted with Non-Disadvantaged Status

ASD
(1)

Non-ASD
Disability

(2)

Dependent Variable:

Cognitive Emotional
Disability Disability

(3)

(4)

Speech
Disability

(5)

0.00009
(0.00026)

-0.00050
(0.00110)

-0.00038
(0.00030)

0.00040
(0.00031)

0.00012
(0.00061)

Learning
Disability

(6)

-0.00113*
(0.00069)

Other Health

Disability

(7)

0.00056
(0.00039)

Independent
Variable
Non-disadv*
Post-2012

Independent
Variable
Non-disadv*
Post-2012

0.010

0.133

0.011

Incidence Rate
Notes: Authors’ estimates of equation (3.1) as described in the text using data on students in grades 2-8 from the 2009-2010 to the 2014-
2015 school years. Each column is a separate regression; N=4,970,113. “Non-disadv” is an indicator for whether the student is eligible for
free/reduced price lunch. All regressions include the full set of controls listed in equation (3.1), including whether a student is white, male, and
limited English proficient as well as school and grade-by-year fixed effects. Estimates in Panel B include the control for a linear time trend
interacted with non-disadvantage status. Standard errors are clustered at the school district level: *,**,*** indicate significance at the 10, 5, and
1 percent level, respectively.

0.016

0.056

0.009

0.035

191

Table C.1.16: The Effect of the ASD Insurance Mandate on Test Scores - No Sample Exclusion
Based on Disadvantaged Status

Panel A: Including All Students Who are Ever Observed Disadvantaged

Control Group:

Non-ASD Non-Sped

Non-ASD Non-Sped

Math

(2)

Non-ASD

Sped
(3)

-0.023
(0.024)

All

(4)
0.014
(0.026)

Reading

(5)
0.010
(0.026)

Non-ASD

Sped
(6)

-0.001
(0.026)

Non-disadv*Post-
2012*ASD

-0.011
(0.023)

-0.009
(0.023)

All

(1)

All

(1)

Observations

2,158,249

1,940,574

230,304

2,151,497

1,940,733

223,320

Panel B: Including All Students Who are Ever Observed Disadvantaged

and Linear Time Trend Interacted with Non-Disadvantage Status
Reading

Math

Control Group:

Non-ASD Non-Sped

Non-ASD Non-Sped

Non-ASD

All

(2)

Sped
(3)

(4)
0.014
(0.026)

(5)
0.010
(0.026)

Non-ASD

Sped
(6)

-0.001
(0.026)

Non-disadv*Post-
2012*ASD

-0.012
(0.023)

-0.009
(0.023)

-0.023
(0.024)

230,304

2,158,249

1,940,574

2,151,497

Observations
Notes: Authors’ estimates of equation (3.2) as described in the text using data on students in grades 2-8 from the
2009-2010 to the 2013-2014 school years, except where specified. The sample includes all students. Each column
is a separate regression. “Non-disadv” is an indicator for whether the student is eligible for free/reduced price lunch
and “ASD” is an indicator for whether the student has an ASD diagnosis in that year. All regressions include the full
set of controls listed in equation (3.2), including whether a student is white, male, and limited English proficient as
well as school and grade-by-year fixed effects. The estimates in Panel B include the linear time trend interacted with
non-disadvantaged status. Estimates also include controls for lagged test score. Standard errors are clustered at the
school district level: *,**,*** indicate significance at the 10, 5, and 1 percent level, respectively.

1,940,733

223,320

192

BIBLIOGRAPHY

193

BIBLIOGRAPHY

Allen, J. (2019). Connecticut Community Colleges Offer SNAP Scholarship. https://www.wnpr.

org/post/connecticut-community-colleges-offer-snap-scholarship.

Altonji, J. G., P. Arcidiacono, and A. Maurel (2016). The Analysis of Field Choice in College and
Graduate School: Determinants and Wage Effects. Handbook of the Economics of Education 5,
305–396.

Altonji, J. G., E. Blom, and C. Meghir (2012). Heterogeneity in Human Capital Investments: High

School Curriculum, College Major, and Careers. Annual Review of Economics 4, 185–223.

Anderoni, J. and A. A. Payne (2011). Is Crowding Out Due Entirely to Fundraising? Evidence

from a Panel of Charities. Journal of Public Economics 95(5-6), 334–343.

