THREE ESSAYS ON ENVIRONMENTAL QUALITY WITH POLLUTING

SECTORS: MINING, ELECTRICITY, AND TRANSPORTATION

By

Nathaly Macarena Rivera Casanoba

A DISSERTATION

Submitted to

Michigan State University

in partial fulfillment of the requirements

for the degree of

Agricultural, Food, and Resource Economics—Doctor of Philosophy

2018

ABSTRACT

THREE ESSAYS ON ENVIRONMENTAL QUALITY WITH POLLUTING SECTORS:

MINING, ELECTRICITY, AND TRANSPORTATION

By

Nathaly Macarena Rivera Casanoba

Mining, electricity, and transportation are currently among the top the economic sectors

contributing to environmental degradation around the world. This dissertation delves into

recurring environmental problems associated to these polluting sectors through the appli-

cation of several quasi-experimental methods and environmental valuation techniques. The

first two essays elicit the value of proximity to environmental disamenities, namely, resource

extraction sites and fossil fuel power stations. The third essay, evaluates the effectiveness

of a command-and-control policy aimed at curbing pollution from automotive sources. To-

gether, these three essays address relevant topics in non-market valuation and environmental

regulation, and in settings that are pertinent to both developed and developing countries.

Essay one addresses whether resource extraction sites, mainly copper mines, represent an

environmental disamenity to households. The opening of new extraction sites brings several

economic benefits to the localities hosting the resource extraction. Yet, these sites generally

impose detrimental effects on local environmental quality, in which case households may

be willing to pay to avoid home locations near these sites. In this essay, I study whether

the high-pollution potential of mining outweighs its local economic benefits in emerging

economies using evidence from the housing market in Chile. I compare rental prices in lo-

cations hosting new openings to rental prices in places without mining, before and after the

openings. Results indicate that resource extraction sites constitute a disamenity to house-

holds in emerging economies, who are willing to pay to avoid proximity to these facilities.

These effects are higher among long-term residents, which represents evidence of taste-based

sorting of households across space. The pollution potential of some of these economic sectors

can be mitigated with the adoption of cleaner production technologies. Essay two delves into

this idea by studying the market impacts of the coal-to-gas conversion process taking place

in the United States during recent years. Given that natural gas burns cleaner than coal,

switching the primary fuel source used in the electricity generation process should improve

local air quality in the neighborhood of converting power plants. In this essay, I examine

whether these local air quality improvements due to fuel switching are capitalized into higher

housing prices of nearby homes. I use evidence of property transactions of several homes

located at certain distances from both switching and coal-fired power plants, and I compare

their sale prices before and after the conversion. Results show that this conversion is capi-

talized into higher housing prices in the vicinity of a unit that switched fuel. These market

capitalizations follow immediately after the closure of coal-fired electricity generators, and

not necessarily after the gas-fired units start operations.

Overall, the first two essays in this dissertation provide empirical evidence on how costly

environmental pollution may be to the public. These costs are translated into dead-weight

losses to society whenever they are not internalized by polluting units when interacting

in the market. Environmental regulation, however, can offset these inefficiencies serving

as an instrument to restore social equilibria. Essay three examines the efficiency of an

environmental regulation mechanism used in Chile to reduced airborne pollution from the

transportation sector by restricting driving of light-duty private vehicles. Driving restrictions

constitute the main policy to regulate emissions from mobile sources in Chile. These bans

are intensified during days of critical air pollution with the official issuance of 24-hour air

quality episodes. I evaluate the effectiveness of this quantity-based instrument in curbing

mobile source pollution using evidence from several pollutants from car emissions. I also

look at the effects of these restrictions on car trips, and on the use of alternative modes of

transportation. Results indicate that this mechanism effectively curtails the number of cars

on the roads, thereby improving local air quality. Evidence suggests that the effectiveness of

this mechanism, however, is conditional to their temporal duration, and to their issuance in

conjunction with air quality episodes informing on the risks of outdoor exposure to pollution.

To my parents, Osvaldo Rivera and Margarita Casanoba.

iv

ACKNOWLEDGEMENTS

First and foremost, I would like to thank my major professor, Dr. Scott Loveridge, for

his support, dedication, leadership, and patience throughout my years in graduate school. I

am also especially grateful for the mentorship and thorough feedback of Dr. Joseph Herriges,

who provided invaluable guidance during the last years of my program. Furthermore, thank

you to my committee members, Dr. Satish Joshi and Dr. Frank Lupi, for their advice and

suggestions that improved my work and research skills. I also thank Dr. Jeffrey Wooldridge

for his direction, comments, and humor. I thank the entire Agricultural, Food, and Resource

Economics (AFRE) faculty and staff for all the support during my time in the program.

Second, I would like to thank all my AFRE and ECON colleagues and friends; to my

study group -Felipe Bardella, Miguel Castro, Adrian Garlati, and Almudena Luna- for the

endless hours of prelims preparation. Thank you to my Environmental Economics classmates,

Miguel Castro and Jungmin Lim, for making the field examination manageable; to the entire

Environmental Economics Reading Group- the field was more interesting and enjoyable after

brainstorming ideas and sharing perspectives with all of you. I also want to thank the many

friends that I made during my studies at MSU, and my entire Ph.D cohort.

Lastly, I wish to thank my family for their love, encouragement and support; to my

parents, Osvaldo and Margarita; to my sisters Leslye and Claudia, and my brother-in-law

Stewart Mason. Their love and encouragement are immeasurable. Sunday’s video calls with

my family were crucial in reminding me to keep going; to my dear friends -Daniela Millacura,

Jacqueline Santander, and Magali Pinat- who were also fundamental in this process. Finally,

I am also grateful to Matheus Capelari, for his relentless support during the last chapter of

my program.

v

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

ESSAY 1. IS MINING AN ENVIRONMENTAL DISAMENITY? EVIDENCE

FROM RENTAL PRICES IN A DEVELOPING COUNTRY . .
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2. Theoretical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3. Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
1.3.1. Resource Extraction Sites
1.3.2. Household information . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.3. Air and Water Pollution Perceptions
. . . . . . . . . . . . . . . . . .
1.4. Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1. Households’ Perceptions of Air and Water Quality . . . . . . . . . . .
1.4.2. Spatial Difference-in-Difference Hedonic Price Estimation . . . . . . .
1.4.3. Spatial Difference-in-Difference Nearest-Neighbor Matching . . . . . .
1.4.4. Other Matching Procedures
. . . . . . . . . . . . . . . . . . . . . . .
1.5. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.1. The Effects of Extraction Site Openings on Rental Prices . . . . . . .
1.5.2. Heterogeneous Effects Among Short-term and Long-term Residents
.
1.5.3. Average Marginal Willingness to Pay Measures
. . . . . . . . . . . .
1.5.4. Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ESSAY 2. THE U.S. COAL-TO-GAS POWER PLANTS CONVERSION
PROCESS AND ITS EFFECTS ON HOUSING PRICES . . . .
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2. Emissions from Fossil-Fueled Power Plants and the Health Risks of Exposure
2.3. Analytical Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1. A Brief on the Hedonic Price Model . . . . . . . . . . . . . . . . . . .
2.3.2. A Localized Change in Air Quality . . . . . . . . . . . . . . . . . . .
2.4. Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . .
2.5. The Coal-To-Gas Power Plant Conversion and its Effects on Housing Prices .
2.5.1. Housing Price Impacts of the Coal-to-Gas Conversion using a Near-Far
Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.4.1. Power Plants
2.4.2. Property Transactions

1
1
2
7
9
9
11
12
14
14
17
18
21
21
21
26
30
32
36

38

47

54
54
55
59
62
62
64
65
65
69
69

72

vi

2.5.2. Housing Price Impacts of the Coal-to-Gas Conversion using a Near-
Gas/Near-Coal Approach . . . . . . . . . . . . . . . . . . . . . . . .
2.5.3. Housing Price Impacts of the Coal-to-Gas Conversion using a Triple-
Differences Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.4. Announcement Effects . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

APPENDIX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

75

77
79
80

82

95

ESSAY 3. THE EFFECTIVENESS OF AIR QUALITY WARNINGS AND
TEMPORARY DRIVING BANS: EVIDENCE FROM AIR POL-
LUTION AND URBAN TRANSIT FLOWS IN SANTIAGO . . 100
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.1.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.2. Driving Restrictions and Environmental Episodes in the Literature
. . . . . 105
3.3. Background Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.3.1. Air Pollution Prevention and Clean-Up Plan . . . . . . . . . . . . . . 111
3.3.2. Air Quality Warnings . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.3.3. Temporary Driving Restrictions . . . . . . . . . . . . . . . . . . . . . 116
3.4. Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.4.1. Air Quality Warnings . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.4.2. Air Pollution and Weather Variables
. . . . . . . . . . . . . . . . . . 118
3.4.3.
Indexes of Air Quality from Particulates (ICAPs) . . . . . . . . . . . 120
3.4.4. Urban Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.5. Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.5.1. Pooled Fuzzy Regression Discontinuity Design . . . . . . . . . . . . . 125
3.5.2. Multi-Cutoff Fuzzy Regression Discontinuity Design . . . . . . . . . . 127

3.6. The Effectiveness of Air Quality Episodes and their Short-Term Driving Re-

strictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.6.1. The Overall Effect of Episodes and their Driving Bans - Pooled FRD 132
3.6.2. Heterogeneous Effectiveness of Episodes - Multi-Cutoff FRD . . . . . 137
3.6.3. Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

3.6.3.1 Monitoring Station Outside the Restricted Area as a Control

Station . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
3.6.3.2 False ICAP . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
3.7. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
APPENDIX A: SUPPLEMENTAL MATERIAL . . . . . . . . . . . . . . . . 150
APPENDIX B: EXAMPLE OF THE CASSMASSI FORECAST MODEL . 159
APPENDIX C: ADDITIONAL ESTIMATIONS . . . . . . . . . . . . . . . . 160

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

vii

LIST OF TABLES

Table 1.1.

Extraction Sites Over Time . . . . . . . . . . . . . . . . . . . . . . . .

10

Table 1.2.

Number of Cities and Households by Group . . . . . . . . . . . . . . .

12

Table 1.3.

Balance Table of Mean Pre-Treatment (2011) Characteristics

. . . . .

13

Table 1.4.

Descriptive Statistics on Cities’ Average Pollution Perceptions . . . . .

16

Table 1.5.

Balance Table of Mean Pre-Treatment (2011) Characteristics After Nearest-
Neighbor Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

Table 1.6.

Spatial DID Estimation Results . . . . . . . . . . . . . . . . . . . . . .

23

Table 1.7.

Spatial DIDNNM Estimation Results

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

25

Table 1.8.

Rental Prices, Wages, and Education in Mining Cities by Type of Resident 28

Table 1.9.

Heterogenous Average Treatment Effects by Type of Resident

. . . . .

29

Table 1.10. Average Marginal Willingness to Pay Measures

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

31

Table 1.11. Robustness Check: Different Matching Procedures

. . . . . . . . . . .

33

Table 1.12. Robustness Check: Placebo Group . . . . . . . . . . . . . . . . . . . .

34

Table 1.13. Robustness Check: Certificates of Occupancy . . . . . . . . . . . . . .

35

Table A1.1. Ridits on Air and Water Pollution Perceptions

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

41

Table A1.2. Average Marginal Effects on Pollution Perceptions

. . . . . . . . . . .

41

Table A1.3. Spatial DID Full Estimation Results

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

42

Table A1.4. Spatial DIDNNM Full Estimation Results

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

44

Table 2.1.

Coal-to-Gas-Fired Power Plants (Treated Plants)

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

68

Table 2.2.

Descriptive Statistics on Main Covariates . . . . . . . . . . . . . . . . .

70

Table 2.3.

Coal-to-Gas Impact on Property Prices - Homes Far from a Coal-to-
Gas-Fired Station as Controls . . . . . . . . . . . . . . . . . . . . . . .

74

viii

Table 2.4.

Coal-to-Gas Impact on Property Prices - Homes Near to a Coal-Fired
Station as Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

Table 2.5.

Coal-to-Gas Impact on Property Prices - Triple Differences Estimator .

78

Table 2.6.

Announcement Effects - Homes Far from a Coal-to-Gas Station as Controls 79

Table 2.7.

Announcement Effects - Homes Near to a Coal-Fired Station as Controls 80

Table A2.1. Prime Mover Description - Gas-Fired EGUs Proposed During 2009-2016. 83

Table A2.2. Prime Mover Description - Coal-Fired EGUs Retired Over 2009-2016. .

83

Table 3.1.

Indexes of Air Quality from Particulates (ICAPs) . . . . . . . . . . . . 116

Table 3.2.

Descriptive Statistics on Pollution and Weather Variables.

. . . . . . . 119

Table 3.3.

Pollutant Concentrations During Episodes . . . . . . . . . . . . . . . . 121

Table 3.4.

Descriptive Statistics on Urban Traffic Flows

. . . . . . . . . . . . . . 123

Table 3.5.

Urban Flows During Episodes . . . . . . . . . . . . . . . . . . . . . . . 124

Table 3.6.

Episodes Impact on Hourly Average Pollution . . . . . . . . . . . . . . 133

Table 3.7.

Episodes Impact on Hourly Average Vehicle Trips

. . . . . . . . . . . 135

Table 3.8.

Episodes Impact on Daily Average Mass-Transit System Trips . . . . . 137

Table 3.9.

Alerts and Pre-emergencies Impact on Hourly Average Pollution Con-
centrations.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

Table 3.10. Alerts and Pre-emergencies Impact on Hourly Average Vehicle Trips

. 140

Table 3.11. Alerts and Pre-emergencies Impact on Daily Average Mass-Transit Sys-

tems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

Table 3.12. Robustness Check - Using Talagante as a Control Station

. . . . . . . 144

Table 3.13. Robustness Check - Using ICAP10,t−7 as the Running Variable

. . . . 146

Table A3.1. 2015 Protocols in Environmental Episodes . . . . . . . . . . . . . . . . 153

Table A3.2. Number of License Plate Digits Restricted by Permanent and Temporary

Driving Restrictions. 1990-2015 . . . . . . . . . . . . . . . . . . . . . . 154

ix

Table A3.3. 2016 Calendar of Driving Restrictions for Santiago

. . . . . . . . . . . 155

Table A3.4. Historical Issuance of Environmental Episodes. 1997-2015 . . . . . . . . 156

Table A3.5. Expected Percentage of Light-Duty Private Cars Affected by Restric-

tions Over 2001-2015. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

Table C3.1. Episodes Impact on Hourly Maximum Pollution . . . . . . . . . . . . . 160

Table C3.2. Episode Impact on Hourly Max Vehicle Trips

. . . . . . . . . . . . . . 161

Table C3.3. Episodes Impact on Daily Max Mass-Transit System Trips . . . . . . . 162

Table C3.4. Alerts and Pre-emergencies Impact on Hourly Average Pollution Con-

centrations.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

Table C3.5. Alerts and Pre-emergencies Impact on Hourly Max Vehicle Trips

. . . 164

Table C3.6. Alerts and Pre-emergencies Impact on Daily Max Mass-Transit System

Trips

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

x

LIST OF FIGURES

Figure 1.1. Overview of the Common Trends Assumption - Rental Prices . . . . .

19

Figure 1.2.

Distribution of Households’ Perceptions of Environmental Pollution in
Mining Cities by Type of Resident . . . . . . . . . . . . . . . . . . . .

27

Figure A1.1. Resource Extraction Site Openings Between 2011-2016 . . . . . . . . .

39

Figure A1.2. Example of Treated, Control, and Placebo Cities, and Treatments

. .

40

Figure A1.3. Overview of the Common Trends Assumption - Certificates of Occupancy 46

Figure 2.1.

U.S. Electric Power Industry Emissions Estimates 1990-2016

. . . . .

55

Figure 2.2.

Bid Curves and the Hedonic Price Function in a Hedonic Market for
Local Air Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure 2.3.

Power Plants with Proposed Gas-Fired and Retired Coal-Fired Gener-
ators
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

66

Figure 2.4.

Treated and Control Power Plants . . . . . . . . . . . . . . . . . . . .

68

Figure 2.5.

Properties in Treated and Control Areas . . . . . . . . . . . . . . . . .

70

Figure A2.1. Differences-in-Differences Design #1 - Houses Near and Far from Coal-
. . . . . . . . . . . . . . . . . . . . . . . .

to-Gas-Fired Power Plants

Figure A2.2. Differences-in-Differences Design #2 - Houses Near Coal-to-Gas-Fired
and Coal-Fired Power Plants . . . . . . . . . . . . . . . . . . . . . . .

Figure A2.3. Triple Differences - Houses Near and Far from Coal-to-Gas-Fired Power
Plants and Houses Near and Far from Coal-Fired Power Plants . . . .
Figure A2.4. Parallel Trends Assumption for Houses Near and Far (d ≥ 6mi) from
Coal-to-Gas-Fired Power Plant - First Retired Coal-Fired EGU as Treat-
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ment
Figure A2.5. Parallel Trends Assumption for Houses Near and Far (d ≥ 8mi) from
Coal-to-Gas-Fired Power Plant - First Retired Coal-Fired EGU as Treat-
ment
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

84

85

86

87

88

xi

Figure A2.6. Parallel Trends Assumption for Houses Near and Far (d ≥ 10mi) from
Coal-to-Gas-Fired Power Plant - First Retired Coal-Fired EGU as Treat-
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ment
Figure A2.7. Parallel Trends Assumption for Houses Near and Far (d ≥ 6mi) from
Coal-to-Gas-Fired Power Plant - First Operative Gas-Fired EGU as
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Treatment
Figure A2.8. Parallel Trends Assumption for Houses Near and Far (d ≥ 8mi) from
Coal-to-Gas-Fired Power Plant - First Operative Gas-Fired EGU as
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Treatment
Figure A2.9. Parallel Trends Assumption for Houses Near and Far (d ≥ 10mi) from
Coal-to-Gas-Fired Power Plant - First Operative Gas-Fired EGU as
Treatment
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Figure A2.10. Parallel Trends Assumption for Homes Near Coal-to-Gas-Fired and
. .

Coal-Fired Stations - First Retired Coal-Fired EGU as Treatment

Figure A2.11. Parallel Trends Assumption for Homes Near Coal-to-Gas-Fired and
Coal-Fired Stations - First Operative Gas-Fired EGU as Treatment . .

89

90

91

92

93

94

Figure 3.1.

Hourly Average Pollution and Vehicle Trips by Season . . . . . . . . . 110

Figure 3.2.

Hourly Average Pollution by Type of Day -Winter- . . . . . . . . . . . 113

Figure 3.3.

Discontinuity Plot on Hourly Average Pollution Concentrations . . . . 131

Figure 3.4.

Discontinuity Plot on Hourly Average Traffic Flow . . . . . . . . . . . 132

Figure A3.1. Spatial Location of Santiago’s Monitoring Stations . . . . . . . . . . . 150

Figure A3.2. Daily Average Pollutant Concentrations . . . . . . . . . . . . . . . . . 151

Figure A3.3. Daily Maximum ICAP10 and ICAP2.5 during Winter 2015 . . . . . . . 157

Figure A3.4. Air Quality Index Densities During Episodes

. . . . . . . . . . . . . . 158

xii

ESSAY 1. IS MINING AN ENVIRONMENTAL DISAMENITY? EVIDENCE

FROM RENTAL PRICES IN A DEVELOPING COUNTRY

Abstract

Resource extractive industries are often challenged by nearby communities due to the envi-

ronmental impacts of the activity. If proximity to mining represents a disamenity to house-

holds, the opening of new mines should lead to a decrease in housing prices. Due to the high

pollution potential of this activity, mine openings could also affect households’ willingness

to pay (WTP) for environmental quality improvements. This study addresses whether the

high-pollution potential of mining outweighs its economic benefits, as capitalized into rental

prices, using Chile as a case study. Repeated cross-sectional data on resource extraction site

openings is combined with households’ information for over 20,000 rental housing units to

estimate a hedonic price equation on rental prices. The identification strategy relies on a

spatial difference-in-difference (DID) design, where houses in cities hosting the site openings

are compared with houses in cities without mining, before and after the new sitings. A

spatial DID nearest-neighborhood matching estimator strengths this strategy. Results show

that mining represents an environmental disamenity to households. Renters who live near

mining get compensated with rental prices that are around 14-25% lower. This represents an

average willingness to pay of USD52-94 a month to avoid proximity to the concentration of

mines. Further distinction among types of residents suggest that this market capitalization

is higher among long-term residents, which constitutes evidence of a taste-based sorting of

households across space.

1

1.1.

Introduction

Extraction sites, mining methods, and mineral processing waste practices adversely

affect human health and environmental quality (U.S. Environmental Protection Agency,

1995). Common mining activities, such as blasting, hauling, and drilling, release high levels of

both particulate matter and gaseous pollutants such as sulfur dioxide (SO2), nitrogen oxides

(NOX), and carbon monoxide (CO), all contributing to local air pollution. Water quality

is also threatened. The high concentration of chemicals and materials on land surrounding

mines, in combination with large amounts of water from mining processes, increase the

likelihood of mine discharges into streams and rivers. Tailing management, as well as noise,

aesthetic, and visual pollution also constitute negative impacts, which in combination with

soil degradation can threaten both agricultural production and human settlement.1

Environmental degradation increases the risks to human health (Chay and Green-

stone, 2003; Currie and Neidell, 2005; Chen et al., 2013), and imposes substantial health

costs on developing countries (Yu, 2011; Greenstone and Jack, 2015). Proximity to the pollu-

tion source magnifies these risks, making these externalities likely important in communities

near the polluting facility. If proximity to mines represents a disamenity to households, new

activity to extract mineral deposits should generally lead to a decrease in prices of nearby

houses. The opening of new mines could also affect households’ willingness to pay (WTP)

for environmental quality improvements, particularly when polluting industries are weakly

monitored.2 This paper studies these ideas using evidence from the rental housing mar-

ket in Chile. Though many scholars address housing market capitalization of proximity to

environmental disamenities, few of them study this relationship in emerging nations.3 Min-

1For more on these impacts see Malm et al. (1990), Ghose and Majee (2000), Ghose and Majee (2001),

Milu, Leroy, and Peiffert (2002), Kitula (2006), Bebbington et al. (2008), and Kemp et al. (2010).

2Broner, Bustos, and Carvalho (2012) suggest that countries holding comparative advantage in polluting
industries might have laxer environmental regulations due the lobbying exerted by these firms to prevent the
enactment of stringent standards. This is in line with the 2016 The Economist article titled ’From Conflict
to Co-operation’, which highlights the mistrust existing among locals in Latin-American countries regarding
the stringency of environmental impact assessments that are submitted by local mining projects. Source:
From Conflict to Co-operation, The Economist, Online, accessed June 12, 2017.

3See for instance Kohlhase (1991), Mendelsohn et al. (1992), Kiel and Williams (2007), Greenstone

2

ing provides these countries with several local economic benefits (Aragon and Rud, 2013;

Loayza and Rigolini, 2016). Yet, recent social and environmental movements have increased

social opposition to mine sitings (e.g. Bebbington et al. (2008) and Urkidi (2010)), casting

doubts on the real benefits of these projects. The case of Chile provides an opportunity to

explore to whether the high-pollution potential of mining outweighs its economic benefits

as captured by housing markets. Chile has a substantial copper mining sector that strongly

contributes to its economic development, albeit with high local pollution. Copper smelters

in this country emit 33% and 58% of annual PM10 and SO2 releases, respectively; and soil

studies in hosting areas have shown high concentrations of copper, arsenic and antimony in

many valleys of the country (De Gregori et al., 2003; Ministry of Environment, 2011). By

using Chile as a focus of the study, this work adds to the scarce literature on environmental

disamenities in developing countries.

Due to Chile’s long mining tradition, this paper avoids comparing rental prices in

areas with and without mining. Instead, it follows Davis (2011) and studies the effects that

recent mine openings have on the rental housing market. Using 2011 and 2016 data from

the National Service of Geology and Mining, this study defines three types of mine openings

that constitute three treatments: all-size mine openings; small -, medium- and large-scale

(SMLs) mine openings; and artisanal mine openings that groups small-scale mines that

operate under rudimentary conditions. It is important to distinguish between SMLs mining

and artisanal mining, as the latter is generally associated with smaller economic impacts

relative to large-scale mining, and with less sustainable production practices. Artisanal

mining may therefore have higher net detrimental impacts than SMLs mining. Cities that

host the exclusive opening of one of these sites are part of a treated group of cities, cities

and Gallagher (2008), and Gamper-Rabindran and Timmins (2013) for hazardous waste sites or Superfund
sites in the US; Kiel and McClain (1995b) and Kiel and McClain (1995a) for waste incinerators in the US;
Boxall, Chan, and McMillan (2005), Muehlenbachs, Spiller, and Timmins (2013), and Muehlenbachs, Spiller,
and Timmins (2015) for oil and shale gas wells in Canada and the US; Currie et al. (2015) for industrial
plants; Gamble and Downing (1982) for nuclear plants; and Davis (2011) for power plants, all in the US. To
the author’s knowledge, only one previous study has looked at the impact of proximity to disamenities in
emerging economies. See Deng, Hernandez, and Xu (2014) on the market impacts of power plants in China.

3

with mine records and zero openings are part of a placebo group of cities, while locations

with no mining records are part of a control group of cities.4 This information is merged

to repeated cross-sectional data on households and rental prices from the 2011 and 2015

versions of the National Socioeconomic Characteristics Survey (CASEN), adding to more

than 24,000 renting households. Assuming no major differences in mine openings between

2015 and 2016, no announcement effects, and no future expectations on these openings,

rental prices for 2015 therefore capture most of the effect of new mine sitings on the rental

market.5

The econometric strategy relies on two quasi-experimental techniques for identifica-

tion. First, a spatial difference-in-difference (DID) hedonic price equation compares house-

holds in treated cities to households in control cities before and after the openings.6 The sec-

ond approach combines the previous spatial DID estimation with a nearest-neighbor match-

ing procedure that finds a match for houses in treated cities from the set of houses in control

cities, and compares them before and after the sitings. This spatial DID nearest-neighbor

matching estimator (DIDNNM) uses distance between the entire set of covariates included

in the hedonic price estimation, but unaffected by the treatment, to construct a group of

comparable houses in untreated cities. This procedure improves the balance of observables

across treated and untreated cities, reducing any potential bias from observable confounders

that can affect the original spatial DID estimation.

Results strongly suggest that mine openings represent an environmental disamenity

to households. Overall, the concentration of extraction site openings outweighs their local

economic benefits on the housing market, leading to a 14-19 percent reduction in monthly

4Chile is administratively divided into 345 communes -or cities- grouped into 15 regions.
5In the case of positive mine openings between 2015 and 2016, estimates using 2015 rental prices provide

a lower bound of the true causal effect.

