THREE ESSAYS ON THE ECONOMICS OF EDUCATION
By
Elizabeth Quin

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Economics – Doctor of Philosophy
2013

ABSTRACT
THREE ESSAYS ON THE ECONOMICS OF EDUCATION
By
Elizabeth Quin

Community colleges are a large part of the nation’s higher education system and provide
an important access point to post-secondary education for many students. Transfer to a four-year
institution is one of the many functions served by community colleges. Despite the importance
of the transfer function, the process of transferring between higher education institutions can be
confusing for students. In order to reduce the uncertainty surrounding the transfer process, states
have formalized and expanded pre-existing institutional transfer agreements to provide clearer
linkages between two-year and four-year institutions of higher education, and many schools also
maintain institution-to-institution agreements. Chapters 1 and 2 provide a closer look at
institution-to-institution policies, and changes in state policies, respectively.
Chapter 1 explores the effects of the transfer admission guarantees (TAG) between
California Community Colleges and some University of California (UC) campuses. Specifically,
I investigate the impact of TAG policies on transfer to and bachelor’s degree completion at UC
campuses. These analyses indicate that TAG is positively related to transfer rates and bachelor’s
degree attainment, but not to the rate at which transfer students graduate. There is no association
between TAG policies and the grade point average (GPA) attained by transfer students.
Chapter 2 adds to a growing literature examining the relationship between state postsecondary transfer and articulation policies and the final educational attainment of students who
begin at public two-year institutions. Researchers have used both the National Education
Longitudinal Study of 1988 and various cohorts of the Beginning Postsecondary Students

Longitudinal Study (BPS). Previous studies use cross-sectional differences in state policies to
investigate the effect of such policies on educational outcomes. Most of these studies conclude
there is little cross-sectional relationship between state articulation policies and education
outcomes such as transfer, credit accumulation, persistence, and degree attainment. In Chapter 2
I build on the existing literature by using multiple cohorts of the BPS, allowing for the
examination of changes in state policies over time. I find no evidence of a relationship between
state transfer policies and either transfer or degree attainment for beginning public two-year or
public four-year students. However, these results are sensitive to the sample used, as well as the
policy definition.
Chapter 3 contributes to the important but small body of research on the role of private
schools in Indian education. It uses a household dataset from India with a rich set of household
covariates and student performance data on reading, writing, and mathematics. For both rural and
urban India the results from regression analyses indicate that private school students perform
better on tests controlling for covariates. In both contexts, however, the private school benefit
becomes largely, statistically, insignificant after conducting multivariate analysis on data
balanced using the propensity score matching technique. The paper also makes an initial attempt
to identify ‘low-fee’ private schools; within the regression framework it finds that children in
such schools may perform no better than their public school counterparts. The data and methods
used in this paper are not without limitations; however these analyses call into question the claim
that private school effect may be unequivocally positive and highlights the potential
heterogeneity in private school performance in the Indian context.

Copyright by
ELIZABETH QUIN
2013

ACKNOWLEDGMENTS

I would like to thank Todd Elder and Steven Haider for their guidance and support,
without which this dissertation would not have been written. I would also like to thank Stacy
Dickert-Conlin and Marilyn Amey for serving on my guidance committee. Additional thanks to
Amita Chudgar, with whom I co-wrote Chapter 3, and who was a wonderful mentor.
I am grateful for the financial assistance provided by the Michigan State University
Economics of Education program, supported by the Institute for Education Sciences. I am also
grateful to all of the faculty and students who participated in this program, and who taught me so
much about both Economics and Education.
Finally, I am extremely grateful for the love and support of my family and friends. I
would like to especially thank my parents for their constant encouragement throughout my life.

v

TABLE OF CONTENTS

LIST OF TABLES …………………………………………………………………………… viii
LIST OF FIGURES ……………………………………………………………………………... x
CHAPTER 1
GUARANTEED TRANSFER POLICIES AND POST-SECONDARY OUTCOMES ………... 1
Introduction ……………………………………………………………………………… 1
Background ……………………………………………………………………………… 3
Transfer in California ……………………………………………………………. 3
Transfer Admission Guarantee (TAG) ………………………………………….. 5
Relevant Literature ………………………………………………………………. 7
Data ……………………………………………………………………………………….9
Methods ………………………………………………………………………………… 11
Empirical Specification ………………………………………………………… 11
Results ………………………………………………………………………………….. 13
Baseline Results ………………………………………………………………... 13
Further Analysis Using Additional Community College Level Data ………….. 17
Event History Analysis Using TAG with the Closest UC Campus ……………. 19
Additional Community College Outcomes …………………………………….. 21
Conclusion and Discussion …………………………………………………………….. 23
APPENDICES …………………………………………………………………………………. 26
APPENDIX A: Policy Appendix ………………………………………………………. 27
APPENDIX B: Additional Community College Level Tables ………………………… 28
APPENDIX C: Pair-level Tables ………………………………………………………. 31
REFERENCES ………………………………………………………………………………… 33
CHAPTER 2
EFFECTS OF STATE-LEVEL TRANSFER POLICIES ON POST-SECONDARY
OUTCOMES …………………………………………………………………………………... 35
Introduction …………………………………………………………………………….. 35
Background …………………………………………………………………………….. 36
State-Level Transfer and Articulation Policies ………………………………… 36
Literature Review ………………………………………………………………. 38
Data …………………………………………………………………………………….. 40
Analysis Sample ………………………………………………………………... 42
Methodology …………………………………………………………………………… 42
Results ………………………………………………………………………………….. 43
Descriptive Statistics …………………………………………………………… 43
Pooled Cohort Analysis ………………………………………………………... 47
Differences-in-differences ……………………………………………………... 48
vi

Additional Analysis ……………………………………………………………. 50
Discussion ……………………………………………………………………………… 55
Conclusion ……………………………………………………………………………... 57
APPENDICES …………………………………………………………………………………. 58
APPENDIX A: Policy Appendix ………………………………………………………. 59
APPENDIX B: Data Appendix ………………………………………………………… 61
APPENDIX C: Additional Tables ……………………………………………………... 63
REFERENCES ………………………………………………………………………………… 69
CHAPTER 3
RELATIONSHIP BETWEEN PRIVATE SCHOOLING AND ACHIEVEMENT: RESULTS
FROM RURAL AND URBAN INDIA ………………………………………………………... 72
Introduction …………………………………………………………………………….. 72
Indian education, relevant background ………………………………………… 72
Private school performance: Existing research from India …………………….. 74
Materials and Methods …………………………………………………………………. 77
Dataset and key variables ………………………………………………………. 77
Dataset ………………………………………………………………….. 78
Definitions of private school …………………………………………… 79
An imperfect effort to identify low-fee private schools ………………... 80
Dependent variables ……………………………………………………. 80
Methods ………………………………………………………………………… 81
Ordinal Logit and Ordinary Least Square Regressions ………………... 81
Propensity score matching ……………………………………………... 83
Results ………………………………………………………………………………….. 88
Regression analysis: Ordinary Least Square and Ordinal Logit models ………. 90
Propensity score matching ……………………………………………………... 93
Limitations and conclusion …………………………………………………………….. 97
Discussion …………………………………………………………………………….. 100
REFERENCES ……………………………………………………………………………….. 102

vii

LIST OF TABLES

Table 1.1 Descriptive statistics of the analysis sample .………………………………………... 14
Table 1.2 Baseline estimates of the TAG effect on transfer to and bachelor’s degrees from UC
campuses ……………………………………………………………………………. 15
Table 1.3 Further community college level analysis using restricted years …………………… 18
Table 1.4 Event history analysis using a balanced panel ………………………………………. 21
Table 1.5 The relationship between TAG and additional community college level outcomes ... 23
Table 1.6 Number of community college campuses with a TAG agreement with each UC campus
in selected years ……………………………………………………………………... 27
Table 1.7 Baseline results using negative binomial specification ……………………………... 28
Table 1.8 Further community college level analysis using restricted sample …………………. 29
Table 1.9 Additional community college level outcomes using negative binomial
specification …………………………………………………………………………. 30
Table 1.10 Pair-level estimates of the TAG effect on transfers to and bachelor’s degrees from UC
campuses ………………………………………………………………………….... 31
Table 1.11 Negative binomial pair-level estimates …………………………………………….. 32
Table 2.1 Descriptive statistics of the analysis sample of first-time beginning Public two-year
and four-year students ……………………………………………………………….. 44
Table 2.2 Student background characteristics and five-year postsecondary outcomes by initial
degree program for first-time beginning public two-year students (means in %) …... 46
Table 2.3 Pooled cohort analysis of the relationship between state articulation policy and
postsecondary outcomes …………………………………………………………….. 48
Table 2.4 Baseline estimates of the state policy effect on postsecondary outcomes …………... 49
Table 2.5 Additional analysis using a lagged policy variable …………………………………. 51
Table 2.6 Sensitivity analysis using two policy definitions with the latter two BPS cohorts ….. 52
Table 2.7 Analysis using students beginning at public four-year institutions …………………. 55
viii

Table 2.8 State-level transfer and articulation policy year by source ………………………….. 59
Table 2.9 BPS variables ………………………………………………………………………... 61
Table 2.10 Sample sizes (unweighted) for each BPS cohort …………………………………... 62
Table 2.11 Descriptive statistics for all first-time beginning public two-year and four-year
students by BPS cohort (means in %) ……………………………………………… 63
Table 2.12 Cross-sectional analysis of the relationship between transfer policies and
postsecondary outcomes …………………………………………………………… 64
Table 2.13 Baseline analysis without sample weights …………………………………………. 65
Table 2.14 Descriptive statistics for first-time beginning public two-year students in Table
2.6 …………………………………………………………………………………... 65
Table 2.15 Sensitivity analysis using two policy definitions with the latter two BPS cohorts (six
year outcomes) ……………………………………………………………………... 67
Table 2.16 Baseline results restricted to the states used in Table 2.6 ………………………….. 68
Table 3.1 Means for the full sample, for dependent variables, propensity scores and selected
b
matching variables, by rural, urban and by different school groups ……………….. 89
Table 3.2 Maximum Likelihood Estimates of a set of Ordinal Logistic Regressions (PROF) and
Ordinary Least Square Regressions (SCORE, READ, WRITE, MATH) on full sample,
by Rural, Urban and by Private school definitions ………………………………….. 91
Table 3.3 Differences in means for the matched sample, for dependent variables, propensity
scores and selected matching variables, by Rural, Urban and by Private school
definitions …………………………………………………………………………… 94
Table 3.4 Maximum Likelihood Estimates of a set of Ordinal Logistic Regressions (PROF) and
Ordinary Least Square Regressions (SCORE, READ, WRITE, MATH) on matched
sample, by Rural, Urban and by Private school definitions …………………………. 96

ix

LIST OF FIGURES

Figure 1.1 Transfers from the California Community Colleges to the UC campuses over time ... 4

x

CHAPTER 1
GUARANTEED TRANSFER POLICIES AND POST-SECONDARY OUTCOMES
Introduction
Community colleges educate a growing number of students, with enrollment growing
from roughly 27 percent of undergraduate post-secondary students in 1970 to over 36 percent by
2010 (U.S. Department of Education 2011). Over the past several decades, policymakers have
paid increasing attention to student transfer between two-year and four-year institutions. State
governments and postsecondary institutions enacted policies aimed at easing transfer in response
to perceived low levels of transfer. The effectiveness of these policies has implications for both
individuals and state governments, because a year of education at a community college is less
expensive for both students and states than a year of education at a four-year public university.
In this paper, I study a transfer admission guarantee (TAG) for students transferring from
California community colleges to particular University of California (UC) campuses. A
February 1996 Research Synopsis from UC Davis discusses the admission guarantee policy as it
relates to both students and the university:
“Transfer Admission Agreements [later TAG] benefit both students and the campus. By
concentrating on a specific set of courses, students can reduce the time spent preparing
for transfer. The campus gains by enrolling students with more focused preparation for
upper division major coursework; such preparation could lead to improved student
performance and reduce the time needed to complete a degree.” (p. 1)
It is clear from this statement that the administrators and campuses who put this policy in
place believed that it would benefit transfer students in two ways. First, the TAG program would
presumably reduce time-to-transfer by focusing students on specific courses and a required grade
point average (GPA). Second, transfer students would perform better at the four-year campus
1

and graduate more quickly due to better pre-transfer preparation. The goal of this paper is to
investigate whether these expectations are met.
Previous studies in this area concentrated on state-level policies and generally found little
association between transfer policies and postsecondary student outcomes. Focusing instead on
detailed data from California allows for an analysis of institution-to-institution policies, which
may be particularly relevant to students. California is an especially rich source of data on
transfer policies, as its 112 community colleges form the largest higher education system in the
nation, serving almost 750,000 full-time equivalent (FTE) students in 2008. California has a
long history of transfer agreements that guarantee the option to transfer to selected University of
California (UC) campuses for community college students who have completed a required
number of credits and maintained a minimum GPA. I study the effects of these TAG policies on
transfer between two-year and four-year institutions, and on bachelor’s degree outcomes of
transfer students.
I find evidence that TAG policies increase transfer to UC campuses and bachelor’s
degrees completed by transfer students at those campuses. Campuses may worry that the quality
of these transfers might decline if the marginal students affected by the policy are less prepared
for upper-level coursework. However, the data suggest that transfer student quality, as measured
by graduation rate or junior-year GPA, does not decline. The magnitude of the transfer and
bachelor’s degree effects are similar, suggesting that graduation rates among transfer students do
not change. Further analysis confirms that graduation rates for students who transferred from
community colleges to the UC system are not related to TAG, implying that the average quality
of transferring students does not decline in the presence of TAG policies even though the number

2

of transfers increases. Additionally, I find no change in transfer student grade point average
(GPA) at UC campuses.
Section 2 gives background on higher education in California, discusses the TAG policy,
and reviews the literature. The data used in the analysis are presented in Section 3. Section 4
discusses the methods used to evaluate this particular policy. Section 5 gives the results and
Section 6 provides a discussion and conclusion.
Background
Transfer in California
California public higher education consists of three systems: the University of California
(UC), California State University (CSU), and the California Community Colleges (CCC). There
1

are currently 9 UC, 23 CSU, and 112 CCC campuses serving undergraduates in California .
Figure 1.1 shows transfers to the UC system over time. Transfers rose dramatically
between 1989 and 2010. The downward trend in transfers to the UC system beginning in 1993
prompted the UC campuses to agree to try to increase transfer students, which resulted in a 1997
policy change. Figure 1.1 also graphs the average number of UC TAG agreements per
community college. The number of TAG agreements per community college rises between 1997
and 2009. I only have complete policy information on TAGs after 1997.

1

See the working version of this paper for maps of the UC and CCC systems.
3

14000

5

Transfers

12000
4

10000

3

8000
6000

2

4000
1

2000
0
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009

0

Average TAGs per community college

6

16000

Year
Total transfers

Average TAGs

Figure 1.1 Transfers from the California Community Colleges to the UC campuses over time

In 1997 a Memorandum of Understanding (MOU) re-asserted the transfer role and set
transfer targets. The signing of this agreement set off a new wave of expansion of the transfer
guarantee programs in California. UC campuses that already had transfer agreements in place,
such as UC Davis, UC San Diego and UC Santa Cruz extended those agreements to more
community colleges. Other campuses initiated a transfer guarantee: for example, UC Irvine and
UC Santa Barbara began their programs after 1997.
There were several other transfer-related policies in place in California during this time.
One is the Intersegmental General Education Transfer Curriculum (IGETC), which fulfills
lower-division general education requirements at both UC and CSU campuses. The “Access to
Transfer Information for Community College Students Act”, passed by the California State
Legislature in 2000, required community colleges to publicize IGETC so that students would
4

know what courses and credits were transferable. The passage of this act suggests that the
IGETC may not have been effective prior to 2000 because students may not have known about
its existence.
The Articulation System Stimulating Interinstitutional Student Transfers (ASSIST)
website (www.assist.org) lists all of the course articulation agreements between each community
college and four-year campus in California. While the ASSIST website guides students about
particular courses that transfer, the admission guarantee policy is much broader in that it is a
guarantee of admission if a student meets certain requirements.
Transfer Admission Guarantee (TAG)
The Transfer Admission Guarantees (TAG) in California, begun at UC Davis in the mid1980’s, expanded to other UC campuses during the 1990s and early 2000s. The TAG policies
have also been called Transfer Admission Agreements (TAA), Guaranteed Admission for
Transfer Entry (GATE – at UC Santa Cruz), and Preliminary Admission in the Field (PAIF – at
UC Irvine). Students typically sign these agreements at the beginning of their second year of
community college to apply for admission to a UC campus in the following fall. Students using
TAG are considered junior-level transfers. In order to sign the agreement, students generally
must have completed 30 transferable semester (45 quarter) units. In addition, many TAG
agreements require a minimum grade point average (GPA), which may vary by campus and by
major within campus. The GPA requirements ranged from 2.8 to 3.2 during the years I consider
below. Students must maintain the minimum GPA and complete a specified number of credits
by the spring before they transfer.
Students do not need to sign a TAG agreement in order to be admitted to a UC campus.
The UC campuses give priority to junior-level community college applicants over other transfer

5

applicants, including students from other four-year institutions. However, signing a TAG allows
for several additional benefits besides priority consideration, such as early review of student
records and a guarantee of admission to the campus. In general, the requirements for signing a
TAG are more stringent than those needed for regular transfer admission. For example, the GPA
minimum for TAG agreements is higher than regular transfer admission GPA. Nonetheless, the
benefits of guaranteed admission are sufficient to encourage a non-trivial portion of transfer
students to sign a TAG.
The number of students signing TAGs varies from campus to campus. According to the
UC Davis Research Synopsis reports from 1996 and 2000, the number of TAGs (then called
TAAs) signed at UC Davis was 202 in 1987-88, 792 in 1994-95, and 716 in 1998-99. These
agreements accounted for 23%, 44%, and 40%, respectively, of entering community college
transfer students at UC Davis in those academic years. In 1994-95, 35% of all entering transfer
students signed an agreement. The number of TAGs submitted to UC Davis for review in 200910 exceeded 3,000. However, the number of TAGs signed was much lower at UC Merced, with
only around 200 students signing a TAG in 2009-10. Students who sign TAGs are more likely to
enroll in UC Davis than transfer students admitted without signing a TAG. For example, in
1998-99, 62% of TAG signers enrolled at UC Davis, compared to 50% of other advanced
standing applicants.
In 2007, the seven UC schools that use admission guarantees agreed to a common name –
Transfer Admission Guarantee (TAG). In addition, the UC campuses decided to use a common
TAG application form for all campuses. The TAG application became available on-line in the
summer of 2010 for students seeking fall 2011 admission. While UC Berkeley and UCLA do
not participate in the TAG program, they do offer priority admission to CCC transfer students.