Andrews, R. J., S. A. Imberman, and M. F. Lovenheim (2017). Risky Business? The Effect
of Majoring in Business on Earnings and Educational Attainment. NBER Working Paper No.
23575.

Armona, L., R. Chakrabarti, and M. F. Lovenheim (2018). How Does For-Profit College Atten-
dance Affect Student Loans, Defaults and Labor Market Outcomes? NBER Working Paper No.
25042.

Bahr, P. R., S. Dynarski, B. Jacob, D. Kreisman, A. Sosa, and M. Wiederspan (2015). Labor
Market Returns to Community College Awards: Evidence from Michigan. EPI Working Paper
01-2015.

Baio, Jon, et al. (2018). Prevalence of Autism Spectrum Disorder Among Children Aged 8 Years
Autism and Developmental Disabilities Monitoring Network. Morbitity and Mortality Weekly
Report 67(6), 1–23.

Baker, R., E. Bettinger, B. Jacob, and I. Marinescu (2018). The Effect of Labor Market Information

on Community College Students’ Major Choice. Economics of Education Review 65, 18–30.

Barry, C. L., A. J. Epstein, S. C. Marcus, A. Kennedy-Hendricks, M. K. Candon, M. Xie, and
D. S. Mandell (2017). Effects of State Insurance Mandates on Health Care Use and Spending
for Autism Spectrum Disorder. Health Affairs 36(10), 1754–1761.

Beffy, M., D. Fougere, and A. Maurel (2012). Choosing the Field of Study in Postsecondary
The Review of Economics and Statistics 94(1),

Education: Do Expected Earnings Matter?
334–347.

Belfield, C. and T. Bailey (2017). The Labor Market Returns to Sub-Baccalaureate College: A

Review. CAPSEE Working Paper.

Bellemare, M. F. and C. J. Wichman (2019). Elasticities and the Inverse Hyperbolic Sine Trans-

formation. Oxford Bulletin of Economics and Statistics.

194

Benson, C. (2018). Is Special Education a Pathway to Supplemental Security Income for Children?

Working Paper.

Bergstrom, T., L. Blume, and H. Varian (1986). On the Private Provision of Public Goods. Journal

of Public Economics 29(1), 25–49.

Betts, J. R. and L. L. McFarland (1995). Safe Port in a Storm: The Impact of Labor Market
Conditions on Community College Enrollments. The Journal of Human Resources 30, 741–
765.

Blank, R. (1989). The Effect of Medical Need and Medicaid on AFDC Participation. Journal of

Human Resources 24(1), 54–87.

Buckles, K. S. and D. M. Hungerman (2018). The Incidental Fertility Effects of School Condom

Distribution Programs. Journal of Policy Analysis and Management 37(3), 464–492.

Buescher, A. V., Z. Cidav, M. Knapp, and D. S. Mandell (2014). Costs of Autism Spectrum

Disorders in the United Kingdom and the United States. JAMA Pediatrics 168(8), 721–728.

Buis, M. (2017). Fmlogit: Stata module fitting a fractional multinomial logit model by quasi

maximum likelihood.

Burbidge, J. B., L. Magee, and A. L. Robb (1988). Alternative Transformations to Handle Extreme
Values of the Dependent Variable. Journal of the American Statistical Association 83(401), 123–
127.

Bureau of Labor Market Information and Strategic Initiatives (2014). Youth and Young Adults
and the Michigan Labor Market. http://milmi.org/Portals/198/publications/Youth and Youth
Adults and the Michigan Labor Market.pdf.

Caetano, G. and H. Macartney (2014). Quasi-Experimental Evidence of School Choice through

Residential Sorting. Working Paper.

Candon, M. K., C. L. Barry, S. C. Marcus, A. J. Epstein, A. Kennedy-Hendricks, M. Xie, and D. S.
Mandell (2019). Insurance Mandates and Out-of-Pocket Spending for Children With Autism
Spectrum Disorder. Pediatrics 143(1).

Card, D. (1995). Using Geographic Variation in College Proximity to Estimate the Return to
Schooling. In L. Christofides, E. K. Grant, and R. Swindinsky (Eds.), Aspects of Labour Eco-
nomics: Essays in Honour of John Vanderkamp. University of Toronto Press.