6In an ideal setting, effects of proximity to the disamenity would be measured by restricting the sample to
the number of houses within some specific radius from each mine siting. Unfortunately, houses’ geographical
location is unavailable in the data. Still, findings in here bring to light the net impacts of cities’ concentration
of extraction sites on the value of housing units for rent. Under the holistic understanding that properties
for rent are part of a comparable housing market, these findings approximate the effects of mining sites on
the housing market as well.

4

rental prices of nearby housing units. This represents a monthly average willingness to

pay of USD 38-52 to avoid mining. When distinguishing by type of mine, results weakly

suggests that new SMLs sites lead to a 10-15 percent decrease in rental prices, while robust

evidence indicates that the opening of artisanal sites impose a price reduction of 17-25

percent, associated with a monthly average willingness to pay of USD 64-94 to avoid this

type of mining. Negative net effects of mine concentration on rental prices suggest that the

environmental impacts of extraction sites discourage households’ location near this activity,

mostly when it comes to artisanal mining. To the author’s knowledge, this is the first study

estimating market capitalizations of proximity to mining in emerging economies.

Given mining’s high pollution potential, this paper also derives implicit prices for

both air and water quality using households’ perceptions of environmental pollution.7 Re-

spondents in CASEN report their pollution perceptions, which are used to construct each

city’s index of pollution, and compared later with objective measures. Results show that

households are willing to pay around USD 10-17 for a one standard-deviation improvement

in the air quality index, while weak evidence suggests an average willingness to pay of USD

6-22 for a similar improvement in the water quality index.8 By providing implicit prices of

air and water quality, this work also adds to the growing field of environmental valuation in

developing countries (Greenstone and Jack, 2015).

Earlier research has shown evidence of a taste-based sorting of individuals across space

7Previous evidence shows how subjective measures outperform objective indicators on environmental
amenities (Berezansky, Portnov, and Barzilai, 2010; Chasco and Gallo, 2013) by reducing the upward bias
from their potential mismatch (Michael, Boyle, and Bouchard, 2000). Two other reasons advocate for the use
of subjective measures of environmental quality. First, individuals might have difficulties when interpreting
objective measures, and so their location might be based on their own insights of local environmental quality
instead of more technical measures (Chasco and Le Gallo, 2015). Second, departures from rationality re-
garding individuals’ real estate market decisions (Kahneman, 2003) suggest that subjective measures might
suit better in a revealed preferences setting to eliciting implicit prices.

8Other works on implicit prices for air quality in Chile suggest a WTP of USD 3-6 per µg/m3 reduction
in PM10 concentrations (Lav´ın, Dresdner, and Aguilar, 2011). These previous works, however, omit the
potentially high correlation between PM10 and other airborne pollutants such as PM2.5, SO2, NOX and
O3 (Ilabaca et al., 1999). If renters are capable of distinguishing among these pollutants, previous WTP
calculations for Chile are upward biased. However, if renters are unable of making such distinctions, these
previous results are likely reflecting implicit prices for air quality in a broad sense rather than being pollutant-
specific.

5

(Chay and Greenstone, 2005). Tastes for amenities differ among individuals, and so house-

holds that downplay environmental disamenities are more likely to sort into less pleasant

places, adding bias to marginal willingness-to-pay (MWTP) measures of proximity to these

disamenities. This paper studies the existence of sorting in Chile by testing heterogeneous

effects between short- and long-term residents of treated and non-treated cities. Evidence of

significant differences in average treatment effects among these groups represents a simple

test of a preference-based sorting of households. Findings reveal that market capitaliza-

tions of the disamenity are higher among long-term residents of mining cities. This suggests

that households moving into cities that host new extraction sites are likely minimizing the

disamenity effect of mining.

The robustness of these findings relies on the geographically-predetermined allocation

of minerals throughout the country, which rules out the possibility of unobservables affect-

ing both sitings and rental prices.9 Yet, unobservables might still affect rental prices and

firms’ decision of opening a new site, or once opened, whether they operate or remain inac-

tive. Despite the empirical consideration of several dummies at the regional level to control

for time-invariant and time-variant unobserved heterogeneity, time-invariant unobservables

common at the local level might still add some bias to the estimation. For instance, com-

modity price booms might influence firms’ decisions regarding opening a new mine as well as

rental prices. This study considers that whenever true, time-invariant unobservables affect

the outcome variable and so firms’ opening decisions both in the same direction, leading

to upward biased estimates. However, as the average net effect of mine openings on rental

prices is expected to be negative, statistically significant DID and DIDNNM estimates are

9Former studies show that the location of unpleasant facilities might not follow a random process. Davis
(2011) shows that power plants in the U.S. tend to locate near neighborhoods with different housing and
demographic characteristics, while Muehlenbachs, Spiller, and Timmins (2013) and Muehlenbachs, Spiller,
and Timmins (2015) provide similar arguments for the location of shale gas wells. When the location of these
facilities is not at random, correlated unobservables might confound identification of the causal effect. This
situation is unlikely to affect identification in this setting for two main reasons. First, mine sitings in Chile
depend almost fully on the resource location, which indeed follows an exogenous process. Second, because
mineral resources in Chile are owned by the national government as opposed to the U.S. where they are
owned by private individuals, mining companies do not have to engage in leasing contracts with households
as a manner of initiating the extraction.

6

sufficient to show significance of the true causal effect.

Notwithstanding the above, a different argument might suggest that both the DID and

the DIDNNM could be downward biased as well. If the opening of polluting facilities occurs

in cities with poorer economic conditions, omitted variables will then affect the probability of

mine openings and rental prices in the opposite direction. Under a downward bias situation,

statistically significant estimates are not enough to show significance of the true effect. This

situation is in any case ruled out by the DID specification whenever these socioeconomic

characteristics are common among households at the treatment level. Further in the paper,

full estimation results show no evidence of differences in average rental prices among treated

and control cities during the pre-treatment period, suggesting the existence of similar initial

economic conditions in both groups.

The remainder of this work proceeds as follows. Section 1.2 outlines the hedonic price

model, while section 1.3 describes the data. Section 1.4 describes the empirical strategy, and

section 1.5 presents the results. Section 1.6 concludes.

1.2. Theoretical Model

An important set of empirical works relies on hedonic price theory to elicit implicit

prices of non-market (dis)amenities.10 Dating back to Ridker and Henning (1967), Ridker and

Henning (1967), and Rosen (1974), the hedonic price method starts with the idea that goods

are valued for their utility-bearing attributes, and their prices are therefore a function of their

characteristics. In the case of a house, its price depends on its physical structure (e.g. size,

rooms, construction material), on the provision of public goods or specific characteristics

of the neighborhood in which the house is located (e.g. parks, open space, local schools,

crime rates), and on the location-specific environmental amenities, such as environmental

quality or distance from mines, to which the house has immediate access (Mendelsohn and

10See Mendelsohn and Olmstead (2009) and Freeman III, Herriges, and Kling (2014) for reviews on the

hedonic price applications.

7

Olmstead, 2009). All these characteristics have a price, implicit in the observed price of the

house (Rosen, 1974).

The model starts with a class of differentiated commodities described by a vector of

characteristics, z=(z1, z2, ..., zn). Two important assumptions follow: the area under study

corresponds to a single housing market, characterized by a continuum of sizes and types

of available housing with different characteristics. Market prices related to each of these

products, and determined by the equilibrium interaction between home buyers and sellers,

are associated to a fixed value of vector z, so that p(z) = p(z1, z2, ..., zn) relates prices and

characteristics. This function p(z) is the equivalent of the hedonic price function, which

gives the minimum price of any package of house characteristics (Rosen, 1974). The partial
derivative of p(·) with respect to the nth house characteristic, ∂p/∂zn, is referred to as the

marginal implicit price of zn, implicit in the overall price of house z.

Consider now an individual that derives utility from occupying house j. Let this utility

be u = u(x, zj), where x is a Hicksian composite good of price 1. The individual maximizes
u(·) subject to a budget constraint M − p(zj) − x = 0, where M represents individual’s

income. The first order condition for the optimal choice of the nth characteristic is equal to:

∂u/∂zn
∂u/∂x

=

∂p
∂zn

.

(1.1)

Now suppose zjn is a measure of local air quality. Equation (1.1) says that, in equilib-

rium, individuals’ marginal willingness to pay (MWTP) for environmental quality is equal to

the marginal implicit price of that characteristic. The marginal implicit price of air quality

reflects the additional amount that households must pay in order to buy or rent a house

located in a neighborhood with better air quality, ceteris paribus. Therefore, it represents

the equilibrium differential that allocates individuals across locations, and compensates in-

dividuals for living in areas with poor air quality. Locations that exhibit high levels of air

pollution must have lower prices to attract homeowners, or tenants, to these locations. Equi-

librium levels on the hedonic price schedule are therefore tangent to the marginal valuation

8

that some specific individual places on air quality, so that the marginal implicit price of air

pollution is equal to that individual’s MWTP for that level of the environmental amenity.

1.3. Data

1.3.1. Resource Extraction Sites

This study focuses on rental market capitalizations of extraction site openings taking

place in Chile between 2011 and 2016. Data on these openings come from inventories of

mineral sites prepared by the National Service of Geology and Mining (Servicio Nacional

de Geologa y Minera -SERNAGEOMIN) for 2011 and 2016. These inventories contain de-

tailed information on the extraction sites located in Chile, including the type of mine (e.g.

open-pit, underground), whether the mine is active, its geographical location, and the type

of mineral extracted from the site. This paper studies the openings of two main types of

extraction sites as defined by SERNAGEOMIN: active mines and artisanal mines. An ac-

tive mine is a small-, medium- or large-scale (SMLs) extraction site that is under formal

and continuous operation over time.11 Artisanal mines are instead a segment of small-scale

mines but operating under informal and rudimentary conditions, and mostly on a sporadic

base depending on the ore price (D´ıaz Tobar, 2015). Many these mines also tend to lack of

monitoring systems, which impede the proper tracking of their environmental impacts and

preclude their proper regulation. Furthermore, artisanal mines’ lack of financial capacity

to invest in abatement technologies (S´anchez and Enr´ıquez, 1996) hampers the application

of environmental best practices, which puts them at a disadvantage in terms of corporate

environmental responsibility when compared to SLMs mines. Since 2011, more than 3,500

new SMLs sites and around 3,000 new artisanal sites were opened in the country (see Table

11SERNAGEOMING uses a number-of-workers criteria to define the scales of production in mining. A
small-scale mine has less than 80 hired workers; a medium-scale mine has between 80 and 400 workers; and
a large-scale mine has for more than 400 workers.

9

1.1).12 During this period, however, many old sites closed operations too. From Table 1.1,

a total of 4,041 sites shut down, with most of these closings coming from artisanal sites, the

most vulnerable segment in this industry.

Table 1.1. Extraction Sites Over Time

Type of Site

Year

Number of

2011

2016 Openings Closings

SMLs Sites

Open-Pit
Underground
Tailings
Others

Artisanal Sites

1,324
249
703
58
314
3,920

4,093
615
1,504
634
1,340
3,590

Total

5,244

7,683

3,537
496
1,275
593
1,173
2,943

6,480

768
130
474
17
147
3,273

4,041

Notes: Other SMLs Sites include mineral processing sites and exploration deposits. Both SMLs and Artisanal
Sites exclude administration facilities. Source: SERNAGEOMIN (2011) and SERNAGEOMIN (2016)

The former site distinction is used to define three treatments at the city-level to

study the impacts of mine openings on rental prices. The first treatment is defined as a

general opening of either SMLs mines, artisanal mines, or both, and it is labeled as “all-

mine openings”; the second treatment is defined as “only SMLs site openings”; while the

third treatment constitutes “only artisanal site openings”.13 Using this distinction, cities

are classified into one of these treatment groups depending on the type of openings taking

place over time, or into a control group of cities if they have no records of mining activity

during this period. Provided that the opening of new sites is truly affecting rental prices in

hosting cities, then cities with a constant number of mines should exhibit no price impacts

attributable to this characteristic. With this rationale, cities with mining activity during

2011 but with no additional openings are part of a placebo group of cities, compared later

with the control group as a falsification test. Cities that simultaneously host site openings

12For a spatial location of the openings see Figure A3.1 in the Appendix.
13The “only SMLs sites” group excludes artisanal mines.

10

and closings are dismissed from the analysis. Figure A1.2 (appendix) illustrates this group

construction.

1.3.2. Household information

Data on extraction sites is combined with information on rental houses, their prices

and their structural characteristics, all from the 2011 and 2015 versions of the National So-

cioeconomic Characteristics Survey (CASEN). Households in this survey describe the struc-

tural characteristics of their dwellings, while those declaring themselves as tenants report

their monthly rent payment. Renting households in this survey add up to approximately

12,000 households each year.14

Previous information on housing units is merged with cities’ attributes that might

simultaneously affect the equilibrium price in the rental market. This information comes

from two main sources. First, we consider data on local public finances and other city-level

characteristics that come from the National System of Municipal Information (Sistema Na-

cional de Informacin Municipal -SINIM). Variables such as the poverty rate and population

density are used to control for the role of the local economy and agglomeration economies,

respectively. Other variables such as the number of public parks and squares, and per capita

expenditures on waste disposal and pick-up are all considered to approximate a city’s quality

of life. Additionally, information on crime rates per 1,000 inhabitants from the Crime Pre-

vention Sub-Secretary (Subsecretara de Prevencin del Delito -SPD) is also considered. The

final dataset contains over 18,000 households over 2 years and 264 cities, classified into one

of the treatment groups, into a control group, or into a placebo group. Cities and households

by group are depicted in Table 1.2.

Table 1.3 displays the mean characteristics by groups of main covariates for the 2011

pre-treatment period. Panel A shows dwellings’ characteristics, while panel B displays cities’

attributes. Columns (1)-(4) display means by group, while the last three columns display

14This includes renting households living in both houses and apartments.

11

Table 1.2. Number of Cities and Households by Group

Period Group

Unit

All-Mine
Openings Mine Openings Mine Openings

Only SMLs

Only Artisanal Placebo

2011

2015

Treated

Control

Treated

Control

Cities

Households

Cities

Households

Cities

Households

Cities

Households

52

1,669

176
4,589

52

2,343

176
6,329

13
208

176
4,589

13
210

176
6,329

12
667

176
4,589

12

1,023

176
6,329

11
655

176
4,589

11
387

176
6,329

Notes: All-mine openings consider cities with only opening of SMLs mines, only opening of artisanal mines,
and the simultaneous opening of both sites (see Figure A1.2). By construction, therefore, this treatment
includes more cities than the last two groups together.

p-values from mean comparisons among the treatments and the control group. Mean com-

parisons in columns (5), (6) and (7) in Table 1.3 reject the null hypothesis of equal means

across groups for most the covariates, revealing their imbalance across groups. Covariate

imbalance among the treatment and control groups might add bias to the difference-in-

difference estimation, as units in treated and control cities may not necessarily have the

same distribution over observed variables. In this case, an adjustment at the group level can

remove this concern. More details on this adjustment are in the next section.

1.3.3. Air and Water Pollution Perceptions

Households’ perceptions of air and water pollution come from the 2015 version of

CASEN. Two indexes of pollution at the city-level are constructed from these perceptions,

which are related later to objective measures of environmental quality. By linking these

indexes to several objective indicators of pollution, any multicollinearity problem generally

associated with different types of pollutants or environmental quality measures is reduced

(Boyle and Kiel, 2001).

Objective indicators on air pollution come from the Emissions and Pollutant Trans-

12

Table 1.3. Balance Table of Mean Pre-Treatment (2011) Characteristics

Variables

Means

p-values

All-Mine
Only Artisanal Control
Openings Mine Openings Mine Openings Group

Only SMLs

(1) vs. (4)

(2) vs. (4)

(3) vs. (4)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Panel A. House-Level Characteristics
# of Bedrooms
# of Bathrooms
Dwelling Type (Base = substandard)

Proportion Row Units
Proportion Single-Family Units
Proportion Apts. (elev.)
Proportion Apts. (no elev.)

Walls Material (Base = substandard)

Proportion Reinforced Concrete
Proportion Masonry
Proportion Drywall

Walls Condition (Base = bad)

Proportion Regular
Proportion Good

Floor Material (Base = substandard)

Proportion Wood
Proportion Tile
Proportion Carpet
Proportion Cement

Floor Condition (Base = bad)

Proportion Regular
Proportion Good

Roof Material (Base = substandard)

Proportion Roof Tiles
Proportion Concrete
Proportion Zincstrips
Proportion Clinkstone

Roof Condition (Base = bad)

Proportion Regular
Proportion Good

Dimension (Base = < 30m2)

Proportion 30-40m2
Proportion 41-60m2
Proportion 61-100m2
Proportion 101-150m2
Proportion +150m2

Panel B. City-Level Attributes
Density (km2)
Crime (per 1,000/inab.)
Poverty (%)
Waste Disposal (%)
# of Parks (p.c.)
# of Public Squares (p.c.)

2.280
1.172

0.382
0.308
0.132
0.158

0.220
0.473
0.244

0.673
0.248

0.364
0.412
0.084
0.054

0.653
0.260

0.091
0.208
0.697
0.004

0.686
0.226

0.247
0.351
0.217
0.058
0.016

2.303
1.038

0.394
0.587

-

0.019

0.053
0.514
0.356

0.558
0.341

0.308
0.466
0.019
0.019

0.505
0.356

0.029
0.010
0.962

-

0.567
0.322

0.236
0.413
0.149
0.063
0.014

2.119
1.266

0.302
0.245
0.294
0.116

0.335
0.344
0.228

0.699
0.227

0.546
0.245
0.120
0.040

0.680
0.251

0.114
0.303
0.580
0.003

0.728
0.207

0.230
0.314
0.262
0.064
0.018

2.338
1.112

0.408
0.445
0.040
0.090

0.111
0.389
0.479

0.609
0.295

0.451
0.359
0.077
0.044

0.609
0.296

0.106
0.074
0.819
0.000

0.645
0.259

0.260
0.353
0.197
0.050
0.015

372.8
3,064.3
15.25
0.195
0.120
4.007

110.1
2,246.4
17.09
0.161
0.096
11.590

978.9
4,029.5
14.45
0.207
0.299
1.023

1,348.3
2,768.9
19.39
0.250
0.120
0.681

0.03
0.00

0.06
0.00
0.00
0.00

0.00
0.00
0.00

0.00
0.00

0.00
0.00
0.39
0.08

0.00
0.01

0.08
0.00
0.00
0.00

0.00
0.01

0.29
0.90
0.09
0.21
0.93

0.03
0.34
0.00
0.04
1.00
0.04

0.59
0.02

0.68
0.00
0.00
0.00

0.01
0.00
0.00

0.14
0.15

0.00
0.00
0.00
0.09

0.00
0.06

0.00
0.00
0.00
0.76

0.02
0.04

0.43
0.07
0.09
0.42
0.91

0.18
0.28
0.34
0.06
0.86
0.00

0.00
0.00

0.00
0.00
0.00
0.04

0.00
0.03
0.00

0.00
0.00

0.00
0.00
0.00
0.62

0.00
0.02

0.52
0.00
0.00
0.02

0.00
0.00

0.10
0.05
0.00
0.13
0.58

0.69
0.04
0.05
0.39
0.33
0.22

Notes: The crime variable considers criminal offenses of strong social connotation. p.c. = per capita.

fers Register (Registro de Emisiones y Transferencia de Contaminantes -RETC) containing

records on annual pollutant emissions for the entire country. In particular, this study con-

siders 2015 ton/year emissions for the three air quality criteria pollutants widely known for

causing smog and a broad range of adverse health effects (U.S. Environmental Protection

13

Agency, 2013): fine particulate matter (PM2.5), carbon monoxide (CO), and nitrogen diox-

ide (NOX). Although not necessarily perceptible by humans, carbon dioxide (CO2) is also

considered as an overall measure of air pollution.15 Similarly, 2015 ton/year water emis-

sions of total suspended solids (TSS) from RETC is used for water pollution. Due to data

availability, measures other than emissions are also considered to approximate water pollu-

tion. These are population density (to proxy for the local economy’s dynamism), and the

regional monthly average streamflow of main rivers (to proxy for the availability of recre-

ational water reservoirs). After examining the 2015 relationship between these measures and

pollution perceptions, the empirical strategy considers 2011 cities’ average prediction of pol-

lution perceptions using the 2011 versions of objective indicators. Details on this imputation

are provided in the next section.

1.4. Empirical Strategy

1.4.1. Households’ Perceptions of Air and Water Quality

From a four-point scale that ranges from “never” to “always”, households in 2015

CASEN state the frequency with which they observe air pollution in their neighborhoods and

water pollution in nearby lakes, estuaries, and rivers. Drawing on the ordinal but subjective

characteristic of these variables, this work assigns weights to each of these categories using

ridit analysis (Bross, 1958). Ridit scoring is a non-parametric tool utilized to compare

more than two datasets with ordered qualitative data.

It uses the observed distribution

of households to construct a numerical quantity (“ridit”) that in this case will serve as an

index of pollution perception. Let x1, x2, ..., xn be the ordered categories on the perception

variable, and p be the probability function defined with respect to the reference category as

pi = P rob(xi) with i = 1, ..., n. A ridit is calculated as follows:

15Objective measures on air pollution are restricted by the extent of its national coverage. One example
is SO2. Despite its availability in the RETC website, information on SO2 emissions is available for only 70%
of cities.

14

(cid:40) 0.5pi +(cid:80)

k<i pk

0.5pi

if

i > 1

if

i = 1.

Riditi =

(1.2)

Intuitively, a ridit works as a modified notion of a percentile in the sense that a low

ridit value for category i can be interpreted as only few households choosing a category k

such that k < i. Table A1.1 (Appendix) reports the ridits calculation for 2015 households’

perceptions of air and water pollution, with column (5) containing the final score. For

instance, column (5) indicates that more people perceive air pollution in their neighborhoods

rather than water pollution, as the ridit for the “never” category is higher in the latter

case. After obtaining these ridits, they are averaged at the city level to get two indexes of

perceptions of air and water pollution.

After the calculation of these indexes, they are regressed on several objective measures

of environmental quality. Under a time-invariant partial relationship assumption, estimation

coefficients from these regressions are used to input 2011 city-level averages. To do so, let

Yj be a fractional random variable representing city j’s average pollution perception taking

realizations yj on the unit interval [0, 1]. The conditional probability of Yj being equal to yj

is defined as follows:

P (Yj = yj|Xj) = E(Yj|Xj) =

exp(Xjγ)

1 + exp(Xjγ)

,

(1.3)

where E(Yj|Xj) is the conditional mean function defined as the logistic function, and Xj is

a vector of covariates (including a constant). As explained in the data section, the vector Xj

includes variables on PM2.5, CO, NOX, and CO2 emissions whenever Yj indicates air pollution

perceptions; and variables on TSS, average streamflow in rivers, and density whenever Yj

represents water pollution perceptions. Because realizations of Yj take values between 0

and 1, estimation of equation 1.3 applies findings in Papke and Wooldridge (1996) by using

a Bernoulli Quasi-Maximum Likelihood Estimator (QMLE) fractional logit regression with

2015 information. After getting the vector of estimated coefficients, ˆγ, predictions of 2011

15

cities’ average pollution perceptions are calculated using 2011 versions of the covariates.

Results for the average marginal effects in equation (1.3), and their significance, are

omitted from the results section, but are displayed in the Appendix (see Table A1.2). Re-

sults for air pollution in this Table indicate that PM2.5, NOX, and CO2 emissions are all

positively related with households’ perceptions of bad air quality. The estimated coefficient

for carbon monoxide (CO) emissions, however, takes an unexpected sign. CO emissions

positively correlate with urbanization levels, which might coincide with the amount of mech-

anisms available to regulate air pollution as well. Public awareness of these mechanisms

might affect households’ perceptions of air quality as they may feel that these policies are

effectively curbing air pollution. In terms of water pollution perceptions, results indicate

a counterintuitive sign for TSS, although this estimated coefficient is small in magnitude.

Regarding stream flow and cities’ density, higher values for these variables are negatively re-

lated to water pollution perceptions. Higher stream flows nearby may be negatively related

to water pollution perceptions because streams assimilative capacity (i.e. the stream capac-

ity to reduce pollutant concentrations) increase with the streamflow magnitude. A higher

population density, instead, might be negatively related to water pollution perceptions due

to a lower variety of water reservoirs available for recreation in high dense areas, which might

affect their perceptions on the contamination of these waterbodies. Table 1.4 displays the

summary statistics of the resulting indexes for each year, which are later transformed into a

standard z-score to ease their interpretation.

Table 1.4. Descriptive Statistics on Cities’ Average Pollution Perceptions

Variable

Period Mean Std. Dev. Min Max N Cities

Air Pollution

Water Pollution

2011
2015

2011
2015

0.467
0.473

0.528
0.509

0.012
0.111

0.011
0.071

0.341
0.304

0.483
0.444

0.493
0.926

0.538
0.960

317
320

282
320

16

1.4.2. Spatial Difference-in-Difference Hedonic Price Estimation

If the resource extraction activity represents a disamenity to households, rental prices

are expected to decrease in cities hosting new mine openings relative to prices in these cities

had mine sites not been opened. Given that this control is not available for observational

data, this study tackles this question using a spatial difference-in-difference (DID) approach

that considers pre- and post-treatment data for a group of treated cities and a group of

control cities. This DID design allows the identification of average treatment effects of mine

openings by capturing group-level and time-level omitted variables that generally confound

the identification of causal effects (Angrist and Pischke, 2008). Additionally, the approach

reduces the omitted time-invariant variable bias that is widely known to affect hedonic price

estimations (Kuminoff, Parmeter, and Pope, 2010). The spatial DID hedonic price equation

is defined as follows:

rhjt = α + δ11[treated]j + δ21[2016] + δ31[treated]j × 1[2016] + Hhjtζ + φj + 
hjt,

(1.4)

where rhjt is the natural log of the rental price paid by household h in city j during year t,

with h = 1, ..., H, j = 1, ..., J, and H > J; 1[treated]j is an indicator variable taking 1 if city

j belongs to a treatment group, and 0 if j belongs to the counterfactual group; 1[2016] is an

indicator variable taking 1 if year t = 2016, and 0 otherwise; Hhjt is a vector of covariates on

housing units and city-level characteristics, including air and water pollution; and 
hjt is an

idiosyncratic effect. Equation (1.4) also includes dummies by region, φj (ggeneralized later
into dummies by region × year, φjt), aimed at controlling for time-invariant unobservables

(and time-variant unobservables at the macro level) that might simultaneously affect the

outcome and the covariates of interest. This strategy constitutes a first alternative to an

instrumental variable estimation (Anderson and West, 2006; Muehlenbachs, Spiller, and

Timmins, 2015). Equation (1.4) is estimated using an ordinary least square (OLS) estimator.

The key parameter in equation (1.4) is δ3 that represents the causal effect on rental

17

prices of new openings in mining areas, or in other words, the average treatment effect on

the treated (ATT) (Blundell and Dias, 2009). The ATT captures the net effect on rental

prices of an increasing concentration of mines in the treatment group, and is expected to

have a negative and significant sign whenever the disamenity effect of these sites exceeds

the economic benefits of increasing mining operations that are capitalized into rental prices.