6

In many ways the California experience with Transfer Admission Guarantees (TAG)
provides ideal variation to study the effect of policies on transfer students’ experiences. Since
1997, the TAG agreements have grown in two ways. First, they rolled-out across the
participating UC campuses over time. Second, for some UC campuses the agreements generally
began regionally and then expanded to include community colleges across the state of California.
For example, UC Davis expanded its program from 56 community colleges in 2000 to 94 partner
colleges by fall 2008. UC Santa Cruz expanded its TAG program from 20 community colleges
in 2000 to 102 by 2008. The empirical analysis in this paper uses both of these sources of
variation in exposure to the TAG policy to identify the effects of TAG on post-secondary
outcomes.
Relevant Literature
Several papers examine whether there is an association between state transfer policies
and student outcomes. The most common outcomes studied are the probability of transfer and,
conditional on transfer, the probability of receiving a bachelor’s degree as well as time-todegree. So far, the bulk of research concludes that the presence of a state policy does not
increase the transfer rate between 2-year and 4-year institutions. The datasets used are the
National Education Longitudinal Study (NELS) 88/2000, and the Beginning Postsecondary
Students (BPS) 89/94 longitudinal study.
The studies that use the NELS:88/2000 are Goldhaber and Gross (2009), Roksa and Keith
(2008) and Reynolds (2007). Goldhaber and Gross (2009) attempt to classify ‘strong’ and
‘weak’ articulation policies. However the authors find only small effects on transfer. Gross and
Goldhaber also conclude that state articulation policies are associated with higher odds of
transfer for Hispanic students but not for other minority groups or first generation college

7

attendees. Roksa and Keith (2008) use a simple indicator for whether a state has a transfer
policy to investigate the outcomes of transfer, bachelor’s degree attainment, and time-to-degree.
They find no effect of a state transfer policy on these outcomes. Reynolds, in a 2007
dissertation, looks at the effect of state policies on students by using propensity score matching.
He matches students who have similar characteristics on the outcome of living in a state with a
transfer policy. He also runs his analysis separately for men and women. Reynolds’ finds that
articulation agreements raise educational attainment for male college attendees, but not overall
attainment for the cohort of high school graduates.
Anderson, Sun, and Alfonso (2006) use the BPS 89/94 to look at transfer rates between
two-year and four-year institutions. They define their policy as presence of a legislated transfer
policy in a state by 1991. They find no effect of presence of a transfer policy on transfer in a
state.
Two studies examine the relationship between California’s transfer agreements and early
post-secondary outcomes of transfer and junior-year GPA at the transfer UC campus. The first
study finds a relationship between a community colleges transfer rate and the use of Transfer
Admission Agreements (TAAs) and Transfer Admission Guarantees (TAGs) (Transfer Velocity
Project, RP Group, 2010). In particular, the Transfer Velocity Project showed a positive
association between a community colleges transfer rate and the number of students signing
TAAs or TAGs with a UC or CSU institution.
The second study by Dupraw and Michael (1995) studies the early outcomes of TAG
transfer students at UC San Diego (UCSD). They compare junior GPA at UCSD for students
who transferred with a TAG to community college students who transferred without a TAG, and
to students who began their studies as freshman at UCSD (known as native students). Their data

8

covers three cohorts of transferring students, from fall 1988 to spring 1991. This period was in
the very early stages UCSD’s TAG program with only a few local community colleges
participating. The authors find that both types of transfer students obtain roughly the same GPA,
and that this GPA is only slightly lower than that received by students who entered the university
as freshman. Transfer students who earned a higher community college GPA were less likely to
face academic probation at UCSD due to poor academic performance. The authors relate this
higher level of academic success to the increase in the GPA requirement for TAG students from
2.4 in fall 1988 to 2.8 in fall 1990. This paper expands on these two studies by relating the
expansion of the TAG policy to transfer and bachelor’s degree outcomes, as well as several other
post-secondary outcomes.
None of the studies listed above is able to take advantage of policy changes over time,
which may be one reason why they find little relationship between state transfer policies and
post-secondary outcomes. Another reason these studies may find little effect on post-secondary
outcomes is that state-level policies generally supplement existing institution-to-institution
agreements. Studying institutional transfer policies may reveal more about what types of
policies can affect post-secondary outcomes for students.
Data
The data used in the analysis come primarily from publicly available data at the California
Postsecondary Education Commission (CPEC) website (http://www.cpec.ca.gov). The data
2

include transfers between each two-year and four-year public institution in California , and
bachelor’s degree outcomes for transfer students at the four-year campuses. The transfer data

2

Data on transfers between community colleges, and from community colleges to in-state
private or out-of-state institutions is not available for all years and missing for some institutions.
As a result, this transfer data will not be used in the analysis.
9

cover both fall-term and full-year transfers from each community college to each UC campus.
Transfer data are coded as occurring in the fall of the academic year.
Bachelor’s degree data consist of the number of bachelor’s degrees received each year at
each public four-year institution by transfer students from each sending community college.
Unlike the transfer data, the bachelor’s degree data is coded as occurring in the spring of the
academic year. That is, students who receive a bachelor’s degree in the 1998-1999 academic
year are coded as receiving that degree in 1999. The data publicly available from CPEC provide
3

snapshots of transfer and graduation, but does not follow cohorts of students over time.

I compile the policy variable mainly from information in the Answers for Transfers
publication from the University of California. Other sources, including campus reports, email
correspondence with Admissions and TAG representatives at the UC campuses, and on-line
searches supplemented the Answers for Transfers information. I consider a TAG policy in
effect the fall of the first academic year that transfer students were accepted using TAG. This
paper analyzes the implementation of the TAG policy between 1997 and 2006. Appendix A
provides information on the number of community colleges that had a TAG with each UC
campus over time.
Other covariates account for possible outside labor market opportunities in the county where
the community college is located. These include county-level employment rates, median
household income, county population, county population growth rate, percent male, percent
white, and percent in age categories zero to 14, 15 to 29, and 30 to 49.

3

Pair-level bachelor’s degree data reports the number of transfer students from community
college j that received a bachelor’s degree at four-year campus h in year t, but does not note
when students transferred. Data linking students over time was not available.
10

Currently, there are 112 community colleges in California. I restrict the analysis to the 107
community colleges that were open for the entire period of the study. There were eight UC
campuses open during the entire study period (UC Merced opened in 2005-06). UC Berkeley
and UCLA never had a TAG policy while UC Riverside had a TAG with all California
community colleges by 1997. Therefore, the policy variation comes from schools added to the
TAG program at UC Davis, Irvine, San Diego, Santa Barbara, and Santa Cruz. Of these five UC
campuses, UC Davis had the most agreements, with 56 community colleges, as of 1997. On the
other hand, UC Irvine and UC Santa Barbara did not have a guaranteed transfer program in place
in 1997. By 2009, the UC campuses with a TAG program, with the exception of UC Santa Cruz,
had added all community colleges.
Methods
I now turn to the empirical specification, which uses the expansion of the TAG policy to
estimate a differences-in-differences model.
Empirical Specification
I perform the analysis at the community college level by relating the expansion of UC
TAGs at the community college to transfer and bachelor’s degree outcomes for students from
4

that college. For each community college the outcomes of interest are aggregate transfers to
UC campuses, and aggregate bachelor’s degrees given to transfer students at UC campuses in

4

I also conducted several analyses examining the impact of TAG between a community college
– UC campus pair on transfers between the pair and bachelor’s degree outcomes of transfer
students at the UC campus. These analyses can be found in Appendix C.
11

each academic year. Transfer and bachelor’s degree outcomes at UC Merced are not included in
this analysis.

5

When the outcome is transfers, Yjt is the log of transfer students from community college
j to the UC system in year t. The year t is the fall of the academic year in which the student
transferred. Specifically, I estimate
(1)

Yjt = α + β#TAGjt + ηXjt + λt + θj + εjt,

The variable of interest, #TAGjt, defines exposure to TAG as the number of TAG agreements the
community college has with UC campuses. Xjt is a set of county labor market and demographic
characteristics defined in the data section above. Equation (1) also includes a set of year fixed
effects, λt, community college fixed effects, θj, and a random error term εjt.
Transfer is a direct outcome, but the goal of transfer policies is to help students attain
bachelor’s degrees. Therefore, degree completion is perhaps a better way to evaluate transfer
policies. To analyze the effect of the TAG policy on graduation with a bachelor’s degree I
estimate
(2)

Yjt = α + β#TAGj,t-3 + ηXjt + λt + θj + εjt,

where Yjt is the log bachelor’s degrees obtained by transfer students from community college j in
the UC system in year t. The variable of interest, #TAGj,t-3, is a three-year lag of the TAG
policy variable in (1). For example, the three-year lag means that students obtaining a bachelor’s

5

For example, when aggregating transfers or bachelor’s degrees at UC campuses, I include all
UC campuses except UC Merced. The same applies to aggregating the independent variable for
all UC campuses.
12

degree in the 1999-2000 school year (coded as 2000), are given the value of the TAG policy in
1997. All other variables are defined above in equation (1).
Equations (1) and (2) relate the number of TAG agreements to community college level
outcomes. However, there may be differential impacts of the TAG policy based on how far the
community college is from the UC campus. Therefore, I define TAGclosestjt as an indicator
variable for having a TAG with the closest UC campus, while also controlling for the number
TAG agreements with other UC campuses, #TAGnotclosestjt. Then, I estimate
(3)

Yjt = α + βTAGclosestjt + γ#TAGnotclosestjt + ηXjt + λt + θj + εjt.
6

Yjt again represents transfer and bachelor’s degree outcomes at either the closest UC campus or
in the UC system. When analyzing bachelor’s degrees, the two TAG policy variables are lagged
by three years as shown above in equation (2). All other variables are defined as above in
equation (1).
Results
Baseline Results
Table 1.1 contains descriptive statistics for the analysis sample. The average number of
annual transfers from a community college to the UC system is 115.29, while there are 56.04
transfers to the closest UC campus. The corresponding averages for bachelor’s degrees received
by transfer students range from 52.11 at the closest UC campus to 106.77 in the UC system.

6

The closest UC campus is defined without considering UC Merced.
13

Table 1.1 Descriptive statistics of the analysis sample
Observations
(1)
Transfers
Year
1,040
UC system
1,040
Closest UC campus
1,003

Mean
(2)

SD
(3)

2001.49
115.29
56.04

2.87
137.63
78.61

Bachelor's degrees
UC system
Closest UC campus

1,040
1,003

106.77
52.11

130.03
74.52

Policy Variables
#TAG
TAG with closest
#TAG not closest

1,040
1,040
1,040

2.76
0.53
2.24

1.19
0.50
1.17

Other Community College Outcomes
Associate's
First-year UC GPA for transfers
1-year UC system persistence
2-year UC system graduation rate
CSU system transfers

1,040
716
716
716
1,040

658.93
2.90
90.71
45.07
471.94

394.25
0.19
6.34
11.13
324.92

Community College County Demographics
Employment rate
Total population (in 10,000s)
% Male
% White
% 0-14 years old
% 15-29 years old
% 30-49 years old
% population change
Median household income

1,040
1,040
1,040
1,040
1,040
1,040
1,040
1,040
1,040

93.93
291.90
50.01
79.52
22.48
21.47
30.53
1.086
48,500

2.713
344.20
1.295
9.459
2.873
2.096
2.161
1.068
11,643

7

Table 1.2 contains baseline estimates of specifications (1)-(3). Multiplying the reported
coefficients by 100 gives an approximation to the percent change in the corresponding dependent
7

Table 1.7 in Appendix B contains baseline results using a negative binomial specification.
14

variable. Column (1) implies that a community college adding one more TAG policy with a UC
campus is associated with a seven percent increase in transfers to the UC system.
Table 1.2 Baseline estimates of the TAG effect on transfer to and bachelor’s degrees from UC
campuses
Ln(BA
Ln(Transfers Ln(BA Ln(Transfers Ln(BA
Ln(Transfers
at
to UC
at UC
to UC
at UC
to closest
closest
system)
system)
system)
system)
UC)
UC)
(1)
(2)
(3)
(4)
(5)
(6)
#TAG

0.07***
[0.02]

#TAG (lag 3)
TAG w/ closest
#TAG not
closest

0.08***
[0.02]
0.02
[0.05]

0.13*
[0.07]

0.08***
[0.02]

0.11***
[0.03]

TAG w/ closest
(lag 3)
#TAG not
closest
(lag 3)

-0.01
[0.05]

0.14**
[0.07]

0.09***
[0.02]

0.07***
[0.03]

Observations
1,040
1,040
1,040
1,040
1,003
1,003
R-sq
0.15
0.15
0.15
0.15
0.11
0.08
#CC
106
106
106
106
105
105
Note. -- All models include controls for county employment rate, median household income,
percent male, percent White, total population, population growth, and population aged 0 to 14,
15 to 29, and 30 to 49. All models also include year and community college fixed effects.
Standard errors, clustered at the community college level, are in brackets.
*** p<0.01, ** p<0.05, * p<0.1

An additional TAG agreement is associated with an eight percent increase in bachelor’s
degrees, as shown in column (2). Taken together, the estimates in columns (1) and (2) suggest
that TAG is associated with a similar percent increase in transfers and bachelor’s degrees,
implying no change in the graduation rate of transfer students.
15

Columns (3) and (4) repeat the analysis in columns (1) and (2), but split the policy
variable into two parts: whether the community college has a TAG with the closest UC campus
and the number of TAG agreements with the other UC campuses. The coefficient on having a
TAG with the closest UC campus is not statistically significant in either the transfer or
bachelor’s degree regression. An additional TAG with a UC campus not closest to your
community college is associated with an eight percent increase in transfers to the UC system, and
a nine percent increase in bachelor’s degrees. Therefore, after accounting for the number of
other TAG agreements, adding a TAG with the closest UC campus does not affect transfers to
the UC system.
Columns (5) and (6) of Table 1.2 consider slightly different dependent variables –
specifically transfer and bachelor’s degree outcomes at only the closest UC campus. Adding a
TAG with the closest campus is associated with a 13% increase in transfers to the closest UC
campus and a 14% increase in bachelor’s degrees. The coefficient in the transfer regression is
only statistically significant at the ten percent level, while the coefficient in the bachelor’s degree
regression is significant the five percent level. The coefficient on the number of TAG
agreements with other UC campuses is positive and statistically significant in both columns (5)
and (6). It is not clear why the number of other TAG agreements significantly impacts transfers
to and graduation rates at the closest UC campus after accounting for the TAG with that campus.
One possible explanation is that the number of TAG agreements captures the amount of publicity
of the TAG program at the community college campus. Therefore, even conditional on a TAG
with the closest UC campus, students will be more informed about transfer requirements to UC
campuses. This broader awareness of transfer options may lead to the positive and statistically
significant coefficient on the number of other TAG agreements.

16

The baseline results suggest that additional TAG agreements increase both transfers and
bachelor’s degrees at UC campuses. The associated percent changes in transfers and bachelor’s
degrees are larger when looking at the TAG agreement with the closest UC campus and
outcomes at that campus. Overall, the results suggest similar percent changes in transfers and
bachelor’s degrees, implying no change in the graduation rate of transfer students at UC
campuses.
Further Analysis Using Additional Community College Level Data
The University of California StatFinder contains publicly available persistence,
graduation, and grade point average (GPA) information for transfer students from each
community college to the UC system (UC StatFinder http://statfinder.ucop.edu/default.aspx).
One-year persistence rate and one-year GPA are measured from 2000 to 2007, and two-year
8

graduation rates from 2000-2006. Panel A of Table 1.3 uses the number of TAG agreements as
the independent variable of interest, while Panel B uses TAG with the closest UC campus and
number of other TAG agreements as the policy variables. Columns (1) and (2) reproduce the
baseline results of Table 1.2, but only include the years for which the StatFinder data is
available. Column (1) of Panel A shows a statistically significant seven percent increase in
transfers to the UC system related to adding one more TAG agreement with a UC campus,
identical to the baseline result in Table 1.2. However, the impact on bachelor’s degrees is
different from the baseline results. However, the impact on bachelor’s degrees is different from
the baseline results. The coefficient on TAG in column (2) is three percent, smaller than the
eight percent from Table 1.2, Column (2) and no longer statistically significant. The coefficients
8

Table 1.8 in Appendix B contains additional results using logs the three- and four-year
graduation rate, and the graduation GPA as outcomes. The three-year graduation rate is
measured from 2000-2005, and the four-year graduation rate and graduation GPA are measured
from 2000-2004.
17

Table 1.3 Further community college level analysis using restricted years
Ln(Transfers Ln(BA
to UC
at UC
system)
system)
(1)
(2)
#TAG

0.07***
[0.02]

#TAG (lag 3)

TAG w/ closest
#TAG not
closest
TAG w/ closest
(lag 3)
#TAG not
closest
(lag 3)

Ln(1-yr
Ln(1-yr GPA
persistence
at UC
at UC system)
system)
(3)
(4)
A. #TAG
0.01
<0.01
[0.01]
[0.01]

Ln(2-yr
graduation
rate from
UC system)
(5)
0.02
[0.02]

0.03
[0.03]
B. Closest TAG
0.01
-0.01
[0.02]
[0.02]

0.07***
[0.02]

-0.02
[0.05]

<0.01
[0.01]

0.03
[0.06]

0.03
[0.02]

<0.01
[0.01]

-0.11
[0.09]
0.04
[0.03]

Observations
716
716
716
716
716
#CC
106
106
106
106
106
Note. -- All models include controls for county employment rate, median household income, percent
male, percent White, total population, population growth, and population aged 0 to 14, 15 to 29, and
30 to 49. All models also include year and community college fixed effects. Standard errors,
clustered at the community college level, are in brackets.
*** p<0.01, ** p<0.05, * p<0.1

on the TAG with closest UC variable from columns (1) and (2) in Panel B of Table 1.3 match in
sign and significance with the baseline results from columns (3) and (4) of Table 1.2. The
coefficient on the TAG with closest UC variable in column (2) is more negative than that in
Table 1.2, Column (4) although neither coefficient is statistically different from zero.
Column (3) reports results using the log of the one-year persistence rate of transfer
students in the UC system as the dependent variable. The coefficient on the #TAG variable

18

suggests that there is no association between persistence and the number of TAG agreements.
Similarly, the estimates in column (3) of Panel B show that there is no statistically significant
relationship between adding a TAG with the closest UC campus and changes in one-year
persistence. Columns (4) and (5) report results from regressions where the dependent variable is
the log of first-year UC GPA and the log of two-year UC graduation rate, respectively. None of
the coefficients on the TAG policy variables in Panels A or B are statistically different from
zero. The results in column (5) support the results from the baseline analysis that finds no
change in the graduation rate. Column (4) suggests there is no relationship between additional
TAG agreements and changes in the quality of transfer students to the UC system as measured
by GPA. Overall, Table 1.3 confirms the baseline results suggesting that the graduation rate
among transfer students does not change in response to changes in TAG policies.
Event History Analysis Using TAG with the Closest UC Campus
To provide a more complete picture of the effect of the TAG policies, I also estimate
event history models that trace out the time path of the effects of TAG:
(4)

Yjt = α + ∑πkDj1(t – Tj = k) + ηXjt + λt + θj + εjt,

where Yjt measures log transfers from community college j either to the UC system or to the
closest UC campus in year t. Dj is a dummy variable equal to one if the community college ever
got a TAG agreement with its closest UC campus, and equal to zero otherwise. The indicator
function 1(t – Tj = k) equals one if the community college is k years from the enactment of the
TAG agreement with the closest UC campus. The omitted category is two years prior to when
TAG is enacted between the community college and its closest UC campus. Community

19

colleges are observed two years pre- and post-policy. All specifications include year and
community college fixed effects.
Table 1.4 presents the results from the event history specification given in equation (4).
Columns (1) and (2) use the log of transfers to the UC system as the outcome variable, while
columns (3) and (4) use transfers to the closest UC. Columns (2) and (4) present results from
equation (4). Columns (1) and (3) present estimates of the TAG policy variables from using the
specification in equation (3) on the event-history sample. Columns (1) and (2) show positive but
statistically insignificant variables related to TAG with the closest UC campus. The imprecise
estimates are likely a result of the small number of community colleges and observations in the
event history sample. These results mirror the findings in column (3) of Table 1.2 that a TAG
with the closest UC campus is not related to transfers to all UC campuses. However, the
coefficient on the number of other TAG agreements is negative but statistically insignificant
here, while it was positive and statistically significant in the baseline results in Table 1.2. Taken
together, these results suggest that adding a TAG with the closest UC campus does not impact
transfers from the community college to the UC system as a whole. However, when looking at
the impact on transfers to the closest UC campus, column (4) shows positive and statistically
significant coefficients beginning the first year the community college gets a TAG with the
closest UC campus. These estimates are larger in magnitude but have the same sign as the
results from column (5) of Table 1.2. Column (3) shows a negative but statistically insignificant
coefficient on the TAG with closest UC campus variable. Overall, the results in Table 1.4 are
broadly consistent with the findings in Table 1.2. However, the small number of community
colleges in the event study sample results in imprecise estimates, limiting what can be learned
from these specifications.