Carruthers, C. K. and W. F. Fox (2016). Aid for All: College Coaching, Financial Aid, and Post-

secondary Persistence in Tennessee. Economics of Education Review 51, 97–112.

Cellini, S. R. (2009). Crowded Colleges and College Crowd-Out: The Impact of Public Subsidies

on the Two-Year College Market. American Economic Journal: Economic Policy 1, 1–30.

Cellini, S. R. and C. Goldin (2014). Does Federal Student Aid Raise Tuition? New Evidence on

For-Profit Colleges. American Economic Journal: Economic Policy 6, 174–206.

195

Cellini, S. R. and N. Turner (2018). Gainfully Employed? Assessing the Employment and Earn-
ings of For-Profit College Students Using Administrative Data. Journal of Human Resources,
forthcoming.

Centers for Disease Control (2018). Autism Spectrum Disorder (ASD). https://www.cdc.gov/

ncbddd/autism/facts.html.

Charles, K. K., E. Hurst, and M. J. Notowidigdo (2018). Housing Booms and Busts, Labor Market

Opportunities, and College Attendance. American Economic Review 108, 2947–2994.

Chatterji, P., S. L. Decker, and S. Markowitz (2015). The Effects of Mandated Health Insurance
Benefits for Autism on Out-of-Pocket Costs and Access to Treatment. Journal of Policy Analysis
and Management 34(2), 328–353.

Chetty, R., J. N. Friedman, E. Saez, N. Turner, and D. Yagan (2017). Mobility Report Cards: The

Role of Colleges in Intergenerational Mobility. NBER Working Paper, No. 23618.

Choi, D., D. Lou, and A. Mukherjee (2018). The Effect of Superstar Firms on College Major

Choice. Working Paper.

Christensen, Deborah L., et al. (2016). Prevalence and Characteristics of Autism Spectrum Dis-
order Among 4-year-old Children in the Autism and Developmental Disabilities Monitoring
Network. Journal of Developmental & Behavioral Pediatrics 37(1), 1–8.

Cidav, Z., S. C. Marcus, and D. S. Mandell (2012). Implications of Childhood Autism for Parental

Employment and Earnings. Pediatrics 129(4), 617–623.

Citizens Research Council of Michigan (2018).

Exploring Michigan’s Urban/Rural
https://www.michiganfoundations.org/sites/default/files/resources/rpt400 Exploring

Divide.
Michigans Urban-Rural Divide.pdf.

Cohodes, S. R., D. S. Grossman, S. A. Kleiner, and M. F. Lovenheim (2016). The Effect of Child
Health Insurance Access on Schooling: Evidence from Public Insurance Expansions. Journal
of Human Resources 51(3), 727–759.

Corsello, C. M. (2005). Early Intervention in Autism. Infants & Young Children 18(2), 74–85.

Cowen, J. M., B. Creed, and V. A. Keesler (2015). Dynamic participation in inter-district open
enrollment: Evidence from Michigan 2006-2013. Michigan State University Education Policy
Center Working Paper No. 49.

Currie, J. and E. Moretti (2003). Mother’s Education and the Intergenerational Transmission of
Human Capital: Evidence from College Openings. The Quarterly Journal of Economics 118,
1495–11532.

Cutler, D. M. and J. Gruber (1996). Does Public Insurance Crowd out Private Insurance? The

Quarterly Journal of Economics 111(2), 391–430.

Dawson, Geraldine, et al. (2010). Randomized, Controlled Trial of an Intervention for Toddlers

with Autism: The Early Start Denver Model. Pediatrics 125(1).

196

Degrow, B. (2017). How School Funding Works in Michigan Mackinac Center for Public Policy.

https://www.mackinac.org/school-funding.

Deming, D. and S. Dynarski (2010). College Aid. In P. B. Levine and D. J. Zimmerman (Eds.),
Targeting Investments in Children: Fighting Poverty When Resources are Limited. University of
Chicago Press.

Deming, D. J., C. Goldin, and L. F. Katz (2012). The For-Profit Postsecondary School Sector:

Nimble Critters or Agile Predators? Journal of Economic Perspectives 26(1), 139–164.