The identification of this causal effect comes from the counterfactual assumption that the

trend in the rent prices for the control group is equivalent to the trend in the rent prices for

the treatment group in absence of the treatment. This assumption, commonly known as the

common trends assumption, constitutes the core of the DID tool as it compares trends across

groups while considering pre-treatment differences in the levels of the response variable.

To examine the fit of the previous assumption, Figure 1.1 shows the average rental

price for each of the treatment groups and the control group using pre-treatment data from

the 2009 and 2011 versions of CASEN. During pre-treatment periods, houses located in

cities affected by either treatment one (all-mine openings) or treatment three (artisanal

openings) have consistently had on average higher rental prices than houses in the control

group, particularly cities in the latter case. Cities affected by treatment two (SMLs mine

openings), however, have experienced lower rental prices. Consistent with the common

trends assumption, these prices evolved in a parallel fashion for cities that experienced only

SMLs mine openings and only artisanal mine openings before the treatment period. For

cities experiencing all-mine openings this assumption seems to hold but not as strongly as

in the other two cases.

1.4.3. Spatial Difference-in-Difference Nearest-Neighbor Match-

ing

Covariates’ mean characteristics displayed in Table 1.3 show a significant imbalance

across groups, highlighting the importance of controlling for these characteristics in the esti-

mations. Yet, imbalance of observables across groups might bias the OLS estimator, leading

18

Figure 1.1. Overview of the Common Trends Assumption - Rental Prices

Notes: Considering tenants with a single family nucleus. Source: Ministerio de Desarrollo Social (2009) and
Ministerio de Desarrollo Social (2011).

to inaccurate conclusions about the impacts of mineral site concentrations on rental prices.

This work’s empirical strategy considers an additional estimation procedure to improve this

balance. A nearest-neighbor matching (NNM) estimator enhances groups’ comparability by

using a vector Hh of covariates, unaffected by the treatment, to match comparable house-

holds in the control group to households in the treatment group, discarding all the units not

selected as matchings. In doing this, control units are expected to be statistically similar to

units in the treatment group (Rosenbaum and Rubin, 1985).

The NNM procedure works as follows. Let Tl be the set of households in treatment

group l = {1, 2, 3}; C the set of counterfactual households; and rTl

h and rC

h be the observed

rental prices (in logs) for treated and control households respectively. For each household h

in treatment group l, define a neighborhood ΩH

m(hl). NNM sets

19

m(hl) = {k1, k2, ..., kmh|h ∈ Tl,||Hh − Hk||s < ||Hh − Hi||s, i (cid:54)= k},
ΩH

(1.5)

where ||Hh − Hk||s = {(Hh − Hk)

(cid:48)

S−1(Hh − Hk)}1/2; S is a symmetric, positive-

definite distance matrix; and m is the smallest number such that the number of elements

in ΩH

m(hl) is (at least) the desired number of matches. Once counterfactual households are

matched, the NNM estimates the ATT as follows:

(cid:88)

h∈Tl

(cid:88)

k∈Ω(hl)

rC
k ),

(1.6)

ˆτAT Tl =

1
nl

h − 1
(rTl
mh

where Ω(hl) = ΩH

m(hl), and nl is the number of households in group Tl. When more than one

period is available, it is possible to apply the Difference-in-Difference Nearest-Neighborhood

Matching (DIDNNM) estimator defined as follows:

(cid:88)

h∈Tlt

1
nlt

h − 1
(rTlt
mh

(cid:88)

k∈Ω(hl)t

k ) − 1
rCt
nlt(cid:48)

(cid:88)

h∈Tlt(cid:48)

h − 1
(rTlt(cid:48)
mh

(cid:88)

k∈Ω(hl)t(cid:48)

rCt(cid:48)
k ),

(1.7)

ˆτAT T−DIDl =

where Tlt, Tlt(cid:48), Ct, and Ct(cid:48) denote the treatment group l and control group in years t and t(cid:48) re-

spectively. Due to its bias reduction potential, this paper implements a DIDNNM procedure

that matches with replacement at least one neighbor to each h in the treatment group (i.e.

m = 1) (Dehejia and Wahba, 2012). Additionally, it considers all the covariates included in

vector H in equation (1.4), except variables on environmental pollution; it requires an exact

match for all the discrete covariates; and it sets S to be the Mahalanobis distance matrix

for continuous covariates.

Table 1.5 displays mean characteristics by group after the NNM procedure for the

pre-treatment period. As in Table 1.4, panel A shows dwellings’ characteristics, while panel

B exhibits cities’ attributes. Columns (1)-(3) show covariates’ means for houses in cities

with all-mine openings and its matched control group, and the p-value of their comparison.

20

Similarly, columns (4)-(6) display the information for houses in cities with only SMLs mine

openings, while columns (7)-(9) for houses in cities with only artisanal mine openings. Pre-

treatment covariates show now an improvement in terms of balance across groups relative

to similar comparisons in Table 1.4. After reducing the control group to only comparable

houses, there is no strong evidence of pre-treatment observables being different between

households in the treatments and in the control groups, particularly for variables in panel

A.

1.4.4. Other Matching Procedures

To check the robustness of the results, three other additional matching options are

considered as alternatives to the NNM procedure. A straightforward modification is to

set the number of minimum matches, m, greater than one. This option means that each

observation h in the treatment group has a set of nearest-neighbors of size m > 1. A second

modification considers the distance in terms of continuous variables between observations in
the treatment group and their matches, ||Hh − Hk||s, to be less or equal than some constant

c called the caliper, while the third modification considers the Euclidean distance matrix as

the S distance matrix.

1.5. Results

1.5.1. The Effects of Extraction Site Openings on Rental Prices

Results for the effects of mine openings on rental prices are displayed in Table 1.6 for

the spatial DID estimator, and in Table 1.7 for the spatial DIDNNM estimator. Columns for

each treatment represent different specifications of equation (1.4), which are used throughout

the remainder of this section. Full results are displayed in Tables A1.3 and A1.4 in the

Appendix section.

21

Table 1.5. Balance Table of Mean Pre-Treatment (2011) Characteristics After
Nearest-Neighbor Matching

Variables

Panel A. House-Level Characteristics
# of Bedrooms
# of Bathrooms
Dwelling Type (Base = substandard)

Proportion Row Units
Proportion Single-Family Units
Proportion Apt. (elev.)
Proportion Apt. (no elev.)

Walls Material (Base = substandard)

Proportion Reinforced Concrete
Proportion Masonry
Proportion Drywall

Walls Condition (Base = bad)

Proportion Regular
Proportion Good

Floor Material (Base = substandard)

Proportion Wood
Proportion Tile
Proportion Carpet
Proportion Cement

Floor Condition (Base = bad)

Proportion Regular
Proportion Good

Roof Material (Base = substandard)

Proportion Roof Tiles
Proportion Concrete
Proportion Zincstrips
Proportion Clinkstone

Roof Condition (Base = bad)

Proportion Regular
Proportion Good

Dimension (Base = < 30m2)

Proportion 30-40m2
Proportion 41-60m2
Proportion 61-100m2
Proportion 101-150m2
Proportion +150m2

Panel B. City-Level Attributes
Density (km2)
Crime (per 1,000/inab.)
Poverty (%)
Waste Disposal (%)
# of Parks (p.c.)
# of Public Squares (p.c.)

Means

Means

Means

All-Mine Matched
Openings Control

p-values

Only SMLs

Matched
Mine Openings Control

p-values

Only Artisanal Matched
Mine Openings Control

p-values

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

2.430
1.205

0.434
0.339
0.078
0.147

0.161
0.567
0.267

0.743
0.215

0.344
0.484
0.065
0.039

0.743
0.214

0.083
0.164
0.753

-

0.760
0.190

0.219
0.377
0.255
0.056
0.013

2.352
1.191

0.447
0.315
0.078
0.159

0.145
0.601
0.249

0.775
0.189

0.323
0.522
0.080
0.040

0.774
0.190

0.072
0.180
0.747

-

0.790
0.167

0.200
0.408
0.246
0.051
0.011

255.5
2,784.5

1,362.6
2,878.9

15.0
0.202
0.125
4.087

18.5
0.244
0.082
0.754

0.02
0.50

0.57
0.24
0.98
0.44

0.16
0.11
0.34

0.08
0.13

0.30
0.08
0.39
0.89

0.10
0.16

0.36
0.32
0.76

-

0.10
0.18

0.29
0.15
0.62
0.61
0.68

0.01
0.76
0.01
0.14
0.25
0.09

2.301
1.044

0.449
0.551

-
-

0.044
0.596
0.353

0.596
0.331

0.324
0.529

-
-

0.544
0.353

0.007

-

0.993

-

0.596
0.324

0.235
0.390
0.169
0.066
0.007

2.371
1.036

0.473
0.527

-
-

0.036
0.617
0.341

0.647
0.293

0.281
0.599

-
-

0.599
0.317

0.006

-

0.994

-

0.653
0.281

0.198
0.455
0.150
0.078
0.006

110.1
2,223.3

17.7
0.217
0.104
12.266

484.3
2,623.4

19.9
0.217
0.052
0.752

0.43
0.80

0.67
0.67

-
-

0.72
0.71
0.83

0.36
0.48

0.43
0.23

-
-

0.34
0.52

0.88

-

0.88

-

0.31
0.43

0.43
0.25
0.65
0.70
0.88

0.31
0.28
0.32
0.26
0.22
0.02

2.464
1.423

0.309
0.361
0.227
0.103

0.254
0.436
0.309

0.780
0.189

0.543
0.275
0.117
0.027

0.780
0.189

0.137
0.261
0.601

-

0.794
0.165

0.144
0.309
0.388
0.089
0.021

2.348
1.332

0.310
0.358
0.235
0.097

0.245
0.435
0.319

0.784
0.187

0.555
0.271
0.113
0.026

0.898
0.184

0.129
0.265
0.606

-

0.803
0.158

0.152
0.313
0.387
0.084
0.019

530.2
2,833.9

12.0
0.218
0.299
1.023

1,224.8
3,045.6

17.5
0.257
0.074
0.754

0.12
0.06

0.99
0.94
0.80
0.80

0.80
0.98
0.79

0.91
0.95

0.77
0.91
0.88
0.90

0.83
0.87

0.76
0.93
0.90

-

0.77
0.82

0.80
0.92
0.98
0.81
0.91

0.44
0.76
0.05
0.47
0.01
0.36

Notes: Matching using 1:1 nearest-neighbor matching procedure that requires an exact match for all the
categorical variables. p.c. = per capita.

Estimates of the spatial DID coefficients in Table 1.6 are all negative and statisti-

cally significant, which strongly suggests that the concentration of mine openings leads to

a decrease in rental prices of houses nearby. Proximity to the health risks and local envi-

22

Table 1.6. Spatial DID Estimation Results

Only SMLs

Mine Openings

Only Artisanal
Mine Openings

(8)

(9)

(7)

-0.107∗
(0.048)
-0.030∗∗
(0.008)
-0.046
(0.028)

×

-0.093∗∗
(0.038)
-0.035∗∗∗
(0.006)
-0.016
(0.019)

×
×

-0.084∗
(0.042)
-0.040∗∗
(0.011)
-0.006
(0.029)

×
×
9,927
1,335
8,592
0.67

Treatments

1[Treated] × 1[2016]

Air Pollution Index

Water Pollution Index

Controls
Region FE
Region × Year FE
Observations
Treatment
Control

R2

(1)

-0.099∗∗
(0.028)
-0.033∗∗
(0.011)
-0.051∗
(0.025)

×

All-Mine
Openings

(2)

(3)

-0.087∗∗
(0.036)
-0.035∗∗∗
(0.006)
-0.033
(0.019)

×
×

-0.098∗
(0.048)
-0.041∗∗∗
(0.009)
-0.026
(0.026)

×
×

(4)

-0.152∗
(0.069)
-0.033∗∗
(0.007)
-0.021
(0.025)

×

(5)

-0.170∗
(0.078)
-0.027∗∗∗
(0.004)
-0.011
(0.015)

×
×

(6)

-0.250∗∗
(0.053)
-0.039∗∗
(0.008)
0.025
(0.020)
×
×
8,906
314
8,592
0.66

11,966
3,374
8,592
0.64

11,966
3,374
8,592
0.65

11,966
3,374
8,592
0.65

8,906
314
8,592
0.65

8,906
314
8,592
0.66

9,927
1,335
8,592
0.66

9,927
1,335
8,592
0.67

Notes: Using rental prices in logs as the response variable. Air and water pollution are the z-scores of their
respective pollution indexes. Controls include house-level characteristics and city-level attributes described
in table 1.3. Clustered standard errors by region in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05,
∗∗∗p < 0.001.

ronmental degradation associated to an expansion of mining activities requires renters to be

compensated with lower rental prices to attract them to these places. Indeed, findings in

columns (1), (2) and (3) reveal that rental prices are on average 9-10 percent lower in cities

experiencing all-mine openings. In other words, renters are compensated for living near the

environmental disamenity with respect to the type of mine located nearby. When distin-

guishing between different site openings, results indicate that this compensation is subject

to the scale of production of the extraction site. Results in columns (4), (5) and (6) show

that the concentration of SMLs mine openings lead to rental prices that are 15-25 percent

lower, while results in columns (7), (8) and (9) indicate that artisanal mine openings reduces

rental prices in around 8-11 percent.

Results for average MWTP for environmental quality improvements are also displayed

in Table 1.6. Capitalizations of the air pollution index into rental housing prices have the

expected sign and are similar in magnitude across the different treatments and specifications.

Estimated coefficients for the air quality index strongly suggest that air quality represents

an amenity to renters. On average, a one standard-deviation improvement in the air quality

23

index increases rental prices by around 2.7-4.1 percent. Capitalizations of the water pollution

index into rental prices also produce the expected sign, however, they are all statistically

insignificant. Weak evidence for the water pollution index suggest that a one standard-

deviation improvement in the water pollution index is related to a 0.6-5.1 percent increase

in rental prices. Limited awareness across households regarding recreational water pollution,

and irrelevance of this characteristic when it comes to households’ location decisions (Boyle

and Kiel, 2001), might both explain the statistically insignificance of these coefficients.

Estimates for the average treatment effect of extraction site openings using a spatial

DIDNNM estimator in Table 1.7 consistently suggest that the concentration of these openings

constitutes an environmental disamenity for households. These estimates are all statistically

significant at the 5% or less for all-mine openings and for artisanal mine openings. Results

in column (1) suggest that rental prices of houses located in areas that concentrate mine

openings of any type are on average 33.5 percent lower. Once the hedonic framework is in-

troduced in columns (2), (3) and (4), together with measurements of environmental quality,

the fit of the regression improves (R2 = 0.7) and the estimated effect of all-mine openings

is reduced to a 14-19 percent. Results for SMLs mine openings in column (5) reveal that

proximity to the concentration of these sites reduces rental prices by 15.7 percent, yet the

significance of this effect disappears after considering a rich set of controls, potentially be-

cause of the small sample size. In any case, weak evidence of lower rental prices in areas

with SMLs site openings is likely explained by the higher economic impacts that these mines

tend to generate locally (Aragon and Rud, 2013; Loayza and Rigolini, 2016), and whose

capitalization into rental prices might to a large extent offset the disamenity effect of their

proximity.

The previous situation is not necessarily the case, however, of proximity to artisanal

mine openings. Results in columns (9)-(12) suggest that the net effect of these openings

constitutes a disamenity to households. The estimated average treatment effect in column (9)

indicates that proximity to these openings reduces rental prices by 47.5 percent on average.

24

Table 1.7. Spatial DIDNNM Estimation Results

All-Mine Openings

Only SMLs Mine Openings

Only Artisanal Mine Openings

Treatments
1[Treated] × 1[2016]

Air Pollution Index

(1)

-0.335∗∗
(0.084)

Water Pollution Index

(2)

-0.151∗∗
(0.036)
-0.045∗∗
(0.018)
-0.057∗∗
(0.020)

Controls
Region FE
Region × Year FE
Observations
Treatment
Control

R2

7,854
2,666
5,188
0.22

7,771
2,649
5,122
0.72

(3)

(4)

-0.139∗∗
(0.044)
-0.051∗∗
(0.012)
-0.033
(0.026)

×
×

7,771
2,649
5,122
0.73

-0.187∗∗
(0.056)
-0.048∗∗
(0.017)
-0.048
(0.045)

×
×
7,771
2,649
5,122
0.73

(5)

-0.157∗
(0.075)

-0.158
(0.105)
-0.036∗
(0.016)
-0.028
(0.042)

×

570
256
314
0.16

547
250
297
0.46

(6)

(7)

(8)

-0.123
(0.139)
-0.037∗
(0.018)
-0.029
(0.038)

×
×

547
250
297
0.50

-0.103
(0.142)
-0.075∗∗
(0.015)
-0.046
(0.059)

×
×
547
250
297
0.51

(9)

-0.475∗∗
(0.137)

(10)
-0.171∗∗
(0.047)
-0.044∗∗
(0.018)
-0.034
(0.030)

×

3,891
1,061
2,830
0.19

3,863
1,060
2,803
0.74

(11)
-0.173∗∗
(0.050)
-0.062∗∗
(0.013)
-0.011
(0.027)

×
×

3,863
1,060
2,803
0.76

(12)

-0.252∗∗∗
(0.029)
-0.045∗
(0.023)
-0.058
(0.035)

×
×
3,863
1,060
2,803
0.76

Notes: Using rental prices in logs as the response variable. Matching on every period using 1:1 nearest-neighbor matching procedure that requires
an exact match for all the categorical variables included in the original hedonic-price estimation. Air and water pollution are the z-scores of their
respective pollution indexes. Controls include house-level characteristics and city-level attributes described in table 1.5. Clustered standard errors
by region in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001.

25

The disadvantage of artisanal mines relative to SMLs mines in terms of environmental best

practices together with their sporadic local economic impacts explain the magnitude of this

effect. Once the hedonic framework and measures on environmental quality are introduced,

the magnitude of this effect decreases although results in columns (10), (11) and (12) still

suggest a decrease in rental prices due to these openings. On average, households living near

artisanal sites are compensated with rental prices that are around 17 and 25 percent lower.

In terms of average MWTP measures, results in Table 1.7 are consistent with the

results in Table 1.6. They have the expected sign, although they are statistically significant

only for the air quality index. Average MWTP estimates for air pollution suggest that a

one standard-deviation improvement in the air quality index increases rental prices around

3.6-7.5 percent, while a similar improvement in the water quality index increases rental prices

around 1.1-5.7 percent, although this last effect is not statistically different from zero.

1.5.2. Heterogeneous Effects Among Short-term and Long-term

Residents

To dismiss the possibility of deriving biased MWTP estimates due to the sorting of

households across cities as a response to differences in their preferences for environmental

disamenities, households are split into long- and short-term residents in both treated and

control cities. Households in 2015 CASEN report their 2010 city of residence, which allows

the distinction of households into either long-term residents (i.e. households living in their

current city for at least five years or more), or short-term residents (i.e. households that

during the last five years moved into their current city). This distinction among residents

allows to capture whether short-term residents of mining areas moved there because of their

26

reduced sensitivity to environmental quality problems. Figure 1.2 shows the distribution of

perceptions of both air and recreational water pollution reported by long- and short-term

residents of cities in the first treatment group. Long-term residents seem to have lower per-

ceptions of frequent air quality problems in their neighborhoods, while they seem to be more

sensitive to the frequency of water quality problems, although differences across group are

not highly clear.

Figure 1.2. Distribution of Households’ Perceptions of Environmental Pollution in Mining
Cities by Type of Resident

Notes: Using 2015 households that declare to rent a house.

Table 1.8 shows average rental prices, wage, and years of education of both long-term

27

and short-term residents in mining cities. Short-term residents of cities experiencing new

mine sitings report higher wages and higher rental prices than long-term residents. On one

hand, if households moving into less pleasant areas have strong preferences for environmental

amenities, they need to be wage-compensated for moving into these places (Roback, 1982).

This idea is consistent, at first, with results in Table 1.8, although results for educational

levels for this group in the same table confounds the previous explanation. On the other

hand, if households exhibit lower preferences for environmental amenities, they can move

into unpleasant areas without necessarily being rent-compensated, as suggested by Table

1.8. Whether this is the case, it implies that the estimation of the rent gradient including

these households underestimate the true rent difference for those with strong dislike for

environmental disamenities. To test this idea, Table 1.9 presents the DID and DIDNNM

results allowing for the treatment effect to interact with long-term residents.

Table 1.8. Rental Prices, Wages, and Education in Mining Cities by Type of Resident

Rental Price (CH$)
Wage (CH$)
Education (years)

Long-term Residents Short-term Residents p−value
0.000
0.000
0.000

160,446
769,592
11.99

208,297
1,009,162

13.79

N

1,739

591

Notes: Using 2015 households that declare to rent a house. Mining cities as defined by treatment 1. p-values
on the difference in means across groups.

Estimates of heterogeneous average treatment effects across long- and short-term

residents of mining cities indicate that long-term residents need higher rent compensations

than short-term residents for living near mining. These results are strongly statistically

significant for residents of cities experiencing all-mine openings and artisanal mine openings,

28

Table 1.9. Heterogenous Average Treatment Effects by Type of Resident

All-Mine Openings

Only SMLs Mine Openings

Only Artisanal Mine Openings

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Treatments
1[Treated] × 1[2016] × 1[Long-term]

Air Pollution Index

Water Pollution Index

Controls
Region FE
Region × Year FE
Estimation
Observations
R2

-0.119∗∗∗
(0.023)
-0.041∗∗∗
(0.009)
-0.027
(0.028)

×
×
DID
11,954
0.66

-0.096∗∗
(0.029)
-0.045∗∗
(0.017)
-0.072∗∗
(0.020)

-0.085∗∗
(0.024)
-0.047∗∗
(0.014)
-0.068
(0.039)

-0.083∗∗
(0.024)
-0.049∗∗∗
(0.011)
-0.049∗
(0.025)

-0.125
(0.126)
-0.040∗∗
(0.008)
0.024
(0.021)
×
×
DIDNNM DIDNNM DIDNNM DID
8,894
0.66

6,636
0.70

6,636
0.72

6,636
0.72

×
×

×
×

×

-0.030
(0.163)
-0.038∗
(0.017)
-0.041
(0.040)

-0.037
(0.155)
-0.036
(0.023)
-0.066
(0.040)

-0.039
(0.150)
-0.076∗∗
(0.018)
-0.102
(0.072)

-0.073∗∗
(0.015)
-0.041∗∗
(0.012)
-0.006
(0.031)

×

×
×

×
×
DIDNNM DIDNNM DIDNNM DID
9,917
0.68

520
0.53

520
0.51

520
0.47

×
×

(10)
-0.068∗
(0.035)
-0.035
(0.023)
-0.044
(0.030)

×

(11)
-0.068∗∗
(0.026)
-0.052∗∗
(0.017)
-0.013
(0.027)

×
×

(12)
-0.071∗∗
(0.029)
-0.041
(0.025)
-0.060
(0.034)

×
×

DIDNNM DIDNNM DIDNNM

2,989
0.75

2,989
0.76

2,989
0.76

Notes: DID = Difference-in-Difference; DIDNNM = Difference-in-Difference Nearest Neighborhood Matching. Using rental prices in logs as the
response variable. Matching on every period using 1:1 nearest-neighbor matching procedure that requires now an exact match by type of resident.
Air and water pollution are the z-scores of their respective pollution indexes. Controls include house-level characteristics and city-level attributes
described in table 1.5. Clustered standard errors by region in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001.

29

yet not for residents of cities in SMLs openings where the sample is smaller. Table 1.9 shows

that market capitalizations of proximity to mining sites are between 3-10 percentage points

higher among households residing in these cities for a longer time. This might be because

these households are more aware of the changes in their neighborhoods’ environmental quality

after new extraction sites are opened, or because households that are less sensitive to the

environmental disamenity sort into these cities without requiring a rent compensation at the

same level as long-term residents. In either case, results in Table 1.9 reveal that the rent

gradient previously estimated in Tables 1.7 and 1.6 is an average rent gradient, averaged

across different subgroups of population.

1.5.3. Average Marginal Willingness to Pay Measures

Monthly average MWTP measures for local air and water quality improvements are

displayed in Table 1.10, estimated using findings in columns (4), (8), and (12) of the spatial

DIDNNM estimation previously shown in Table 1.7. As natural resource extraction sites are

widely known not only for their negative impacts on local air and water quality, but also for

disrupting aesthetics, polluting soil sediments, and disturbing the local social environment,

this work also calculates MWTP measures to avoid proximity to these sites as an overall

measure of the local impact of mining going beyond air and water quality degradation.

Estimates in Table 1.10 suggest that renters are monthly compensated in around

USD 52-94 (CH$ 36,240-65,661) for living in the vicinity of new mining activity. In other

words, individuals in Chile are willing to monthly pay a premium to avoid proximity to

mining. These results are statistically significant at the 5% level or less. In terms of implicit

prices of environmental quality, results in Table 1.10 consistently show environmental quality

30

Table 1.10. Average Marginal Willingness to Pay Measures

Treatments

Proximity to Mines (CH$)

Only SMLs

All-Mine
Only Artisanal
Openings Mine Openings Mine Openings
-65,661.62∗∗∗
-36,240.71∗∗

-9,730.54

USD

Air Quality Improvement (CH$)

-51.8

-9,250.07∗∗∗

-13.9

-7,020.90∗∗∗

USD

Water Quality Improvement (CH$)

USD

-13.2

-9,378.56

-13.4

-10.0

-4,302.93

-6.1

-93.8

-11,817.65∗∗

-16.9

-15,025.81

-21.5

Notes: Calculations using estimates in columns (4), (8), and (12) of table 1.7 for each of the treatments.
Monthly rental price calculated with the total number of observations used in obtaining the estimates. Dollar
value: CH$700. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001.

represents an amenity to renters. Individuals are willing to pay between USD 10-17 (CH$

7,021-11,818) a month for a unit improvement in the air quality index, and between USD

6-22 (CH$ 4,303-15,026) for a unit improvement in the water quality index, although these

calculations are only statistically significant for the air quality amenity. Implicit prices for

air quality are found higher in areas subject to the opening of artisanal extraction sites,

which is an indication that households living near artisanal mines need a higher compensa-

tion for unpleasant environmental quality relative to households living near other extraction

sites. The lower rental compensation is obtained in areas near SMLs mines. An air quality

improvement in these areas is associated with an increase of USD 10 (CH$ 7,021) in rental

prices, while in areas near artisanal sites this capitalization is expected to be around USD

17 (CH$ 11,818). This differential suggests that households are willing to pay 70 percent

more for air quality improvements when living near artisanal mining.