20

Table 1.4 Event history analysis using a balanced panel
Ln(Transfers to Ln(Transfers to
UC system)
UC system)
(1)
(2)

Ln(Transfers to
closest UC)
(3)

Ln(Transfers to
closest UC)
(4)

Excluded: -2
-1
0
1
2
#TAG not closest
TAG closest

-0.12
[0.08]
0.31
[0.22]

-0.01
[0.25]
0.38
[0.29]
0.27
[0.39]
0.44
[0.44]
-0.13
[0.08]

-0.10
[0.13]
-0.02
[0.20]

0.26
[0.26]
0.74**
[0.27]
0.88*
[0.43]
1.53**
[0.59]
-0.12
[0.14]

Observations
113
113
105
105
R-sq
0.43
0.45
0.53
0.55
#CC
24
24
24
24
Note. -- The event study analysis is based on when the community college gets a TAG with the
closest UC campus. All models include controls for county employment rate, median household
income, percent male, percent White, total population, population growth, and population aged 0
to 14, 15 to 29, and 30 to 49. All models also include year and community college fixed effects.
Standard errors, clustered at the community college level, are in brackets.
*** p<0.01, ** p<0.05, * p<0.1

Additional Community College Outcomes
Table 1.5 examines other community college level outcomes to look for evidence of TAG
9

impacts outside of the intended target of transfers to and graduation from UC campuses. It is
possible that the estimated increase in transfers to UC campuses in Table 1.2 does not measure
an overall increase in transfers from each community college, but rather a shift in transfers from
9

Table 1.9 in Appendix B contains results using a negative binomial specification.
21

CSU campuses to UC campuses. I use additional data on transfers and bachelor’s degrees at
CSU campuses from the CPEC website to investigate the possible diversion effects of the UC
TAG policy. The CSU transfer and bachelor’s outcomes are aggregated for each community
college to the CSU campus system. Panel A shows the coefficient on the #TAG variable while
Panel B reports results from regressions using the closest TAG.
Columns (1) and (2) use the log of transfer and bachelor’s degree outcomes in the CSU
system as the dependent variables. Panels A and B show similar small, positive coefficients on
the TAG policy variables. None of the coefficients are statistically significant. The results using
outcomes in the CSU system suggest no impact of the UC TAG policy on transfer, bachelor’s
degree completion, or graduation rates at the CSU campuses. It should be noted that some CSU
campuses also had an admission guarantee during this period. I was not able to get full policy
information for the CSU campuses to conduct an analysis of CSU TAG. However, it appears
that the expansion of the UC TAG policy did not simply divert transfers from the CSU campuses
to UC campuses.
The UC TAG policy may also affect associate’s degree completion at the community colleges.
While students do not need to complete an associate’s degree to take advantage of TAG, the
TAG course and GPA requirements may affect associate’s degree attainment. Information on
the number of associate’s degrees granted by each community college in each year comes from
the CPEC website. Column (3) of Table 1.5 shows the coefficients on the policy variables using
the log of associate’s degrees as the outcome. Panel A shows no statistically significant
relationship between the number of TAG agreements and associate’s degrees awarded at the
community college. There are no statistically significant coefficients in Panel B. The results

22

using associate’s degrees suggest no major impact of TAG on associate’s degrees received at the
community college.

Table 1.5 The relationship between TAG and additional community college level outcomes
Ln(Transfers to CSU
Ln(BA at
Ln(AA from
system)
CSU system)
community college)
(1)
(2)
(3)
A. #TAG
#TAG
0.02
-0.00
[0.01]
[0.02]
#TAG (lag 3)
0.01
[0.01]
B. Closest TAG
TAG w/ closest
0.02
0.03
[0.03]
[0.05]
#TAG not closest
0.01
-0.01
[0.01]
[0.03]
TAG w/ closest
0.02
(lag 3)
[0.03]
#TAG not closest
0.01
(lag 3)
[0.01]

Observations
1,040
1,040
1,040
#CC
106
106
106
Note. -- All models include controls for county employment rate, median household income,
percent male, percent White, total population, population growth, and population aged 0 to 14,
15 to 29, and 30 to 49. All models also include year and community college fixed effects.
Standard errors, clustered at the community college level, are in brackets.
*** p<0.01, ** p<0.05, * p<0.1

Conclusion and Discussion
This paper assesses the effects of inter-institution guaranteed transfer policies on postsecondary outcomes for students who begin their postsecondary education at a community
college. Estimates suggest that TAG is associated with increases in transfers to and bachelor’s
degree completions at UC campuses. However, the similar magnitude of the transfer and degree
23

impacts suggests no change in the graduation rates of transfer students. In addition, TAG has
little effect on persistence or GPAs of transfer students into the UC system. Overall, these results
suggest that TAG is associated with increases the number of students transferring and attaining
bachelor’s degrees with little change in the quality of these transfer students.
What aspects of the UC TAG program are likely to be responsible for these patterns in
the data? From a student perspective, TAG provides both information and certainty in the
transfer process. Information comes from both the specified course and GPA requirements, as
well as early review of student records to ensure the student is on track to meet those
requirements prior to transfer. The GPA and course requirements may also be related to
maintaining the quality of transfer students, as measured by GPA and graduation rate, even
though the number of transfers increases.
In addition, TAG provides certainty to the admissions directors UC campuses. Students
may only sign a TAG with one UC campus. This rule was de facto enforced during my study
period due to each UC campus having separate TAG forms. UC Davis noted that TAG signers
were more likely to enroll than transfer students admitted without signing a TAG. After the
study period, the UC campuses experimented with letting students sign multiple TAGs through a
common application form or an online form. When students could sign TAGs with multiple UC
campuses, it lowered the probability that a TAG applicant would enroll at a particular UC
campus. While the TAG application remains online, the UC campuses now enforce that students
may only sign a TAG with one UC campus, although students can still apply to multiple
campuses. The one-TAG policy likely ensures that students who sign TAG agreements are more
likely to enroll than other advanced-standing transfer applicants. Therefore, the TAG program

24

provides greater certainty to UC campuses about the number of students they will enroll in the
following year.
While I can speculate about the potential mechanisms of the TAG policy, future research
should explore other transfer policies to shed light on policy components that are particularly
effective. Do other institution-to-institution or state-level guaranteed transfer policies lead to the
same outcomes? What is the effect of transfer policies that do not contain a guarantee? In
addition to examining other transfer policies, future research may benefit from longitudinal
student-level data. First, student-level data would allow for construction of cohort transfer and
graduation rates for each pair of campuses, which is not possible with the current publicly
available data used in this paper. Second, student data would allow for a more in-depth analysis
of how students from different racial, ethnic, or socioeconomic groups respond to the policy.

25

APPENDICES

26

APPENDIX A

Policy Appendix
Table 1.6 Number of community college campuses with a TAG agreement with each UC campus
in selected years
UCB

UCD

UCI

UCLA

UCM

UCR

UCSD

UCSB

UCSC

1986

None

3

None

None

.

None

None

None

None

1988

None

25

None

None

.

None

3

None

None

1995

None

56

None

None

.

?

?

None

?

1997

None

56

None

None

.

All

14

None

17

1998

None

56

None

None

.

All

14

None

17

1999

None

56

None

None

.

All

14

None

17

2000

None

56

16

None

.

All

15

None

20

2001

None

60

16

None

.

All

16

None

92

2002

None

70

19

None

.

All

17

3

92

2003

None

81

19

None

.

All

17

9

94

2004

None

81

19

None

.

All

24

10

97

2005

None

82

22

None

All

All

26

All

99

2006

None

90

22

None

All

All

27

All

99

2007

None

90

29

None

All

All

33

All

101

2008

None

94

29

None

All

All

33

All

101

2009

None

All

29

None

All

All

All

All

103

2010

None

All

All

None

All

All

All

All

103

27

APPENDIX B

Additional Community College Level Tables
Table 1.7 Baseline results using negative binomial specification
UC System
Transfers
BA
Transfers
BA
(1)
(2)
(3)
(4)
#TAG
#TAG (lag 3)
TAG w/ closest
#TAG not closest

Closest UC
Transfers
BA
(5)
(6)

0.06***
[0.01]
0.06***
[0.01]
0.03
[0.03]
0.06***
[0.01]

TAG w/ closest
(lag 3)
#TAG not closest
(lag 3)

0.25***
[0.04]
0.06***
[0.01]
0.03
[0.02]
0.06***
[0.01]

0.17***
[0.04]
0.05***
[0.01]

Observations
1,062
1,055
1,062
1,055
1,070
1,060
#CC
107
106
107
106
107
106
Year FE
X
X
X
X
X
X
Note. -- All models include controls for county employment rate, median household
income, percent male, percent White, total population, population growth, and population
aged 0 to 14, 15 to 29, and 30 to 49. All models also include year and community college
fixed effects. Standard errors, clustered at the community college level, are in brackets.
*** p<0.01, ** p<0.05, * p<0.1

28

Table 1.8 Further community college level analysis using restricted sample
UC Grad. Rate
Ln(3-year)
Ln(4-year)
(1)
(2)
A. #TAG
#TAG
0.01
0.01
[0.01]
[0.01]

TAG w/ closest
#TAG not closest

B. Closest TAG
-0.01
[0.03]
0.01
[0.01]

0.00
[0.04]
0.01
[0.01]

Graduation
Ln(GPA)
(3)
-0.00
[0.00]

-0.01
[0.01]
-0.00
[0.01]

Observations
615
510
510
R-squared
0.06
0.05
0.08
#CC
106
105
105
Note. -- All models include controls for county employment rate, median household income,
percent male, percent White, total population, population growth, and population aged 0 to 14,
15 to 29, and 30 to 49, as well as year and community college fixed effects. Standard errors,
clustered at the community college level, are in brackets. *** p<0.01, ** p<0.05, * p<0.1

29

Table 1.9 Additional community college level outcomes using negative binomial specification
Associate's
CSU System
AA
Transfers
BA
(1)
(2)
(3)
#TAG

0.02**
[0.01]

0.03***
[0.01]

#TAG (lag 3)

TAG w/ closest
#TAG not closest

-0.01
[0.01]
0.02
[0.02]
0.02**
[0.01]

TAG w/ closest
(lag 3)
#TAG not closest
(lag 3)

0.02
[0.02]
0.02**
[0.01]
-0.06***
[0.02]
<0.01
[0.01]

Observations
1,062
1,062
1,055
#CC
107
107
107
Year FE
X
X
X
Note. -- All models include controls for county employment rate, median household income,
percent male, percent White, total population, population growth, and population aged 0 to 14,
15 to 29, and 30 to 49, as well as year and community college fixed effects. Standard errors,
clustered at the community college level, are in brackets. *** p<0.01, ** p<0.05, * p<0.1

30

APPENDIX C

Pair-level Tables
Table 1.10 Pair-level estimates of the TAG effect on transfers to and bachelor's degrees from UC campuses
Ln(Transfers
Ln(Transfer
Ln(Transfers
)
Ln(BA)
s)
Ln(BA)
)
Ln(BA)
(1)
(2)
(3)
(4)
(5)
(6)
TAG
0.04
0.02
0.12***
[0.03]
[0.03]
[0.04]
TAG (lag 3)
0.06*
0.03
0.14***
[0.03]
[0.03]
[0.04]
# other TAG
0.05***
0.05***
0.03
[0.02]
[0.01]
[0.02]
# other TAG
0.08***
0.08***
0.05***
(lag 3)
[0.02]
[0.01]
[0.02]

Observations
R-squared
#Pair
Pair FE
UC
campus*year
FE

4,723
0.06
607
X

4,723
0.05
607
X

6,502
0.07
811
X

6,502
0.05
811
X

4,723
0.12
607
X

4,723
0.09
607
X

Ln(Transfer
s)
(7)
0.12***
[0.04]

Ln(BA)
(8)

0.14***
[0.04]
0.04**
[0.01]
0.06***
[0.01]

6,502
0.12
811
X

6,502
0.09
811
X

X
X
X
X
TAG
TAG
TAG
TAG
UC campuses
campuses
campuses
All
All
campuses
campuses
All
All
Note. -- All models include controls for county employment rate, median household income, percent male, percent White, total
population, population growth, and population aged 0 to 14, 15 to 29, and 30 to 49, as well as year and pair fixed effects. Standard
errors, clustered at the pair level, are in brackets. *** p<0.01, ** p<0.05, * p<0.1
31

Table 1.11 Negative binomial pair-level estimates
Ln(Transfers)
Ln(BA)
(1)
(2)
TAG
0.10***
[0.02]
TAG (lag 3)
0.10***
[0.02]
# other TAG
0.05***
[0.01]
# other TAG
0.07***
(lag 3)
[0.01]
Observations
#Pair
Pair FE

Ln(Transfers)
(3)
0.10***
[0.02]

Ln(BA)
(4)

0.09***
[0.02]
0.05***
[0.01]
0.06***
[0.01]

6,183
6,183
8,273
8,273
619
619
828
828
X
X
X
X
TAG campuses
TAG campuses
All
All
Note. -- All models include controls for county employment rate, median household income,
percent male, percent White, total population, population growth, and population aged 0 to 14,
15 to 29, and 30 to 49, as well as year and pair fixed effects. Standard errors, clustered at the
pair level, are in brackets. *** p<0.01, ** p<0.05, * p<0.1

32

REFERENCES

33

REFERENCES

Anderson, G., Sun, J. C., and Alfonso, M. (2006) “Effectiveness of Statewide Articulation
Agreements on the Probability of Transfer: A Preliminary Policy Analysis.” Review of Higher
Education, 29(3), 261-291.
Chapman, C., Laird, J., and KewalRamani, A. (2010) “Trends in High School Dropout and
Completion Rates in the United States: 1972-2008.” (NCES 2011-012) National Center for
Education Statistics, Institute of Education Sciences, U.S. Department of Education.
Washington, DC. http://nces.ed.gov/pubs2011/2011012.pdf
Dupraw, C. and Michael, W. (1995) “Community College Transfer Students: Comparing
Admission and Success.” College and University, 71(2), 10-18.
Ehrenberg, R. and Smith, C. (2004) “Analyzing the Success of Student Transitions from 2- to 4year Institutions within a State.” Economics of Education Review, 23(1), 11-28.
Gross, B. and Goldhaber, D. (2009) “Community College Transfer and Articulation Policies.”
Center on Reinventing Public Education. CRPE Working Paper # 2009_1.
http://survey.csuprojects.org/uploads/2-/4X/2-4XaUlUV9-55g-XS_ldcQ/Community-CollegeTransfer-and-Articulation-Policies.pdf
Roksa, J. and Keith, B. (2008) “Credits, Time, and Attainment: Articulation Policies and Success
After Transfer.” Educational Evaluation and Policy Analysis, 30(3): 236-254.
Reynolds, C. “Estimating the Effects of Articulation Agreements Using Re-Weighting
Techniques,” mimeo, University of Michigan, Ann Arbor, 2007.
Transfer Velocity Project: Key Findings on Student Transfer in California Community Colleges.
The Research and Planning Group for California Community Colleges: Center for Student
Success. March 2010. http://www.rpgroup.org/sites/default/files/TVPBrief.pdf’
UC Davis Student Affairs Research & Information. (1996). Enrollment and Graduation Patterns
of Undergraduates Transferring to UC Davis: 1980-1995. No. 66: 1-12.
http://www.sariweb.ucdavis.edu/downloads/167ResearchSyn1980to95EnrollGradPatternsOfUnd
ergradTransfersToUCD.pdf
U.S. Department of Education, National Center for Education Statistics. (2012) “Chapter 3:
Postsecondary Education: Degree-Granting.” (Table 199). In Digest of Education Statistics.
Washington, D.C.: U.S. Department of Education.
Public use data downloaded from the California Postsecondary Education Commission (CPEC).
http://www.cpec.ca.gov/

34

CHAPTER 2
EFFECTS OF STATE-LEVEL TRANSFER POLICIES ON POST-SECONDARY
OUTCOMES
Introduction
Community colleges are the access point to postsecondary education for millions of
students due to low tuition and open enrollment policies (American Association of Community
Colleges http://www.aacc.nche.edu/AboutCC/Trends/Pages/default.aspx). These institutions
serve many purposes, providing workplace training and remedial education, as well as preparing
students for transfer to other postsecondary institutions. While the transfer mission has remained
important, in the late 1980s falling transfer rates from community colleges caused concern for
policymakers (Barry & Barry, 1992). Therefore, over the past several decades, states have
instituted a series of policies aimed at making higher education seamless; that is, students can
begin at any public postsecondary institution and transfer to other institutions without major loss
of credit. The effectiveness of statewide articulation policies has implications for students as it
could change the cost of a bachelor’s degree, as well as labor market outcomes.
This paper extends the literature on the effects of statewide transfer and articulation
policies. To do this, I use data from three cohorts of the Beginning Postsecondary Students
(BPS) longitudinal study. I supplement the BPS data with state articulation policy information
from two different sources. The BPS tracks students who first begin postsecondary education in
1989-90, 1995-96, and 2003-04, allowing me to look at state policy changes over time. Previous
studies look at the cross-sectional relationship between articulation agreements and
postsecondary outcomes, generally finding little impact on transfer and degree attainment
outcomes.

35

Baseline estimates suggest that state articulation policies are not related to transfer or
degree attainment outcomes for students who begin at public two-year or public four-year
institutions. These results are sensitive to the BPS cohorts used in the analysis, as well as to the
policy definition. The latter two BPS cohorts show a negative impact on certificate completion,
and a positive, marginally significant impact on bachelor’s degree attainment. Using a transfer
policy definition focused on the transfer of academic credit does show a positive, statistically
significant impact on transfer.
Section 2 gives background on state-level transfer and articulation policies, and reviews
the literature. The data used in the analysis are presented in Section 3. Section 4 discusses the
methodology used in this paper, and Section 5 provides results. Section 6 provides a discussion,
and Section 7 concludes.
Background
State-Level Transfer and Articulation Policies
States have a variety of policies pertaining to articulation and transfer between postsecondary
institutions. A report from the Education Commission of the States (ECS) in 2010 describes the
types of transfer policies, and indicates whether the state had that policy by 2010. Policies
include cooperative agreements, transfer data reporting, incentives and rewards for transfer
students, statewide articulation guides, common general education core requirements, and
common course numbering. For example, Alabama has a common core, but no common course
numbering system. Additionally, states may have legislated policies or policies adopted by
higher education governing boards that apply statewide. Associate’s degree transfer policies –
where associate’s degree completion ensures transfer or credit acceptance at public four-year

36

institutions in the same state – are generally included as part of legislation. Ignash and
Townsend (2000) also discuss types of state transfer policies.
As the ECS report highlights, there is considerable variation in the presence of state policies.
This variation extends to the timing of these policies, with some states having long-standing
transfer agreements, while other states have only recently added state articulation policies. I will
be exploiting this policy variation to identify the effect of state articulation policies on
postsecondary outcomes of students.
To define state transfer/articulation policies, I use data from several sources. The first source
is a report on state transfer and articulation policies from ECS (2010). I define policies based on
the information from the ‘Statewide Policy’ column in the report. These policies include
legislation as well as policies put in place by higher education systems. I use the ECS policy
data for my baseline analysis because I can define state policies from the ECS for all three BPS
cohorts. The second policy data source is from a Government Accountability Office report in
2005. State policies are defined from the year of the state legislation listed in Appendix II of the
report. These policies are related to the transfer of academic credit. I use state policies from the
GAO report as a sensitivity check on the definition of the state policy. I can only define state
policies from the GAO accurately for the latter two BPS cohorts. Previous studies use policy
data from Ignash and Townsend (2000), who define transfer policies based on a survey of the
states. However, since the policy data only go through 1999, these data do not cover the needed
time span for this study.
Table 2.8 in the Policy Appendix contains information on the policy year from each data
source. In some states, the data sources agree on the policy year. However, in other states there
is disagreement on the policy year, and sometimes disagreement on the whether a statewide

37

articulation policy exists. There are 12 states where the ECS and GAO data disagree on the
timing of the policy; in half of those states the ECS policy comes before the GAO policy.
Additionally, there are six states where the ECS and GAO policy data disagree about the
presence of a state articulation policy; in one of these states the disagreement occurs because the
ECS policy is put in place after the GAO report in 2005. Some of these differences in the timing
and presence of state policies can be explained by the different policy definitions used by the
ECS and GAO reports. The GAO study looks strictly at policies related to the transfer of
academic credit, while the ECS policy definition is broader and includes legislation and higher
education board policies governing transfer of credit as well as transfer of associate’s degrees.
These alternate definitions of state transfer policies seem to impact different postsecondary
outcomes, as discussed in the Results section below.
Literature Review
A small literature examines whether there is an association between state transfer policies
and postsecondary outcomes for students. The main outcomes of interest are the probability of
transfer and, conditional on transfer, the probability of receiving a bachelor’s degree as well as
time-to-degree. The conclusion of the literature so far is that the presence of a state policy does
not increase the transfer rate between two-year and four-year institutions for students who begin
postsecondary at a two-year institution. Past research has used student data from the National
Education Longitudinal Study (NELS) 88/2000, and the Beginning Postsecondary Students
(BPS) 89/94 longitudinal study to investigate this research question.
The paper closest to this one is Anderson, Sun, and Alfonso (2006) because they use the
first cohort of the BPS. Their study looks at transfer rates between two-year and four-year
institutions for students initially enrolled at public two-year colleges. The state articulation

38

policy is defined as the presence of a legislated transfer policy in the state by 1991. They find no
effect of presence of a transfer policy on transfer in a state.
Other studies use NELS:88/2000 to investigate the relationship between articulation
policies and transfer (Goldhaber and Gross (2009), Roksa and Keith (2008) and Reynolds
(2007)). Similar to Anderson, Sun, and Alfonso (2006), Roksa and Keith (2008) use a simple
indicator for whether a state has a transfer policy to investigate the outcomes of transfer,
bachelor’s degree attainment, and time-to-degree. They find no relationship between state
transfer policy and these outcomes. Goldhaber and Gross (2009), in addition to using an
indicator for the presences of a state policy, also attempt to classify ‘strong’ and ‘weak’
articulation policies. However the authors find only small effects on transfer. The authors also
use indicators for various types of state policies (automatic transfer of associate’s degree,
common course numbering across institutions). Again, there is little relationship between type
of transfer policy and postsecondary outcomes. Gross and Goldhaber also conclude that state
articulation policies are associated with higher odds of transfer for Hispanic students but not for
other minority groups or first generation college attendees. Reynolds (2007) looks at the effect
of state policies on students by using propensity score matching. He matches students who have
similar predicted probabilities of living in a state with a transfer policy, running his analysis
separately for men and women. Reynolds finds that articulation agreements raise educational
attainment for male college attendees, but not overall attainment for the cohort of high school
graduates. He also studies an associate’s degree policy in North Carolina using data from the
Integrated Postsecondary Education Data System (IPEDS). He finds increases in associate’s
degree completion, but no impact on bachelor’s degree attainment.