Denning, J. T. (2017). College on the Cheap: Consequences of Community College Tuition Re-

ductions. American Economic Journal: Economic Policy 9(2), 155–188.

Domina, T., N. Pharris-Ciurej, A. M. Penner, E. K. Penner, Q. Brummet, S. R. Porter, and
Is Free and Reduced-Price Lunch a Valid Measure of Educational Dis-

T. Sanabria (2010).
advantage? Educational Researcher 47(9), 539–555.

Eckert, F., T. C. Fort, P. K. Schott, and N. J. Yang (2020). Imputing Missing Values in the Census

Bureau’s County Business Patterns. NBER Working Paper, No. 26632.

Eldevik, Sigmund, et al. (2009). Meta-analysis of Early Intensive Behavioral Intervention for

Children with Autism. Journal of Clinical Child & Adolescent Psychology 38(3), 439–450.

Elwell, J. (2018). The Effects of Expansions of Children’s Medicaid Eligibility on Program Par-

ticipation and Labor Supply. Working Paper.

Ersoy, F. Y. (2019). Reshaping Aspirations: The Effects of the Great Recession on College Major

Choice. Working Paper.

Ferreri, S. and S. Bolt (2011). Educating Michigan’s Students with Autism Spectrum Disorder
(ASD): An Initial Exploration of Programming ‘The ASD Michigan Project’. Education Policy
Center Special Report.

Foote, A. and M. Grosz (2019). The Effect of Local Labor Market Downturns on Postsecondary

Enrollment and Program Choice. Education Finance and Policy, forthcoming.

Fountain, C., M. D. King, and P. S. Bearman (2011). Age of Diagnosis for Autism: Individual
and Community Factors across 10 Birth Cohorts. Journal of Epidemiology and Community
Health 65, 503–510.

Foxx, R. M. (2008). Applied Behavioral Analysis Treatment of Autism: The State of the Art.

Child and Adolescent Psychiatric Clinics of North America 17(4), 821–834.

Fronstin, P. (2012). Self-Insured Health Plans: State Variation and Recent Trends by Firm Size.

Employee Benefit Research Institute Notes 33(11), 2–11.

Ganz, M. L. (2007). The Lifetime Distribution of the Incremental Societal Costs of Autism.

Archives of Pediatrics & Adolescent Medicine 161(4), 343–349.

Genniaoli, N. and A. Shleifer (2010). What Comes to Mind. Quarterly Journal of Economics.

197

Goodman, J., M. Hurwitz, and J. Smith (2017). Access to 4-Year Public Colleges and Degree

Completion. Journal of Labor Economics 35(3), 829–867.

Goodman, S. F. and A. H. Volz (2019). Attendance Spillovers between Public and For-Profit Col-
leges: Evidence from Statewide Variation in Appropriations for Higher Education. Education
Finance and Policy, forthcoming.

Gordon, N. (2004). Do Federal Grants Boost School Spending? Evidence from Title I. Journal of

Public Economics 88(9-10), 1771–1792.

Granpeesheh, Doreen, et al. (2009). The Effects of Age and Treatment Intensity on Behavioral
Intervention Outcomes for Children with Autism Spectrum Disorders. Research in Autism Spec-
trum Disorders 3(4), 1014–1022.

Grosz, M. (2018). Do Postsecondary Training Programs Respond to Changes in the Labor Market?

Working Paper.

Gruber, J. and D. M. Hungerman (2007). Faith-based Charity and Crowd-out During the Great

Depression. Journal of Public Economics 91(5-6), 1043–1069.

Gruber, J. and K. Simon (2008). Crowd-out 10 Years Later: Have Recent Public Insurance Expan-

sions Crowded Out Private Health Insurance. Journal of Health Economics 27(2), 201–217.

Ham, J. C. and L. D. Shore-Sheppard (2005). Did Expanding Medicaid Affect Welfare Participa-

tion? Industrial and Labor Relations Review 58(3), 452–470.

Hansen, S. N., D. E. Schendel, and E. T. Parner (2015). Explaining the increase in the prevalence of
autism spectrum disorders: the proportion attributable to changes in reporting practices. JAMA
Pediatrics 169(1), 56–62.

Hastings, J., C. A. Neilson, and S. D. Zimmerman (2015). The Effects of Earnings Disclosure on

College Enrollment Decisions. NBER Working Paper, No. 21300.