Estimated implicit prices of air quality are in line with previous estimates for PM10

found in Lav´ın, Dresdner, and Aguilar (2011), although three times higher. Lav´ın, Dresdner,

and Aguilar (2011) derive implicit prices for PM10 using objective measures of aggregated

31

coarse particulate matter concentrations, while this study uses households’ perceptions of air

pollution in general. Therefore, any potential difference in the magnitude of these estimates

is likely due to a different intensity with which households perceive air contamination in

their neighborhoods, which generally adds an upward bias to calculations of implicit prices

of environmental amenities using objective measures.

1.5.4. Robustness Checks

Previous results for the effects of mine openings on rental prices are robust to several

specification checks. Table 1.11 displays estimates of causal impacts using different matching

techniques. Consistent with results in Table 1.7, estimates for average treatment effects of

mine openings using additional matching methods are found all statistically significant and

with the expected sign for all-mine openings and artisanal mine openings. This reveals that

the capitalization of the disamenity effect of either a general opening or an artisanal opening

on rental prices is not conditional to the choice of the matching method. Rental prices

are estimated to be once again around 14-19 percent lower in areas near general openings,

while they are estimated to be 19-25 percent lower in areas near artisanal openings. Results

for SMLs mine openings, however, are not robust to other matching procedures. In terms

of environmental quality, capitalizations of air quality also produce the expected sign and

are statistically significant in most of the cases.

In contrast, estimates for water quality

improvements are weak, and with opposite signs in some cases.

If lower rental prices in areas with extraction site openings reflect the disamenity

effect of the concentration of new mining activity, rental properties in areas with a constant

number of mines should have a price vector that shows no significant variations during this

32

Table 1.11. Robustness Check: Different Matching Procedures

All-Mine Openings

Only SMLs Mine Openings

Only Artisanal Mine Openings

Treatments

1[Treated] × 1[2016]

Air Pollution Index

Water Pollution Index

Controls
Region × Year FE
Observations
Treatment
Control

R2

Five

Neighbors
-0.148∗∗
(0.064)
-0.050∗∗
(0.013)
0.039
(0.034)
×
×

14,125
1,840
12,285
0.71

Five

-0.179∗
(0.084)
-0.037∗∗
(0.011)
-0.076
(0.055)

Caliper Euclidean
(0.0001) Distance Neighbors
-0.241∗
-0.187∗∗
(0.118)
(0.056)
-0.048∗∗
-0.066∗∗
(0.019)
(0.017)
0.001
-0.048
(0.045)
(0.032)
×
×
1,015
165
850
0.58

×
×
9,820
2,649
7,171
0.78

×
×
7,771
2,649
5,122
0.73

Five

-0.103
(0.142)
-0.075∗∗
(0.015)
-0.046
(0.059)

Caliper Euclidean
(0.0001) Distance Neighbors
-0.193∗∗
(0.039)
-0.053∗∗
(0.023)
0.050
(0.037)
×
×
7,026
763
6,263
0.71

-0.241∗
(0.107)
-0.041
(0.061)
0.063
(0.129)
×
×
533
250
283
0.54

×
×
547
250
297
0.51

Caliper Euclidean
(0.0001) Distance
-0.244∗∗
-0.252∗∗∗
(0.050)
(0.029)
-0.045∗
0.042
(0.026)
(0.023)
-0.101∗
-0.058
(0.048)
(0.035)

×
×
3,863
1,060
2,803
0.76

×
×
6,304
1,060
5,244
0.74

Notes: Matching on every period. All matching procedures require an exact match in all categorical variables included in the original hedonic-price
estimation. Both Caliper and Euclidean distance matching require a 1:1 match. Clustered standard errors by region in parentheses. Significance
levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001.

33

Table 1.12. Robustness Check: Placebo Group

1[Constant#Mines] × 1[2016]

Air Pollution Index

Water Pollution Index

Controls
Region × Year FE
Observations
Treatment
Control

R2

Original

Matched Sample

Sample

-0.046
(0.052)
-0.044∗∗
(0.011)
0.037
(0.021)
×
×
9,564
972
8,592
0.65

One

Five

Neighbor Neighbors

Caliper

Euclidean
Distance

-0.131
(0.084)
-0.056∗
(0.024)
0.032
(0.063)
×
×
1,627
759
868
0.58

-0.091
(0.074)
-0.048∗∗
(0.016)
0.083∗∗
(0.031)

×
×
2,920
481
2,439
0.59

-0.131
(0.084)
-0.056∗
(0.024)
0.032
(0.063)
×
×
1,627
759
868
0.58

-0.032
(0.042)
-0.037
(0.020)
0.051
(0.070)
×
×
1,686
759
927
0.52

Notes: Matching on every period. All matching procedures require an exact match in all categorical variables
included in the original hedonic-price estimation. Both Caliper and Euclidean distance matching require a
1:1 match. Clustered standard errors by region in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05,
∗∗∗p < 0.001.

period relative to properties located in control cities, ceteris paribus. Using this rationale,

a falsification test in which cities with a constant number of mines during the study period

are considered as treated cities to be compared against control cities as an additional check.

Results for this placebo test are displayed in Table 1.12. Estimates for the effect of a constant

number of extraction sites over time are found to be negative but statistically insignificant in

all the cases. These results constitute strong evidence about the validity of previous findings

as they indicate that the effect of a placebo treatment on rental prices is not statistically

different from zero.

An additional concern in this study is that the concentration of mine openings is

boosting new housing construction and leading to depressed prices of the older rental housing

stock. To rule out the possibility that this study’s findings are merely reflecting a new

34

Table 1.13. Robustness Check: Certificates of Occupancy

Treatments

1[Treated] × 1[2015]
1[Treated] × 1[2016]

Cities FE
Region × Year FE
Observations

All-Mine
Openings

Only SMLs

Only Artisanal
Mine Openings Mine Openings

(1)

(2)

(3)

(4)

(5)

(6)

-208.4
(132.4)

-111.6
(139.3)

-98.87
(124.1)

-67.33
(62.62)

×
×
435

×
×
442

-51.30
(46.25)

×
×
363

70.37
(122.0)
×
×
360

×
×
365

×
×
367

Notes: Using the # of certificates of occupancy as the response variable. Clustered standard errors by region
in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001.

housing price vector due to a supply expansion, Table 1.13 displays the results for a spatial

DID estimation regression of the number of certificates of occupancy issued during the study

period on all the treatments, the post-treatment period, and a time-invariant unobserved

heterogeneity. Certificates of occupancy verify the compliance of new housing with applicable

housing codes, indicating whether these new properties are suitable conditions for occupancy.

Thus, the number of these certificates issued every year represents a proxy of the stock of

housing in each locality. Findings for the average treatment effect of mine openings on the

number of issued certificates are all statistically insignificant, irrespective of whether year

2015 or 2016 is considered. This means that there is not statistical evidence that the number

of buildings suitable for occupancy before and after the opening of new mines is different

in treated cities relative to control cities. Results in Table 1.13 dismiss the possibility that

a new housing price vector resulting from an expansion of the housing supply due to new

mining activity is confounding identification. For the validity of this robustness exercise, an

overview of the common trends assumption for the number of these certificates issued during

35

pre-treatment years is displayed in Figure A1.3 in the Appendix section.

1.6. Conclusions

This work documents rental market capitalizations of proximity to the concentration

of new extraction sites in emerging economies using Chile as an example. Using spatial

variation in the concentration of new extraction sites, rental prices of houses in cities hosting

new mining sitings are compared to rental prices of houses in cities with no mining records,

before and after the sitings. Results show that the disamenity effect of mine sitings out-

weighs the benefits that hosting economies get from new mining activity as capitalized into

rental prices. Findings strongly suggest that households living near new mining activity get

compensated with lower rental prices, indicating that households are willing to pay higher

housing prices to avoid proximity to these facilities. Further analyses show that this com-

pensation is higher for long-term residents of cities hosting new mine activity, a suggestion

of a preference-based sorting of households across space.

Implicit prices for environmental quality are also elicited in this study. Air quality and

water quality indexes are constructed using households’ perceptions of environmental pollu-

tion, which are subsequently related to objective measures on ambient degradation. Findings

strongly suggest that air quality is capitalized into higher rental prices, while evidence on

market capitalizations of recreational water quality is not conclusive.

Previous studies on the local impacts of the natural resource extraction are generally

centered around local labor market effects, poverty levels, and real income levels that in-

habitants of local economies hosting the extraction get access to. Yet, findings in this study

36

bring to light some of the negative environmental aspects of mining operations. Mining is

generally associated with increased environmental risks, aesthetics and visual impacts, and

soil degradation. Moreover, many of the resource extraction sites in developing countries

are left abandoned after operations cease due to the lack of a more stringent environmental

regulation or simply a weak law enforcement, increasing the risks of environmental disasters

such as mine drainages. One of the main contributions of this research is to show that

this activity represents an environmental disamenity in these countries as their unpleasant

effects are capitalized into market home rental prices. This suggests the need of more re-

search on the welfare effects of increased environmental risks due to mining operations when

quantifying the local economic impacts of this activity.

This study can be improved in several ways. Data availability on properties sold

during the treatment period can provide insights on the long-term effects of mine openings

by examining the housing market impacts instead of rental market effects. Availability of

rental prices for different periods can also approximate the previous idea by allowing the

study of duration effects of mine openings. Additional data on housing will allow future

scholars to establish whether negative net impacts of new mining activity persist over time.

37

APPENDIX

38

Figure A1.1. Resource Extraction Site Openings Between 2011-2016

Notes: Sites plotted with ArcMap 10.1 using information from Albers (2012), SERNAGEOMIN (2011) and
SERNAGEOMIN (2016). Red dots symbolize SMLs site openings and blue dots are for Artisanal sites.

39

Figure A1.2. Example of Treated, Control, and Placebo Cities, and Treatments

(a) Treated, Control, and Placebo Cities

(b) Treatments Example

Notes: Panel (a) depicts different scenarios that allow the classification of cities as either treated, control, or
placebo cities, or cities to dismiss from the analysis. Type-A cities are cities that over time have a constant
number of mining sites, and so they are part of a placebo group of cities. Type-B cities are cities that over
time experienced the closure and the opening of different types of mines, and therefore, they are dismissed
from the analysis. Type-C cities are cities that for each year show no records of mine sitings. These cities
belong to the control group of cities. Finally, type-D cities are cities that experienced the siting of a new
mine, and therefore, they are part of a treated group of cities. Panel (b) displays the treatments considered
in the analysis. Type-A cities are cities that host the opening of an artisanal mine only. Type-B cities are
cities that host the opening of a new SMLs mine and a new artisanal mine, that is, “all-mine openings”.
Type-C cities are control cities, while type-D cities are cities that host the opening of a SMLs mine only.

40

Table A1.1. Ridits on Air and Water Pollution Perceptions

Variable

Category

(1)

(2)

Air Pollution

Water Pollution

Never

Few times
Frequently

Always
Total

Never

Few times
Frequently

Always
Total

7,913
2,578
1,419
1,088
12,998

11,533

753
373
339

12,998

3,957
1,289
710
544

5,767
377
187
170

(3)

0

7,913
10,491
11,910
12,998

0

11,533
12,286
12,659
12,998

(4)

(5)

3,957
9,202
11,201
12,454

0.3044
0.7080
0.8617
0.9581

5,767
11,910
12,473
12,829

0.4436
0.9163
0.9596
0.9870

Notes: Based on Bross (1958), the approach assigns the continuous index number as half the share of the
observations in each response category’s upper and lower cumulative percentiles. For example, column (1)
has the frequency distribution of the identified distribution. In this case, the identified distribution considers
the national population of households renting a house. Column (2) has one-half of the corresponding entry
in column (1). Column (3) has the cumulate of column (1) displaced one category upward. Column (4) =
column (2) + column (3). Entries in column (5) represent the ridits equivalent to column (4) divided by
grand total in (1).

Table A1.2. Average Marginal Effects on Pollution Perceptions

Response Variable

Covariates

Air Pollution (N = 311)

PM25

CO

NOX

CO2

Coeff.
.00411∗

-.00075∗∗

.00040∗

.000003∗

Water Pollution (N = 312)

TSS

-.000014∗∗∗
Streamflow -.064879∗
-.003497∗∗∗
Density

p-value

0.08

0.03

0.09

0.05

0.00
0.08
0.00

Notes: Results from a cross-section Bernoulli QMLE fractional logit regression for 2015 on cities’ average
pollution perceptions. p-values calculated using standard errors robust to a misspecification of the Bernoulli
generalized linear model variance assumption. Variables on pollutant emissions are measured in thousand of
tons/year. Streamflow refers to the regional monthly average streamflow of main rivers measured in L/seg.
Density is measured in #thousand of inab./km2

41

Table A1.3. Spatial DID Full Estimation Results

Treatments

All-Mine Openings

Only SMLs Mine Openings

Only Artisanal Mine Openings

(1)

(2)

(3)

(4)

(5)

(6)

(7)

0.006
(0.034)
0.411∗∗∗
(0.026)
-0.086∗∗
(0.036)
-0.035∗∗∗
(0.006)
-0.033
(0.019)
0.091∗∗∗
(0.014)
0.293∗∗∗
(0.029)
0.144∗∗∗
(0.014)
0.143∗∗∗
(0.023)
0.367∗∗∗
(0.059)
0.077∗
(0.038)
0.182∗∗∗
(0.039)
0.117∗∗
(0.039)
-0.017
(0.055)
0.111∗∗
(0.027)
0.039∗
(0.021)
0.235∗∗∗
(0.026)
0.219∗∗∗
(0.019)
0.325∗∗∗
(0.023)
0.156∗∗∗
(0.0226)
0.044∗∗
(0.0160)
0.120∗∗∗
(0.011)
0.505∗
(0.246)
0.438
(0.255)
0.362
(0.247)
0.293
(0.294)
0.059∗∗
(0.015)

0.048
(0.081)
0.375∗∗∗
(0.020)
-0.170∗
(0.078)
-0.027∗∗∗
(0.004)
-0.011
(0.015)
0.094∗∗∗
(0.015)
0.285∗∗∗
(0.029)
0.151∗∗∗
(0.014)
0.145∗∗∗
(0.021)
0.325∗∗
(0.068)
0.094∗∗
(0.041)
0.115∗∗
(0.039)
0.059
(0.048)
-0.062
(0.062)
0.130∗∗
(0.031)
0.052∗∗
(0.021)
0.207∗∗∗
(0.034)
0.221∗∗∗
(0.025)
0.327∗∗∗
(0.040)
0.159∗∗∗
(0.0308)
0.042∗∗
(0.0137)
0.122∗∗∗
(0.005)
0.763∗∗∗
(0.083)
0.715∗∗∗
(0.084)
0.626∗∗∗
(0.069)
0.264∗∗
(0.054)
0.055∗∗
(0.017)

0.071
(0.078)
0.829∗∗∗
(0.076)
-0.250∗∗
(0.053)
-0.039∗∗
(0.008)
0.025
(0.020)
0.095∗∗∗
(0.015)
0.284∗∗∗
(0.028)
0.144∗∗∗
(0.015)
0.135∗∗∗
(0.023)
0.313∗∗
(0.067)
0.079∗
(0.041)
0.121∗∗
(0.038)
0.065
(0.048)
-0.055
(0.062)
0.130∗∗
(0.032)
0.052∗∗
(0.021)
0.206∗∗∗
(0.034)
0.219∗∗∗
(0.027)
0.321∗∗∗
(0.040)
0.155∗∗
(0.0342)
0.045∗∗
(0.0152)
0.124∗∗∗
(0.005)
0.768∗∗∗
(0.077)
0.723∗∗∗
(0.075)
0.626∗∗∗
(0.062)
0.275∗∗∗
(0.048)
0.054∗∗
(0.017)

0.069
(0.070)
0.412∗∗∗
(0.030)
-0.107∗
(0.048)
-0.030∗∗
(0.008)
-0.046
(0.028)
0.089∗∗∗
(0.015)
0.290∗∗∗
(0.033)
0.142∗∗∗
(0.021)
0.144∗∗∗
(0.026)
0.387∗∗∗
(0.074)
0.121∗∗
(0.046)
0.151∗∗
(0.038)
0.090∗∗
(0.035)
-0.014
(0.070)
0.121∗∗
(0.028)
0.051∗
(0.024)
0.282∗∗∗
(0.032)
0.267∗∗∗
(0.031)
0.387∗∗∗
(0.050)
0.208∗∗∗
(0.0243)
0.039∗∗
(0.0136)
0.120∗∗∗
(0.010)
0.946∗∗∗
(0.113)
0.898∗∗∗
(0.110)
0.791∗∗∗
(0.094)
0.761∗∗
(0.265)
0.063∗∗
(0.014)

0.011
(0.041)
0.619∗∗∗
(0.078)
-0.098∗
(0.048)
-0.041∗∗∗
(0.009)
-0.026
(0.026)
0.090∗∗∗
(0.014)
0.294∗∗∗
(0.029)
0.142∗∗∗
(0.014)
0.141∗∗∗
(0.024)
0.362∗∗∗
(0.0579)
0.073∗
(0.038)
0.181∗∗∗
(0.038)
0.117∗∗
(0.039)
-0.016
(0.054)
0.112∗∗
(0.027)
0.040∗
(0.021)
0.231∗∗∗
(0.025)
0.216∗∗∗
(0.020)
0.319∗∗∗
(0.022)
0.151∗∗∗
(0.0228)
0.045∗∗
(0.0164)
0.119∗∗∗
(0.011)
0.510∗
(0.237)
0.448∗
(0.244)
0.367
(0.237)
0.302
(0.283)
0.058∗∗
(0.015)

-0.028
(0.112)
0.394∗∗∗
(0.027)
-0.152∗
(0.069)
-0.033∗∗
(0.007)
-0.021
(0.025)
0.095∗∗∗
(0.014)
0.290∗∗∗
(0.030)
0.160∗∗∗
(0.017)
0.154∗∗∗
(0.017)
0.345∗∗
(0.078)
0.126∗∗
(0.047)
0.122∗∗
(0.046)
0.061
(0.051)
-0.026
(0.082)
0.129∗∗
(0.029)
0.055∗∗
(0.021)
0.239∗∗∗
(0.030)
0.249∗∗∗
(0.027)
0.361∗∗∗
(0.047)
0.183∗∗∗
(0.0301)
0.041∗∗
(0.0161)
0.119∗∗∗
(0.006)
0.892∗∗∗
(0.110)
0.824∗∗∗
(0.113)
0.746∗∗∗
(0.096)
0.436∗∗∗
(0.067)
0.059∗∗
(0.015)

42

(8)
0.068∗
(0.030)
0.405∗∗∗
(0.027)
-0.093∗∗
(0.038)
-0.035∗∗∗
(0.006)
-0.016
(0.019)
0.090∗∗∗
(0.016)
0.275∗∗∗
(0.032)
0.138∗∗∗
(0.018)
0.140∗∗∗
(0.026)
0.347∗∗∗
(0.059)
0.095∗
(0.046)
0.172∗∗
(0.036)
0.108∗∗
(0.034)
-0.023
(0.049)
0.115∗∗
(0.030)
0.044∗
(0.024)
0.250∗∗∗
(0.034)
0.240∗∗∗
(0.028)
0.352∗∗∗
(0.040)
0.183∗∗∗
(0.0298)
0.047∗∗
(0.0128)
0.127∗∗∗
(0.007)
0.783∗∗∗
(0.089)
0.747∗∗∗
(0.086)
0.642∗∗∗
(0.069)
0.626∗
(0.290)
0.062∗∗
(0.017)

(9)
0.062∗
(0.028)
0.326∗∗∗
(0.006)
-0.084∗
(0.042)
-0.040∗∗
(0.011)
-0.006
(0.029)
0.090∗∗∗
(0.016)
0.275∗∗∗
(0.032)
0.138∗∗∗
(0.018)
0.139∗∗
(0.028)
0.341∗∗∗
(0.057)
0.089∗
(0.048)
0.173∗∗
(0.035)
0.109∗∗
(0.034)
-0.020
(0.050)
0.117∗∗
(0.030)
0.045∗
(0.023)
0.247∗∗∗
(0.034)
0.237∗∗∗
(0.028)
0.344∗∗∗
(0.039)
0.177∗∗∗
(0.0321)
0.048∗∗
(0.0141)
0.126∗∗∗
(0.007)
0.787∗∗∗
(0.085)
0.757∗∗∗
(0.077)
0.645∗∗∗
(0.065)
0.638∗
(0.289)
0.061∗∗
(0.017)

Dwelling Type (Base = Precarious)

Walls Material (Base = Precarious)

Drywall

Floors Material (Base = Precarious)

1[Treated]

Row Units

Regular Units

Apts. (elev.)

Apts. (no elev.)

# of Bedrooms

# of Bathrooms

Air Pollution Index

Concrete

Masonry

Water Pollution Index

1[2016]
1[Treated] × 1[2016]

0.018
(0.044)
0.419∗∗∗
(0.027)
-0.099∗∗
(0.028)
-0.033∗∗
(0.011)
-0.051∗
(0.025)
0.089∗∗∗
(0.014)
0.304∗∗∗
(0.029)
0.150∗∗∗
(0.017)
0.142∗∗∗
(0.023)
0.389∗∗∗
(0.067)
0.092∗∗
(0.037)
0.187∗∗
(0.044)
0.120∗∗
(0.039)
0.020
(0.067)
Walls Condition (Base = Bad)
0.115∗∗∗
(0.024)
0.041∗
(0.021)
0.267∗∗∗
(0.024)
0.246∗∗∗
(0.023)
0.370∗∗∗
(0.032)
0.188∗∗∗
(0.0244)
Floors Condition (Base = Bad)
0.036∗
(0.0193)
0.113∗∗∗
(0.014)
0.614∗
(0.292)
0.535
(0.300)
0.468
(0.290)
0.487
(0.324)
0.058∗∗
(0.013)

Roof Condition (Base = Bad)

Carpet

Cement

Regular

Good

Good

Regular

Wood

Tile

Roof Tiles

Concrete

Zincstrips

Clinkstone

Good

Roof Material (Base = Precarious)

Table A1.3 (cont’d)

All-Mine Openings

Only SMLs Mine Openings

Only Artisanal Mine Openings

Regular

(1)

0.036∗∗
(0.0148)

(2)

0.036∗∗
(0.0144)

(3)

0.035∗∗
(0.0144)

(4)

0.048∗∗
(0.0163)

(5)

0.046∗∗
(0.0173)

(6)

0.044∗∗
(0.0173)

(7)
0.041∗
(0.0183)

(8)
0.042∗
(0.0188)

(9)
0.041∗
(0.0182)

Dimension (Base = < 30m2)

30-40m2

41-60m2

61-100m2

101-150m2

> 150m2

Density

Waste Disposal

# of Parks

# of Public Squares

Crime

Poverty

Constant

0.0569
(0.0355)
0.148∗∗∗
(0.023)
0.251∗∗∗
(0.024)
0.333∗∗∗
(0.029)
0.369∗∗∗
(0.036)

0.0646
(0.0431)
0.152∗∗∗
(0.026)
0.240∗∗∗
(0.027)
0.321∗∗∗
(0.029)
0.347∗∗∗
(0.045)

-0.000001

0.0583
(0.0445)
0.150∗∗∗
(0.026)
0.234∗∗∗
(0.026)
0.316∗∗∗
(0.027)
0.347∗∗∗
(0.046)
0.00001

0.080
(0.053)
-0.065
(0.049)
-0.0003
(0.0007)

0.0659
(0.0424)
0.153∗∗∗
(0.026)
0.242∗∗∗
(0.027)
0.322∗∗∗
(0.029)
0.345∗∗∗
(0.044)
0.000008

0.0512
0.0564
0.0525
0.0544
(0.0367)
(0.0353)
(0.0348)
(0.0345)
0.142∗∗∗
0.146∗∗∗
0.152∗∗∗
0.153∗∗∗
(0.023)
(0.022)
(0.019)
(0.020)
0.246∗∗∗
0.250∗∗∗
0.253∗∗∗
0.255∗∗∗
(0.024)
(0.023)
(0.019)
(0.021)
0.341∗∗∗
0.333∗∗∗
0.343∗∗∗
0.339∗∗∗
(0.029)
(0.029)
(0.026)
(0.027)
0.374∗∗∗
0.366∗∗∗
0.364∗∗∗
0.365∗∗∗
(0.034)
(0.036)
(0.032)
(0.033)
-0.000002 -0.00002∗∗∗ -0.00002∗∗
-0.0000008 -0.00001∗∗∗ -0.00001∗∗
(0.000007) (0.000003) (0.000004) (0.000006) (0.000004) (0.000005) (0.000007) (0.000003) (0.000004)
0.185∗∗
0.148∗
(0.067)
(0.074)
-0.081∗∗
0.076
(0.082)
(0.026)
0.014
-0.0003
(0.0003)
(0.009)

0.195∗
(0.085)
-0.011
(0.051)
-0.001
(0.0007)
(0.009)
0.00002∗∗∗ 0.00002∗∗∗ 0.00002∗∗∗ 0.00004∗∗∗ 0.00004∗∗∗ 0.00004∗∗∗ 0.00002∗∗ 0.00002∗∗ 0.00002∗∗∗
(0.000004) (0.000004) (0.000004) (0.000004) (0.000005) (0.000004) (0.000005) (0.000004) (0.000003)
-0.016∗∗
-0.019∗∗
(0.005)
(0.006)
9.62∗∗∗
9.53∗∗∗
(0.17)
(0.35)

0.108
(0.077)
-0.032
(0.057)
-0.0005
(0.0005)

-0.010∗∗
(0.003)
8.86∗∗∗
(0.10)

0.111
(0.095)
-0.044
(0.056)
-0.00005
(0.0005)

-0.010∗∗
(0.004)
8.81∗∗∗
(0.10)

0.0545
(0.0349)
0.154∗∗∗
(0.021)
0.255∗∗∗
(0.021)
0.338∗∗∗
(0.027)
0.366∗∗∗
(0.033)

0.183∗∗
(0.070)
-0.044
(0.038)
-0.0005
(0.0003)

-0.016∗∗
(0.006)
10.13∗∗∗
(0.31)

-0.016∗∗
(0.006)
10.04∗∗∗
(0.31)

-0.012∗∗
(0.004)
9.17∗∗∗
(0.13)

0.053
(0.086)
0.179∗∗
(0.068)
-0.001

0.179∗∗
(0.076)
-0.015
(0.051)
0.00006
(0.009)

-0.015∗∗
(0.005)
9.47∗∗∗
(0.15)

-0.019∗∗
(0.006)
9.30∗∗∗
(0.19)

Region FE
Region × Year FE
Observations
R2

×

11,966
0.636

11,966
0.652

×

11,966
0.653

8,906
0.647

×

8,906
0.656

×
8,906
0.657

9,927
0.659

×

9,927
0.672

×
9,927
0.673

Notes: Using rental prices in logs as the response variable. Air and water pollution are the z-scores of their
respective pollution indexes. Clustered standard errors by region in parentheses. ∗p < 0.10, ∗∗p < 0.05,
∗∗∗p < 0.001

43

Table A1.4. Spatial DIDNNM Full Estimation Results

Treatments

All-Mine Openings

Only SMLs Mine Openings

Only Artisanal Mine Openings

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

1[Treated]

1[2016]
1[Treated] × 1[2016]

0.023
(0.042)
0.452∗∗∗
(0.033)
-0.151∗∗
(0.036)
-0.045∗∗
Air Pollution Index
(0.018)
Water Pollution Index -0.057∗∗
(0.020)
0.096∗∗∗
(0.016)
0.234∗∗∗
(0.031)

# Bathrooms

# Bedrooms

0.004
(0.033)
0.457∗∗∗
(0.039)
-0.138∗∗
(0.044)
-0.051∗∗
(0.012)
-0.033
(0.026)
0.094∗∗∗
(0.015)
0.229∗∗∗
(0.030)

Dwelling Type (Base = substandard)

Walls Material (Base = substandard)

Masonry

Drywall

Concrete

Row Units

Apts. (elev.)