39

None of the studies listed above is able to take advantage of policy changes over time,
which may be one reason why they find little relationship between state transfer policies and
post-secondary outcomes. The aim of this paper is to extend the analysis on state-level transfer
policies by looking at changes over time, so I use three cohorts of the BPS to examine the impact
of transfer policies on students who begin at public two-year institutions.
Data
The primary data source for this paper is the Beginning Postsecondary Students (BPS)
longitudinal study. The BPS follows cohorts of students who have enrolled in postsecondary
education for the first time. The sample does not include students concurrently enrolled in high
school. I will use all three waves: BPS: 1990-94, 1996-01, and 2004-09.
Each initial cohort comes from the National Postsecondary Student Aid Study (NPSAS),
a large nationally representative sample of postsecondary students and institutions. The first
wave follows students for five years (through 1994), while the last two waves follow students for
six years (through 2001 and 2009, respectively). Students in each BPS cohort are interviewed
three times: at the end of their first year in postsecondary education, two years later, and then
again two or three years later depending on the BPS cohort. So, students in the BPS: 90-94
cohort are first interviewed in the spring of 1990 at the end of the 1989-90 academic year.
The BPS tracks students postsecondary outcomes, including degree attainment, transfer
between institutions, and stopout. I use this information to create indicators for whether students
ever transferred or attained a degree during the study. Specifically, I create an indicator for
whether a student ever transferred, and ever transferred to a four-year institution. Similarly, I
create indicators for whether students ever receive an associate’s degree, or ever receive a
bachelor’s degree. These measures of transfer and degree attainment are the main outcomes of

40

interest. The baseline estimates use all three cohorts of the BPS, so outcomes are defined five
years after first entering post-secondary. For analysis that only includes the latter two cohorts of
the BPS, outcomes are defined six years after entering post-secondary.
The main independent variables are policy indicators for each state. These policy
variables indicate whether a state had any articulation policy in a particular year. The three data
sources for the policy variables are listed in the previous section. I code a state as having a
transfer policy if it is in place by December 1989 for BPS:90/94, December 1995 for BPS:96/01,
and December 2003 for BPS:04/09.
The BPS datasets also contain information on student demographic and background
characteristics such as sex, race/ethnicity, age when first enrolled, parental education, dependent
status, marital status, presence of children, and whether the student received a high school
diploma. Additionally, I include the student’s degree program, attendance intensity, degree goal
when first enrolled, and indicators for having a job while enrolled, taking any remedial courses,
whether they receive financial aid, and whether the student is in the top or bottom income
quartile (excluded category is the middle 50% of income). These variables are commonly used
in other papers in the literature. Table 2.9 describes the variables from the BPS used in this
study.
Other studies also include state-level variables, in addition to the state transfer policy
indicator, to account for different state environments. I use variables from Reynolds (2007)
dataset, which he generously shared with me. State-level covariates include, unemployment rate,
appropriations, financial aid, per capita personal income, population distribution (percent 5-17,
18-24, and 65 and over), tuition at universities, other four-year institutions, and two-year

41

institutions, and the number of public and private two-year and four-year institutions (per 18-24
population). All dollar values are in 2004 dollars.
Analysis Sample
Table 2.10 in Appendix A shows the sample size for each BPS cohort as I impose sample
restrictions. The BPS:90-94 is the smallest cohort with a starting sample size of 7,253 students.
The BPS:96-01 and BPS:04-09 cohorts contain 12,085 and 16,684 students respectively. After
conditioning on beginning postsecondary studies at a public two-year campus, the sample sizes
fall to 704, 1516, and 5549 in each of the three BPS cohorts. The analogous sample sizes for
beginning at a public 4-year institution are 1,613, 5,161, and 4,643 in each of the BPS cohorts.
The next cut to the data keeps only students who responded to both the first and last BPS
interview, which is needed to observe postsecondary attainment outcomes. Additionally, I drop
students with missing information on background characteristics or postsecondary outcomes.
Finally, the analysis sample only contains students who attend their first institution in a state that
is present in two of the three BPS cohorts. The final sample sizes for beginning public two-year
students is 622 in BPS:90/94, 742 in BPS:96/01, and 5378 in BPS:04/09, which forms the
baseline analysis sample.
The BPS provides analysis weights to account for the complicated survey design. I use
analysis weights in all descriptive tables and regression analysis in this paper. Additionally, I
cluster standard error by state in all regression analysis.
Methodology
My research question is what is the effect of state articulation policies on education
outcomes of students who begin their postsecondary studies at public two-year institutions? I
use variation over time in state policies on articulation to identify this relationship.

42

I use a differences-in-differences strategy as shown in the equation below:
(1)

Yist+5 = α + ηXist + β4Policyst + λt + θs + εist,

where i indexes individual students, s indexes state, and t indexes BPS cohort. Xist is a vector of
individual-level covariates. Yist+5 denotes the transfer or degree attainment outcome five years
after the student begins post-secondary education. I include only those students who begin their
postsecondary education at two-year institutions. Policyst refers to one of the state transfer and
articulation policies described above. All policy and background demographic characteristics are
defined as of the student’s first year in post-secondary education. Background characteristics
include the student’s sex, race, completion of a high school diploma, dependent status, and
whether the student received any financial aid. Additionally, I include the beginning degree
program in which the student enrolled and an indicator variable for whether the student had a
goal of obtaining a bachelor’s degree. These variables are common in other studies in this
literature.
Results
Descriptive statistics
Table 2.1 shows descriptive statistics by BPS cohort for students who began
1

postsecondary at either a public two-year or four-year institution . The table shows demographic
changes in beginning postsecondary students with higher percentages of female and minority
students in the latter cohort. Student’s goals are also changing over time; around 68% of
beginning public two-year students want to get a bachelor’s degree or higher in the first two BPS

1

Descriptive statistics for students with no missing information in all states is in Appendix Table
2.11.
43

Table 2.1 Descriptive statistics of the analysis sample of first-time beginning Public two-year
and four-year students
1989-1990
1995-96
2003-04
2-yr
4-yr
2-yr
4-yr
2-yr
4-yr
Student Background
Female
50.6%
53.5%
54.1%
54.6%
56.2% 55.1%
White, non-Hispanic
75.9%
82.6%
71.8%
72.5%
60.3% 71.6%
Black, non-Hispanic
8.0%
8.4%
11.9%
11.5%
14.2%
8.9%
Hispanic
11.8%
4.1%
11.3%
9.1%
15.9%
8.5%
Other race
4.3%
4.9%
5.0%
7.0%
9.6%
11.0%
HS diploma
92.4%
98.5%
87.7%
98.4%
85.6% 96.1%
Certificate program
11.6%
10.0%
4.4%
Assoc's degree program
68.8%
72.9%
79.6%
Bachelor's program
76.8%
92.2%
91.7%
Age 22 or less
76.8%
95.9%
68.6%
94.6%
69.9% 95.1%
Dependent
50.0%
82.9%
59.7%
91.7%
62.9% 92.7%
Received any aid
27.5%
48.3%
50.4%
74.9%
53.9% 76.7%
Goal: BA or higher
68.1%
94.8%
68.1%
82.4%
81.3% 98.5%
Parent ed. BA or higher
26.3%
42.3%
25.6%
43.7%
27.9% 54.6%
Attend full-time
37.2%
76.2%
47.0%
87.6%
48.8% 90.3%
Married
19.6%
3.6%
22.3%
3.4%
17.5%
2.9%
Have a child
18.1%
3.3%
24.7%
4.5%
22.2%
3.5%
Work while enrolled
82.7%
78.8%
79.7%
61.3%
77.7% 59.0%
Take remedial courses
18.9%
16.2%
24.0%
17.0%
29.4% 19.4%
Income (bottom
quartile)
26.0%
20.3%
28.7%
25.2%
25.1% 18.5%
Income (top quartile)
17.3%
27.7%
17.1%
24.0%
25.5% 30.9%
Postsecondary Outcomes
Ever transfer
43.0%
28.1%
39.6%
25.5%
36.3% 24.2%
Attain certificate
13.0%
3.8%
10.4%
3.1%
8.0%
1.9%
Attain associate's
22.2%
6.1%
19.5%
4.6%
15.0%
3.9%
Attain bachelor's
6.5%
46.6%
7.4%
46.6%
5.8%
48.9%
State characteristics
Transfer policy (ECS)
22.0%
14.9%
54.2%
32.5%
74.6% 68.5%
Transfer policy (GAO)
.
.
64.5%
44.3%
87.4% 82.5%
Unemployment rate
5.3%
5.4%
5.6%
5.3%
6.1%
5.8%
Sample size
(unweighted)
622
1509
742
3230
5378
4281
Source: Author calculations from BPS:90/94, BPS:96/01 and BPS:04/09.
Note. -- Student background characteristics are defined for the first year in postsecondary while
all degree and transfer variables are defined as of five years after beginning postsecondary. BPS
analysis weights are used. States in at least two of three BPS cohorts.

44

cohorts, while that rises to over 81% in the latter BPS cohort. Beginning public four-year
students show a different pattern, with almost 95% of students aiming for at least a bachelor’s
degree in 1989-90, dropping to 82.4% in 1995-96, and rising again to 98.5% in 2003-04.
Despite these changes in student’s aspirations, five-year transfer and degree attainment outcomes
change little over time. Only around six percent of beginning public two-year students
completes a bachelor’s degree within five years. While five-year bachelor’s degree completion
rates are higher for beginning public four-year students, they are still below 50%. Finally, states
are adding articulation policies over time, so more students attend school in states with transfer
policies in the later BPS cohorts. Only 22.0% of beginning public two-year students in 1989-90
attends an institution in a state with an articulation policy (as defined by ECS), while that
number rises to 74.6% in 2003-04. More students attend institutions in states with policies as
defined by the GAO.
Table 2.2 expands on the descriptive statistics in Table 2.1 by dividing beginning public
two-year students by their beginning degree program: no degree program, certificate, and
associate’s degree. Students who begin at public 2-year institutions are transferring to other
institutions. Between 37.1-44.5% of students who begin in an associate’s degree program
transfer to another institution within five years. Students who begin in a certificate program have
the lowest rates of transfer. In addition, the percentage of students who begin at public 2-year
institutions and want to get a bachelor’s degree is quite high (with the exception of students
enrolled in certificate programs), but actual rates of (five-year) bachelor’s degree attainment are
much lower. While this may be optimal (students realize they do not need a bachelor’s degree
for their chosen career path, or are not prepared academically to complete a bachelor’s degree),

45

Table 2.2 Student background characteristics and five-year postsecondary outcomes by initial
degree program for first-time beginning public two-year students (means in %)
1989-1990
1995-96
2003-04
None Cert. Assoc. None Cert. Assoc. None Cert. Assoc.
Student Background
Female
45.9 48.5
52.3
61.3 52.7
52.6
52.6 55.2
57.0
White, non-Hispanic
69.3 75.2
77.9
70.5 66.7
72.8
61.1 62.1
60.0
Black, non-Hispanic
7.6
8.8
7.9
11.6 20.4
10.8
7.5
20.6
15.2
Hispanic
13.0 10.9
11.6
9.4
12.0
11.6
21.1 11.1
15.1
Other race
10.1
5.2
2.6
8.6
0.9
4.8
10.3
6.2
9.6
HS diploma
88.6 92.8
93.5
85.1 74.9
90.0
80.7 78.8
87.0
Age 22 or less
67.1 69.0
80.9
54.2 48.6
74.7
61.0 44.9
73.1
Dependent
41.9 36.7
54.6
46.8 27.3
67.2
56.5 38.8
65.6
Received any aid
17.3 27.6
30.4
41.6 62.1
50.9
40.8 67.3
55.7
Goal: BA or higher
59.8 60.2
71.8
62.2 40.3
73.4
74.3 45.1
84.7
Parent ed. BA or
higher
26.6 17.3
27.8
25.7
9.6
27.8
28.0 12.9
28.7
Attend full-time
20.0 33.8
42.7
28.6 35.3
52.9
40.0 46.0
50.7
Married
32.0 19.2
16.1
31.2 34.6
18.6
21.8 35.2
15.6
Have a child
23.2 19.6
16.4
31.8 53.2
19.1
25.7 38.8
20.6
Work while enrolled
85.8 83.2
81.8
86.7 70.5
79.3
78.6 72.2
77.8
Take remedial courses
21.7 14.6
18.8
11.6 10.4
28.8
22.8 28.0
30.8
Income (bottom
quartile)
27.1 26.3
25.6
26.9 32.0
28.7
18.8 23.3
26.5
Income (top quartile)
17.5
6.1
19.1
25.1 12.3
15.9
32.4 30.8
23.8
Postsecondary
Outcomes
Ever transfer
Attain certificate
Attain associate's
Attain bachelor's

37.1
15.4
15.3
5.8

44.2
19.4
14.4
2.8

44.5
11.2
25.6
7.4

33.0
13.0
11.6
2.7

18.9
26.6
9.9
0.9

44.0
7.6
22.7
9.4

35.7
8.4
6.2
4.8

24.5
49.2
5.1
1.2

37.1
5.7
17.3
6.3

State characteristics
Transfer policy ECS
Transfer policy (GAO)
Unemployment rate

18.4
.
5.3

18.6
.
5.5

23.5
.
5.2

53.3
59.5
5.6

57.6
73.2
5.3

54.0
64.4
5.6

84.8
90.0
6.2

60.3
67.8
5.7

73.3
88.0
6.1

Sample size
(unweighted)
110
83
429
114
60
568
667
Source: Author calculations from BPS:90/94, BPS:96/01 and BPS:04/09.

296

4415

Note: Student background characteristics are defined for the first year in postsecondary while
all degree and transfer variables are defined as of five years after beginning postsecondary.
46

BPS analysis weights are used. States in at least two of three BPS cohorts (37 states).
state transfer policies are often put in place to address the gap between degree goal and degree
attainment. Is there a relationship between the state transfer policy and degree attainment?
Pooled Cohort Analysis
Table 2.3 presents a pooled cohort analysis using all three BPS cohorts

2,3

. Each column

uses a different dependent variable: transfer to another institution, attainment of a certificate,
associate’s degree, and bachelor’s degree. Results are from a linear probability model. State
articulation policies are positively related certificate and bachelor’s degree attainment, and
negatively related to transfer and associate’s degree completion. However, none of the
coefficients are statistically different from zero. Conversely, several other variables of interest
are significantly related to postsecondary outcomes. Specifically, a student’s goal to attain at
least a bachelor’s degree is positively related to both transfer and bachelor’s degree completion
with coefficients significant at the 1% level. A beginning public two-year student who wants to
attain a bachelor’s degree is 19.8% more likely to transfer than students without those
aspirations. A bachelor’s degree goal lowers a student’s probability of receiving a certificate by
just over four percent. Full-time attendance in the first year of postsecondary increases a
student’s probability of transferring, as well as of attaining an associate’s or bachelor’s degree.
Younger students are more likely to transfer than older students, but no more likely to attain a
degree.

2

Coefficients are only shown for selected variables. Complete regression analyses are available
from the author upon request.
3
Cross-sectional analysis for each BPS cohort is presented in Appendix Table 2.12.
47

Table 2.3 Pooled cohort analysis of the relationship between state articulation policy and
postsecondary outcomes
(1)
(2)
(3)
(4)
Transfer
Certificate
AA
BA
Transfer policy (ECS)
Attend full-time
Goal: BA or higher
Age 22 or less

-0.010
(0.025)
0.105***
(0.017)
0.198***
(0.018)
0.089**
(0.040)

0.004
(0.013)
-0.001
(0.013)
-0.041**
(0.016)
-0.023
(0.019)

-0.000
(0.016)
0.098***
(0.015)
0.035*
(0.019)
0.029
(0.021)

0.015
(0.012)
0.060***
(0.010)
0.052***
(0.007)
0.011
(0.010)

R-squared
0.145
0.056
0.080
0.070
N
6742
6742
6742
6742
Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma,
dependent status, degree program in the first year, receipt of any financial aid, marital status,
indicator for having a child, having a job while enrolled, taking remedial courses parent's
education, and being in the top or bottom income quartile. Models also include state-level
controls for unemployment rate, appropriations, tuition at universities, other four-year
campuses, and two-year campuses, personal income per capita, population demographics, state
aid, and the number of public and private two-year and four-year institutions, and year fixed
effects. * p<0.10, ** p<0.05, *** p<0.01

Differences-in-differences
Multiple BPS cohorts allow for the investigation of whether changes in state transfer
policies impact student postsecondary outcomes. Table 2.4 presents results from Equation
(1)

4,5

. Each column represents a different postsecondary outcome. There is a negative but

4

Appendix Table 2.13 contains results from a differences-in-differences specification not using
the BPS sample weights.
5
Additional analyses, available upon request, run separate regressions for males and females, as
well as different racial subgroups. Some results differ from the baseline analysis: there is a large
positive, though marginally significant, relationship between state transfer policy and transfer for
48

statistically insignificant coefficient on the policy variable in columns (1) and (2) relating to
student transfer and attainment of a certificate within five years. State transfer policies are
positively related to associate’s and bachelor’s degree attainment, but the coefficients are not
statistically different from zero. It is possible that five years is not a long enough time to detect a
bachelor’s degree impact, given the lengthening amount of time students take to complete
bachelor’s degrees, even when they begin at four-year institutions.
Table 2.4 Baseline estimates of the state policy effect on postsecondary outcomes
(1)
(2)
(3)
Transfer
Certificate
AA
Transfer policy (ECS)
Goal: BA or higher

-0.014
(0.047)
0.199***
(0.018)

-0.008
(0.023)
-0.039**
(0.017)

0.034
(0.032)
0.039**
(0.019)

(4)
BA
0.025
(0.016)
0.051***
(0.007)

R-squared
0.160
0.066
0.092
0.075
N
6742
6742
6742
6742
Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma,
dependent status, degree program in the first year, receipt of any financial aid, marital status,
indicator for having a child, having a job while enrolled, taking remedial courses parent's
education, and being in the top or bottom income quartile. Models also include state-level
controls for unemployment rate, appropriations, tuition at universities, other four-year
campuses, and two-year campuses, personal income per capita, population demographics, state
aid, and the number of public and private two-year and four-year institutions. In addition,
models include state and year fixed effects. * p<0.10, ** p<0.05, *** p<0.01