Hershbein, B. and M. Kearney (2014). Major Decisions: What College Graduates Earn Over Their
Lifetimes. https://www.hamiltonproject.org/papers/major decisions what graduates earn over
their lifetimes/.

Hilliard, T. (2016). Autonomy and Innovation: Systemic Change in a Decentralized State. Tech-

nical report, Jobs for the Future.

Hillman, N. and T. Weichman (2016). Education Deserts: The Continued Significance of “Place”
in the Twenty-First Century. Technical report, American Council on Education, Washington,
DC.

Hillman, N. W. and E. L. Orians (2013). Community Colleges and Labor Market Conditions: How
Does Enrollment Demand Change Relative to Local Unemployment Rates? Research in Higher
Education 54(7), 765–780.

198

House Fiscal Agency (2017).

Four-Year Degree Offerings at Michigan’s Community Col-
https://www.house.mi.gov/hfa/PDF/CommunityColleges/CC FourYearDegrees memo

leges.
Oct17.pdf.

Howlin, P., I. Magiati, and T. Charman (2009). Systematic Review of Early Intensive Behavioral
Interventions for Children with Autism. American Journal of Intellectual and Developmental
Disabilities 1141, 23–41.

Hubbard, D. (2018). The Impact of Local Labor Market Shocks on College Choice: Evidence

from Plant Closings in Michigan. Working Paper.

Hungerman, D. M. (2005). Are Church and State Substitutes? Evidence from the 1996 Welfare

Reform. Journal of Public Economics 89(11-12), 2245–2267.

Huttunen, K. and K. Riukula (2019). Parental Job Loss and Children’s Careers. IZA Discussion

Paper No. 12788.

J¨abrink, K. (2007). The Economic Consequences of Autistic Spectrum Disorder Among Children

in a Swedish Municipality. Autism 11(5), 453–463.

Jacobson, J. W., J. A. Mulick, and G. Green (1998). Cost-benefit Estimates for Early Intensive Be-
havioral Intervention for Young Children with Autism - General Model and Single State Case.
Behavioral Interventions: Theory & Practice in Residential & Community-Based Clinical Pro-
grams 13(4), 201–226.

Keele, L., S. Lorch, M. Passarella, D. Small, and R. Titiunik (2017). An Overview of Geographi-
cally Discontinuous Treatment Assignments with an Application to Children’s Health Insurance.
In M. D. Cattaneo and J. C. Escanciano (Eds.), Regression Discontinuity Designs: Theory and
Applications, Volume 38 of Advances in Econometrics. Emerald Group Publishing.

Knapp, M., R. Romeo, and J. Beecham (2009). Economic Cost of Autism in the UK. Autism 13(3),

317–336.

Krolikowski, P. M. and K. G. Lunsford (2020). Advance Layoff Notices and Labor Market Fore-

casting. Federal Reserve Bank of Cleveland Working Paper No. 20-03.

Lapid, P. A. (2017). Expanding College Access: The Impact of New Universities on Local Enroll-

ment. Job Market Paper.

Lavelle, Tara A., et al. (2014). Economic Burden of Childhood Autism Spectrum Disorders. Pedi-

atrics 133(3), e520–e529.

Levine, P. B. and D. Schanzenbach (2009). The Impact of Children’s Public Health Insurance

Expansions on Educational Outcomes. Forum for Health Economics & Policy 12(1).

Lindo, J. M., J. Schaller, and B. Hansen (2018). Caution! Men not at work: Gender-specific labor

market conditions and child maltreatment. Journal of Public Economics 163, 77–98.

Liu, S., W. Sun, and J. V. Winters (2018). Up in STEM, Down in Business: Changing College

Major Decisions with the Great Recession. Contemporary Economic Policy.

199

Long, M. C., D. Goldhaber, and N. Huntington-Klein (2015). Do completed college majors re-

spond to changes in wages? Economics of Education Review 49, 1–14.

Lovenheim, M. F., R. Reback, and L. Wedenoja (2016). How Does Access to Health Care Affect
Teen Fertility and High School Dropout Rates? Evidence from School-based Health Centers.
NBER Working Paper No. 22030.