Regular Units

Apts. (no elev.)

0.087
(0.112)
0.113
(0.124)
0.365∗
(0.193)
-0.056
(0.116)
0.159∗∗∗
(0.032)
0.061∗
(0.028)
-0.073
(0.051)
Walls Condition (Base = bad)
0.199∗∗
(0.047)
0.094∗∗
(0.028)
0.281∗∗∗
(0.033)
0.214∗∗∗
(0.032)
0.312∗∗∗
(0.064)
0.223∗∗
(0.055)
Floors Condition (Base = bad)

Wood

Tile

Carpet

Cement

Good

Regular

0.111
(0.107)
0.144
(0.120)
0.363∗
(0.180)
-0.035
(0.115)
0.169∗∗
(0.038)
0.070∗∗
(0.024)
-0.108∗
(0.054)
0.200∗∗
(0.049)
0.098∗∗
(0.029)
0.252∗∗∗
(0.031)
0.198∗∗∗
(0.028)
0.269∗∗∗
(0.052)
0.185∗∗
(0.049)

Floors Material (Base = substandard)

0.029
(0.035)
0.947∗∗∗
(0.206)
-0.187∗∗
(0.056)
-0.048∗∗
(0.017)
-0.048
(0.045)
0.096∗∗∗
(0.015)
0.229∗∗∗
(0.030)

0.111
(0.107)
0.144
(0.119)
0.355∗
(0.182)
-0.029
(0.118)
0.162∗∗
(0.037)
0.068∗∗
(0.024)
-0.104∗
(0.053)
0.200∗∗
(0.050)
0.096∗∗
(0.030)
0.238∗∗∗
(0.031)
0.186∗∗∗
(0.026)
0.256∗∗∗
(0.047)
0.171∗∗
(0.053)

0.095
(0.055)
0.026
(0.038)

Good

Regular

0.093
(0.058)
0.023
(0.040)

0.095
(0.055)
0.028
(0.040)

Roof Material (Base = substandard)

Concrete

-0.072
(0.053)
-0.239∗∗∗
(0.030)
Roofs Condition (Base = bad)

Zincstrips

Good

Regular

0.018
(0.021)
0.044
(0.026)

Dimension (Base = < 30m2)
-0.002

30-40m2

-0.070
(0.049)
-0.229∗∗∗
(0.030)

0.031
(0.017)
0.058∗∗
(0.022)

-0.072
(0.047)
-0.229∗∗∗
(0.029)
0.037∗
(0.017)
0.061∗∗
(0.019)

0.055
(0.045)
0.493∗∗∗
(0.058)
-0.171∗∗
(0.046)
-0.044∗∗
(0.018)
-0.034
(0.030)
0.085∗∗∗
(0.016)
0.176∗∗
(0.038)

0.066
(0.084)
0.155
(0.086)
0.339∗
(0.175)
0.058
(0.117)

0.074
(0.071)
0.013
(0.055)
-0.160∗∗∗
(0.032)
0.424∗∗∗
(0.063)
0.322∗∗
(0.074)
0.441∗∗∗
(0.065)
0.389∗∗∗
(0.047)
0.432∗∗∗
(0.084)
0.341∗∗
(0.107)

-0.062
(0.054)
-0.132∗∗
(0.045)
-0.070∗∗
(0.024)
-0.264∗∗∗
(0.044)

0.043
(0.097)
0.300∗∗∗
(0.056)
-0.158
(0.105)
-0.036∗
(0.016)
-0.028
(0.042)
0.111∗∗
(0.025)
0.370∗∗∗
(0.073)

0.012
(0.094)
0.292∗∗
(0.066)
-0.123
(0.139)
-0.038∗
(0.018)
-0.029
(0.038)
0.104∗∗
(0.023)
0.362∗∗∗
(0.061)

0.009
(0.100)
0.295
(0.298)
-0.103
(0.142)
-0.075∗∗
(0.015)
-0.046
(0.059)
0.109∗∗
(0.023)
0.378∗∗∗
(0.064)

-
-

-
-

-
-

-0.049
(0.043)

-0.056
(0.049)

-0.066
(0.047)

-
-

-
-

-
-

-0.084
(0.297)

-0.060
(0.326)

-0.078
(0.331)

-0.047
(0.270)
-0.018
(0.174)
-0.044
(0.215)
0.251∗∗∗
(0.047)
0.093
(0.106)

0.022
(0.063)
0.110∗
(0.050)
0.164
(0.489)
-0.022
(0.153)

0.037
(0.061)
-0.019
(0.066)

-0.032
(0.265)
-0.009
(0.173)
-0.037
(0.211)
0.255∗∗
(0.052)
0.102
(0.106)

0.023
(0.076)
0.101∗
(0.054)
0.072
(0.462)
0.022
(0.158)

0.062
(0.053)
-0.009
(0.058)

-0.215
(0.123)

-0.252
(0.138)

-0.011
(0.264)
-0.045
(0.180)
-0.042
(0.210)
0.208∗∗
(0.058)
0.081
(0.116)

0.021
(0.053)
0.134∗∗
(0.052)
0.188
(0.521)
-0.065
(0.199)

0.040
(0.055)
-0.019
(0.051)

-0.263∗
(0.139)

0.047
(0.059)
0.106
(0.081)

0.007
(0.057)
0.543∗∗∗
(0.067)
-0.173∗∗
(0.050)
-0.062∗∗
(0.013)
-0.011
(0.027)
0.089∗∗∗
(0.017)
0.162∗∗
(0.034)

0.060
(0.063)
0.136∗
(0.069)
0.293∗
(0.147)
0.018
(0.090)

0.089
(0.076)
0.043
(0.054)
-0.156∗∗
(0.040)
0.403∗∗∗
(0.058)
0.298∗∗
(0.069)
0.352∗∗
(0.078)
0.317∗∗
(0.067)
0.335∗∗
(0.096)
0.262∗
(0.131)

-0.047
(0.053)
-0.101∗∗
(0.0436)
-0.055∗∗
(0.022)
-0.229∗∗∗
(0.0386)

0.059
(0.041)
0.056
(0.065)
0.091∗

0.068
(0.048)
0.313∗∗∗
(0.012)
-0.252∗∗∗
(0.029)
-0.045∗
(0.023)
-0.058
(0.035)
0.090∗∗∗
(0.017)
0.165∗∗
(0.037)

0.067
(0.067)
0.143∗
(0.071)
0.299∗
(0.153)
0.033
(0.095)

0.081
(0.076)
0.032
(0.059)
-0.153∗∗
(0.046)
0.409∗∗∗
(0.054)
0.305∗∗
(0.067)
0.341∗∗
(0.078)
0.299∗∗
(0.068)
0.320∗∗
(0.094)
0.248
(0.137)

-0.042
(0.060)
-0.101∗
(0.052)
-0.066∗∗
(0.022)
-0.234∗∗∗
(0.039)

0.050
(0.039)
0.049
(0.062)

0.080

0.023
(0.040)
0.112
(0.067)

0.009
(0.045)
0.112
(0.075)

0.044
(0.054)
0.055
(0.076)

-0.012

-0.005

0.026

0.053

0.056

0.069

44

Treatments

41-60m2

61-100m2

101-150m2

> 150m2

Density

Waste Disposal

# of Parks

# of Squares

Crime

Poverty

Constant

Table A1.4 (cont’d)

All-Mine Openings

Only SMLs Mine Openings

Only Artisanal Mine Openings

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(0.052)
0.104∗
(0.052)
0.182∗∗
(0.051)
0.285∗∗
(0.064)
0.239∗∗
(0.088)

-0.000005
(0.00001)

(0.048)
0.102∗
(0.049)
0.179∗∗
(0.046)
0.280∗∗
(0.062)
0.239∗∗
(0.089)
-0.00001
(0.00001)

(0.045)
0.111∗∗
(0.047)
0.188∗∗
(0.041)
0.288∗∗∗
(0.057)
0.244∗∗
(0.086)
-0.00002
(0.00001)

(0.061)
(0.068)
0.135∗∗
0.109
(0.044)
(0.059)
0.236∗∗∗
0.216∗∗∗
(0.040)
(0.040)
0.416∗∗∗
0.323∗∗∗
(0.063)
(0.057)
0.225
0.191
(0.136)
(0.127)
0.00003
-0.00001
(0.00002) (0.00003)

(0.039)
0.166∗∗
(0.043)
0.233∗∗∗
(0.030)
0.341∗∗∗
(0.038)
0.297∗∗
(0.072)

(0.044)
(0.063)
0.183∗∗
0.140∗∗
(0.044)
(0.042)
0.250∗∗∗
0.230∗∗∗
(0.032)
(0.040)
0.332∗∗∗
0.411∗∗∗
(0.041)
(0.072)
0.292∗∗
0.288∗∗
(0.079)
(0.111)
-0.00002 -0.0000009 -0.00004∗∗ -0.00005∗∗
(0.00001)
(0.0001)

(0.041)
0.190∗∗
(0.042)
0.260∗∗∗
(0.027)
0.344∗∗∗
(0.043)
0.295∗∗
(0.070)

(0.00002)

(0.00001)

0.249
-0.040
(0.218)
(0.112)
-1.210∗∗
0.199∗∗
(0.442)
(0.060)
-0.001∗∗
-0.001∗∗
(0.0004)
(0.001)
0.00002∗∗ 0.00002∗∗∗ 0.00002∗∗∗ 0.00004∗∗
(0.000004) (0.000004)
-0.021∗∗
-0.025∗∗
(0.008)
(0.008)
10.66∗∗∗
11.03∗∗∗
(0.374)
(0.329)

0.196
(0.267)
-0.009
(0.111)
0.003
(0.035)
0.00003∗∗
(0.00002) (0.00003) (0.00002) (0.000008) (0.000006) (0.000007)
-0.018∗∗
(0.004)
10.61∗∗∗
(0.276)

0.215
(0.243)
-0.126
(0.089)
-0.007
(0.037)
0.00002∗∗
-0.018∗∗
(0.004)
10.41∗∗∗
(0.250)

0.263
(0.240)
0.098∗
(0.047)
0.006
(0.030)
0.000003
-0.032∗∗
(0.008)
10.57∗∗∗
(0.325)

0.145
(0.229)
-0.696
(0.416)
-0.001
(0.001)
0.00004
-0.011∗∗
(0.004)
10.14∗∗∗
(0.249)

(0.000004)
-0.021∗∗
(0.007)
10.86∗∗∗
(0.302)

0.194
(0.210)
-0.637
(0.429)
-0.0002
(0.0005)
0.00003

-0.009
(0.006)
10.13∗∗∗
(0.254)

0.140
(0.116)
-0.077
(0.078)
-0.001∗∗
(0.0003)

0.162
(0.134)
-0.171∗∗
(0.069)
-0.001∗
(0.0003)

-0.009
(0.005)
10.40∗∗∗
(0.289)

Region FE
Region × Year FE
Observations
R2

7,771
0.717

×

7,771
0.731

×
7,771
0.733

×

547
0.464

547
0.497

×
547
0.507

3,863
0.743

×

3,863
0.756

×
3,863
0.759

Notes: Using rental prices in logs as the response variable. Matching on every period using 1:1 nearest-
neighbor matching procedure that requires an exact match for all the categorical variables included in the
original hedonic-price estimation. Air and water pollution are the z-scores of their respective pollution
indexes. Clustered standard errors by region in parentheses. ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

45

Figure A1.3. Overview of the Common Trends Assumption - Certificates of Occupancy

Notes: Excluding two cities with abnormalities in their reports: Huasco (all-mine openings) and Vitacura
(control). Source: SINIM (2017).

46

REFERENCES

47

REFERENCES

Albers, C. (2012). Coberturas SIG para la Ense˜nanza de la Geograf´ıa de Chile. Ed. by Uni-

versidad de la Frontera.

Anderson, Soren T and Sarah E West (2006). “Open Space, Residential Property Values,

and Spatial Context”. In: Regional Science and Urban Economics 36.6, pp. 773–789.

Angrist, Joshua D and J¨orn-Steffen Pischke (2008). Mostly Harmless Econometrics: An Em-

piricist’s Companion. Princeton University Press.

Aragon, Fernando M and Juan Pablo Rud (2013). “Natural Resources and Local Communi-
ties: Evidence From a Peruvian Gold Mine”. In: American Economic Journal: Economic
Policy 5.2, pp. 1–25.

Bebbington, A et al. (2008). “Mining and Social Movements: Struggles over Livelihood and
Rural Territorial Development in the Andes”. In: World Development 36.12, pp. 2888–
2905.

Berezansky, B, B Portnov, and B Barzilai (2010). “Objective vs. Perceived Air Pollution as
a Factor of Housing Pricing: A Case Study of the Greater Haifa Metropolitan Area”. In:
Journal of Real Estate Literature 18.1, pp. 99–122.

Blundell, R and M C Dias (2009). “Alternative Approaches to Evaluation in Empirical

Microeconomics”. In: Journal of Human Resources 44.3, pp. 565–640.

Boxall, Peter C, Wing H Chan, and Melville L McMillan (2005). “The Impact of Oil and
Natural Gas Facilities on Rural Residential Property Values: A Spatial Hedonic Analysis”.
In: Resource and Energy Economics 27.3, pp. 248–269.

Boyle, M and K Kiel (2001). “A Survey of House Price Hedonic Studies of the Impact of

Environmental Externalities”. In: Journal of Real Estate Literature.

Broner, Fernando, Paula Bustos, and Vasco Carvalho (2012). Sources of Comparative Advan-
tage in Polluting Industries. Tech. rep. Cambridge, MA: National Bureau of Economic
Research, Cambridge, MA.

Bross, I.D.J. (1958). “How to Use Ridit Analysis”. In: Biometrics 14.1, pp. 18–38.

Chasco, C and J Le Gallo (2015). “Heterogeneity in Perceptions of Noise and Air Pollution:
A Spatial Quantile Approach on the City of Madrid”. In: Spatial Economic Analysis 10.3,
pp. 317–343.

48

Chasco, Coro and Julie Le Gallo (2013). “The Impact of Objective and Subjective Measures
of Air Quality and Noise on House Prices: A Multilevel Approach for Downtown Madrid”.
In: Economic Geography 89.2, pp. 127–148.

Chay, K Y and M Greenstone (2005). “Does Air Quality Matter? Evidence from the Housing

Market”. In: Journal of Political Economy 113.2, pp. 376–424.

Chay, Kenneth Y and Michael Greenstone (2003). “The Impact of Air Pollution on Infant
Mortality: Evidence from Geographic Variation in Pollution Shocks Induced by a Reces-
sion”. In: The Quarterly Journal of Economics 118.3, pp. 1121–1167.

Chen, Yuyu et al. (2013). “Evidence on the Impact of Sustained Exposure to Air Pollution
on Life Expectancy from China’s Huai River Policy.” In: Proceedings of the National
Academy of Sciences of the United States of America 110.32, pp. 12936–12941.

Currie, Janet and Matthew Neidell (2005). “Air Pollution and Infant Health: What Can We
Learn from California’s Recent Experience?” In: The Quarterly Journal of Economics
120.3, pp. 1003–1030.

Currie, Janet et al. (2015). “Environmental Health Risks and Housing Values: Evidence from
1,600 Toxic Plant Openings and Closings.” In: The American Economic Review 105.2,
pp. 678–709.

Davis, L W (2011). “The Effect of Power Plants on Local Housing Values and Rents ”. In:

Review of Economics and Statistics 93.4, pp. 1391–1402.

De Gregori, Ida et al. (2003). “Monitoring of Copper, Arsenic and Antimony Levels in
Agricultural Soils Impacted and Non-Impacted by Mining Activities, from Three Regions
in Chile”. In: Journal of Environmental Monitoring 5.2, pp. 287–295.

Dehejia, Rajeev H and Sadek Wahba (2012). “Causal Effects in Nonexperimental Studies:
Reevaluating the Evaluation of Training Programs”. In: Journal of the American Statis-
tical Association.

Deng, Guoying, Manuel A Hernandez, and Shu Xu (2014). “When Power Plants Leave Town:
Environmental Quality and the Housing Market in China”. In: SSRN Electronic Journal.

D´ıaz Tobar, Julio C´esar (2015). “Dise˜no de una Estrategia de Intervenci´on Basada en el

Desarrollo de Capital Humano para la Miner´ıa Artesanal”. In:

Freeman III, A Myrick, Joseph A Herriges, and Catherine L Kling (2014). The Measurement

of Environmental and Resource Values: Theory and Methods. Routledge.

Gamble, H B and R H Downing (1982). “Effects of Nuclear Power Plants on Residential

Property Values”. In: Journal of Regional Science 22.4, pp. 457–478.

49

Gamper-Rabindran, Shanti and Christopher Timmins (2013). “Does Cleanup of Hazardous
Waste Sites Raise Housing Values? Evidence of Spatially Localized Benefits”. In: Journal
of Environmental Economics and Management 65.3, pp. 345–360.

Ghose, M K and S R Majee (2000). “Assessment of the Impact on the Air Environment
Due to Opencast Coal Mining — an Indian Case Study”. In: Atmospheric Environment
34.17, pp. 2791–2796.

— (2001). “Air Pollution Caused by Opencast Mining and its Abatement Measures in India”.

In: Journal of Environmental Management 63.2, pp. 193–202.

Greenstone, Michael and Justin Gallagher (2008). “Does Hazardous Waste Matter? Evidence
from the Housing Market and the Superfund Program”. In: The Quarterly Journal of
Economics 123.3, pp. 951–1003.

Greenstone, Michael and B Kelsey Jack (2015). “Envirodevonomics: A Research Agenda for

an Emerging Field”. In: Journal of Economic Literature 53.1, pp. 5–42.

Ilabaca, M et al. (1999). “Association between Levels of Fine Particulate and Emergency
Visits for Pneumonia and other Respiratory Illnesses among Children in Santiago, Chile
”. In: Journal of the Air & Waste Management Association (1995) 49.9 Spec No, pp. 154–
163.

Kahneman, D (2003). “Maps of Bounded Rationality: Psychology for Behavioral Economics”.

In: American Economic Review 93.5, pp. 1449–1475.

Kemp, Deanna et al. (2010). “Mining, Water and Human Rights: Making the Connection”.

In: Journal of Cleaner Production 18.15, pp. 1553–1562.

Kiel, K A and K T McClain (1995a). “The Effect of an Incinerator Siting on Housing

Appreciation Rates”. In: Journal of Urban Economics 37.3, pp. 311–323.

Kiel, K A and M Williams (2007). “The Impact of Superfund Sites on Local Property Values:

Are All Sites the Same?” In: Journal of Urban Economics 61.1, pp. 170–192.

Kiel, Katherine A and Katherine T McClain (1995b). “House Prices during Siting Decision
Stages: The Case of an Incinerator from Rumor through Operation”. In: Journal of
Environmental Economics and Management 28.2, pp. 241–255.

Kitula, A G N (2006). “The Environmental and Socio-Economic Impacts of Mining on Lo-
cal Livelihoods in Tanzania: A Case Study of Geita District”. In: Journal of Cleaner
Production 14.3-4, pp. 405–414.

Kohlhase, Janet E (1991). “The Impact of Toxic Waste Sites on Housing Values”. In: Journal

of Urban Economics 30.1, pp. 1–26.

50

Kuminoff, Nicolai V, Christopher F Parmeter, and Jaren C Pope (2010). “Which Hedonic
Models Can We Trust to Recover the Marginal Willingness to Pay for Environmental
Amenities?” In: Journal of Environmental Economics and Management 60.3, pp. 145–
160.

Lav´ın, F V, J Dresdner, and R Aguilar (2011). “The Value of Air Quality and Crime in
Chile: a Hedonic Wage Approach”. In: Environment and Development Economics 16.03,
pp. 329–355.

Loayza, Norman and Jamele Rigolini (2016). “The Local Impact of Mining on Poverty and
Inequality: Evidence from the Commodity Boom in Peru”. In: World Development 84.C,
pp. 219–234.

Malm, Olaf et al. (1990). “Mercury Pollution Due to Gold Mining in the Madeira River

Basin, Brazil”. In: Ambio 19.1, pp. 11–15.

Mendelsohn, Robert and Sheila Olmstead (2009). “The Economic Valuation of Environ-
mental Amenities and Disamenities: Methods and Applications”. In: Annual Review of
Environment and Resources 34, pp. 325–347.

Mendelsohn, Robert et al. (1992). “Measuring Hazardous Waste Damages with Panel Mod-

els”. In: Journal of Environmental Economics and Management 22.3, pp. 259–271.

Michael, H J, K J Boyle, and R Bouchard (2000). “Does the Measurement of Environmental
Quality Affect Implicit Prices Estimated from Hedonic Models”. In: Land Economics
76.2, p. 283.

Milu, V, J Leroy, and C Peiffert (2002). “Water Contamination Downstream from a Copper
Mine in the Apuseni Mountains, Romania”. In: Environmental Geology 42.7, pp. 773–
782.

Ministerio de Desarrollo Social (2009). Encuesta de Caracterizaci´on Socioecon´omica Nacional
CASEN. Retrieved from http://observatorio.ministeriodesarrollosocial.gob.
cl/casen/basededatos_historico.php. Accessed: 2016-03-01.

— (2011). Encuesta de Caracterizaci´on Socioecon´omica Nacional CASEN. Retrieved from
http://observatorio.ministeriodesarrollosocial.gob.cl/casen/basededatos_
historico.php. Accessed: 2016-03-01.

Ministry of Environment (2011). Official Environment Status Report. Retrieved from http:

//mma.gob.cl.

Muehlenbachs, Lucija, Elisheba Spiller, and Christopher Timmins (2013). “Shale Gas De-
velopment and the Costs of Groundwater Contamination Risk”. In: SSRN Electronic
Journal.

51

Muehlenbachs, Lucija, Elisheba Spiller, and Christopher Timmins (2015). “The Housing
Market Impacts of Shale Gas Development”. In: The American Economic Review 105.12,
pp. 3633–59.

Papke, Leslie E and Jeffrey Wooldridge (1996). “Econometric Methods for Fractional Re-
sponse Variables with an Application to 401 (k) Plan Participation Rates”. In: Journal
of Applied Econometrics 11, 619?632.

Ridker, R G and J A Henning (1967). “The Determinants of Residential Property Values
with Special Reference to Air Pollution ”. In: The Review of Economics and Statistics
49.2, p. 246.

Roback, Jennifer (1982). “Wages, Rents, and the Quality of Life”. In: The Journal of Political

Economy 90.6, pp. 1257–1278.

Rosen, Sherwin (1974). “Hedonic Prices and Implicit Markets: Product Differentiation in

Pure Competition”. In: Journal of Political Economy 82.1, pp. 34–55.

Rosenbaum, Paul R and Donald B Rubin (1985). “Constructing a Control Group Using
Multivariate Matched Sampling Methods that Incorporate the Propensity Score”. In:
The American Statistician 39.1, pp. 33–38.

S´anchez, J M and S M Enr´ıquez (1996). Impacto Ambiental de la Peque˜na y Mediana Miner´ıa

en Chile. Tech. rep. Universidad de Chile.

SERNAGEOMIN (2011). Atlas de Faenas Mineras. Retrieved from http://biblioteca.

sernageomin.cl/opac/index.asp.

— (2016). Instalaciones Mineras de Chile. Retrieved from https://www.sernageomin.cl.

Accessed: 2016-05-12.

SINIM (2017). Datos Municipales. Retrieved from http://datos.sinim.gov.cl. Accessed:

2017-08-10.

Urkidi, Leire (2010). “A Glocal Environmental Movement Against Gold Mining: Pascua–Lama

in Chile”. In: Ecological Economics 70.2, pp. 219–227.

U.S. Environmental Protection Agency (1995). Human Health and Environmental Damages
from Mining and Mineral Processing Wastes. Technical Background Document Support-
ing the Final Rule Applying Phase IV Land Disposal Restrictions to Newly Identified
Mineral Processing Wastes.

— (2013). America’s Children and the Environment. Retrieved from https://www.epa.

gov/sites/production/files/2015-06/documents/ace3_2013.pdf.

52

Yu, Fei (2011). “Indoor Air Pollution and Children’s Health: Net Benefits from Stove and
Behavioral Interventions in Rural China”. In: Environmental and Resource Economics
50.4, pp. 495–514.

53

ESSAY 2. THE U.S. COAL-TO-GAS POWER PLANTS CONVERSION

PROCESS AND ITS EFFECTS ON HOUSING PRICES

Abstract

Recent fuel-switching power plant projects in the United States promise several en-

vironmental gains at both the global and the local level. Using records on more than

1,000,000 property transactions, I elicit local property value impacts from the coal-to-natural

gas switching experienced by properties located near the fuel-switching power plants.

I

adopt a spatial difference-in-difference approach using property transactions near, and far,

from switched plants, and comparable coal-fired power plants. A triple difference estimator

strengthens these estimations. Results indicate that property values increase in the immedi-

ate vicinity of the fuel-switching plants. Most of these impacts occur immediately after the

shutdown of a coal-fired power generation unit, revealing the disamenity effect of coal-fired

power plants.

54

2.1.

Introduction

The electric power sector is currently the largest source of greenhouse emissions in the

United States. Emissions from this sector account for 34% of current carbon dioxide (CO2)

domestic anthropogenic emissions, and for an important portion of other greenhouse gas

emissions such as methane (CH4) and nitrous oxide (N2O), as well as a significant amount of

air pollutant emissions (U.S. Environmental Protection Agency, 2018). Trends in emissions

from fossil fuel combustion, however, show a stable decrease in both CO2 discharges and

emissions of other local air pollutants from this sector (see Figure 1). These reductions are

the result of multiple factors: (1) substitution from coal to natural gas in power generation,

(2) the incorporation of non-fossil energy sources in electricity generation, (3) environmental

regulation, and (4) technology adoption. How these factors affect individuals’ welfare remains

yet unanswered. Using evidence from the U.S. residential housing market, this paper provides

an answer to the previous question by studying welfare implications of the coal-to-gas fuel

conversion process revealed through prices in this market.