The patterns in this analysis are similar to, though not as strong as, what Reynolds (2007)
found when looking at the associate’s degree transfer policy in North Carolina. His analysis
showed an increase in associate’s degrees, but no detectable change in bachelor’s degree receipt
following the policy. One common policy implemented by states is to designate the associate’s
degree as a transfer degree. It is possible that these policies are impacting associate’s degree
Black students. Additionally, there is a large positive and statistically significant impact of state
transfer policy on associate’s degree completion for Hispanic students.
49

completion. However, it is difficult to pin down when associate’s degree policies are added
using the policy data sources I have. As a result, it is not possible to distinguish the impact of
specific transfer policy elements (specifically relating to associate’s degrees) versus the presence
of any policy in the state. If I were able to accurately define associate’s degree transfer policies,
I might be able to find similar results to Reynolds (2007).
As in the pooled cohort analysis above, the student’s goal of receiving at least a
bachelor’s degree is positively associated with both transfer to other institutions and eventual
associate’s or bachelor’s degree receipt for beginning public, two-year students. However, this
goal is negatively related to attaining a certificate. Having a high school diploma is also
positively and statistically significantly related to receiving an associate’s degree within five
years.
Additional Analysis
While the baseline results do not show any relationship between state transfer policies
and post-secondary outcomes, I investigate alternate specifications to look for evidence of policy
impacts. Specifically, I analyze the sensitivity of the baseline results to the timing of the
policies, different policy definitions, and a different sample of students.
First, I look at the impact of defining whether a state has a policy at a different time. In
the baseline analysis in Table 2.4, a state is considered to have a policy if it is in place by
December of the first academic year for the BPS cohort. To check the timing, I now consider a
state to have a policy if it is in place by December of the second academic year; specifically, by
December of 1990, 1996, and 2004 for each BPS cohort respectively (compared to December of
1989, 1995, and 2003 for Table 2.4). The effect of this different policy classification is to define
some states as having a policy in place in an earlier cohort than previously. This specification

50

does not test for endogeneity of the policy in the traditional sense, but it is a sensitivity check on
the timing of the policy.
Table 2.5 shows the results in the differences-in-differences framework using the lagged
policy definition. The coefficients on the lagged policy variable are all small and statistically
insignificant. The coefficients on the lagged policy variable have the same sign and significance
as the coefficients on the policy variable in Table 2.4. Therefore, the impact of state policies on
associate’s degree completion is not sensitive to when the policy is defined.
Table 2.5 Additional analysis using a lagged policy variable
(1)
(2)
Transfer
Certificate
Transfer policy (ECS lagged)
Goal: BA or higher

-0.023
(0.049)
0.199***
(0.018)

-0.027
(0.022)
-0.039**
(0.017)

(3)
AA

(4)
BA

0.007
(0.033)
0.039*
(0.019)

0.023
(0.016)
0.051***
(0.007)

R-squared
0.160
0.066
0.091
0.075
N
6742
6742
6742
6742
Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma,
dependent status, degree program in the first year, receipt of any financial aid, marital status,
indicator for having a child, having a job while enrolled, taking remedial courses parent's
education, and being in the top or bottom income quartile. Models also include state-level
controls for unemployment rate, appropriations, tuition at universities, other four-year
campuses, and two-year campuses, personal income per capita, population demographics, state
aid, and the number of public and private two-year and four-year institutions. In addition,
models include state and year fixed effects. * p<0.10, ** p<0.05, *** p<0.01

Second, I check the sensitivity of the analysis using the ECS policy definition, to using
policies defined from the GAO report. Unfortunately, I cannot define GAO policies reliably for
the BPS:90/94 cohort, so this analysis only uses the last two BPS cohorts. As a result of using a
slightly different sample, I use both the ECS and GAO policy definitions in the differences-in-

51

Table 2.6 Sensitivity analysis using two policy definitions with the latter two BPS cohorts
(1)
(2)
(3)
(4)
(5)
Trans.
Cert.
AA
BA
Trans.
Transfer policy (ECS)

-0.005
(0.075)

-0.095**
(0.044)

-0.009
(0.058)

R-squared
N

(7)
AA

(8)
BA

-0.031
(0.061)
-0.026
(0.016)

0.009
(0.068)
-0.002
(0.024)

0.059*
(0.031)
0.040***
(0.008)

0.089
6028

0.096
6028

0.081
6028

0.044*
(0.022)

Transfer policy (GAO)
Goal: BA or higher

(6)
Cert.

0.164***
(0.018)

-0.025
(0.016)

-0.002
(0.023)

0.040***
(0.007)

0.206**
(0.085)
0.162***
(0.017)

0.171
6028

0.091
6028

0.096
6028

0.081
6028

0.173
6028

Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma, dependent status, degree program in the
first year, receipt of any financial aid, marital status, indicator for having a child, having a job while enrolled, taking remedial courses
parent's education, and being in the top or bottom income quartile. Models also include state-level controls for unemployment rate,
appropriations, tuition at universities, other four-year campuses, and two-year campuses, personal income per capita, population
demographics, state aid, and the number of public and private two-year and four-year institutions. In addition, models include state
and year fixed effects. * p<0.10, ** p<0.05, *** p<0.01
6

differences analysis in Table 2.6 . Columns (1) thru (4) use the ECS definition, while columns (5) thru (8) use GAO. Using the new
sample, there is now a positive, and marginally significant (at the 10% level), relationship between the ECS state policy variable and
bachelor’s degree attainment. Additionally, there is a negative and statistically significant relationship between attaining a certificate

6

Appendix Table 2.14 contains descriptive statistics for the sample used in Table 2.6, and Table 2.15 provides an analysis using sixyear transfer and degree attainment outcomes.
52

and the ECS policy variable. The negative but statistically insignificant relationship between
ECS policy and transfer remains. The impact on associate’s degree completion is now negative,
but not statistically different from zero.
The coefficients on certificate and bachelor’s degree attainment in Table 2.6 are larger
than those in Table 2.4, while the coefficients on transfer and associate’s degree completion are
smaller. There may be two reasons for the difference in coefficient size. First, Table 2.6 drops
the first BPS cohort. Second, there are fewer states present in the analysis sample in Table 2.6.
In each analysis, I kept states present in at least two BPS cohorts, which kept 37 states in the
sample in Table 2.4, but only 34 states in the sample in Table 2.6. Using a restricted sample of
7

states gives results similar to the baseline analysis in Table 2.4 . Therefore, it appears that
dropping the first BPS cohort drives the change in estimated coefficients in Table 2.6 as
compared to the baseline results. Next, I investigate the impact of changing the policy definition
on the results.
Column (5) shows a large, positive and statistically significant relationship between the
state policy variable as defined from the GAO reports and transfer. Adding a state transfer
policy as defined by the GAO increases a student’s probability of transfer by 20.6%.
Additionally, the coefficient on the GAO policy is positive and marginally significant in the
bachelor’s degree regression in column (8). This matches the result using the ECS policy in
column (4) of Table 2.6. Notably, there is no statistically significant relationship between the
GAO state articulation policy and the probability a student completes an associate’s degree
within five years. However, the GAO policies are related to the transfer of academic credit.
Easing transfer of credit between institutions does not necessarily encourage students to
7

See Appendix Table 2.16 for a specification of the baseline results using only the states present
in the sample in Table 2.6.
53

complete an associate’s degree. The strong relationship between transfer and GAO policy is
likely explained by the GAO policy focus on transfer of academic credit. The ECS definition of
statewide policies is broader than that of GAO and seems to include policies that are related to
students completing associate’s degrees.
Finally, I investigate the relationship between state articulation policies and
postsecondary outcomes for students who begin at public four-year institutions. Much of the
previous literature focuses solely on students who begin at public two-year institutions.
However, many state transfer policies, while aimed at easing transfer for public two-year
students, cover all public postsecondary students in the state. For example, general education
policies which require that general education credits transfer between public institutions in the
state can impact transfer and degree outcomes for both two-year and four-year beginners. In
Table 2.7, I use all three BPS cohorts in a differences-in-differences analysis with beginning
public four-year students (instead of public two-year students as in Table 2.4). Adding a state
transfer policy is related to a 3.5% increase the probability of transfer. However, the coefficient
is not statistically different from zero. There is a positive relationship between state transfer
policies and completion of a certificate or associate’s degree, and a negative relationship for
bachelor’s degree completion. However, none of these coefficients are statistically different
from zero. These results match those in Reynolds (2007), who found no statistically significant
relationship between state transfer policy and degree attainment for four-year students.

54

Table 2.7 Analysis using students beginning at public four-year institutions
(1)
(2)
(3)
Transfer
Certificate
AA
Transfer policy (ECS)
Goal: BA or higher

R-squared
N

(4)
BA

0.035
(0.027)
0.027
(0.026)

0.012
(0.009)
-0.018
(0.012)

0.018
(0.013)
-0.041***
(0.013)

-0.020
(0.027)
0.087***
(0.023)

0.022
9020

0.013
9020

0.052
9020

0.152
9020

Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma,
dependent status, degree program in the first year, receipt of any financial aid, marital status,
indicator for having a child, having a job while enrolled, taking remedial courses parent's
education, and being in the top or bottom income quartile. Models also include state-level
controls for unemployment rate, appropriations, tuition at universities, other four-year
campuses, and two-year campuses, personal income per capita, population demographics, state
aid, and the number of public and private two-year and four-year institutions. In addition,
models include state and year fixed effects. * p<0.10, ** p<0.05, *** p<0.01

Discussion
This paper, as well as much of the previous literature, finds little relationship between state
transfer policies and postsecondary outcomes. However, several papers, focusing on one policy
in one state, do find stronger impacts on specific outcomes. As mentioned previously, Reynolds
(2007) found a positive relationship between an associate’s degree policy in North Carolina and
associate’s degree completion. Quin (2013) showed that adding a transfer guarantee policy at
community colleges in California increased transfer and bachelor’s completion. However, Quin
(2013) found no impact on associate’s degree completion. There was no associate’s degree
component to the admission guarantee. Taken together, these findings suggest that the
postsecondary outcomes impacted depend on the type of transfer policy put in place. Therefore,
combining different policies that may affect different outcomes may not produce strong results.

55

Using a single indicator for whether a state has any policy may be one reason why the
literature has found little relationship between state transfer policies and postsecondary
outcomes. State transfer policies are quite varied, as displayed in ECS (2010). A state may have
none of these policies, or may combine components of several different types of policies. The
patterns of types of state policies in place are difficult to classify. Therefore, using an indicator
for whether a state has a policy without distinguishing the type of policy may mask the true
impact of different policy components. The one exception to using a single policy indicator is
Gross and Goldhaber (2009), who investigate the impact of specific transfer policy components.
They include indicators for each type of transfer policy, such as automatic transfer of an
associate’s degree, but find no impact on transfer. However, they do not look at associate’s
degree attainment as an outcome.
Moving forward, researchers should focus on thinking about what outcomes they expect
particular policies to impact. Researchers should also consider whether these policies have
impacts prior to transfer; for example, do state transfer policies impact who goes to college, or
what college they initially choose. Reynolds (2007) found that state policies did not affect
college attendance, but were related to an increase in the probability of women attending twoyear colleges. I conducted a similar exercise using first-time enrollment data from the Integrated
Postsecondary Education Data System (IPEDS), but found no relationship between state transfer
8

policy and initial college type. Thinking about expected outcomes may help researchers begin
to classify policies into clear groups. Additionally, it may help identify particular policies that
may be working counter to each other.

8

Results are available upon request to the author.
56

Conclusion
This paper expands on previous research investigating the relationship between state
articulation policies and postsecondary outcomes by looking at policy changes over time.
Estimates suggest that state articulation policies are not related to transfer or degree completion
for either beginning public two-year or public four-year students. However, these results are
sensitive to the definition of the policy used, as well as the number of BPS cohorts in the
analysis. Aspirations to receive a bachelor’s degree or higher is positively associated with
bachelor’s degree completion in all specifications.
This paper uses data from three cohorts of the BPS, which has several strengths and
weaknesses. One advantage of the BPS data is that it tracks multiple cohorts of entering
postsecondary students, allowing for the investigation of trends over time. In addition, the BPS
includes older students; this is more representative of the postsecondary population, particularly
for two-year institutions. However, the tradeoff is that there is no high school to college link.
As a result it is not possible, using the BPS, to investigate the effect of state transfer policies on
whether students attend college, or possible shifting between four-year and two-year institutions.
In addition, the BPS only tracks students for five to six years. This is a relatively short period to
look at bachelor’s degree outcomes, especially for transfer students, as even students who begin
at four-year institutions may now take six years to graduate.

57

APPENDICES

58

APPDENDIX A

Policy Appendix
Table 2.8 State-level transfer and articulation policy year by source
State
Alabama
Alaska
Arizona
Arkansas
California
Colorado
Connecticut
Delaware
Florida
Georgia
Hawaii
Idaho
Illinois
Indiana
Iowa
Kansas
Kentucky
Louisiana
Maine
Maryland
Massachusetts
Michigan
Minnesota
Mississippi
Missouri
Montana
Nebraska
Nevada
1
New Hampshire
New Jersey
New Mexico
New York
North Carolina
North Dakota
Ohio

2

3

GAO (2005)
1994

ECS (2010)
1994

1996
1989
1988
1985
1996

1996
1989
1991
1985
1996

1975

1975

1973
1987

1973
1987

1991
1997
1999
1985
1996
by 1991

1991
July, 1990
1999

1985

1996
1988
1999
2001

by 1994
1991
1997
1994
1995
by 1987
1997
2003
59

1994
2008

1995
1997
2007
2003

Table 2.8 (cont’d)
Oklahoma
Oregon
Pennsylvania
Rhode Island
South Carolina
South Dakota
Tennessee
Texas
Utah
Vermont
Virginia
Washington
West Virginia
Wisconsin
Wyoming

1995
1987
2000
1987
1994
1998
2000
1997
2001
2004
1983
2001
1973
1997

1992
1983
2000

39

States with articulation policy

1995
1987
2006
1979
2009
1998
2000
1997
1997, repealed 2001

36

1997

1

States with no policy year do not have a state policy, with the exception of New Hampshire,
for which reliable data about the state policy year was not available.
2
Policy year is the year of the state legislation listed in Appendix II.
3

Policy year is the year of the legislation/policy listed in the "Statewide Policy" column.

60

APPENDIX B

Data Appendix
Table 2.9 BPS variables
Label
Sex
Age when first enrolled
Race/ethnicity
Any remedial education participation
Dependency status
Receive any financial aid
Attend full-time
Degree goal
Parents highest education
Degree program during first year
Have a high school diploma
Marital status
Have a child
Work while enrolled
Income
First institution type
First institution state
Ever transfer institutions
Ever transfer to a four-year institution
Ever obtain a certificate
Ever obtain associate's degree
Ever obtain bachelor's degree
BPS analysis weights

My variable name
BPS:90/94 variable
BPS 96/01 variable
BPS 04/09 variable
female
H_GENDR
SBGENDER
GENDER
age
AGE
AGE
AGE
race
H_RACE
SBRACE
RACE
remedial
REMEDIAL
RMANYY1
REMETOOK
dependent
EC_DEPE1
SBDEP1Y1
DEPEND
anyaid
ANYAID89
AIDANY1
TOTAID
fulltimeattend
ATTNSTAT
ATTNSTAT
ATTNSTAT
goalbaorhigher
ASPIRE
EPHDEGY1
HIGHLVEX
paredbaorhigher
RPARED
PBEDHI3
PAREDUC
degreeprogram
PROGTYP
DGPGMY1, PGM2Y1
UGDEG
hsdiploma
H_HSDIP
HSDEG
HSDEG
married
MARITAL
SBMAFAY1
SMARITAL
child
KIDS8990
SBMRCHY1
DEPCHILD
job
EMWKHR3
J1HOURY1
JOBENR2
topquart, bottomquart
FAMINCPR
PCTALL2
PCTALL
beginatpublic2yr
OFCO8990
ITNPSAS
FSECTOR
firstinststate
FIPS
INSTATE
INSTSTAT
evertransfer
TRANTO
TRINTY2B
TFNUM6Y
transferto4yr
EVER4YR
TRINTY2B
IT4Y6Y
getcert
RECDCT
DGDTCT2B
ATCTDT6Y
everaa
DEGASTAT
DGRETY2B
ATTYPE6Y
everba
DEGASTAT
DGRETY2b
ATTYPE6Y
weight1and3
BPS94AWT
B01LWT2
WTA000
61

Table 2.10 Sample sizes (unweighted) for each BPS cohort
BPS:90/94
BPS:96/01
Beginning sample
size
7253
12085

BPS:04/09
16684

Public 2yr

Public 4yr

Public 2yr

Public 4yr

Public 2yr

Public 4yr

Students beginning
at institution sector

704

1613

1516

5161

5549

4643

Responded to first
and last interview
wave

689

1603

1061

3916

5546

4541

No missing
information

629

1509

749

3230

5401

4301

First institution
state present in 2
of 3 BPS cohorts
622
1509
742
3230
5378
4281
Source: Author calculations from BPS:90/94, BPS:96/01, and BPS:04/09
Note: There are 37 states with beginning public 2-year students in two of the three BPS cohorts,
and 43 states with beginning public 4-year students in two of the three BPS cohorts.

62

APPENDIX C

Additional Tables
Table 2.11 Descriptive statistics for all first-time beginning public two-year and four-year
students by BPS cohort (means in %)
1989-1990
1995-96
2003-04
Public
Public
Public
Public
Public
Public
2-year
4-year
2-year
4-year
2-year
4-year
Student Background
Female
51.2
53.2
51.2
53.9
56.3
55.2
White, non-Hispanic
76.6
82.1
72.8
73.5
60.4
71.1
Black, non-Hispanic
8.0
8.6
11.5
10.7
14.1
8.8
Hispanic
11.1
3.9
11.0
8.9
15.9
8.5
Other race
4.2
5.4
4.6
7.0
9.6
11.5
HS diploma
92.2
98.6
88.2
98.5
85.6
96.2
Certificate program
11.6
8.8
4.4
Associate's degree program
67.4
72.9
79.7
Bachelor's program
75.9
92.5
91.7
Age 22 or less
75.1
95.9
72.5
94.6
69.9
95.2
Dependent
47.9
83.0
65.5
91.8
62.8
92.7
Received any aid
27.6
48.2
42.5
68.7
53.9
76.4
Goal: BA or higher
67.0
94.6
67.2
81.4
81.1
98.5
Parent ed. BA or higher
25.8
42.5
26.2
46.4
27.9
54.4
Attend full-time
36.2
76.3
46.8
86.6
48.9
90.1
Married
20.4
3.8
21.7
3.6
17.5
2.9
Have a child
19.1
3.2
21.0
4.2
22.3
3.6
Work while enrolled
82.7
78.6
81.8
60.7
77.6
59.0
Take remedial courses
18.2
16.1
23.2
16.5
29.4
19.3
Income (bottom quartile)
26.2
20.4
28.5
23.4
25.1
18.6
Income (top quartile)
17.3
27.6
16.9
27.1
25.4
30.8
Postsecondary Outcomes
Ever transfer
42.7
28.1
39.5
25.3
36.3
24.1
Transfer to 4-year
24.5
12.6
.
.
.
.
Attain certificate
13.1
3.9
11.0
3.0
8.0
1.9
Attain associate's
21.6
6.2
19.3
4.3
15.1
3.9
Attain bachelor's
6.1
46.7
6.9
47.6
5.8
48.7
State characteristics
Transfer policy (ECS)
21.1
14.5
55.1
33.0
74.3
67.0
Transfer policy (GAO)
.
.
66.2
44.2
87.3
81.1
Unemployment rate
5.2
5.3
5.6
5.4
6.1
5.8
63

Table 2.11 (cont’d)
Sample size (unweighted)
663
1575
980
3852
5546
4541
Source: Author calculations from BPS:90/94, BPS:96/01 and BPS:04/09.
Note: Student background characteristics are defined for the first year in postsecondary while
all degree and transfer variables are defined as of five years after beginning postsecondary.
BPS analysis weights are used. States in at least two of three BPS cohorts.