Ma, J. and S. Buam (2016). Trends in Community Colleges: Enrollment, Prices, Student Debt,

and Completion. College Board Research Brief .

Mandell, D. S., C. L. Barry, M. Xie, K. Shea, K. Mullan, and A. J. Epstein (2016). Effects of
Autism Spectrum Disorder Insurance Mandates on the Treated Prevalence of Autism Spectrum
Disorder. JAMA Pediatrics 170(9), 887–893.

Mandell, D. S., K. H. Morales, M. Xie, L. J. Lawer, A. C. Stahmer, and S. C. Marcus (2010).
Age of Diagnosis Among Medicaid-Enrolled Children with Autism, 2001-2004. Psychiatric
Services 61(8), 822–829.

Mari-Bauset, Salvador et al. (2014). Evidence of the Gluten-Free and Casein-Free Diet in Autism

Spectrum Disorders: A Systematic Review. Journal of Child Neurology 29(12), 1718–1727.

Matson, J. L. and A. M. Kozlowski (2011). The increasing prevalence of autism spectrum disor-

ders. Research in Autism Spectrum Disorders 5(1), 418–425.

McFarlin, I., B. McCall, and P. Martorell (2018). How Much Do Tuition Subsidies Promote Col-

lege Access? Evidence from Community College Taxing Districts. Working Paper.

Michelmore, K. and S. Dynarski (2017). The Gap Within the Gap: Using Longitudinal Data to

Understand Income Differences in Educational Outcomes. AERA Open 3(1), 1–18.

Michigan Center for Educational Performance & Information (2017). Michigan Community Col-
leges Activities Classification Structure (ACS) 2016-2017 Data Book & Companion. http:
//www.michigancc.net/acs/ACS%202016-17.pdf.

Michigan Community College Association (2019). Fast Facts. https://www.mcca.org/fast-facts.

Michigan House Fiscal Agency (2017). Budget Briefing: Community Colleges. https://www.

house.mi.gov/hfa/PDF/Briefings/CC BudgetBriefing fy17-18.pdf.

Moffitt, R. and B. Wolfe (1992). The Effect of the Medicaid Program on Welfare Participation and

Labor Supply. The Review of Economics and Statistics 74(4), 615–626.

Montmarquette, C., K. Cannings, and S. Mahseredjian (2002). How Do Young People Choose

College Majors? Economics of Education Review 21, 543–556.

Mountjoy, J. (2019). Community Colleges and Upward Mobility. Working Paper.

Mullainathan, S. (2002). A Memory-Based Model of Bounded Rationality. Quarterly Journal of

Economics.

200

Natanson, H.

(2019).

Gov. Northam proposes making community college free for
https://www.washingtonpost.com/local/education/

some
gov-northam-proposes-making-community-college-free-for-some-job-seekers-in-virginia/
2019/12/12/8f2a25fa-1cdc-11ea-8d58-5ac3600967a1 story.html.

job-seekers

in Virginia.

National Center for Education Statistics (2011). Guidelines for Using the CIP to SOC Crosswalk.

https://nces.ed.gov/ipeds/cipcode/resources.aspx?y=55.

National Center for Education Statistics (2018). Digest of Education Statistics, 2016, Table 308.10.

https://nces.ed.gov/programs/digest/d17/tables/dt17 308.10.asp?current=yes.

National Student Clearinghouse Research Center (2017). Snapshot Report – Contribution of
https:

Two-Year Public Institutions to Bachelor’s Completions at Four-Year Institutions.
//nscresearchcenter.org/snapshotreport-twoyearcontributionfouryearcompletions26/.

Oreopoulos, P. and K. G. Salvanes (2011). Priceless: The Non-Pecuniary Benefits of Schooling.

Journal of Economic Perspectives 25(1), 159–184.

Page, L. C. and J. Scott-Clayton (2016). Improving College Access in the United States: Barriers

and Policy Responses. Economics of Education Review 51, 4–22.

Papke, L. E. and J. M. Wooldridge (1996). Econometric Methods for Fractional Response Variables
with an Application to 401(K) Plan Participation Rates. Journal of Applied Econometrics 11,
619–632.

Patterson, R. W., N. G. Pope, and A. Feudo (2019). Timing Is Everything: Evidence from College

Major Decisions. IZA Discussion Paper, No. 12069.