Figure 2.1. U.S. Electric Power Industry Emissions Estimates 1990-2016

Notes: U.S. Energy Information Administration (2018).

55

Domestic electricity generation comes primarily from the combustion of fossil fuels.

Coal-fired combustion dominated this power generation until 2016, the year in which natural

gas surpassed coal in this process, and became the leading fuel source in the electric power

industry (U.S. Energy Information Administration, 2018). Over the last decade, a decline in

natural gas prices due to an increased supply from shale gas production led to an increasing

share of natural gas-fired power generation in lieu of the coal-fired production. Decreasing

variable costs of this fossil fuel and an efficient CO2-to-energy ratio, motivated fuel switch-

ing by several electricity generation units (EGUs) across the country, which replaced coal

as their primary fuel source. As natural gas burns cleaner than coal, this coal-to-gas con-

version process is expected to enhance environmental quality at both a global and a local

scale, particularly through a greenhouse gas emissions reduction and lower discharge of local

airborne pollutants. Access to cleaner air is a benefit to households located near a polluting

facility through the possibility of a better health, clearer views, and more enjoyment from

outdoor activities (Grainger, 2012). This paper aims at answering whether these benefits

are capitalized into property values of houses located near a switching plant.

To estimate causal impacts, I use an extensive dataset of property transactions that

took place in the neighborhood of several fuel-switching power plants, which switched their

primary fuel source between 2009 and 2016. To identify the exact date of the switching, I

distinguish between the year in which the first coal-fired EGU stopped generation, and the

year in which the first natural gas-fired unit started operations. The inclusion of several

switching power plants leads to the selection of a large dataset of property transactions

throughout the country, which allows me to control further in the estimations for any county-

level macroeconomic effects affecting housing prices (Muehlenbachs, Spiller, and Timmins,

2015).

For identification, I use a spatial differences-in-differences (DID) design that compares

sales prices of properties located near switching power plants to sales prices of properties

located farther away from these facilities, before and after the conversion. After the inclusion

56

of several fixed effects, the spatial DID estimator provides a reliable estimation of the causal

impact of proximity to a coal-to-gas-fired power plant on housing prices. To separate out

the effect of a local air quality improvement from the general disamenity effect of proximity

to a (former) coal-fired power plant, I estimate a second DID design in which properties

near switching plants are compared with properties near traditional coal-fired power plants.1

This strategy derives an estimator of the causal effect that is consistent in the presence

of unobservable factors common to neighborhoods hosting the siting of a power generation

facility. Findings from the empirical estimations are finally strengthened with a triple-

differences (DDD) estimator that is robust to unobservables varying with proximity to power

plants in general.

Previous works on valuation and power plants expose the disamenity effect of prox-

imity to these facilities. Blomquist (1974) estimates the total property value impact of the

Winnetka power plant in Illinois, showing that damages are measurable up to 1 mile from

the plant, with the disamenity factor disappearing at farther distances. The author calcu-

lates an elasticity measure of property values with respect to the distance to the plant at

0.09, which indicates an increase of 9 percent in property values at a one-mile distance from

the facility. On the housing impacts of ethanol plants openings in Michigan, Hodge (2011)

shows an 18 percent drop in prices of houses located at a 2-mile distance or less from a new

ethanol facility. In a large-scale study on power plants, Davis (2011) evaluates the effect of

1993-2000 power plants siting decisions on local housing values and rents across the country.

The author shows that property values decrease by 4-7 percent points in houses located near

a new plant. Currie et al. (2015) extends Davis (2011)’s work by examining market and en-

vironmental impacts of proximity to 1,600 power plants in the U.S. Their results show that

prices of house properties that are located within 0.5 miles of a toxic’s plant site decrease

by 11 percent points after the opening of a new plant, while they find no evidence of plant

closure effects on housing prices. Deng, Hernandez, and Xu (2014) use microtransaction data

1See Davis (2011) for a broad description of negative externalities associated with proximity to coal-fired

power plants.

57

on housing values to assess the effects of air quality improvements after the relocation of two

major power plants in China. The authors show that power plant closings are associated

with a positive and significant 6 percent increase in the price of new apartment units, with

a 65-108 percent increase in transaction volumes. I contribute to the previous literature by

studying property value impacts of local air quality improvements from the fuel-switching of

power plants. A distinctive characteristic of this study is the use of a comprehensive dataset

on housing transactions across the country, which allows rich specifications of the hedonic

price equations. As far as the author’s knowledge, this is the first study documenting market

value impacts of the power plants fossil fuel switching.

Results in this study indicate positive and significant property value impacts of the

coal-to-gas power plant conversion process. Findings suggest that shutdowns of coal-fired

unit increase the value of nearby properties by 8-13 percent relative to properties located

farther away. These effects are significant only within a 0.6-mile distance from the facility.

When properties that are located near a switching facility are compared to properties near a

coal-fired power station, this effect increases up to 20-37 percentage points. The magnitude of

this effect decreases for a larger buffer, although the impact is still positive and significant.

Results from the triple-differences estimator in the immediate vicinity corroborate these

impacts. Findings for the DDD estimator suggests that the inactivity of a coal-fired EGU

increases property values by around 26-40 percent points.

When the treatment is defined as the start of operations of natural gas-fired EGUs,

results indicate positive effects on property values but only when properties that are located

near coal-fired power plants are used as a control group. Evidence indicates that property

value impacts increase by 40-62 percentage points. These results represent market capital-

izations of the coal-to-gas fuel switching after teasing out the disamenity effects of proximity

to fossil-fuel power plants in general. Yet, results for the DDD estimator suggest that these

effects are not perceptible.

The remainder of this paper proceeds as follows. Section 2.2 describes the type of

58

emissions from fossil-fueled power plants, and the risks of exposure to this pollution. Section

2.3 briefly discusses the analytical framework, while Section 2.4 documents the data. Section

2.5 presents the empirical strategy and the main results. Section 2.6 concludes.

2.2. Emissions from Fossil-Fueled Power Plants and

the Health Risks of Exposure

Fossil-fueled power generation is responsible for a large share of greenhouse gas emis-

sions, particularly carbon dioxide (CO2), the primary gas emitted from fossil fuel combustion

in the United States (U.S. Environmental Protection Agency, 2018). Among fossil fuels,

coal is the most harmful to burn due to its high carbon content, which places the coal-fired

generation as the technique with the highest CO2 emissions per energy output, and far worse

than generation based on natural gas, which possesses the lowest CO2-to-energy ratio content

(U.S. Energy Information Administration, 2017).2 Coal burning also releases an important

amount of other pollutants such as sulfur dioxide, nitrogen dioxides, mercury, and particulate

matter; all associated to several adverse health effects, mortality risks, and threats to life

expectancy (e.g. Chay and Greenstone (2003a), Chay and Greenstone (2003b), Currie and

Neidell (2005), Bateson and Schwartz (2007), Chen et al. (2013), Arceo, Hanna, and Oliva

(2016), and Knittel, Miller, and Sanders (2016)). Coal combustion also affects water qual-

ity, as coal ash released after the combustion can end up in water reservoirs, contaminating

waterways and drinking water sources (Carlson and Adriano, 1993).

This section briefly summarizes some evidence on the detrimental effects of exposure

to the main pollutants derived from the coal combustion, and explores the expected changes

in local air quality from its displacement by natural gas in the electricity generation. The

evidence suggests that its displacement by natural gas is expected to release less particulate

2Black et al. (2010) shows that natural gas emits around 50% less CO2 when burned in new efficient

natural gas-fired power plants compared to coal-fired power plants.

59

matter and mercury pollution. Some reductions in sulfur dioxide and nitrogen dioxide can

also be expected.

Sulfur dioxide (SO2) is an invisible gas that belongs to the family of sulfur oxide

(SOX) gases, formed when fuel containing sulfur is burned, during metal smelting, or other

industrial processes (U.S. Environmental Protection Agency, 2014). The electric power

sector is currently responsible for 44% of SO2 emissions, which places this sector as the

largest anthropogenic source of emissions for this pollutant in the U.S., with coal-fired power

plants contributing to more than 90% of these emissions (U.S. Environmental Protection

Agency, 2018). Exposure to high concentrations of this pollutant are associated with eyes,

nose, and throat irritation, infectious complications of chronic obstructive pulmonary disease,

and to increments in hospital admissions due to asthma (World Health Organization, 2006).

When this gas enters in the atmosphere, it also reacts with other compounds to form fine

particulate matter. Particulate matter (PM) is the general term used to describe solid

particles, dust and drops found in the air, with different composition and sizes. Evidence

about the health effects of exposure to heavy (PM10) and fine particulate matter (PM2.5),

the two PM size ranges widely monitored, suggest significant impacts on infant mortality

(Chay and Greenstone, 2003a; Chay and Greenstone, 2003b; Knittel, Miller, and Sanders,

2016; Arceo, Hanna, and Oliva, 2016).

Nitrogen dioxides (NOX) are a group of reactive gases that include nitrogen diox-

ide, nitrous acid, and nitric acid. Although mobile sources are responsible for the highest

domestic anthropogenic release of NOX into the atmosphere, stationary fossil fuel combus-

tion represents 29% of the annual domestic NOX emissions (U.S. Environmental Protection

Agency, 2018). Evidence about outdoor exposure to NOX suggests an increase in asthma

and bronchitis diagnoses in children (Pershagen et al. (1995), Orehek et al. (1976), Chauhan

et al. (2003), and Gauderman et al. (2005)). Also, NOX can react to the presence of heat and

sunlight in the atmosphere to create ground-level ozone (smog), a pollutant associated with

lung diseases, and premature deaths (Bell et al., 2004; Bell, Dominici, and Samet, 2005).

60

Mercury (Hg) is a toxic metal present in rocks, including coal in its natural state.

The coal combustion releases mercury into the air, which eventually settles into water bod-

ies through atmospheric deposition. Once in the water, mercury is transformed by aquatic

microbes into methylmercury, a poisonous form of mercury that accumulates in fish.

In-

takes of contaminated fish are associated with cardiovascular diseases (Salonen et al., 1995;

Guallar et al., 2002), and central nervous system damage in unborn babies (Clarkson, 1990).

Currently, coal-fired power plants represent the largest source of domestic anthropogenic

emissions of this global air pollutant (U.S. Environmental Protection Agency, 2015).

Pollution discharges from the U.S. power sector have notably decreased during the

last decade, in part because of the coal-to-gas fuel-switching in the power generation mix.

Natural gas consists mostly of methane (CH4) and small proportions of hydrocarbon gas

liquids and nonhydrocarbon gases. The promotion of natural gas as cleaner burning fossil

fuel comes mostly from of its lower carbon content per unit of energy relative to coal and

diesel, as well as the negligible amounts that its combustion emits of other pollutants such as

Hg, PM, NOX, SO2. Some evidence, however, suggest that more research is needed in terms

of NOX and SO2 comparative emissions between natural gas and coal (Jaramillo, Griffin,

and Matthews, 2007). Additionally, natural gas can also emit carbon monoxide (CO) and

volatile organic compounds (VOCs) when used in motor vehicles, although at lower discharge

rates than diesel.

Emissions leakages from transmission and distribution pipelines, however, can un-

dermine the benefits of natural gas as a cleaner fuel replacement in the power generation

described so far. The drilling and extraction of natural gas from wells, and its transportation

in pipelines results in leakage of methane, a greenhouse gas that is roughly 30 times stronger

than CO2 at trapping heat in the atmosphere. Evidence shows that this leakage can reach

up to 9% of this fuel total life-cycle emission (Tollefson, 2013). Yet, life-cycle greenhouse

gas emissions of natural gas have been shown to be lower relative to the diesel and coal

combustion (Jaramillo, Griffin, and Matthews, 2007; Burnham et al., 2011).

61

2.3. Analytical Framework

2.3.1. A Brief on the Hedonic Price Model

This study uses the hedonic price model to elicit average marginal willingness to pay

measures for a nonmarginal improvement in local air quality due to coal-to-gas-fired power

plant conversions. Studies that apply hedonic price modeling to associate housing prices

and quantities of environmental quality date back to Ridker (1967) and Ridker and Henning

(1967), under the reasoning that the value of a house can be considered a function of its

characteristics, such as structural, neighborhood, and environmental characteristics (Free-

man, 1979).3 This section briefly describes the hedonic price method and its implications

when evaluating a local improvement in an environmental amenity. The model predicts that

housing prices increase in response to an increase in environmental quality.

The hedonic price theory considers that an item h can be valued by a vector z of

characteristics (z1, z2, ..., zj). In the case of a house, these characteristics include structural

and neighborhood characteristics, and local environmental quality. The price of house h,

therefore, can be considered as the sum of each of its homogeneous attributes in a price

function described as follows:

Ph(z) = P (z1, z2, ..., zj).

(2.1)

This function Ph(z) is referred to as the hedonic price function and indicates the

amount that an individual must pay for a bundle with characteristics z. The partial deriva-
tive of Ph(·) with respect to zj, ∂Ph(·)/∂zj, gives the marginal implicit price of characteristic

zj. Utility-maximizing individuals derive at the same time utility from a housing item .

In the housing market equilibrium from the interactions of buyers and sellers, individuals’

marginal willingness to pay (WTP) for characteristic zj equals the marginal implicit price

of that characteristic.

3For a more comprehensive review on hedonic price applications, see Mendelsohn and Olmstead (2009)

and Freeman III, Herriges, and Kling (2014).

62

Rosen (1974) distinguishes the price hedonic function Ph(z) from the bid function θi =
θ(M, zj, z−j, u∗), which represents what an individual i is willing to pay for different values of

characteristic zj, holding income M , other characteristics z−j, and utility constant at a level
u∗. Heterogeneous individuals’ preferences and income lead to different bid functions, and

so to different chosen quantities of characteristic zj. The relationship between the hedonic

price function Ph(z), and the bid functions θ1 and θ2 for individuals 1 and 2, respectively,

for characteristic zj is depicted in Figure 2.2.

Figure 2.2. Bid Curves and the Hedonic Price Function in a Hedonic Market for Local Air
Quality

Suppose zj is a measure of local air quality. Figure 2.2 shows that both bid functions

exhibit diminishing marginal willingness to pay for zj, and that given the hedonic price

function, individuals 1 and 2 choose levels of air quality where their marginal WTP for zj
(cid:48)(cid:48)
j ,

equals the marginal implicit price determined by the hedonic price function at z

(cid:48)
j and z

respectively. Given the market equilibrium, individuals’ utilities would be lower at sites with

higher or lower levels of air quality.

63

2.3.2. A Localized Change in Air Quality

This study aims to capture property value impacts of the recent coal-to-gas conversion

process in the U.S. The mechanism behind these impacts is the nonmarginal improvement in

local air quality derived from the use of a cleaner input in the power generation near specific

power stations. To establish welfare effects of this nonmarginal change, it is assumed that the

change in environmental quality is a localized change (Palmquist, 1992), and therefore, the

hedonic price function does not shift in response to this change. This is a valid assumption

as the number counties under analysis in this study represents only a small portion of the

entire U.S. housing market (see Section 2.4). Hence, any air quality improvement in these

counties is not enough to force a significant relocation of individuals that could lead to a

new hedonic price equation.

Consider now that the coal-to-gas conversion process improves air quality in a neigh-

borhood from z

(cid:48)
j to z

(cid:48)(cid:48)
j . From Figure 2.2, this nonmarginal change is expected to increase

the price of house h. For individual 1 originally consuming z

(cid:48)
j, the new price for air quality

exceeds her WTP for the air quality amenity. This individual can decide to relocate to a

place with lower air quality and restore the equilibrium, or to stay, in which case she would

be better off due to an increase in wealth. If the individual relocates, the welfare effect of

the nonmarginal change in zj comes from the gain to the new individual that locates in

the neighborhood experiencing the amenity enhancement, represented by her WTP for an

air quality improvement. If the individual stays, the welfare effect is the change in wealth

experienced by the homeowner individual. The implied change in total welfare from an air

quality improvement, therefore, can be obtained by multiplying the observed equilibrium

price differential due to the coal-to-gas conversion by the number of local residential housing

units.

64

2.4. Data

2.4.1. Power Plants

Data on electricity generation units (EGUs) come from the Annual Electric Generator

Reports (Form EIA-860) compiled by the Energy Information Administration (EIA). This

survey form has annual information, at the generator-level, on existing and planned units

belonging to power plants with generation capacity equal or higher than 1MW.4 From this

survey, I obtain information on EGUs capacity, age, primary technology, and primary fuel

source, among other features. Among the key aspects of establishing accurate market impacts

on coal-to-gas-fired power plants on nearby properties is the geographical location of these

plants. This feature is reported by the EIA-860 plant-level data files but only since 2009.

Due to this, the present study conducts the evaluation of property value impacts using

information from 2009 to 2016.

Power stations are generally equipped with more than one EGU and, depending on

the combustion technology; they can have several boilers (Ummel, 2012). This complexity

challenges the classification of these facilities, as EGUs can use more than one fuel source.

Hence, to identify the stations that switched from coal to natural gas in at least one of their

generation units, I classify EGUs by a combination of its primary technology and its primary

fuel. Based on this, I identify the proposed gas-fired EGUs from the set of “proposed EGUs”

listed from 2010 to 2016, and for which natural gas, or any other gas, is highlighted as their

primary fuel source.5 This process leads to 831 EGUs located in 269 power stations across

the country. The spatial location of these power stations is displayed by the top panel of

Figure 2.3.6

4Further information on the Form EIA-860 and the public data available at Form EIA-860’s website:

https://www.eia.gov/electricity/data/eia860/. Dated November 9, 2017. Retrieved April 23, 2018.

5Other gases include blast furnace gas.
6Table A2.1 in the Appendix contains a description of the prime mover (i.e. the engine, turbine, water
wheel, or similar machine) that drivers the proposed gas-fired generator units, as well as their summer and
winter generation capacity.

65

Figure 2.3. Power Plants with Proposed Gas-Fired and Retired Coal-Fired Generators

(a) Power Plants with Proposed Gas-Fired Generators

Notes: Using data from EIA-860 forms.

(b) Power Plants with Retired Coal-Fired Generators

Despite the numerous gas-fired units projected during 2010-2016, some of them are

either still in construction, and many others are not necessarily replacing old coal-fired units.

To establish the set proposed gas-fired EGUs that effectively started operations during the

period under analysis, I match this selection to the set of operating units running under the

same primary fuel source. The matching procedure works as follows. First, generators that

66

operate with fuels other than natural gas, or any other gas, as their primary fuel source are

dismissed. From the remaining set, I rule out power plants that over time report generation

units with a status other than “in operation”, to avoid factors other than the fuel switching

from confounding the analysis. The last step before the matching is the identification of

gas-fired generators by their first operating year: generators for which their first operating

year is before 2009 are also dismissed. This subset of units is then matched to the set of

proposed gas-fired units. The final dataset contains all power plants in the U.S. with one of

more MWs of capacity, which have at least one new gas-fired EGU in operation since 2009.

I pursue a similar procedure to identify retired coal-fired units. I first select units

from the set of “retired EGUs” listed from 2009 to 2016, and for which coal, or any of its

derivatives, is indicated as their primary fuel source.7 This process leads to 481 EGUs, part

of 213 power plants across the country. The spatial location of these power plants is in the

bottom panel of Figure 2.3.8 Plants with retired coal-fired EGUs are later matched with the

set of plants with operable units to determine the exact moment of retirement. The matching

procedure for this case is as follows. Operating generators with a prime fuel source other

than coal are dismissed from the data. Then, I rule out power plants that over time report

generators with a status other than “in operation” to minimize confounding factors. This

resulting subset of plants with operating coal-fired generators is later used in the selection of

control plants. Finally, I select generators by their last operating year; units for which their

last operating year is before 2009 are dismissed from the analysis. The final dataset contains

all power plants in the U.S. with capacity equal or higher than 1MWs, which over 2009-2016

had at least one coal-fired EGU continuously under operation until its later retirement.

The final step in the identification of the coal-to-gas-fired units is the match between

the sets of feasible new gas-fired generators, and the set of retired coal-fired generators. After

performing the match at the plant level, I detect five power plants with the desired features.

7Coal and coal derivatives refer to anthracite coal, bituminous coal, lignite coal, coal-derived synthesis
gas, subbituminous coal, refined coal, anthracite culm, bituminous gob, fine coal, lignite waste, and waste
coal.

8Table A2.2 (Appendix) contains the description of these EGUs’ prime movers and generation capacity.

67

These switched plants locate in Colorado, Georgia, North Carolina (2), and Pennsylvania.

From previously identified subset of operating coal-fired generators, and using a location

criterion, I select the set of control power stations. The final set of power stations includes

five switched plants in the treatment group, and thirty-three coal-fired power plants in the

control group. Table 2.1 displays the list of plants in the treatment group, including the plant

names, their location, and the number of new EGUs introduced. Table 2.1 also displays the

timing of the switching, which I define in two ways: (1) the year in which the first coal-fired

generator stopped operating, and (2) the year in which the first gas-fired unit started oper-

ating. The spatial location of both treated and control power stations is shown in Figure 2.4.

Table 2.1. Coal-to-Gas-Fired Power Plants (Treated Plants)

Plant Name

Utility Name

State

County

New First Coal-Fired First Gas-Fired
Units

Stopped

Started

Cherokee
Jack McDonough
Hunlock Power St.
Dan River
Buck

Public Service Co
Georgia Power Co

UGI Development Co

CO
GA
PA

Adams
Cobb

Luzerne

Duke Energy Carolinas, LLC NC Rockingham
Duke Energy Carolinas, LLC NC

Rowan

3
7
2
4
3

2010
2010
2009
2011
2010

2015
2011
2011
2012
2011

Figure 2.4. Treated and Control Power Plants

Notes: Using data from EIA-860 forms.

68

2.4.2. Property Transactions

Data on property transactions come from Zillow, a national web-based real estate

data provider, with information on buying, selling, renting, and remodeling of more than 110

million homes in the U.S. The Zillow data contain information on structural characteristics

of properties, such as the number of rooms, number of bathrooms, and on the property exact

address and geographical location, the square footage, the year the structure was built, and

the number of stories.

From the full dataset, I select transaction records of properties sold near treated

and control power plants. To that end, I chose only those transactions that took place in

counties where the selected power plants are located, and in their first-contiguity neighbors.

I also chose arm’s-length transactions on single-family residential homes that involve a single

parcel. To minimize data entry error, I trimmed the top and bottom 5% of the data based

on a price-per-square-foot metric, and top and bottom 1% of the data based on additional

metrics such as bathrooms per square foot and bedrooms per square foot. Finally, I discarded

all transactions referring to homes for which the construction year is higher than the sale

year. The final dataset contains more than 1,000,000 records on transactions dated between

2009 and 2016, and in areas near coal-to-gas- and coal-fired power plants. The location of

these properties is in Figure 2.5, while descriptive statistics for the main covariates are in

Table 2.2.

2.5. The Coal-To-Gas Power Plant Conversion and its

Effects on Housing Prices

This study aims to capture market impacts of local air quality improvements on

prices of residential properties located near power plants switching their primary fuel source

away from coal, in at least one of their generators. To this end, the econometric strategy

compares the sales price of properties located at certain distances from the fuel-switching

69

Figure 2.5. Properties in Treated and Control Areas

Table 2.2. Descriptive Statistics on Main Covariates

Variables

N

Mean

S.D.

Min Max

Panel A. Treated Counties
# of Bathrooms
# of Half Bathrooms
# of Bedrooms
Home age
Square feet

102,469
102,469
102,469
102,469
102,469

2.30
0.39
3.16
26.81

2,068.51

0.96
0.51
1.18
20.43
900.71

Panel B. Control Counties
# of Bathrooms
# of Half Bathrooms
# of Bedrooms
Home age
Square feet

732,971
732,971
732,971
732,971
732,971

1.96
0.49
3.07
36.16

0.94
0.65
1.11
29.88

2,166.12

1,007.79

0
0
0
0
768

0
0
0
0
768

9
4
9
189
6,268

12
16
12
215
6,270

plants, before and after the switching. This first difference gives an idea of how prices

of nearby properties change over time after controlling for year-level unobservable factors.

Notwithstanding, time-invariant unobservables common to areas near and far from power

stations may also affect sales prices. A second difference resulting from the comparison of

prices across properties located at different distances from these facilities, provides a price

differential that is consistent to the presence of unobservables varying with proximity to

70

these units. When these two differences are combined in a difference-in-difference (DID)

design, it is possible to derive a consistent estimate of the causal effect of proximity to a

new coal-to-gas-fired power plant. Figure A2.1 in the Appendix describes this near-far DID

strategy.

The previous approach leads to the causal identification of housing price impacts of

proximity to fuel-switching facilities. Yet, as stated before the goal here is the establishment

of property value impacts from localized improvements in air quality that come from the

coal displacement in power generation. To properly disentangle the effect of changes in

local air quality from the general (dis)amenity effects of proximity to fossil-fueled power

plants, I consider an additional approach that allows me to consider unobservable factors

that could be common to neighborhoods facing the siting of these power plants.

In this

second scheme, I replace the previous near-far second difference with the price differential

that emerges between homes located in the vicinity of a coal-to-gas plant, and homes that

locate near power plants that still use coal as their prime fuel source. The joint arrangement

of this near-gas/near-coal differential and the time differential provides a consistent estimate

of causal impacts of a localized air quality improvement from a coal-to-gas switching on

housing prices, while controlling for unobservables common to neighborhoods hosting power

plants. By enhancing the homogeneity of homes across the treatment and the control groups,

this second DID strategy reduces the risks of an omitted variables bias generally common to

hedonic application. Figure A2.2 in the Appendix summarizes the near-gas/near-coal DID

strategy.

A final empirical strategy for causal identification is the combination of the two pre-

vious schemes in a triple-difference (DDD) estimator. If time-variant unobservables that are

common to neighborhoods hosting generator facilities vary simultaneously with proximity

to these installations, they can be factored out with a DDD estimator and with several

fixed effects so that estimations of causal effects are consistent to the presence time-variant

and time-invariant unobservables across these many spatial divisions. This feature makes

71

the DDD estimator a refinement of the two previous estimators that use double differences.

Figure A2.3 in the Appendix exhibits the triple differences strategy.

An important consideration is the timing of the fuel switching. Section 2.4 anticipated

two different approaches to identify the switching time, which in the vocabulary of the

treatment effect literature corresponds to the time of the treatment implementation. Due to

the many factors that take place in the fuel switching, I follow two different approaches for

the proper definition of the treatment implementation period. The first approach considers

the year in which the first coal-fired unit was retired from power generation in each of the

defined switched stations (see Table 2.1, column 6). Whenever the treatment under analysis

corresponds to the localized air quality improvement that arises from the dismissal of a

polluting generator, this definition should allow to capture these causal impacts. However,

in cases when the fuel switching is characterized by a smooth transition between coal-fired

and gas-fired units, the year in which the first gas-fired unit started to generate electricity

can also provide some information of when local air quality improvements are expected to

happen. In this second approach, causal effects are due to local air quality improvements

derived from the smooth transition into a cleaner generation technology. In the remainder

of this section, I present the reduced form equations to estimate and the main results.