Table 2.12 Cross-sectional analysis of the relationship between transfer policies and
postsecondary outcomes
(1)
(2)
(3)
(4)
Transfer
Certificate
AA
BA
A. BPS:90/94 (N = 622)
Transfer policy (ECS)
-0.059*
-0.024
0.147***
0.008
(0.035)
(0.033)
(0.036)
(0.022)
Goal: BA or higher
0.252***
-0.069
0.132***
0.083***
(0.045)
(0.047)
(0.035)
(0.016)
R-squared
0.160
0.013
0.104
0.055
B. BPS:96/01 (N = 742)
Transfer policy (ECS)
0.097**
0.006
-0.018
0.045***
(0.047)
(0.025)
(0.027)
(0.016)
Goal: BA or higher
0.170***
0.021
-0.005
0.032*
(0.035)
(0.034)
(0.028)
(0.017)
R-squared
0.208
0.045
0.114
0.087
C. BPS:04/09 (N = 5378)
Transfer policy (ECS)
-0.008
0.017
-0.054***
-0.006
(0.022)
(0.012)
(0.015)
(0.009)
Goal: BA or higher
0.156***
-0.061***
-0.014
0.045***
(0.022)
(0.013)
(0.023)
(0.005)
R-squared
0.120
0.127
0.064
0.057
Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma,
dependent status, degree program in the first year, receipt of any financial aid, marital status,
indicator for having a child, having a job while enrolled, taking remedial courses parent's
education, and being in the top or bottom income quartile. Models also include state-level
controls for unemployment rate, appropriations, tuition at universities, other four-year
campuses, and two-year campuses, personal income per capita, population demographics, state
aid, and the number of public and private two-year and four-year institutions. * p<0.10, **
p<0.05, *** p<0.01

64

Table 2.13 Baseline analysis without sample weights
(1)
(2)
Transfer
Certificate
Transfer policy (ECS)
Goal: BA or higher

0.007
(0.031)
0.184***
(0.013)

-0.007
(0.019)
-0.062***
(0.009)

(3)
AA

(4)
BA

0.016
(0.030)
0.024*
(0.014)

0.015
(0.017)
0.058***
(0.006)

R-squared
0.133
0.112
0.071
0.069
N
6742
6742
6742
6742
Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma,
dependent status, degree program in the first year, receipt of any financial aid, marital status,
indicator for having a child, having a job while enrolled, taking remedial courses parent's
education, and being in the top or bottom income quartile. Models also include state-level
controls for unemployment rate, appropriations, tuition at universities, other four-year
campuses, and two-year campuses, personal income per capita, population demographics, state
aid, and the number of public and private two-year and four-year institutions. In addition,
models include state and year fixed effects. * p<0.10, ** p<0.05, *** p<0.01

Table 2.14 Descriptive statistics for first-time beginning public two-year students in Table 2.6
1995-96
2003-04
Public 2-year
Public 2-year
Student Background
Female
54.1%
56.1%
White, non-Hispanic
71.8%
59.9.%
Black, non-Hispanic
11.9%
14.4%
Hispanic
11.3%
16.1%
Other race
5.0%
9.7%
HS diploma
87.7%
85.5%
Certificate program
10.0%
4.4%
Associate's degree program
72.9%
79.7%
Age 22 or less
68.6%
70.4%
Dependent
59.7%
63.5%
Received any aid
50.4%
53.3%
Goal to get BA or higher
68.1%
81.6%
Parent education BA or higher
25.6%
28.0%
Attend full-time
47.0%
48.7%
Married
22.3%
17.0%
Have a child
24.7%
21.6%
Work while enrolled
79.7%
77.6%
Take remedial courses
24.0%
29.4%
65

Table 2.14 (cont’d)
Income (bottom quartile)
28.7%
25.2%
Income (top quartile)
17.1%
25.3%
Postsecondary Outcomes (after 5 yrs)
Ever transfer
39.6%
36.5%
Attain certificate
10.4%
8.0%
Attain associate's
19.5%
15.0%
Attain bachelor's
7.4%
5.8%
Postsecondary Outcomes (after 6 years)
Ever transfer
42.3%
40.1%
Transfer to 4-year
30.2%
27.5%
Attain certificate
11.4%
9.6%
Attain associate's
20.7%
18.1%
Attain bachelor's
11.1%
11.4%
State characteristics
Transfer policy (ECS)
54.2%
74.5%
Transfer policy (GAO)
64.5%
87.2%
Unemployment rate
5.6%
6.1%
Sample size (unweighted)
742
5286
Source: Author calculations from BPS:96/01 and BPS:04/09.
Note: Student background characteristics are defined for the first year in postsecondary while all
degree and transfer variables are defined as of five or six years after beginning postsecondary. BPS
analysis weights are used. States in at least two of three BPS cohorts.

66

Table 2.15 Sensitivity analysis using two policy definitions with the latter two BPS cohorts (six year outcomes)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Transfer Certificate
AA
BA
Transfer Certificate
AA
BA
Transfer policy
(ECS)

0.050
(0.069)

-0.084**
(0.038)

0.002
(0.064)

Transfer policy
(GAO)
Goal: BA or higher

0.165***
(0.021)

-0.028
(0.019)

-0.005
(0.025)

0.007
(0.030)
0.243***
(0.075)
0.069*** 0.163***
(0.012)
(0.020)

-0.021
(0.056)
-0.028
(0.019)

0.030
(0.070)
-0.005
(0.025)

0.043
(0.042)
0.069***
(0.012)

R-squared
0.172
0.077
0.085
0.115
0.174
0.076
0.085
0.115
N
6028
6028
6028
6028
6028
6028
6028
6028
Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma, dependent status,
degree program in the first year, receipt of any financial aid, marital status, indicator for having a child, having a
job while enrolled, taking remedial courses parent's education, and being in the top or bottom income quartile.
Models also include state-level controls for unemployment rate, appropriations, tuition at universities, other fouryear campuses, and two-year campuses, personal income per capita, population demographics, state aid, and the
number of public and private two-year and four-year institutions. In addition, models include state and year fixed
effects. * p<0.10, ** p<0.05, *** p<0.01

67

Table 2.16 Baseline results restricted to the states used in Table 2.6
(1)
(2)
Transfer
Certificate
Transfer policy (ECS)
Goal: BA or higher

0.002
(0.044)
0.200***
(0.018)

-0.015
(0.024)
-0.039**
(0.017)

(3)
AA

(4)
BA

0.032
(0.032)
0.039*
(0.020)

0.027
(0.016)
0.052***
(0.007)

R-squared
0.159
0.064
0.092
0.077
N
6636
6636
6636
6636
Note. -- All models include controls for sex, race/ethnicity, receipt of high school diploma,
dependent status, degree program in the first year, receipt of any financial aid, marital status,
indicator for having a child, having a job while enrolled, taking remedial courses parent's
education, and being in the top or bottom income quartile. Models also include state-level
controls for unemployment rate, appropriations, tuition at universities, other four-year
campuses, and two-year campuses, personal income per capita, population demographics, state
aid, and the number of public and private two-year and four-year institutions. In addition,
models include state and year fixed effects. * p<0.10, ** p<0.05, *** p<0.01

68

REFERENCES

69

REFERENCES

Anderson, G., Sun, J. C., and Alfonso, M. (2006) “Effectiveness of Statewide Articulation
Agreements on the Probability of Transfer: A Preliminary Policy Analysis.” Review of Higher
Education, 29(3): 261-291.
Barry, R., and Barry, P. (1992) “Establishing equality in the articulation process.” In B. W.
Dziech and W. R. Vilter (Eds.), Prisoners of elitism: The community college’s struggle for
stature. New Directions for Community Colleges, 78(2): 35-44. San Francisco: Jossey-Bass.
Education Commission of the States (2010) “Transfer and Articulation Policies.”
http://www.ecs.org/clearinghouse/90/70/9070.pdf
Government Accountability Office (2005) “Transfer students: Postsecondary institutions could
promote more consistent consideration of coursework by not basing determinations on
accreditation.” (No. GAO-06-22). Washington, DC.
Gross, B. and Goldhaber, D. (2009) “Community College Transfer and Articulation Policies.”
Center on Reinventing Public Education. CRPE Working Paper # 2009_1.
http://survey.csuprojects.org/uploads/2-/4X/2-4XaUlUV9-55g-XS_ldcQ/Community-CollegeTransfer-and-Articulation-Policies.pdf
Ignash, J., and Townsend, B. (2000) “Evaluating State-Level Articulation Agreements
According to Good Practice.” The Community College Review, 28(3): 1-21.
Quin, E. (2013) “Guaranteed Transfer Policies and Post-Secondary Outcomes.” Working paper.
Michigan State University, East Lansing.
Roksa, J. and Keith, B. (2008) “Credits, Time, and Attainment: Articulation Policies and Success
after Transfer.” Educational Evaluation and Policy Analysis, 30(3): 236-254.
Reynolds, C. “Estimating the Effects of Articulation Agreements Using Re-Weighting
Techniques,” mimeo, University of Michigan, Ann Arbor, 2007.
U.S. Department of Education. National Center for Education Statistics. Beginning
Postsecondary Students Longitudinal Study Second Follow-up (BPS:90/94) Final Technical
Report, NCES 96-153, by Daniel J. Pratt, Roy W. Whitmore, Jennifer S. Wine, Karen M.
Blackwell, Barbara H. Forsyth, Timothy K. Smith, Elizabeth A. Becker, Kurt J. Veith, Marisa
Mitchell, and Geoffrey D. Borman. Larry G. Bobbitt, project officer. Washington DC: 1996.
U.S. Department of Education, National Center for Education Statistics. Beginning
Postsecondary Students Longitudinal Study 1996-2001 (BPS:1996-2001) Methodology Report,
NCES 2002-171, by Jennifer S. Wine, Ruth E. Heuer, Sara C. Wheeless, Talbic L. Francis, Jeff
70

W. Franklin, and Kristin M. Dudley, RTI. Paula R. Knepper, Project Officer. Washington, DC:
2002.
Wine, J., Janson, N., and Wheeless, S. (2011). 2004/09 Beginning Postsecondary Students
Longitudinal Study (BPS:04/09) Full-scale Methodology Report (NCES-2012-246). National
Center for Education Statistics, Institute of Education Sciences, U.S. Department of Education.
Washington, DC. Retrieved [date] from http://nces.ed.gov/pubsearch.

71

CHAPTER 3
RELATIONSHIP BETWEEN PRIVATE SCHOOLING AND ACHIEVEMENT: RESULTS
FROM RURAL AND URBAN INDIA
Introduction
Several recent developments in Indian education have led to a growing interest in the
performance of private schools, especially in their relative ability to improve student
achievement compared to public schools. In the western literature, which focuses primarily on
the United States, several researchers have investigated this issue using a range of datasets and
econometric techniques. In the Indian context, however, such research is still quite limited, so
this study makes a useful contribution. We use the propensity score matching technique along
with ordinal logit and ordinary least squares regressions on recent, nationally representative,
household data from rural and urban India, and investigate the relative performance of public and
private schools.
Indian education, relevant background
Several interrelated factors are driving the interest in how private schools are performing
in India. In recent years India has enjoyed tremendous success in enrolling millions of
previously un-enrolled children in school (e.g., Government of India, 2009-2010). As more
children enroll in school, academics, policymakers and parents have become more interested in
how well students are performing, regardless of what school they attend.
At the same time, this massive surge in enrollments has placed additional pressure on the
already strained government school system which is also the largest provider of education in
1

India. Several studies point to widespread parental dissatisfaction with the performance of the

1

According to 2008-2009 data at the elementary level 70-80 percent of all rural enrollments are
in government school where as 36-50 percent of all urban enrollments are in government
72

government schools. They argue that government schools are fast losing their enrollments to
private providers (e.g., Muralidharan & Kremer, 2006). Guided at least in part by this growing
dissatisfaction with the government school system, the nation has seen a steady growth in private
school enrollments (e.g., Kingdon, 2007; Wadhwa, 2009). This growth in the private school
sector is yet another factor that guides the growing interest in their performance compared to that
of public schools.
The phenomenon of the growth of private schooling is also nuanced. Based on recent
research, James Tooley and his colleagues argue that much of the growth in private schooling is
actually occurring in the so-called ‘low-fee’ private school sector: privately-run schools that cost
far less than elite private schools. One of the most important cost-saving strategies among these
schools is lower teacher salaries. According to Tooley and Dixon (2003), at least in certain parts
of urban India such schools may actually dominate and outnumber public schools. This
fascination with a potentially low-cost yet more effective alternative to public schools is the third
factor that has led to an interest in the performance of private schools in India.
Finally, and most recently, the Indian government passed the Right to Education Bill.
Among other things this landmark bill mandates that all private schools must reserve 25% of
their seats for poor and marginalized children, and that the cost of these seats will be paid by the
government. The bill also mandates that the government will regularize the operation of private
schools (Gazette of India, 2009). Once again, this highly debated and discussed bill has
underscored the need to better understand how well private schools are performing in India.
To summarize; general growth in school enrollment, increased performance pressure on
government school systems, growth in the private school sector, the emergence of low-fee
schools. Since about 70 percent of India population still resides in the rural area this makes
government the largest provider of elementary education.
73

private schools, and the added focus on private schools because of the Right to Education bill
have led to great interest in the performance of private schools in India. Do these schools
produce higher achievement and do they produce such outcomes more efficiently (at lower
costs)? These questions have been at the center of several civil society and policy debates.

2

Private school performance: Existing research from India
Little empirical evidence is available on the performance of private schools in India.
Researchers have found that, in general, without accounting for covariates (or in raw terms)
children in private schools out-perform those in public schools (e.g., Wadhwa, 2009). Even after
controlling for covariates, various studies that rely on varied samples and varied methods
generally tend to find a private school advantage (e.g., Kingdon, 2007; Muralidharan & Kremer,
2006; Tooley, Dixon , Shamsan, & Schagen, 2010; Goyal & Pandey, 2009).
However, to truly identify the effect of private schooling it is important not just to control
for covariates in a regression framework but also to account for the selection issue. That is
because children do not enroll in schools randomly; their school participation reflects their
parents’ explicit choices and the value their family places on education, factors that may in turn
be related to the children’s achievement. Conventional data indicate that children ‘sort’ into
school types: typically, the children from better off and better informed families tend to enroll in
private schools (e.g., Goyal & Pandey, 2009). This sorting leads to the possibility that in the

2

For instance, a group of citizens in New Delhi, India’s capital, have formally come together to
demand a more systematic implementation of school choice via school voucher programs. This
School Choice Campaign (SCC) spearheaded the Delhi school voucher program. To give
another example, in 2009 India’s leading Economics and Policy journal Economic and Political
Weekly carried an impassioned exchange spanning several issues between policy scholars,
educational and economic researchers about the role and importance of private schools in Indian
education.
74

regression framework cov (PRIVATE, ei) ≠ 0, where ei is the error term in the regression
equation which in turn would lead to a biased estimation of the ‘private’ school effect.
In India, to our knowledge, four studies have attempted to explicitly account for, or
correct for, the selection issue (Kingdon, 1996; French & Kingdon, 2010; Desai, Dubey,
Vanneman, & Banerji, 2008; Goyal, 2009). In general these studies find a positive effect for
private schools; after appropriate corrections are made for the selection issue, this effect may be
attenuated but it does not disappear. Kingdon (1996) uses the selection approach developed by
Heckman, Ichimura, and Todd (1997) to correct for selection bias, and Goyal (2009) uses a
technique proposed by Altonji, Elder and Taber (2005, cited in Goyal 2009) to measure the
selection bias. The key limitation of these studies is that their findings have only limited
generalizability. Kingdon’s study is based on data from one district in one state, and Goyal’s
study uses data from eight districts in just one Indian state.
French and Kingdon (2010) had access to data from all of rural India, with a set of
household covariates including maternal age, maternal grade level education, and maternal
reading ability. Like Goyal (2009), they measured selection bias using the technique proposed by
Altonji, Elder and Taber. They used household fixed effects to correct for selection and measure
the effect of private schooling at the child level (which is the focus of the present paper). They
similarly used a 3-year panel of village level data to correct for selection and measure the private
school effect at the village level. Using the household fixed effects approach they compared the
performance of two children within the same household where one child was enrolled in private
school and another in public school. This approach allowed them to account for all the observed
and unobserved household covariates. However, the household fixed effects approach is also not
without limitations. The problem with using household fixed effects is that parents often invest
75

differently in children within the family depending on the child’s observed potential (observed
by the parent, not by the researcher) or due to some other form of favoritism (in India for
example preferential investment in the male child is well documented) . Thus a more ‘heavily
invested’ child attending private school may be doing better simply because he or she has more
academic talent to begin with or because he or she is being favored in other ways, in addition to
being ‘selected’ for (rather than being ‘randomly allocated’ to) the more expensive private
education. The other issue potentially with the household fixed effects approach is the
representativeness of the sample which may have implications on the generalizability of these
findings. The households that make different schooling decisions for their children may be a
small subset of the broader population and they may be unique compared to the rest of the
3

population .
Finally, Desai et al. (2008) use the India Human Development Survey (IHDS) data,
which we also use for this study. They also use the Heckman et al. selection approach to first
model private school choice, and then measure the achievement levels associated with private
schooling. Citing Stolzenberg and Relles (1997) they acknowledge that the use of the Heckman
approach is sensitive to the exclusion restrictions; therefore, like French and Kingdon (2010)
3

A descriptive analysis of our nationally representative data (which is different than the data
used by French and Kingdon, 2010) upheld both these concerns. We find that families where
siblings attend different school types do treat public and private school attending siblings
differently. This was reflected in the private school attending sibling receiving additional
educational support at home (in terms of receiving private tuition and longer private tuition
hours) and missing fewer days at school, in spite of attending schools that were much farther
away from home. Also, confirming the male-child bias observed elsewhere in the Indian
literature, private school attending children in families where other siblings attended public
schools were disproportionately more male than the overall proportion of male children attending
private schools. We also found that only 5-6 percent of rural and urban children reside within
such families that engage in such differential decisions for their children. In rural areas such
children came from families that were more educated and economically better off. In both rural
and urban areas such children were disproportionately likely to come from female headed
households.
76

they also use the household fixed effects approach which suffers from the limitations discussed
above. In addition Desai et al. analyze rural and urban data within a single econometric
framework. Given the differential spread and demand for private schooling in rural and urban
India and given the large overall differences between these two settings, such a unified treatment
of rural and urban schools may not be appropriate for such an analysis.
In summary, several existing studies have corrected or attempted to correct for selection
bias and thus present a valuable addition to the literature on private schooling in India; still, each
leaves room for improvement and additional analysis in terms of data, methods and modeling
approaches. Also, despite the recent growing interest in low-fee unaided private schools, none of
these four recent studies has attempted to evaluate their performance. Given this background,
and given the growing interest in the performance of private schools in India among
policymakers, academics, and members of civil society, the present study using nationally
representative data makes an important contribution to the literature. We treat rural and urban
data separately, make an initial attempt to distinguish the performance of low-fee private
schools, and most importantly attempt to correct for the inherent selection bias in private school
choice using propensity score matching approach- a “powerful method for reducing bias” (Imai,
2005, pp.295) when using a dataset with an extensive set of relevant covariates, like the one we
use for this study.
Materials and methods
Dataset and key variables
For our analysis we used the propensity score matching technique (Rosenbaum & Rubin,
1983a) to compare the performance of public and private schools. We supplemented this
analysis with regressions on appropriate covariates.