Payne, A. A. (1998). Does the Government Crowd-out Private Donations? New Evidence from a

Sample of Non-profit Firms. Journal of Public Economics 69(3), 323–345.

Peters, C., K. Lausch, and M. Udow-Phillips (2014). Autism Spectrum Disorder in Michigan.

Center for Healthcare Research and Transformation Issue Brief .

Peters-Scheffer, N., R. Didden, H. Korzilius, and P. Sturmey (2011). A Meta-Analytic Study on
the Effectiveness of Comprehensive ABA-based Early Intervention Programs for Children with
Autism Spectrum Disorders. Research in Autism Spectrum Disorders 5(1), 60–69.

Reback, R. and T. L. Cox (2018). Where Health Policy Meets Education Policy: School-based

Health Centers in New York. Working Paper.

Reynolds, C. L. (2012). Where to attend? Estimating the effects of beginning college at a two-year

institution. Economics of Education Review 31, 345–362.

Rouse, C. E. (1995). Democratization or Diversion? The Effect of Community Colleges on Edu-

cational Attainment. Journal of Business & Economic Statistics 13(2), 217–224.

Ruijs, N. M. and T. T. Peetsma (2009). Effects of inclusion on students with and without special

educational needs reviewed. Educational Research Review 4(2), 67–79.

201

Sentz, R., M. Metsker, P. Linares, and J. Clemans (2018). How Your School Affects Where You

Live. https://www.economicmodeling.com/how-your-school-affects-where-you-live/.

Shu, P. (2016). Innovating in Science and Engineering or ”Cashing In” on Wall Street? Evidence

on Elite STEM Talent. Harvard Business School Working Paper 16-067.

Smith,
side
replicating-tennessees-approach-free-community-college-takes-money-and-more.

In-
https://www.insidehighered.com/news/2017/03/02/

A.
A.
Higher

(2017).
Ed.

Replicating

the

Tennessee

Promise.

Snyder, M. and S. Boelscher (2018). Driving Better Outcomes: Fiscal Year 2018 State Status & Ty-
pology Update. http://hcmstrategists.com/wp-content/uploads/2018/03/HCM DBO Document
v3.pdf.

Solon, G., S. J. Haider, and J. M. Wooldridge (2015). What Are We Weighting For? Journal of

Human Resources 50, 301–316.

Stevens, A. H., M. Kurlaender, and M. Grosz (2018). Career Technical Education and Labor Mar-
ket Outcomes: Evidence from California Community Colleges. Journal of Human Resources,
forthcoming.

United States General Accounting Office (2003). The Worker Adjustment and Retraining Noti-
fication Act: Revising the Act Could Clarify Employer Responsibilities and Employee Rights.
https://www.gao.gov/new.items/d031003.pdf.

U.S. Department of Labor (2019). Plant Closings & Layoffs. https://www.dol.gov/general/topic/

termination/plantclosings.

Viru´es-Ortega, J. (2010). Applied Behavior Analytic Intervention for Autism in Early Childhood:
Meta-analysis, Meta-regression and Dose-response Meta-analysis of Multiple Outcomes. Clin-
ical Psychology Review 304, 387–399.

Weinstein, R. (2019). Local Labor Markets and Human Capital Investments. Working Paper.

Wiswall, M. and B. Zafar (2015). Determinants of College Major Choice: Identification using an

Information Experiment. Review of Economic Studies 82(2), 791–824.

Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.).

Cambridge, MA: The MIT Press.

Xia, X. (2016). Forming wage expectations through learning: Evidence from college major

choices. Journal of Economic Behavior & Organization 132.

Xu, Guifeng, et al. (2019). Prevalence and Treatment Patterns of Autism Spectrum Disorder in the

United States, 2016. JAMA Pediatrics 173(2), 153–159.

Yelowitz, A. (1995). The Medicaid Notch, Labor Supply, and Welfare Participation: Evidence

from Eligibility Expansions. The Quarterly Journal of Economics 110(4), 909–939.

202

Zwaigenbaum, Lonnie, et al. (2015). Early Intervention for Children with Autism Spectrum Dis-
order Under 3 Years of Age: Recommendations for Practice and Research. Pediatrics 136(S1),
s60–s81.

203