2.5.1. Housing Price Impacts of the Coal-to-Gas Conversion using

a Near-Far Approach

The specification of the reduced form hedonic price equation in the near-far DID

approach is as follows:

ln yijt = β0 + β11[N ear/F ar]d≤θ

ijt + β21[Af ter]jt+

β31[N ear/F ar]d≤θ

ijt × 1[Af ter]jt + Xitη + κ + τ + 
ijt,

(2.2)

where ln yijt is the log of the price y of house i in the neighborhood of plant j at time

72

t; 1[Af ter]it is an indicator variable taking 1 if the treatment implementation occurs in plant

j during year t (=0 otherwise); Xit is a vector of house i characteristics during year t; κ is a
spatial fixed-effect (i.e. county, city, state, or county × year); t is a temporal fixed effect; and


ijt is an idiosyncratic error. An important variable in equation (2.2) is the indicator variable
taking the value of 1 if house i, sold in year t, locates at distance d ≤ θ of
1[N ear/F ar]d≤θ

ijt

plant j. Empirical evidence shows that in most cases, pollution is detectable only in a small

radius of distance from the polluting facility, which in general does not exceed 2 miles (Davis

(2011), Hodge (2011), Currie et al. (2015), and Muehlenbachs, Spiller, and Timmins (2015)).

I use this empirical evidence to consider several small radii of distances for θ: 0.6, 0.9, 1.2,

and 1.5 miles. To avoid the transition of houses between the treatment and the control group
when varying the size of the radius, the variable 1[N ear/F ar]d≤θ
takes the value of 0 when
house i locate at distance d ≥ ¯θ of plant j, where ¯θ takes values in the interval of 6 and 10

ijt

miles. With this specification, the coefficient, the coefficient β3 in equation (2.2) measures

the causal impact of proximity to a fuel-switched power plant. Distances to power plants
are calculated using the geodist command in Stata®. All estimations are carried using an
ordinary least square estimator (OLS) with standard errors clustered at the county × year

level.

Near-far DID estimations of property value impacts due to the coal-to-gas conversion

process are displayed in Table 2.3. A standard in most of the following set of result tables will

be the division of the estimation results into a top panel, describing estimation results when

the treatment implementation is defined as the year in which the first coal-fired EGU stops

the generation, and a bottom panel displaying estimation results for when the treatment

adoption is defined as the year of the first gas-fired EGU starts operations. The columns

represent different estimation results and different control groups. The parallel trends as-

sumptions that follow these results are depicted in Figures A2.4, A2.5, and A2.6, for the

first treatment implementation year definition, and in Figures A2.7, A2.8, and A2.9 for the

second one.

73

Table 2.3. Coal-to-Gas Impact on Property Prices - Homes Far from a Coal-to-Gas-Fired
Station as Controls

Proximity

Control ≥ 6mi
(2)
(1)

Control ≥ 8mi
(2)
(1)

Control ≥ 10mi
(2)
(1)

Panel A. First Coal-fired EGU Stopped Operations

Treated ≤ 0.6mi: 1[Near]×1[After]

Treated ≤ 0.9mi: 1[Near]×1[After]

Treated ≤ 1.2mi: 1[Near]×1[After]

Treated ≤ 1.5mi: 1[Near]×1[After]

Panel B. First Gas-Fired EGU Started Operations

Treated ≤ 0.6mi: 1[Near]×1[After]

0.125∗∗
(0.054)

-0.063
(0.066)

-0.057
(0.048)

-0.002
(0.072)

0.110∗∗
(0.048)

-0.078
(0.058)

-0.059
(0.047)

0.088∗∗
(0.037)
-0.140∗∗
(0.048)
-0.078∗
(0.041)

0.078∗∗
(0.029)
-0.147∗∗∗
(0.036)
-0.078∗∗
(0.035)

-0.00001
(0.072)

0.002
(0.080)

0.005
(0.080)

0.101
(0.090)

-0.040
(0.057)

-0.024
(0.049)

0.017
(0.059)

×

791,812
792,263
792,954
793,451

0.088
(0.085)

-0.055
(0.055)

-0.034
(0.048)

0.011
(0.058)

×

791,812
792,263
792,954
793,451

0.100
(0.111)

-0.073
(0.052)

-0.027
(0.052)

0.025
(0.065)

×

756,804
757,255
757,946
758,443

0.081
(0.104)
-0.088∗
(0.047)

-0.040
(0.047)

0.015
(0.062)

×

756,804
757,255
757,946
758,443

0.089∗∗
(0.040)
-0.136∗∗
(0.048)

-0.065
(0.046)

0.031
(0.098)

0.152
(0.130)

-0.053
(0.056)

0.0004
(0.059)

0.059
(0.077)

×

707,435
707,886
708,577
709,074

0.093∗∗
(0.032)
-0.131∗∗∗
(0.038)

-0.057
(0.041)

0.038
(0.095)

0.137
(0.125)

-0.062
(0.052)

-0.006
(0.055)

0.053
(0.074)

×

707,435
707,886
708,577
709,074

Treated ≤ 0.9mi: 1[Near]×1[After]

Treated ≤ 1.2mi: 1[Near]×1[After]

Treated ≤ 1.5mi: 1[Near]×1[After]

County FE
County × Year FE
N (0.6mi)
N (0.9mi)
N (1.2mi)
N (1.5mi)

Notes: All estimations include # of bathrooms, # of half bathrooms, # of bedrooms, house age, house age2,
square feet, square feett2, and fixed effects by year, city, and state. Clustered standard errors at the county
× year level in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

Estimation results in the top panel of Table 2.3 indicate positive and significant

property value impacts of adjacency to a power station that shut down one of its coal-

fired generators. Properties that locate at a near distance (within 0.6 miles) from a plant

dismissing a coal-fired unit experience an increase in prices of around a 11-13 percent points

when compared to properties located beyond a 6-mile distance. These results, although

smaller in magnitude, are consistent through different control group definitions and pairs

74

well with previous evidence on local air quality improvements being less noticeable at farther

distances. When larger radii for near properties is considered, however, the fuel-switching

impact on property prices turns into a negative effect for houses located within 0.9 and 1.2

miles from the plants. A negative impact of a retired coal-fired EGU on homes located near

the facility, but not exactly in its immediate vicinity, could be due to a substitution of home

units, i.e. homes that are adjacent to a plant get more attractive after the switching and

depress prices in substitute homes that are located nearby. In any case, this effect disappears

once a larger buffer of distance is considered. Regarding the effects of the power generation

by a new gas-fired EGU, results for the bottom panel in Table 2.3 indicates no evidence

of housing price impacts when natural gas-fired generators start operations except when a

0.9 mi radius is considered. This effect, however, is weakly significant and not consistent

through different specifications.

2.5.2. Housing Price Impacts of the Coal-to-Gas Conversion using

a Near-Gas/Near-Coal Approach

The second empirical approach uses houses near coal-fired power plants as controls

in lieu of houses located far from a switched plant. Similar to equation (2.2), the hedonic

price equation in this near-gas/near-coal DID scheme is specified as follows:

ln yijt = γ0+γ11[N earGas/N earCoal]d≤θ

ijt + γ21[Af ter]jt+

γ31[N earGas/N earCoal]d≤θ

ijt × 1[Af ter]jt + Xitδ + κ + τ + νijt,

(2.3)

Relative to equation (2.2), this new specification considers now an indicator variable

1[N earGas/N earCoal]d≤θ
that takes the value of 1 if house i, sold in year t, locates at
distance d ≤ θ of plant j (=0 otherwise). Provided that equation (2.3) controls for time-

ijt

invariant and time-variant unobservable factors that are common to neighborhoods near

power plants (either switching fuels or not innovating at all), then coefficient γ3 captures the

75

causal impact of a localized improvement in air quality.

Results for the DID estimation using homes near coal-fired power stations as controls

are in Table 2.4. Figures A2.10 and A2.11 in the Appendix plot a test on the parallel trends

assumption on this specification for the different specifications on treatment implementation

time. Findings in the top panel of Figure 2.4 indicate positive and significant impacts of a

coal-fired generator closure on housing prices, which is consistent with a local improvement

in air quality. Results for the richer specification in column (2) for houses in the immediate

vicinity (within 0.6 miles) indicate a property value increase of 31 percentage points. This

effect becomes weakly significant when larger radii are considered (within 0.9-1.2 miles), yet

they are positive and significant at the 99% level of significance for homes located within 1.5

miles. Results in columns (1) and (2) for a radius of 1.5 miles indicate that housing prices

increase around 28 to 37 percentage points. Similarly, results in the bottom panel consistently

suggest positive and significant property value impacts of the coal-to-gas switching. Results

for the richer specifications in columns (2) indicate that property prices increase by around

40-47 percentage points due to the fuel-switching for distance buffer larger than 0.9 miles.

Notwithstanding, property prices in column (1) are found to decrease for houses located in

immediate vicinity of the plant (within 0.6 miles), although the significance disappears when

a richer specification is considered in column (2).

The fact that impacts from gas-fired units’ start-ups are insignificant when properties

located farther away from a switched plant are used as controls (see bottom panel of Table

3), but positive and significant when properties located near coal-fired power plants are

instead the controls, suggests that the disamenity effect of proximity to the power plant

could be offsetting any capitalization effect from air quality improvements due to a gas-

fired generation. Yet, when this broad disamenity effect is set apart from the environmental

amenity (that is, when properties around a coal-fired station are used as controls), property

value impacts of a local air quality improvement start to emerge as seen in Table 2.4 (bottom

panel). These results in Table 2.4 represent suggestive evidence of positive welfare effects

76

Table 2.4. Coal-to-Gas Impact on Property Prices - Homes Near to a Coal-Fired Station as
Controls

Proximity

Treated & Control Treated & Control Treated & Control Treated & Control

≤ 0.6mi

≤ 0.9mi

≤ 1.2mi

≤ 1.5mi

(1)

(2)

(1)

(2)

(1)

(2)

(1)

(2)

Panel A. First Coal-fired EGU Stopped Operations
0.353∗
(0.208)

0.311∗∗∗
(0.059)

-0.498∗
(0.256)

1[Near] x 1[After]

Panel B. First Gas-Fired EGU Started Operations
0.622∗∗∗
(0.158)

-0.462∗∗
(0.152)

1[Near] x 1[After]

City FE
State × Year FE
N

×
360

-0.015
(0.127)

×
×
360

×
1175

-0.018
(0.045)

0.194∗
(0.113)

0.149
(0.119)

0.279∗∗
(0.126)

0.367∗∗∗
(0.106)

0.452∗∗
(0.144)

×
×
1170

0.493∗∗∗
(0.103)

×
2490

0.467∗∗∗
(0.116)

×
×
2481

0.564∗∗∗
(0.101)

×
3893

0.398∗∗
(0.141)

×
×
3865

Notes: All estimations include # of bathrooms, # of half bathrooms, # of bedrooms, house age, house age2,
square feet, square feett2, and fixed effects by year. Clustered standard errors at the county × year level in
parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

from the coal-to-gas power plant conversion process.

2.5.3. Housing Price Impacts of the Coal-to-Gas Conversion using

a Triple-Differences Estimator

Table 2.5 displays results for the DDD estimator. The top panel exhibit the results for

a treatment implementation time defined as the shutdown of the first coal-fired EGU, while

the bottom panel present the results for a treatment implementation time defined as the

operation starting of the first natural gas-fired EGU. Results in the top panel consistently

indicate that the shutdown of a coal-fired unit generates significant and positive effects on

property values of houses located within a 0.6-mile distance from the fuel-switching plants.

Estimation result in column (2) indicates that reduction of coal combustion in the power

industry increases housing prices in 25-40 percentage points. This effect, however, disappears

at farther distances from the power plants. Results in column (2) for homes located within

0.9 miles from the facility suggest the existence of a negative impact. Yet, this effect is

77

weakly significant and disappears with other specifications. Findings in the bottom panel

suggest no effects on property prices after a gas-fired unit stars the power generation.

Table 2.5. Coal-to-Gas Impact on Property Prices - Triple Differences Estimator

Proximity

Control ≥ 6mi
(2)
(1)

Control ≥ 8mi
(2)
(1)

Control ≥ 10mi
(2)
(1)

Panel A. First Coal-fired EGU Stopped Operations

Treated ≤ 0.6mi: 1[Near]×1[After]

Treated ≤ 0.9mi: 1[Near]×1[After]

Treated ≤ 1.2mi: 1[Near]×1[After]

Treated ≤ 1.5mi: 1[Near]×1[After]

Panel B. First Gas-Fired EGU Started Operations

Treated ≤ 0.6mi: 1[Near]×1[After]

0.391∗∗∗
(0.084)

0.351∗∗∗
(0.083)

0.297∗∗∗
(0.075)

0.043
(0.083)

0.034
(0.069)

0.060
(0.083)

0.021
(0.073)

0.019
(0.064)

0.051
(0.079)

-0.075
(0.066)

-0.012
(0.058)

0.051
(0.089)

0.100
(0.108)

0.013
(0.074)

0.052
(0.074)

0.063
(0.077)

×

791,812
792,263
792,954
793,451

0.091
(0.105)

-0.004
(0.073)

0.036
(0.075)

0.049
(0.076)

×

791,812
792,263
792,954
793,451

0.065
(0.123)

-0.042
(0.067)

0.024
(0.070)

0.053
(0.077)

×

756,804
757,255
757,946
758,443

0.256∗∗∗
(0.072)
-0.094∗
(0.050)

-0.033
(0.050)

0.033
(0.085)

0.051
(0.118)

-0.059
(0.065)

0.005
(0.068)

0.034
(0.076)

×

756,804
757,255
757,946
758,443

0.297∗∗∗
(0.076)

0.277∗∗∗
(0.072)

-0.072
(0.070)

0.005
(0.060)

0.080
(0.104)

0.117
(0.141)

-0.020
(0.070)

0.046
(0.075)

0.082
(0.088)

×

707,435
707,886
708,577
709,074

-0.079
(0.056)

-0.006
(0.053)

0.072
(0.099)

0.113
(0.136)

-0.025
(0.069)

0.041
(0.074)

0.076
(0.086)

×

707,435
707,886
708,577
709,074

Treated ≤ 0.9mi: 1[Near]×1[After]

Treated ≤ 1.2mi: 1[Near]×1[After]

Treated ≤ 1.5mi: 1[Near]×1[After]

County FE
County × Year FE
N (0.6mi)
N (0.9mi)
N (1.2mi)
N (1.5mi)

Notes: All estimations include # of bathrooms, # of half bathrooms, # of bedrooms, house age, house age2,
square feet, square feett2, and fixed effects by year, city, and state. Clustered standard errors at the county ×
year level in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

78

2.5.4. Announcement Effects

This section deals with the possibility that property prices are indeed responding to

households’ expectations instead of capturing capitalizations of local air quality improve-

ments. To analyze this, I replace the year of the treatment implementation by the year in

which the natural gas power generation was first announced in local newspapers of areas

hosting the fuel-switching plants. In all the cases, this year is posterior to the year in which

gas-fired units start power generation. Tables 2.6 and 2.7 present the results of this approach

for the spatial near-far DID approach, and the spatial near-gas/near-coal DID approach, re-

spectively.

Table 2.6. Announcement Effects - Homes Far from a Coal-to-Gas Station as Controls

Proximity

Treated ≤ 0.6mi: 1[Near]×1[After]

Treated ≤ 0.9mi: 1[Near]×1[After]

Treated ≤ 1.2mi: 1[Near]×1[After]

Treated ≤ 1.5mi: 1[Near]×1[After]

County FE
County × Year FE
N (0.6mi)
N (0.9mi)
N (1.2mi)
N (1.5mi)

Control ≥ 6mi
(2)
(1)

Control ≥ 8mi
(2)
(1)

Control ≥ 10mi
(2)
(1)

0.052
(0.083)

-0.052
(0.054)

-0.044
(0.044)

-0.003
(0.054)

×

0.039
(0.077)

-0.067
(0.052)

-0.053
(0.044)

-0.010
(0.053)

×

0.027
(0.098)
-0.089∗
(0.049)

-0.053
(0.044)

0.0003
(0.058)

×

0.008
(0.089)
-0.104∗∗
(0.044)
-0.065∗
(0.039)

-0.009
(0.055)

×

0.049
(0.106)

-0.077
(0.050)

-0.036
(0.049)

0.025
(0.068)

×

0.038
(0.101)
-0.084∗
(0.046)

-0.040
(0.044)

0.021
(0.065)

×

791,812
792,263
792,954
793,451

791,812
792,263
792,954
793,451

756,804
757,255
757,946
758,443

756,804
757,255
757,946
758,443

707,435
707,886
708,577
709,074

707,435
707,886
708,577
709,074

Notes: All estimations include # of bathrooms, # of half bathrooms, # of bedrooms, house age, house age2,
square feet, square feet2, and fixed effects by year, city, and state. Clustered standard errors at the county
× year level in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

Results in Table 2.6 rreject the existence of announcement effects in the immediate

vicinity of a coal-to-gas-fired power plant (within 0.6 miles). Some negative effects are found

79

Table 2.7. Announcement Effects - Homes Near to a Coal-Fired Station as Controls

Proximity

1[Near] × 1[After]

City FE
State × Year FE
N

Treated & Control Treated & Control Treated & Control Treated & Control

≤ 1.5mi

≤ 0.6mi

≤ 0.9mi

≤ 1.2mi

(1)

-0.464∗∗
(0.184)

×

(2)

-0.385∗∗
(0.172)

×
×

(1)

0.477∗∗
(0.163)

×

(2)

0.061
(0.262)

×
×

(1)

0.307∗∗
(0.112)

×

(2)

0.003
(0.185)

×
×

(1)

0.411∗∗∗
(0.119)

×

(2)

0.258
(0.180)

×
×

Notes: All estimations include # of bathrooms, # of half bathrooms, # of bedrooms, house age, house age2,
square feet, square feet2, and fixed effects by year. Clustered standard errors at the county × year level in
parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

for larger buffers (within 0.9 and 1.2 miles). These significant announcement effects could be

explaining the negative results in the top panel of Table 2.5, which lead to a conclusion that

local air quality improvements are capitalized only in the close vicinity (within 0.6 miles)

and that other effects found for larger buffers are merely indicating market expectation

regarding these neighborhoods. The fact that local air pollution is generally perceived at

short distances from the polluting facilities, increases the plausibility of this explanation

as local air quality improvements are probably not perceived by households living beyond

1-mile from these plants. When the near-gas/near-coal approach is adopted, results in Table

2.7 once again suggest announcement effects. Whether the previous findings are driven by

buyers’ and sellers’ expectations instead of by improvements in environmental amenities, can

be answered through the exploration of time-heterogeneous impacts of the fuel switching.

2.6. Conclusions

The recent coal displacement by natural gas as the prime fuel source used in the

power generation promises several environmental improvements at both the global and the

local scale. This paper delves into the local air quality improvements of the coal-to-gas fuel-

switching process carried out by several power stations across the country. I study property

80

value impacts of this process using an extensive dataset of property transactions that took

place in the vicinity of five coal-to-gas-fired power stations across the country during 2009

to 2016. To approach causality, I use two difference-in-difference estimators combined with

a triple differences estimator across the space.

Results from a triple difference estimator indicate positive and significant impacts of

this fuel-switching process within 0.6 miles from the switched power plants. These impacts

are mostly found after the shutdown of coal-fired power generation units, and disappear

once the start up of power generation from the first gas-fired unit is considered. A test on

announcement effects suggest that market expectations could be driving some of the results,

mostly at larger distances from the switched plants.

81

APPENDIX

82

Table A2.1. Prime Mover Description - Gas-Fired EGUs Proposed During 2009-2016.

Prime Mover Description

Freq. Percent

Combined Cycle Steam Part
Combined Cycle Total Unit
Compressed Air Energy Storage
Combined Cycle Single Shaft
Combined Cycle Combustion Turbine Part
Fuel Cell
Combustion (Gas) Turbine
Internal Combustion Engine
Steam Turbine

Total

90
48
2
3
186
21
246
225
10

831

10.83
5.78
0.24
0.36
22.38
2.53
29.6
27.08
1.2

100

Capacity (MW)

Summer Winter

319.83
343.78

312.90
363.22

-
-

180.77
1.98
53.72
10.56
9.15

-
-

191.65
1.98
57.90
10.56
9.10

82.56

86.43

Notes: Average capacity within category. Combined cycle facilities generally consist of one or more gas
turbines (which compress air and mixes it with hot fuel making the gas turbine to spin to generate electricity),
one or more heat recovery steam generators (which capture exhaust heat from the gas turbine creating steam),
and one or more steam turbines (which send its energy to the generator drive shaft, where it is converted
into additional electricity). Compressed air energy storage facilities store energy generated at one point for
use at another time using ambient air compressed and stored under pressure in a underground cavern or
a large size storage tank. During high power demand, the compressed air is retrieved and used to drive a
turbine. Combined cycle single shaft plants comprise a gas turbine and a steam turbine driving a common
generator, while in combined cycle combustion turbine part plants each gas turbine and each steam turbine
have their own generator. Fuel cell facilities use electrochemical cells to convert chemical energy from a fuel
into electricity. Combustion gas turbine plants use an air compressor to feed the combustion chamber, a
combustion system that mix air and steady steam of fuel burning them at high temperatures, and a turbine
that receives the high temperature high pressure gas from the combustion system to spin an array of rotating
blades to produce electricity. Internal combustion engine plants use a combustion chamber as well, however,
combustion is intermittent. Steam turbine plants extract thermal energy from steam, which is sent through
a steam turbine that rotates the shaft of a generator to create electricity.

Table A2.2. Prime Mover Description - Coal-Fired EGUs Retired Over 2009-2016.

Prime Mover Description Freq. Percent

Capacity (MW)

Summer Winter

Steam Turbine

Total

481

481

100

100

122.71

124.05

122.71

124.05

Notes: Average capacity within category.

83

Figure A2.1. Differences-in-Differences Design #1 - Houses Near and Far from
Coal-to-Gas-Fired Power Plants

Notes: The figure depicts the spatial assortment of houses around a coal-to-gas-fired power plant X before
(left-hand side, or year 1) and after (right-hand side, or year 2) the switching. In both cases, a radius of
distance d from power plant X is used to define the set of houses that are located ”near” the switching
facility (gray area). In this double differences design, the first price difference comes from the comparison of
housing units inside the gray area (near X), and housing units outside in the blue area (far X). The second
price difference comes from the comparison of housing units before (year 1) and after (year 2) the switching.

F irst Dif f erence = P ricesN earX − P ricesF arX
Second Dif f erence = P ricesY ear2 − P ricesY ear1

84

Figure A2.2. Differences-in-Differences Design #2 - Houses Near Coal-to-Gas-Fired and
Coal-Fired Power Plants

Notes: The top figure depicts the spatial assortment of houses around a coal-to-gas-fired power plant X,
before (left-hand side, or Year 1) and after (right-hand side, or Year 2) the switching. A radius of distance
d from power plant X is used to define the set of houses that are located ”near” the switching facility (gray
area). The bottom figure depicts the spatial assortment of houses around a coal-fired power plant Z before
(left-hand side, or Year 1) and after (right-hand side, or Year 2) power plant X switched fuels. A radius
of distance d from power plant Z is used to define the set of houses that are located ”near” the control
facility (gray area). In this double differences design, the first price difference comes from the comparison of
housing units inside the gray area that surrounds power plant X, and housing units inside the gray area that
surrounds power plant Z. The second price difference comes from the comparison of housing units before
(year 1) and after (year 2) the switching.

F irst Dif f erence = P ricesN earX − P ricesN earZ
Second Dif f erence = P ricesY ear2 − P ricesY ear1

85

Figure A2.3. Triple Differences - Houses Near and Far from Coal-to-Gas-Fired Power
Plants and Houses Near and Far from Coal-Fired Power Plants

Notes: The top figure depicts the spatial assortment of houses around a coal-to-gas-fired power plant X,
before (left-hand side, or Year 1) and after (right-hand side, or Year 2) the switching. The bottom figure
depicts the spatial assortment of houses around a coal-fired power plant Z before (left-hand side, or Year 1)
and after (right-hand side, or Year 2) power plant X switched fuels. In this triple differences design, housing
units that are located inside the gray area that surrounds power plant X are part of the treatment group,
compared against houses inside the blue area, the gray area surrounding power plant Z, and the orange area.
These houses are compared before and after the switching. In this triple differences design, the first price
difference comes from the comparison of housing units inside the gray area that surrounds power plant X
(near X), and housing units outside in the blue area (far X). The second price difference comes from the
comparison of housing units inside the gray area that surrounds power plant Z (near Z), and housing units
outside in the orange area (far Z). The third price difference comes from the comparison of housing units
before (year 1) and after (year 2) the switching.

F irst Dif f erence = P ricesN earX − P ricesF arX
Second Dif f erence = P ricesN earZ − P ricesF arZ
T hird Dif f erence = P ricesY ear2 − P ricesY ear1

86

Figure A2.4. Parallel Trends Assumption for Houses Near and Far (d ≥ 6mi) from Coal-to-Gas-Fired Power Plant - First
Retired Coal-Fired EGU as Treatment

Notes: Using the year in which the first coal-fired EGU stopped operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≥ 6mi from a coal-to-gas-fired station as controls.

87

Figure A2.5. Parallel Trends Assumption for Houses Near and Far (d ≥ 8mi) from Coal-to-Gas-Fired Power Plant - First
Retired Coal-Fired EGU as Treatment

Notes: Using the year in which the first coal-fired EGU stopped operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≥ 8mi from a coal-to-gas-fired station as controls.

88

Figure A2.6. Parallel Trends Assumption for Houses Near and Far (d ≥ 10mi) from Coal-to-Gas-Fired Power Plant - First
Retired Coal-Fired EGU as Treatment

Notes: Using the year in which the first coal-fired EGU stopped operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≥ 10mi from a coal-to-gas-fired station as controls.

89

Figure A2.7. Parallel Trends Assumption for Houses Near and Far (d ≥ 6mi) from Coal-to-Gas-Fired Power Plant - First
Operative Gas-Fired EGU as Treatment

Notes: Using the year in which the first gas-fired EGU started operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≥ 6mi from a coal-to-gas-fired station as controls.

90

Figure A2.8. Parallel Trends Assumption for Houses Near and Far (d ≥ 8mi) from Coal-to-Gas-Fired Power Plant - First
Operative Gas-Fired EGU as Treatment

Notes: Using the year in which the first gas-fired EGU started operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≥ 8mi from a coal-to-gas-fired station as controls.