77

Dataset
We used the India Human Development Survey 2005 (IHDS), a nationally representative
survey of 41,554 households, conducted jointly by researchers from the University of Maryland,
in the United States and the National Council of Applied Economic Research in India. This
survey is unique: in addition to several standard household variables, variables pertaining to
individuals in the household, and basic descriptors for children’s schools, it also includes
assessments of the children’s reading, writing, and arithmetic skills. Standard test-score data,
which are often used to compare the performances of children in public and private schools, are
collected from children within classrooms and schools. Because such datasets rely on the
children’s reports about their household (parental education, household possessions, household
economic status, etc.) they often cannot describe the students’ home environments completely
accurately. The IHDS data, however, provide both a rich set of household descriptors and testscore data (collected from children in the household aged eight to eleven), making it an ideal
dataset to address the issue of selection into private schools.
To conduct this analysis we worked with data on those children who were enrolled in
school at the time of data collection and for whom the test-score data were available (which
constrained us to the age group eight to eleven). Missing data were not a serious concern for our
study. Less than one percent children in rural and urban data had missing information on one of
the three key outcome variables (discussed in detail below). These observations simply dropped
out of the regression analysis since it is not advisable to impute missing data on the dependent
variables. Even fewer observations if any, were missing for the other variables. Depending on
the definition of private school utilized the final sample covered around 7,000 children from

78

rural India and around 3,000 children from urban India. Various relevant tables provide the
precise sample sizes as applicable.
Definitions of private school
Broadly, India has at least three types of schools. Public schools are the schools owned
and run by the government. The private school category can be divided into two groups: privateaided, and private-unaided. In private-aided schools the government provides all or the majority
of the funding (aid) but the school is run privately. Private-unaided schools are both funded and
run privately. Private-aided schools are more regulated, as most of their decision-making is
overseen by the government, unlike the situation in the private unaided schools. Prominent
researchers have argued that for all practical purposes ‘private aided’ schools should not be
included in the ‘private’ category, as they bear much greater resemblance to regular public
schools in most regards including their fee structure (Kingdon, 2007). For the purpose of this
paper, therefore, we use two different definitions of private school. We define PRIVATE to
4

include only private non-aided schools ; we group both public and private-aided schools in the
‘public’ category. We define PRIVATE_ONLY to include once again private non-aided schools
in the private category, but this time in the public category we include only government schools,
i.e. we compare non-aided private schools with government schools only. Thus using
PRIVATE_ONLY effectively drops the observations where the child is enrolled in private-aided
schools. This will result in not using little over 2 percent of rural observations and close to 6
percent of the urban observations. For the sake of simplicity throughout the rest of the paper we
refer to ‘private non-aided’ as simply ‘private,’ and indicate the reference category by identifying
if we are using the PRIVATE or PRIVATE_ONLY as the independent variable.

4

From our data we are unable identify if a school is recognized by the government or not.
79

An imperfect effort to identify low-fee private schools
In addition to comparing private and public schools, we made an effort, albeit imperfect,
to identify ‘low-fee’ private schools. We did so explicitly to see if our data would provide any
evidence for or against the benefits of low-fee private schools relative to government schools.
Corresponding with the two private school definitions we classified a private school as ‘low-fee
private school’ (LFPVT and LFPVT_ONLY) if its reported fees were lower than the maximum
fees for government schools in the same district for the same grade level (primary or middle) and
for the same setting (rural or urban). The remaining private schools for the same grade level,
setting, and district were identified as ‘high-fee private schools’ (HFPVT and HFPVT_ONLY).
Dependent variables
The key outcome of interest is student performance. The assessment data that we
analyzed were collected using widely used tests developed by the NGO Pratham. The tests are
described in detail by Desai, Dubey, Vanneman, and Banerji (2008). These tests are limited in
their psychometric abilities but they do provide a useful benchmark of student achievement
levels in India where such data has traditionally been hard, if not impossible, to obtain. The
children were given separate reading, math, and writing tasks. Their performance in reading was
graded on a scale of 0 to 4 (from cannot read at all, to can read a whole story), in math on a scale
of 0 to 3 (cannot recognize 2-digit numbers to can divide a 3-digit number by a 1-digit number)
and in writing on a scale of 0 to 1 (cannot write, to can write with 2 mistakes or less).
In our analysis, none of the scores on these three tasks were distributed normally—which
was not unexpected given the nature of the items. We therefore generated two primary dependent
variables. We generated a dependent variable SCORE which simply is the summation of the
student’s performance on the three tasks. Such a continuous variable allows us to harness the

80

maximum amount of variation in child achievement available in the dataset. However, a ‘unit
5

change’ in SCORE has no clear meaning in terms of changes in the child’s skill-levels . So we
additionally generated an average score for each student across the three tasks. We then ranked
students based on where they fall in the national distribution of average proficiency scores. This
dependent variable Proficiency (PROF) is an ordinal variable with three values: 1 = low average
proficiency (a child in the bottom third of the national average proficiency distribution), 2 =
medium average proficiency (a child in the middle third) and 3 = high average proficiency (a
child in the top third). While this dependent variable is more complex to understand, it also
carries a clearer substantive meaning. In addition to these two primary dependent variables, we
conducted all the analysis separately for each of the three test-scores (READ, WRITE, MATH).
Methods
While we are primarily interested in the results from the propensity score matching analysis, we
also conducted regression analysis. We conducted this analysis to compare our findings with a
large body of existing literature that only corrects for selection through covariates in the
regression framework. For the sake of simplicity we discuss the methods with respect to
independent variable PRIVATE, but the same details apply also to independent variable
PRIVATE_ONLY.
Ordinal Logit and Ordinary Least Square Regressions
Equations 1 and 2 express the regressions we estimated. Equation (1) quantifies the
private benefit with appropriate covariates. Equation (2) uses the low-fee, high-fee definition
along with the covariates. We conducted each analysis separately for both the rural and urban
data. Each regression model includes appropriate sample weights and accounts for clustering at

5

A visual inspection reveals that this variable is also not normally distributed.
81

the district level. The covariates are derived from the literature. The HOUSEHOLD variables
are the family’s caste, income, assets, the number of people in the household, and the sex and
education of the household head; the CHILD variables are the child’s gender and age, the years
of education completed by the child, whether the child receives private tuition, and whether or
not the child works in addition to attending school.
Y = β0 + βPVT PRIVATE + βHH HOUSEHOLD + βc CHILD

+e

Y = β0 + βLFPVTLFPVT + βHFPVTHFPVT + βHH HOUSEHOLD + βc CHILD + e

(1)
(2)

We estimate equations (1) and (2) for the dependent variables SCORE, READ, WRITE,
and MATH using OLS. We used ordinal logistic regression for dependent variable PROF. The
ordered logit model assumes that there is an underlying, continuous, latent dependent variable y*
mapped onto the observed variable PROF. In our case, that would mean there is some
underlying ability to read, write, and do arithmetic. Formally, the underlying process looks like
y* = x'β + ε, where y* is the unobserved dependent variable, x is a vector of independent
variables, and β is the vector of regression coefficients. However, we only see the proficiency
group that each student is in. As a result, we can write our observed dependent variable PROF,
which is proficiency, as a function of the underlying y*. This can be written as PROF = 1 if 0 ≤
y* ≤ τ1, and PROF = 2 if τ 1 < y* ≤ τ 2, and PROF=3 if τ 2 < y* where τ’s are called thresholds or
cutpoints. The use of the ordinal logit model also assumes that the error term ε has a logistic
2

distribution with a mean of 0 and a variance of π /3 (Long Scott, 1997).
The results from ordinal logit regressions are presented as odds ratio, where an odds ratio
of greater than 1 associated with PRIVATE (or PRIVATE_ONLY) indicates greater odds of
being ranked in the higher proficiency level nationally for a child attending private school versus
82

the reference category. In the OLS scenario a positive coefficient associated with PRIVATE (or
PRIVATE_ONLY) indicates the units by which the given dependent variable (SCORE, READ,
WRITE, MATH) will increase for a student attending private school versus the reference
category.
Propensity score matching
The problem with regression analysis is that the estimates of βPVT are only as good as the
controls included in the model. And while the results are sufficient to establish associations, they
are not sufficient to generate causal claims. This is because children are not allocated to schools
randomly. The potential family factors associated with sending a child to private school (such as
greater access to resources, greater involvement in the child’s educational experiences, greater
willingness and ability to invest in the child’s education) may also be associated with the child’s
achievement. In the standard regression framework therefore a positive βPVT coefficient may at
least in part be reflecting not the actual effect of attending private schools, but perhaps the effect
of greater parental attention and involvement.
More precisely, let PRIVATE=1 indicate attending private school (treatment) and
PRIVATE =0 indicate attending public school (control). Let Y once again indicate the student
outcome on the test. To estimate the effect of private school on outcomes, ideally, we would like
to see Yi1 – Yi0 for each student i. But we cannot observe both treatment and control test scores
on the same student. We can only see either Yi1 or Yi0 for each child, not both. As a result, we
cannot estimate Y1 – Y0 for our sample. This problem could be addressed if it was possible to
allocate children randomly to different school types. But this is also not feasible.

83

However, with retrospective data we can estimate the average treatment on children in
the treatment (private school) group, (ATT), τATT = E(Y1 – Y0 | X, PRIVATE =1) where X
denotes a set of observed covariates used to calculate propensity score (Smith & Todd, 2001).
With this approach we use an extensive set of covariates X and a standard regression model for
dichotomous outcome to calculate each student’s probability of enrolling in private school
(propensity score). Next we compare students who have the same or similar propensity of being
enrolled in a private school but a group that was enrolled in private schools (treatment) and
group that was not (control). This approach, of comparing students based on their propensity
score rather than matching students on all the confounding covariates, reduces the dimensionality
of the problem (Rubin, 1997).
The use of the propensity score method is predicated on the assumption of ‘strong
ignorability’ (Rosenbaum & Rubin, 1983b). This assumption states that:
(Y0) ⊥ PRIVATE | X

(3)

In other words, conditional on the observed covariates, the unobserved potential outcomes are
independent of treatment assignment. This assumption states that the observed covariates (X)
contain sufficient information for the counterfactual outcome Y0 to be independent of the
treatment. The IHDS dataset are uniquely suited for this purpose since they allow us to observe a
number of relevant covariates that describe the child, their home and family, and their learning
environment including covariates that may be relevant to modeling private school choice.
A second important assumption is overlap or common support. This assumption states
that everyone in the defined population has some chance of being treated or not treated based on
their observed characteristics. Formally:

84

0 < P(PRIVATE=1 | X) <1 for all X

(4)

In the propensity score framework, we instead work with the probability of being assigned to the
treatment, as calculated by a standard model for binary outcomes. This means that the fitted
propensity score values for the treated and untreated groups must overlap.
In addition, Heckman, Ichimura, and Todd (1997) identify several features of such
evaluation studies which if attained will further reduce bias. For both treatment and control
groups these include (1) same distribution on unobserved attributes (2) same distribution on the
observed attributes (3) using the same instrument for data collection; and (4) ensuring that both
groups were placed in the same economic environment. They further suggest that fulfilling
conditions (2) through (4) is relatively more important than fulfilling condition (1).
Our study fulfills criteria (2) and (3). We cannot make claims about the identical
distribution of unobserved attributes. In fact this is one of the key limitations of the propensity
score matching approach. Like several standard statistical techniques propensity score matching
relies on variables that are observed. Matching between treatment and control groups on these
observed variables no matter how perfect is still unable to comment on the matching between the
two groups on unobserved variables. Similarly, we are unable to ensure the same environment
for both groups since states in India vary considerably in their rates of private school
6

enrollment . However, by running the analysis separately for rural and urban India we made at
least a partial attempt to compare treatment and control in not altogether dissimilar
environments.
We conducted this analysis separately for rural and urban data and for the two definitions
of private schools. To calculate the propensity scores we used probit regressions for binary
6

These limitations are not uncommon; a recent paper by Doyle (2009) using propensity score
matching analysis in a different context identifies the same two limitations of his study.
85

outcomes, to predict the probability of a child enrolling in private school given their background.
We follow Ruben and Thomas (1996) and include an extensive set of relevant covariates.
Specifically, this regression included all the controls except ‘age’ and ‘years in school’ that we
mentioned earlier in the regression framework. In addition the probit estimation included three
other sets of attributes. These were, first, attributes related to the child: the number of days the
child was absent from school, and whether the child receives support from a government agency
regardless of school type, hours spent on private tuition. The second group of attributes is
related to the family and the importance it places on educating the child: the highest education
level among adults over age 21, whether the household is below the poverty line, whether the
household head can speak English, and the number of children in the household. The third group
of attributes related to the child’s school experience; that are not due to the school choice but
factors that may influence parental decisions to enroll their children in one type of school versus
another. For example, how far is the school from the child’s home, how early does the school
start teaching English, and does the school provide meals? (The second factor was used by Desai
et al. (2008) and the third by Goyal (2009) as controls in their respective studies). A detailed list
of these variables, their coefficients in the probit model, and associated descriptive statistics for
the control and treatment groups are available upon request.
After fitting the probit regressions, we used the psmatch2 command in Stata to match
observations. We employed nearest neighbor matching without replacement with caliper to
match the data for rural and urban data and for PRIVATE and PRIVATE_ONLY (Guo & Fraser,
2010). We checked for balance in the matched sample using the pstest command at 5% level of
7

significance and we also ensured that the p-value associated with the likelihood-ratio test of the

7

One variable in rural PRIVATE_ONLY balanced at 10% level of significance.
86

joint insignificance of all the regressors after matching was greater than 0.10. Following Guo
and Fraser (2010) we began with checking for balance with a caliper size of 0.25 σPS (where σPS
is the standard deviation of the estimated propensity score) and tried different caliper sizes to
attain balance while maintaining the largest sample possible. The final caliper size for the rural
data were 0.17 σPS and 0.20 σPS and for urban data 0.09 σPS and 0.10 σPS for PRIVATE and
PRIVATE_ONLY respectively. The literature is not clear on the value of using sample weights
while generating a propensity score analysis. We did not use sample weights at this stage since
propensity score analysis is used to match treatment and control groups within the sample
(Zanutto, 2006). Intuitively the matched subsample generated in this manner identifies
comparable children (based on the available covariates) where one group attended private nonaided school and the other attended the reference category schools (depending on whether we
used PRIVATE or PRIVATE_ONLY variable). Comparing these ‘similar’ children who
received different types of schooling provides one way to address the non-random allocation of
children to different types of schools. As may be anticipated, this matched sample is a smaller
subset of the complete dataset (we discuss this reduction in sample size in the results section
below).
To arrive at the estimate or the treatment effect, after the matching, we conducted
regression analysis including ordered logit regression analysis for outcome PROF and OLS for
outcomes SCORE, READ, WRITE and MATH on the matched sample. For these analyses we
controlled for age and year in school. These regression analyses used the sample weights in the
dataset as we were now interested in generating population-level estimates of the treatment effect
(Zanutto, 2006).

87

Results
Table 3.1 presents descriptive data for rural and urban India. The first panel compares
children’s performance on each of the five dependent variables separately for each of the school
groups (public schools including private aided schools, public schools excluding private aided
schools, private schools, low fee private schools, high fee private schools). The second panel
provides a similar comparison on school fee and a selected set of variables used for matching. As
the first panel shows, comparing raw average performance of public-private children on five
different outcome variables across rural and urban areas shows a clear private school advantage.
Private school children significantly outperform their public school counterparts on all the
outcomes consistently. Interestingly, some of the private school heterogeneity is already visible
in this data. When we group children in to low fee private and high fee private school, we find
that the low fee private school children perform less well compared to the high fee private school
children. The second panel reveals that these private school attending children may belong to
systematically different household than their public school counterparts. The descriptive data
show the relatively privileged position of children who attend private schools. These differences
in background and school fees are especially stark when we compare high fee private school
attending children to public school attending children. Their families tend to be well off in terms
of both income and assets and their households tend to be better educated. The Table also reveals
that, overall, children in urban areas tend to belong to better-off families regardless of the school
type, compared to rural children. And in many instances the home background of a child
attending a public school in an urban area is comparable to that of a child attending a private
school in a rural area. This observation underscores the importance of conducting separate rural,

88

Table 3.1 Means for the full sample, for dependent variables, propensity scores and selected matching variables, by rural, urban and
b
by different school groups
Rural
Urban
Public
Public
Public
Public
Low
High
Low
High
(includin (excludin
(includin (excludin
g pvt.
g pvt.
fee
fee
g pvt.
fee
fee
g pvt.
aided)
aided)
Private private private
aided)
aided)
Private private private
Dependent Variables
PROF
1.76
1.75
2.17
2.04
2.24
2.03
2.00
2.32
2.22
2.37
SCORE
4.44
4.41
5.66
5.28
5.86
5.26
5.18
6.05
5.79
6.17
READ
2.42
2.41
2.98
2.81
3.07
2.78
2.76
3.16
3.09
3.20
WRITE
0.63
0.63
0.79
0.74
0.81
0.74
0.74
0.84
0.79
0.86
MATH
1.39
1.37
1.89
1.73
1.98
1.73
1.68
2.05
1.91
2.12
Selected matching variables (and Propensity Score)
a
School fee
481
466
2152
1348
2587
816
742
3564
2649
4023
c

Household income
Household assets
c
Household head's educ
Highest level of educ.
c
for female age > 21
a

Propensity Score
N

35,486
8.84

35,231
8.75

70,793 56,127 78,693
13.47 12.86 13.80

3.86

3.80

6.08

5.36

2.51

2.43

4.94

4.35

0.08
5748

0.07
5574

0.67
1374

481

a

54,876
13.47

52,969
13.27

86,843 78,071 91,220
17.88 17.28 18.17

6.47

5.92

5.80

8.76

8.04

9.12

5.26

4.70

4.51

7.47

6.77

7.83

893

0.25
1458

0.25
1285

0.77
1577

525

1052

The sample sizes for the propensity score and school fee are slightly reduced (127 and 28 less observations respectively in the rural
data, and 74 and 15 less observations respectively in the urban data).
b
All public, private differences statistically significant with p<=0.05. (This includes 12 comparisons; public (including private aided)
vs. private (PRIVATE), low fee private, high fee private, and public (excluding private aided) vs. private (PRIVATE_ONLY), low
fee private, and high fee private for both rural and urban data).
c
Income is measured in Indian Rupees and Education levels are measured in years.

89

urban analysis, especially when trying to generate ‘matched’ data. The Table also highlights the
need to make appropriate corrections for selection in order to discern the true private effect.
Regression analysis: Ordinary Least Square and Ordinal Logit models
Table 3.2 presents the results of the regression results for all the five outcomes for rural
and urban data separately by the two definitions of the private school variable. We conducted
ordinal logit regression analysis for dependent variable PROF and OLS for the other four
dependent variables including SCORE, READ, WRITE and MATH. It is important to remind the
readers that the coefficients across two different regression methods are not directly comparable.
In each panel, columns (1)-(2) correspond to equations (1)-(2). Focusing first on column (1), the
results show that, in agreement with earlier studies, there is indeed a positive association between
attending private school and having a higher test performance. This positive private effect after
controlling for household covariate is somewhat smaller than the raw differences noted in Table
3.1. Also, this effect is consistently positive and significant across all the outcome variables, two
separate definitions of private schools and across rural and urban India. It is worth noting that
across both dependent variables and independent variables specifications, in absolute terms this
association between the school’s status as private and the child’s test performance is higher in
rural areas than in urban ones.
Column (2) adds an additional nuance to this ‘private effect.’ As we mentioned earlier,
we do not believe that our classification of low- and high-fee private schools is by any means
perfect; it is merely an attempt to identify what parents might consider ‘cheap’ private schools
within their area, given the public school costs. But based on this imprecise definition of
‘low/high fee private’ we find a few patterns worthy of note. First, in both rural and urban
settings, we find a strong positive association between a child attending a ‘high fee’ private

90

Table 3.2 Maximum Likelihood Estimates of a set of Ordinal Logistic Regressions (PROF) and Ordinary Least Square Regressions
(SCORE, READ, WRITE, MATH) on full sample, by Rural, Urban and by Private school definitions
Public (excluding pvt. Aided) vs. Private
Public (including pvt. Aided) vs. Private [PRIVATE]
[PRIVATE_ONLY]
Dependent
variable
Rural
Urban
Rural
Urban
(1)
(2)
(1)
(2)
(1)
(2)
(1)
(2)
PROF
Private
2.04
1.38
2.02
1.41
[0.19]***
[0.11]***
[0.20]***
[0.09]***
High fee private
2.66
1.48
2.64
1.51
[0.24]***
[0.10]***
[0.25]***
[0.10]***
Low fee private
1.27
1.22
1.27
1.25
[0.22]
[0.17]
[0.23]
[0.16]*
SCORE
Private
0.73
0.30
0.73
0.34
[0.09]***
[0.07]***
[0.10]***
[0.07]***
High fee private
0.96
0.35
0.96
0.38
[0.08]***
[0.06]***
[0.08]***
[0.06]***
Low fee private
0.33
0.22
0.33
0.26
[0.09]***
[0.13]*
[0.09]***
[0.13]*
READ
Private
0.37
0.16
0.37
0.16
[0.06]***
[0.04]***
[0.06]***
[0.04]***
High fee private
0.47
0.15
0.47
0.15
[0.05]***
[0.04]***
[0.05]***
[0.04]***
Low fee private
0.19
0.18
0.20
0.18
[0.06]***
[0.08]**
[0.06]***
[0.07]**
WRITE
Private
0.08
0.05
0.08
0.04
[0.02]***
[0.02]***
[0.02]***
[0.02]***
High fee private
0.12
0.06
0.12
0.05
[0.01]***
[0.02]***
[0.01]***
[0.02]***
Low fee private
0
0.03
0
0.02
[0.05]
[0.02]
[0.05]
[0.02]
91

Table 3.2 (cont’d)
MATH
Private

0.28
[0.04]***
High fee private

0.09
[0.04]***

0.28
[0.04]***

0.14
[0.04]***

0.36
0.14
0.36
[0.03]***
[0.04]***
[0.03]***
Low fee private
0.14
0.01
0.14
[0.07]*
[0.07]
[0.08]*
N
7122
7122
3035
3035
6948
6948
2862
Note. Standard errors in brackets. * 0.10 significance level. ** 0.05 significance level. *** 0.01 significance level.

0.18
[0.04]***
0.05
[0.06]
2862

school and their test performance. This association is larger in magnitude than simply the ‘private’ effect observed in column (1).
More interesting, however, is the lack of significance of low-fee private schools at 5% level of significance in the rural and urban data
for the ordinal logit model with dependent variable PROF and for dependent variables WRITING and MATH using OLS, regardless
of the public school comparison group used. When we use the OLS approach with dependent variable SCORE, we find that the lowfee private schools in the rural area have a positive significant albeit smaller coefficient associated with them. In the urban area the
low-fee private effect attains significance only at 10% level of significance for both comparisons. Finally for outcome READ both
low-fee and high-fee private school effect is positive and significant for both rural and urban data at 5% level of significance. It is
likely that the overall of significance of SCORE may in part be explained by the positive findings for READ, since READ ranges
from 0-4 and is the single largest component of SCORE.
At the very least, the data seem to indicate that children may not do well simply because they are enrolled in a ‘private’ school,
and especially in urban areas there may be ‘private’ schools identified by some threshold of costs associated with private schooling
92

within a given area, region, or grade level that in fact fail to significantly improve student
achievement outcomes simply because they are ‘private.’ Most important, perhaps, these results
highlight the importance of considering the heterogeneity in private schools rather than treating
them as a single, homogenous type of schooling experience.
Propensity score matching
Table 3.3 presents the results of matching the data using the propensity score technique
separately for two definitions of private schools. After we conducted the matching, the rural and
urban sample reduced as expected. (The rural sample is 1056 and 982 after matching on
PRIVATE and PRIVATE_ONLY and urban sample is now 912 and 794 respectively for the two
definitions of private schools. In other words, the matched rural sample contains 14-15% of the
total rural sample and the matched urban contains 25-30% of the total urban sample). As Table
3.3 indicates, after the matching, in both rural and urban samples the mean observations match
on several of the key home background variables. Figures included in the published version of
the paper, show the distribution of the propensity score and some key variables before and after
matching, for rural and urban data for the two separate definitions of private schools. Once again
the figures reveal that after matching the matched data are almost identically distributed on
several key attributes.
The Table also presents the mean score comparisons on the five outcomes on the matched
data (ATT) before we conduct the regression analysis on the matched data. Both for rural and
urban data and for both definitions of private schools we now find no statistically significant
difference in the performance of children in public and private schools post-matching. In fact, in
a few separate cases in the rural and the urban data we find a positive public school effect at 10%

93

level of significance (tcritical = |1.64|) and even at 5% level of significance in one case (tcritical =
|1.96|).

Table 3.3 Differences in means for the matched sample, for dependent variables, propensity
scores and selected matching variables, by Rural, Urban and by Private school definitions
Rural
Urban
Public
Private
t-test
Public
Private t-test
PRIVATE (Public including pvt. aided vs. Private)
Dependent Variables
PROF
2.14
2.10
-0.64
2.26
2.18
-1.35
SCORE
5.59
5.44
-1.00
5.88
5.71
-1.12
READ
2.94
2.90
-0.47
3.05
3.01
-0.56
WRITE
0.78
0.77
-0.44
0.81
0.81
0.00
MATH
1.87
1.77
-1.55
2.01
1.89
-1.85
Propensity Score and selected matching variables
Propensity score
0.47
0.50
-1.61
0.58
0.60
-1.18
Household income
52,579
54,643
-0.46
71,864
76,391 -0.90
Household assets
12.02
11.69
1.01
16.07
16.71
-1.87
Household head's education
5.43
5.37
0.21
7.59
7.99
-1.20
Highest level of education for
female age>21
4.33
4.15
0.64
6.51
7.10
-1.74
N
528
528
456
456
PRIVATE_ONLY (Public, excluding pvt. aided vs. Private)
Dependent Variables
PROF
2.19
2.09
-1.84
2.22
2.12
SCORE
5.65
5.42
-1.52
5.85
5.60
READ
2.99
2.90
-1.07
3.06
2.95
WRITE
0.79
0.74
-1.65
0.81
0.81
MATH
1.88
1.78
-1.55
1.98
1.84
Propensity Score and selected matching variables
Propensity score
0.47
0.50
-1.84
0.58
0.60
Household income
52,378
55,228
-0.64
70,264
71,929
Household assets
12.09
11.96
0.38
15.92
16.45
Household head's education
5.19
5.49
-0.99
7.23
7.86
Highest level of education for
female age>21
4.19
4.21
-0.07
6.24
6.88
N
491
491
397
397
Note. Income is measured in Indian Rupees and Education levels are measured in years.

94

-1.70
-1.60
-1.31
0.00
-2.03
-1.09
-0.33
-1.46
-1.76
-1.76

While the simple comparison of mean on the matched data presented in Table 3.3 already
indicates the loss of significance associated with private schools in our data, we conclude with a
final set of regression analysis. Table 3.4 presents the results of the ordinary least square and
ordinal logit analysis on the matched data for the five outcome variables controlling for child’s
age and years of completed schooling, as presumably these attributes will have their own
independent influence on how well a child does on the tests. After the matching we found that
the years of education the child had received and in certain cases their age was positively
associated with the child’s performance levels. However, in both the rural and urban data,
regardless of the definition of the independent variable used (PRIVATE or PRIVATE_ONLY)
we find no statistically significant private affect for dependent variables SCORE, WRITE and
MATH at 5% level of significance. In the rural data when using PRIVATE as the definition of
the key independent variable we find a positive and significant effect associated with private
schooling at 5% level of significance for PROF and READ but not at a more stringent 1% level
of significance. Thus evidence from matched data presented in Tables 3 and 4 provide
insufficient evidence to make an unequivocal claim about the superiority of private schools.

95

Table 3.4 Maximum Likelihood Estimates of a set of Ordinal Logistic Regressions (PROF) and
Ordinary Least Square Regressions (SCORE, READ, WRITE, MATH) on matched sample, by
Rural, Urban and by Private school definitions
Public (excluding pvt.
Public (including pvt.
Aided) vs. Private
Aided) vs. Private
[PRIVATE_ONLY]
[PRIVATE]
Dependent variable
PROF
Private

Rural
Urban
Rural
Urban
1.5
1.00
1.23
0.96
[0.29]**
[0.14]
[0.23]
[0.15]
Years of education completed
1.74
1.57
1.92
1.72
[0.13]***
[0.09]***
[0.17]***
[0.12]***
Age
1.11
1.20
1.07
1.06
[0.11]
[0.10]**
[0.11]
[0.09]
SCORE Private
0.45
0.10
0.27
-0.02
[0.24]*
[0.14]
[0.22]
[0.15]
Years of education completed
0.61
0.57
0.75
0.64
[0.07]***
[0.06]***
[0.08]***
[0.07]***
Age
0.19
0.15
0.11
0.06
[0.10]*
[0.10]
[0.12]
[0.09]
Reading Private
0.31
0.06
0.22
0
[0.13]**
[0.08]
[0.12]**
[0.08]
Years of education completed
0.33
0.27
0.39
0.31
[0.04]***
[0.04]***
[0.04]***
[0.04]***
Age
0.05
0.06
0.05
0.02
[0.06]
[0.06]
[0.06]
[0.05]
Writing Private
0.04
0.05
0.06
0.04
[0.04]
[0.03]*
[0.04]
[0.03]
Years of education completed
0.07
0.07
0.09
0.07
[0.01]***
[0.01]***
[0.01]***
[0.01]***
Age
0.04
0
0.01
-0.01
[0.03]*
[0.02]
[0.02]
[0.02]
Math
Private
0.09
-0.02
0.04
-0.06
[0.09]
[0.06]
[0.09]
[0.07]
Years of education completed
0.22
0.24
0.26
0.25
[0.03]***
[0.03]***
[0.03]***
[0.03]***
Age
0.10
0.09
0.05
0.05
[0.04]**
[0.04]**
[0.04]
[0.04]
N
1056
912
982
794
Note. Standard errors in brackets. * 0.10 significance level. ** 0.05 significance level. *** 0.01
significance level.

96

Limitations and conclusion
Our study makes an important contribution to the limited body of empirical knowledge
on private school performance in India. We used recent, nationally representative data from rural
and urban schools and investigated the relative performance of private versus public schools.
Capitalizing on a rich set of home background covariates, we used the propensity score matching
technique to balance the data and arrive at an estimate of the ‘private effect’ for the rural and
urban data.
Before we discuss the main findings from our study we must acknowledge its limitations.
The key outcome variables (SCORE and PROF) used in this study cannot claim to have several
of the desirable psychometric properties that may be available in large scale test-score data sets
such as the Trends in International Math and Science Studies (TIMSS) or PISA. On the other
hand, India unfortunately has not participated in any such study in recent years, and in general
very few sources of nationally representative student performance data from India are publicly
available along with the set of household covariates that would be necessary for such an analysis.
As an attempt to avoid overreliance on these dependent variables, we presented all our analysis
using the three separate test-score variables which suffer from their own limitations.
Second, we must remind our readers that while propensity score matching is an attempt
to minimize the selection bias issue by comparing children who are alike on several attributes,
the estimates generated from this analysis are still only as good as the covariates available to us
and this analysis cannot make claims about matching children or families on their unobserved
traits. Also, the reader should note that creating matched data inherently reduces sample size
available to the researcher. We cannot rule out the possibility that with a larger nationally
representative sample of children beyond this sample that was collected from 41,554 households

97

and a larger matched sample our findings may look different. In this regard too, while we
acknowledge the limitations of our data we must note that to our knowledge this is the only
recent, nationally representative, dataset available from India that provides both achievement
data and extensive home background data.
Moreover, we are not able to comment on the cost-effectiveness of private schools in this
study. In other words, perhaps children at private schools do not perform differently than those in
public schools; however, as Tooley et al. (2010) and others argue, these schools simply produce
the same achievement levels more cheaply. That scenario is possible, but our data do not allow
us to comment on the cost of producing these achievement outcomes from the school’s
perspective.
Finally, we also acknowledge that without the benefit of longitudinal data, it is always
possible to argue that certain aspects of the home environment may be affected by the child’s
school type rather than the other way around; in that case, such variables will not form suitable
controls for our analysis. We have paid careful attention to this potential pitfall of propensity
score matching in which the researcher may include a whole host of covariates, sometimes
without proper justification. Instead, we have tried to ensure that all the covariates we included
in the analysis are supported by the existing literature and/or are appropriate based on the
language used in the questionnaire, the distribution of the variables, etc. For instance, we
excluded from the analysis variables that related to child’s perception of their teacher, or the time
they spent on school work at home. While it is likely that these variables simply reflect the child
and/or their family’s disposition/commitment towards education, it is equally likely that these are
post-treatment variables.

98

Having noted these limitations, our analysis reveals some patterns that we believe make a
valuable contribution to the ongoing conversation on the role of private schooling in India. The
regression results on the full sample show the importance of accounting for even simple child
and home covariates. These results also highlight the importance of conducting separate analyses
for rural and urban India where possible, as the results from these two areas are often divergent.
Overall the regression results on the full sample agreed with the existing literature that has found
a positive association between attending private schools and having better achievement
outcomes. However, when we distinguished private schools by their fee levels (using an
imperfect measure of ‘low/high’ fee private schools) we found that children in such schools may
not always perform better than those in public schools. As noted earlier, this finding says nothing
about the cost-effectiveness of these so-called ‘low-fee’ private schools, but at the very least this
finding highlights the importance of recognizing that the term ‘private school’ in the Indian
context ought not to be treated as a homogenous unit. Rather, possible variations in the quality of
private schools may be associated with student achievement levels. Not surprisingly, these
results show that children in private schools, whose parents can afford higher fees, do end up
doing significantly better than children attending public schools.
After we matched students on a series of covariates using propensity score matching
technique, the regression analysis revealed a different result. We found that in both the rural and
urban data at 5% level of significance the coefficients associated with private schooling was no
longer statistically significant in 18 out of 20 cases we analyzed. We found no negative effect
associated with attending private schools; however, unlike earlier researchers we failed to find
that private school attendance is associated with any systematic and specific benefit in terms of
increased achievement. Overall, the regression results on the matched data failed to produce a

99

clear result that may indicate that children in private schools may be outperforming those in
public school in our dataset.
Discussion
After the impressive strides India has taken to improve school enrollment, there is now
deep interest in understanding the factors associated with improved achievement levels. A
simultaneous and growing dissatisfaction with the public school system has led many researchers
and policy advocates to argue that increased privatization and greater reliance on private schools
may be important to improve school performance in India. The proponents often argue that
privatization in any shape or form (even in terms of low fee/low cost private schools) may be
preferable to public provision of education. While the arguments on either side about the roles
and limitations of public schools are impassioned, we have limited empirical evidence to inform
this debate. Not surprisingly, children from families of higher socio-economic status are more
likely to enroll in private schools; to our knowledge, however, few studies have attempted to
correct for this non-random selection issue in the Indian case. The limited evidence on this
important issue is concerning from the Indian perspective and also from the perspective of other
smaller developing countries facing similar pressures yet even less systematic research evidence.
Our study using data, on children aged eight to eleven from a representative sample of
rural and urban households in India makes a contribution to this limited empirical literature. In
disagreement with some of the other recent studies from India, we find insufficient evidence to
claim that children in private schools outperform those in public schools in India. As we discuss
clearly in our limitations section better data are needed and there is certainly room to improve
analytical approaches beyond those we used in this study. However, our findings generated using
appropriate and sophisticated analyses are important because they call into question the

100

consistency of ‘positive’ private effect. Policymakers in the developed world have for decades
benefited from nuanced, extensive and multi-faceted conversation on the implications of
privatization in education (for instance, Witte (1992), Goldhaber (1996), Rouse (1998), Bettinger
(2005), Cohen-Zada (2009), Zimmer, Gill, Booker, Lavertu, Witte (in press)). In the same
manner, we hope that this study and other similar studies lead to a more robust conversation on
the benefits and limitations of privatization in the developing world.

101

REFERENCES

102

REFERENCES

Bettinger, E. (2005). The effect of charter schools on charter students and public schools.
Economics of Education Review, 24, 133-147.
Cohen-Zada, D. (2009). An alternative instrument for private school competition. Economics of
Education Review, 28, 29-37.
Desai, Sonalde, Amaresh Dubey, B.L. Joshi, Mitali Sen, Abusaleh Shariff, and Reeve
Vanneman. India Human Development Survey (IHDS) [Computer file]. ICPSR22626-v2.
University of Maryland and National Council of Applied Economic Research, New Delhi
[producers], 2007. Ann Arbor, MI: Inter-university Consortium for Political and Social Research
[distributor], 30 June 2009.
Desai, S., Dubey, A., Vanneman, R., & Banerji, R. (2008). Private schooling in India: A new
educational landscape. Brookings-NCAER India Policy Forum.
Doyle, W.R. (2009). The effect of community college enrollment on bachelor’s degree
completion. Economics of Education Review 28, 199–206.
French R., & Kingdon, G. (2010). The relative effectiveness of private and government schools
in Rural India: Evidence from ASER data. DoQSS Working Paper No. 10-03, June 2010.
London: University of London Institute of Education.
Goldhaber, D. (1996). Public and private high schools: Is school choice an answer to the
productivity problem? Economics of Education Review, 15, 93-109.
Government of India. (2009). Annual report 2009-2010. New Delhi: Department of Elementary
Education and Literacy, Department of Secondary and Higher Education, Ministry of Human
Resource Development.
Goyal, S. (2009). ‘Inside the house of learning’: The relative performance of public and private
schools in Orissa. Education Economics, 17(3), 315-327.
Goyal S., & Pandey, P. (2009). How do government and private schools differ? Findings from
two large Indian states. Report No. 30, South Asia Human Development Sector. Washington,
DC: The World Bank.
Guo, S., & Fraser, M. W. (2010). Propensity score analysis: Statistical methods and
applications. Thousand Oaks, CA: Sage Publications.
Heckman, J. J., Ichimura, H., & Todd, P. E. (1997). Matching as an econometric evaluation
estimator: Evidence from evaluating a job training program. The Review of Economic Studies,
64(4), 605–654.
103

Imai, Kosuke. (2005). “Do Get-Out-The-Vote Calls Reduce Turnout? The Importance of
Statistical Methods for Field Experiments” American Political Science Review, Vol. 99, No. 2
(May), pp. 283-300.
Kingdon G. (2007). The progress of school education in India. Oxford Review of Economic
Policy, 23(2), 168–195.
Kingdon, G. (1996). The quality and efficiency of public and private education: A case study of
Urban India. Oxford Bulletin of Economics and Statistics 58(1) 57–82.
Long Scott, J. (1997). Regression models for categorical and limited dependent variables.
Thousand Oaks, CA: Sage Publications.
Mehta A. C. (2011). Elementary Education in India: Analytical Report 2008-2009. New Delhi:
National University of Educational Planning and Administration and Department of School
Education and Literacy Ministry of Human Resource Development Government of India
Muralidharan, K., and M. Kremer. (2006). Public and private schools in rural India. Unpublished
report. Cambridge, MA: Massachusetts Institute of Technology.
Rosenbaum, P. R., & Rubin, D. B. (1983a). The central role of the propensity score in
observational studies for causal effects. Biometrika, 70(1), 41–55.
Rosenbaum, P. R., & Rubin, D. B. (1983b). Assessing sensitivity to an unobserved binary
covariate in an observational study with binary outcome. Journal of the Royal Statistical Society
Series B, 45(2), 212–218.
Rouse, C. (1998). Private School Vouchers and Student Achievement: An Evaluation of the
Milwaukee Parental Choice Program. Quarterly Journal of Economics, 113, 553-602.
Rubin, D. B. (1997). Estimating causal effects from large data sets using propensity scores.
Annals of Internal Medicine, Part 2, 127, 757-763.
Rubin DB, Thomas N. Matching using estimated propensity scores, relating theory to practice.
Biometrics 1996; 52:249–264.
Smith, J. A., & Todd, P. E. (2001). Reconciling conflicting evidence on the performance of
propensity score methods. American Economic Review, 91(2), 112–118.
Stolzenberg, R. M., & Relles, D.A. (1997). Tools for intuition about sample selection
bias and its correction. American Sociological Review, 62(3), 494-507.
The Gazette of India. (2009). The Right of Children to Free and Compulsory Education Act,
2009. http://education.nic.in/Elementary/free%20and%20compulsory.pdf

104

Tooley, J., & Dixon, P. (2003). Private schools in India: A case study from India. UK: CfBT
Research and Development.
Tooley J., Dixon, P, Shamsan, Y., & Schagenb, I. (2010). The relative quality and costeffectiveness of private and public schools for low-income families: A case study in a developing
country. School Effectiveness and School Improvement, 21 (2), 117–144.
Wadhwa W. (2009). Are private schools really performing better than government schools? In
Annual Status of Education Report (Rural) New Delhi: Pratham.
Witte, J. (1992). Private school versus public school achievement: are there findings that should
affect the educational choice debate? Economics of Education Review, 11, 371-394.
Zanutto, E. L. (2006). A comparison of propensity score and linear regression analysis of
complex survey data. Journal of Data Science, 4, 67-91.
Zimmer, R., Gill, B., Booker, K., Lavertu, S., and Witte, J. Examining Charter Student
Achievement Across Seven States. Economics of Education Review. Article in Press.

105