91

Figure A2.9. Parallel Trends Assumption for Houses Near and Far (d ≥ 10mi) from Coal-to-Gas-Fired Power Plant - First
Operative Gas-Fired EGU as Treatment

Notes: Using the year in which the first gas-fired EGU started operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≥ 10mi from a coal-to-gas-fired station as controls.

92

Figure A2.10. Parallel Trends Assumption for Homes Near Coal-to-Gas-Fired and Coal-Fired Stations - First Retired
Coal-Fired EGU as Treatment

Notes: Using the year in which the first coal-fired EGU stopped operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a coal-fired station as controls.

93

Figure A2.11. Parallel Trends Assumption for Homes Near Coal-to-Gas-Fired and Coal-Fired Stations - First Operative
Gas-Fired EGU as Treatment

Notes: Using the year in which the first gas-fired EGU started operations as the treatment year. Homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a
coal-to-gas-fired station as treated homes and homes at d ≤ 0.6, 0.8, 1.2, and 1.5mi from a coal-fired station as controls.

94

REFERENCES

95

REFERENCES

Arceo, Eva, Rema Hanna, and Paulina Oliva (2016). “Does the Effect of Pollution on Infant
Mortality Differ Between Developing and Developed Countries? Evidence from Mexico
City”. In: The Economic Journal 126.591, pp. 257–280.

Bateson, Thomas F and Joel Schwartz (2007). “Children’s Response to Air Pollutants”. In:

Journal of Toxicology and Environmental Health, Part A 71.3, pp. 238–243.

Bell, Michelle L, Francesca Dominici, and Jonathan M Samet (2005). “A Meta-Analysis of
Time-Series Studies of Ozone and Mortality With Comparison to the National Morbidity,
Mortality, and Air Pollution Study”. In: Epidemiology 16.4, pp. 436–445.

Bell, Michelle L et al. (2004). “Ozone and Short-term Mortality in 95 US Urban Communities,

1987-2000”. In: JAMA 292.19, pp. 2372–2378.

Black, James et al. (2010). Cost and Performance Baseline for Fossil Energy Plants Volume

1: Bituminous Coal and Natural Gas to Electricity.

Blomquist, Glenn (1974). “The Effect of Electric Utility Power Plant Location on Area

Property Value”. In: Land Economics 50.1, pp. 97–100.

Burnham, Andrew et al. (2011). “Life-Cycle Greenhouse Gas Emissions of Shale Gas, Natural
Gas, Coal, and Petroleum”. In: Environmental Science & Technology 46.2, pp. 619–627.

Carlson, Claire L and Domy C Adriano (1993). “Environmental Impacts of Coal Combustion

Residues”. In: Journal of Environmental Quality 22.2, pp. 227–247.

Chauhan, A J et al. (2003). “Personal Exposure to Nitrogen Dioxide (NO2) and the Severity

of Virus-Induced Asthma in Children”. In: The Lancet 361.9373, pp. 1939–1944.

Chay, Kenneth Y and Michael Greenstone (2003a). Air Quality, Infant Mortality, and the

Clean Air Act of 1970. Tech. rep. National Bureau of Economic Research.

— (2003b). “The Impact of Air Pollution on Infant Mortality: Evidence from Geographic
Variation in Pollution Shocks Induced by a Recession”. In: The Quarterly Journal of
Economics 118.3, pp. 1121–1167.

Chen, Yuyu et al. (2013). “Evidence on the Impact of Sustained Exposure to Air Pollution
on Life Expectancy from China’s Huai River Policy.” In: Proceedings of the National
Academy of Sciences of the United States of America 110.32, pp. 12936–12941.

Clarkson, Thomas W (1990). “Mercury – An Element of Mystery”. In: The New England

Journal of Medicine 323.16, pp. 1137–1139.

96

Currie, Janet and Matthew Neidell (2005). “Air Pollution and Infant Health: What Can We
Learn from California’s Recent Experience?” In: The Quarterly Journal of Economics
120.3, pp. 1003–1030.

Currie, Janet et al. (2015). “Environmental Health Risks and Housing Values: Evidence from
1,600 Toxic Plant Openings and Closings.” In: The American Economic Review 105.2,
pp. 678–709.

Davis, L W (2011). “The Effect of Power Plants on Local Housing Values and Rents ”. In:

Review of Economics and Statistics 93.4, pp. 1391–1402.

Deng, Guoying, Manuel A Hernandez, and Shu Xu (2014). “When Power Plants Leave
Town: Environmental Quality and the Housing Market in China”. In: Available at SSRN
2502442.

Freeman, A Myrick (1979). “The Hedonic Price Approach to Measuring Demand for Neigh-
borhood Characteristics”. In: The Economics of Neighborhood. Ed. by David Segal. El-
sevier. Chap. 9, pp. 191–217.

Freeman III, A Myrick, Joseph A Herriges, and Catherine L Kling (2014). The Measurement

of Environmental and Resource Values: Theory and Methods. Routledge.

Gauderman, W James et al. (2005). “Childhood Asthma and Exposure to Traffic and Nitro-

gen Dioxide”. In: Epidemiology 16.6, pp. 737–743.

Grainger, Corbett A (2012). “The Distributional Effects of Pollution Regulations: Do Renters

Fully Pay for Cleaner Air?” In: Journal of Public Economics 96.9-10, pp. 840–852.

Guallar, Eliseo et al. (2002). “Mercury, Fish Oils, and the Risk of Myocardial Infarction”.

In: New England Journal of Medicine 347.22, pp. 1747–1754.

Hodge, T. R. (2011). “The Effect of Ethanol Plants on Residential Property Values: Evidence

from Michigan”. In: The Journal of Regional Analysis Policy 41.2, pp. 148–167.

Jaramillo, Paulina, W Michael Griffin, and H Scott Matthews (2007). “Comparative Life-
cycle Air Emissions of Coal, Domestic Natural Gas, LNG, and SNG for Electricity Gen-
eration”. In: Environmental Science & Technology 41.17, pp. 6290–6296.

Knittel, Christopher R, Douglas L Miller, and Nicholas J Sanders (2016). “Caution, Drivers!
Children Present: Traffic, Pollution, and Infant Health”. In: Review of Economics and
Statistics 98.2, pp. 350–366.

Mendelsohn, Robert and Sheila Olmstead (2009). “The Economic Valuation of Environ-
mental Amenities and Disamenities: Methods and Applications”. In: Annual Review of
Environment and Resources 34, pp. 325–347.

97

Muehlenbachs, Lucija, Elisheba Spiller, and Christopher Timmins (2015). “The Housing
Market Impacts of Shale Gas Development”. In: The American Economic Review 105.12,
pp. 3633–59.

Orehek, J et al. (1976). “Effect of Short-Term, Low-Level Nitrogen Dioxide Exposure on
Bronchial Sensitivity of Asthmatic Patients.” In: Journal of Clinical Investigation 57.2,
pp. 301–307.

Palmquist, Raymond B (1992). “Valuing Localized Externalities”. In: Journal of Urban Eco-

nomics 31.1, pp. 59–68.

Pershagen, G¨oran et al. (1995). “Air Pollution Involving Nitrogen Dioxide Exposure and

Wheezing Bronchitis in Children”. In: International Journal of Epidemiology 24.6, pp. 1147–
1153.

Ridker, R G and J A Henning (1967). “The Determinants of Residential Property Values
with Special Reference to Air Pollution ”. In: The Review of Economics and Statistics
49.2, p. 246.

Ridker, Ronald Gene (1967). Economic Costs of Air Pollution, Studies in Measurement.

Frederick A. Praeger, Publishers, New York.

Rosen, Sherwin (1974). “Hedonic Prices and Implicit Markets: Product Differentiation in

Pure Competition”. In: The Journal of Political Economy 82.1, pp. 34–55.

Salonen, Jukka T et al. (1995). “Intake of Mercury from Fish, Lipid Peroxidation, and the
Risk of Myocardial Infarction and Coronary, Cardiovascular, and Any Death in Eastern
Finnish Men”. In: Circulation 91.3, pp. 645–655.

Tollefson, Jeff (2013). “Methane Leaks Erode Green Credentials of Natural Gas”. In: Nature

493.7430, p. 12.

Ummel, Kevin (2012). “Carma Revisited: an Updated Database of Carbon Dioxide Emissions
from Power Plants Worldwide”. In: Available at SSRN: https://ssrn.com/abstract=
2226505.

U.S. Energy Information Administration (2017). How Much Carbon Dioxide is Produced
when Different Fuels are Burned? Retrieved from https://www.eia.gov/tools/faqs/
faq.php?id=73&t=11. Accessed: 2017-06-08.

— (2018). United States Electricity Profile 2016. Retrieved from https://www.eia.gov/

electricity/state/unitedstates/. Accessed: 2018-01-25.

U.S. Environmental Protection Agency (2014). EPA’s Report on the Environment (ROE):
Sulfur Dioxide Emissions. Retrieved from https://cfpub.epa.gov/roe/indicator.
cfm?i=22. Accessed: 2017-12-05.

98

U.S. Environmental Protection Agency (2015). EPA’s Report on the Environment (ROE):
Mercury Emissions. Retrieved from https://cfpub.epa.gov/roe/indicator.cfm?i=
14. Accessed: 2018-01-20.

— (2018). “Inventory of US Greenhouse Gas Emissions and Sinks: 1990-2016”. In: Environ-

mental Protection Agency 2018.

World Health Organization (2006). Air Quality Guidelines. Global Update 2005. World

Health Organization.

99

ESSAY 3. THE EFFECTIVENESS OF AIR QUALITY WARNINGS AND

TEMPORARY DRIVING BANS: EVIDENCE FROM AIR POLLUTION

AND URBAN TRANSIT FLOWS IN SANTIAGO

Abstract

Driving restrictions are common strategies to curb mobile source emissions in many cities of

the world. In this study, I use high frequency data on air pollution and urban transit flows to

evaluate the effectiveness of temporary license plate-based driving bans, triggered with 24-

hour air quality warnings, in curbing local air pollution in Santiago, Chile. Taking advantage

of the discontinuities in the air quality index used to announce these warnings, I estimate

the effect of these incidents, and their driving bans, in reducing ambient concentrations of

major pollutants from car emissions. As effective driving bans lessen mobile source pollution

by getting cars off the roads and pushing drivers towards cleaner forms of transportation,

I also evaluate this policy using data on vehicle trips, and on the use of Santiago’s mass-

transit systems. For identification, I employ a fuzzy regression discontinuity design that uses

the thresholds in Santiago’s air quality index as instruments for the episodes’ announcement.

Results show that on average, temporary driving restrictions reduce car traffic by 4-6% during

peak hours, and around 7-10% during off-peak hours on days with air quality warnings. This

is consistent with reductions in PM, CO, and NOX concentrations mostly during peak hours.

Evidence from alternative transportation modes indicates that drivers substitute primarily

towards the metro, primarily during hours at which the system is not running at full capacity.

Days with exacerbated levels of air pollution, however, show lower increases in mass-transit

system ridership, which suggest an avoidance behavior on affected drivers who stay at home

to avoid outdoor exposure to pollution. Results in this study provide suggestive evidence

that air quality warnings about the risks of outdoor exposure to pollution can help securing

the effectiveness of temporary driving restrictions.

100

3.1.

Introduction

This paper evaluates a command-and-control policy aimed at reducing mobile source

pollution in Santiago during days of critical air pollution. High levels of air pollution severely

affect health (Chay and Greenstone, 2003; Neidell, 2004; Currie and Neidell, 2005; Knittel,

Miller, and Sanders, 2016), educational outcomes (Currie et al., 2009; Stafford, 2015), labor

(Ostro, 1983; Hausman, Ostro, and Wise, 1984; Hanna and Oliva, 2015), and productivity

(Crocker and Horst, 1981; Zivin and Neidell, 2012; Chang et al., 2016; Lichter, Pestel,

and Sommer, 2017), particularly in developing countries (Greenstone and Hanna, 2014;

Greenstone and Jack, 2015; Hanna and Oliva, 2015). In Latin America alone, air pollution

puts at risk the health of more than 80 million inhabitants, generating annual losses of

about 65 million working days (United Nations Environment Programme, 2002). Among

countries in this region, Chile ranks second after Mexico in exposure to coarse particulate

matter (PM10) in urban areas, and first in exposure to fine particulate matter (PM2.5), with

Santiago as one of the most polluted cities in the country (World Health Organization, 2014).

As in many metropolitan areas, poor air quality in Santiago is primarily due to mobile

source pollution. Mobile sources account for 60 to 95 percent of the city’s annual PM10

and PM2.5 emissions, and contribute substantially to the concentration of other pollutants

such as nitrogen oxides (NOX), carbon monoxide (CO), and volatile organic compounds

(VOCs).1 Temporary driving restrictions aimed at curbing this pollution prohibit the driving

in Santiago of a share of light-duty private vehicles based on the last digits of their license

plates.2 As a deviation from similar schemes in other countries, however, these bans affect

both dirty and clean cars, and only during days of severe air contamination, when a set of

mitigation actions is put into place.3 These mitigation actions are part of 24-hour preventive

1Data from the Register on Pollutant Release and Transfer (Registro de Emisiones y Transferencias de

Contaminantes - RETC)’s website: http://www.retc.cl. Retrieved on March 15, 2017.

2Temporary driving restrictions complement a permanent ban that have affected dirty cars in Santiago

since 1990. See further details in Section 3.3.

3Clean cars are those that hold a green sticker, used to designate vehicles that meet specific emissions
standards. These stickers are mandatory self-adhesive sticker added to the windshields of all fuel-efficient
vehicles (e.g. eco-diesel, electric cars, etc.), and vehicles with catalytic converters, which certifies vehicles’

101

measures called “environmental episodes” (hereafter called “air quality warnings” or simply

“episodes”), implemented in Santiago whenever official authorities foresee worrisome levels

for the city’s air quality index. Different index thresholds lead to a three-tier system of

environmental episodes (alerts, pre-emergencies, and emergencies), whose temporary driving

bans further tightens a permanent driving restriction on dirty vehicles, and impose new

restrictions on clean vehicles.4,5 Taking advantage of the discontinuities in the issuance of

these events, I empirically evaluate the effects of these episodes, and their short-term driving

bans, in curbing mobile source pollution in Santiago.

License plate-based driving restrictions constitute a common governmental strategy

to reduce airborne pollution and traffic congestion in many cities of the world (e.g. Athens,

Beijing, Berlin, Bogota, Mexico City, Milan, New Delhi, S˜ao Paulo).6,7 Mixed findings on

their causal impacts, however, still cast doubt on their effectiveness as an instrument to im-

prove local air quality. Empirical evidence shows drivers bypass these bans either by shifting

their driving towards hours or days unaffected by the policy, or by purchasing a second car to

fully avoid these restrictions, are the main caveats to the effectiveness of this policy (Davis,

2008; Gallego, Montero, and Salas, 2013; Bonilla, 2016; Zhang, Lawell, and Umanskaya,

2017). Most of these findings relate to evidence about mobile source pollution, but little is

known about the impacts these bans have on vehicle flows or on alternative modes of trans-

compliance with the emission standards needed to drive in Santiago’s Metropolitan Region. Vehicles earning
a green sticker are brand new vehicles sold after September 1st, 1992, and emitting less than 0.25g/km
of hydrocarbons, less than 2.11g/km of CO, less than 0.62g/km of NOX , and less than 0.125g/km of
particulate matter. Drivers obtain these stickers when getting their vehicle permits, issued after annual
mandatory cars inspections that certify their emissions. During 2005, 30% of Santiago’s vehicles were dirty
vehicles responsible for 54% and 24% of CO and PM10 emissions from mobile sources, respectively (Atal,
2009). Since then, however, the number of dirty vehicles driving in the city has dropped dramatically. Similar
calculations in 2015 numbers reveal that dirty vehicles are responsible for approximately 3.6% of current CO
and 1.6% of current PM10 emissions from mobile sources.

4For instance, clean private vehicles with license plates ending in 0 and 1 may not be used during the

first pre-emergency episode of the year.

5Vintage-specific driving restrictions may have significant impacts on fleet turnover. Barahona, Gallego,
and Montero (2016) show that Santiago’s permanent driving restriction pushes the city’s vehicle composition
toward vehicles with a green sticker, which are not affected by this policy. Indeed, 56% of 2001 registered
cars in Santiago had a green sticker. These cars now represent 98% of registered cars in the city (INE, 2015).
6Other common type of driving restriction is the low emission zone program mostly used to reduce

congestion. Under this program, drivers are banned from entering to some specific zones of the city.

7I use license plate-based driving restrictions and driving restrictions interchangeably throughout.

102

portation. Effective driving bans get cars off the roads and push their drivers towards cleaner

and uncrowded modes of transportation. An effective policy is therefore expected to produce

measurable impacts not only on pollution but also on car and mass-transit system flows. Re-

duced automotive congestion due to the implementation of these bans, however, could also

induce unaffected drivers to drive more, offsetting any initial reductions in pollution and car

flows, yet not necessarily affecting ridership in mass-transit systems. This paper sheds light

into these effects by studying the causal effects that temporary driving bans, triggered with

air quality warnings, generate on pollution, vehicle flows, and mass-transit system ridership

as a substitute of driving. The lifespan of these bans can induce different long-term behaviors

by affected drivers that could deviate from previous evidence and lead to different conclu-

sions regarding the effectiveness of the policy. The fact that driving bans are triggered with

temporary air quality warnings might also encourage an avoidance behavior and persuade

the marginal constrained driver to stay at home instead of substituting her driving. Due to

its policy innovation, and the robust data available, Santiago’s experience provides an im-

portant opportunity to estimate the effectiveness of short-term driving restrictions triggered

with warnings that uncovered the risks of outdoor exposure to pollution.8

I analyze the effectiveness of Santiago’s policy using hourly ambient concentrations

of four major pollutants from automotive emissions: PM10, PM2.5, CO, and NOX. I also

look at hourly vehicle trips recorded by counting stations, and at the number of trips taken

in both Santiago’s Metro and Santiago’s bus rapid transit system Transantiago. To explore

causality, I use a fuzzy regression discontinuity (FRD) design that employs the air quality

index as the variable forcing the probability of an episode. The FRD approach exploits

the arbitrary cutoffs in the air quality index that define the different air quality episodes,

and accounts for potential confounding factors affecting the outcomes by defining a narrow

window around these cutoffs. Using this strategy, I evaluate both the overall episodes’ impact

in a pooled FRD design, and their individual effectiveness in a multiple cutoff FRD design.

8The possibility that temporary driving bans jointly announced with air quality warnings could be indeed

an efficient design was first suggested in Gallego, Montero, and Salas (2013).

103

The time span between the issuance of an episode and the implementation of its abatement

actions rules out potential unobservables simultaneously affecting both the forcing and the

outcome variables. Additionally, the use of hourly data on traffic flows allows me to isolate

the effects of driving restrictions from other abatement actions triggered by these episodes

but targeting stationary source emissions. The empirical evaluation of temporary driving

bans using evidence on urban transit flows constitutes one of the main contributions of this

study.

Results show that temporary driving restrictions effectively keep drivers off the roads

during days of critical air pollution. Estimation results indicate that the issuance of episodes

decreases average vehicle trips by around 4-5 percentage points, and average pollution con-

centrations by around 2-3 percentage points both during peak hours. Average vehicle trips

are also curtailed by 7 percentage points during off-peak hours. Higher reductions in vehi-

cle trips during off-peak hours relative to peak hours suggests that some drivers may take

advantage of a reduced road congestion during restricted hours of days with driving bans.

Results reveal positive effects on daily use of mass-transit systems as well, mostly towards the

metro and during off-peak hours, which indicates that restricted drivers substitute towards

alternative modes of transportation running with some extra capacity.

When allowing for heterogeneous episodes’ effects, results consistently indicate re-

ductions in both hourly average and maximum vehicle trips, and pollution concentrations

during all episodes. Average car trips during peak hours are curtailed by 4-6 percentage

points in days with air quality alerts, and by 5-6 percent during air quality pre-emergencies.

Similar results are found for car trips during off-peak hours in days with the mildest episodes

neglecting the existence of an intertemporal shift in traffic towards unrestricted hours. Re-

ductions during off-peak hours, however, disappear during stronger episodes as drivers may

take advantage of a reduced road congestion when stricter bans are in place. Findings for

mass-transit system ridership indicate daily increase in ridership during days with both alerts

and pre-emergencies, specifically for the metro, and during restricted and unrestricted hours

104

as well. These increments, however, are lower during days with more severe restrictions, in-

dicating that some drivers might stay at home to avoid outdoor exposure to pollution during

days of exacerbated air contamination. In this case, schemes combining driving restrictions

together with air quality warnings would constitute a successful design to temporarily curtail

the number of cars in the roads.

The plan of this paper is laid out as follows. Section 3.2 reviews the literature on

driving restrictions and highlights the contributions of the present study. Section 3.3 provides

background details on Santiago’s air quality episodes and driving restrictions, while Section

3.4 describes the data. Section 3.5 describes the empirical strategy, section 3.6 presents the

results and the robustness checks. Summary and conclusions are in Section 3.7.

3.2. Driving Restrictions and Environmental Episodes

in the Literature

Despite the rich body of studies addressing the impacts of license plate-based driving

restrictions on local air quality, conclusions on their effectiveness are still sparse. Looking at

the effectiveness of the Hoy No Circula (HNC) driving restriction program implemented in

1989 in Mexico City, Eskeland and Feyzioglu (1997) suggest that HNC led to an unintentional

increase in driving, and so on air pollution, due to a greater use of old cars, congestion effects,

and purchasing of a second car after the program implementation. Davis (2008) reinforces

this conclusion through a quasi-experimental approach on impact evaluation of HNC on

several airborne pollutants. Using a regression discontinuity (RD) design around the time of

the HNC adoption, the author finds no evidence of pollution reductions in Mexico City and,

instead, shows indications of a shift in driving towards hours or days unaffected by the policy.

The author revises evidence from gasoline sales and vehicle registrations, and concludes that

HNC led to an increase in the number of dirty vehicles on the roads.9 Gallego, Montero, and

9Salas (2010) revisits Davis (2008)’s estimations and finds that both the sign and significance of the
HNC impacts vary greatly with the time window and the length of the time trend polynomial used in

105

Salas (2013) offer a new overview on HNC by testing spatial and demographic heterogeneous

effects of the policy. Using a similar identification strategy as Davis (2008), the authors find

evidence of a 5-13 percentage point reduction in CO concentrations after HNC, but only

for middle-income groups and during the first month of the policy implementation. These

effects, however, vanish in the long run where the policy is found to lead to increases in CO

concentrations after twelve months of its adoption. Regardless of the evidence from HNC,

Gallego, Montero, and Salas (2013) present the first study shedding light into potential

time-differentiated behavioral effects of these bans on affected drivers.

Similar works highlighting the ineffectiveness of driving bans are Cantillo and De

Dios Ort´uzar (2014), Bonilla (2016), and Zhang, Lawell, and Umanskaya (2017), for the case

of Bogota’s Pico y Placa (PyP) program. PyP was created in 1998 with the intention of

reducing traffic congestion in the city. Yet, authors such as Cantillo and De Dios Ort´uzar

(2014), question its outcomes highlighting the long-run ineffectiveness of this policy. Regard-

ing the different PyP phases and their effects on CO concentrations, Bonilla (2016) finds no

evidence of air quality improvements due to this policy, and suggests instead an increase in

pollution concentrations during the most drastic phase of PyP. Similarly, Zhang, Lawell, and

Umanskaya (2017) suggest that PyP led to an increase in airborne pollution, particularly for

pollutants such as NO2, NOX, and O3. Among the mechanisms explaining this failure, the

authors emphasize the role of the pollutants’ atmospheric chemistry as a factor reducing the

effectiveness of this policy.

In spite of the above, other works underline the effectiveness of driving bans in curbing

mobile source pollution. Carrillo, Malik, and Yoo (2016) study the effects of Quito’s Pico y

Placa (PyP) program in CO ambient concentrations. Using a difference-in-difference (DID)

approach, the authors find a 9-11 percent reduction in CO concentrations and, contrary to the

the RD specification. Particularly, smaller time windows around the HNC implementation reveal positive
and significant effects of the program on several airborne pollutants, which unveil a potential behavioral
distinction between short-term and long-term drivers’ adaptation to this policy. However, the order of the
time trend polynomial used to control for time-varying factors affecting pollution levels compromises all the
HNC’s significant impacts found by these authors.

106

findings for HNC, the authors find no evidence of shifts in traffic towards unrestricted hours.

Similar conclusions are found for Beijing’s driving restriction plan. Viard and Fu (2015)

use both a RD and a DID design to evaluate the every-other-day and the one-day-a-week

restrictions in this city. The authors report 18 and 21 percent reductions in particulate matter

concentrations due to these two policies, respectively; while similar results are reported

by Chen et al. (2013) on the short-term driving bans imposed during the 2008 Olympic

Games. Wolff (2014) obtains similar conclusions for the low-emission zones (LEZ) program

in Germany, which restricts dirty vehicles on specific roads. Using a DID approach with

geographical variation, the author finds 9 percentage point reduction in heavy particulate

matter concentrations due to this program, rejecting the hypothesis that dirty vehicles drive

more outside LEZ. The success of driving bans in these cases, as opposed to the failure of

comparable designs in others, has been linked to either the population’s general propensity

to comply with restrictions, as in the case of Beijing’s bans (Viard and Fu, 2015), and to

specific characteristics of the policy that foster compliance at a higher degree, such as the

emission-levels car differentiation of the LEZ program that incentivizes drivers to adopt

cleaner vehicles (Wolff, 2014).10

For the case of Santiago’s policy, previous studies suggest a significant relationship

between environmental episodes and either air pollution reductions (Atal, 2009; Troncoso,

Grange, and Cifuentes, 2012; Mullins and Bharadwaj, 2015), or the discouraging use of

private cars in the city (Grange and Troncoso, 2011). Particularly, Atal (2009) finds CO

and PM10 reductions during pre-emergency episodes, while Troncoso, Grange, and Cifuentes

(2012) suggests PM10, PM2.5, CO and NOX reductions during the same episodes although,

only for weekdays. Using evidence on urban traffic flows, Grange and Troncoso (2011) show

a 5.3 percent reduction in the number of cars in circulation and a 3.3 percent increase in

metro ridership during pre-emergencies.

While previous evidence on Santiago’s case suggests the effectiveness of the policy,

10Evidence on the incentives of cleaner technologies due to driving bans has also been suggested by

Barahona, Gallego, and Montero (2016) for Santiago’s case.

107

none of the previous works prioritize causality. A sole exception is the work by Mullins

and Bharadwaj (2015), who address causal effects of Santiago’s environmental episodes on

particulate matter concentrations, and mortality rates. The authors use a DID approach

combined with propensity score matching estimators to construct a control group of days with

similar pollution levels before the enforcement of this policy. They find suggestive evidence

that environmental episodes reduce PM10 concentrations by 20 percent, with this reduction

persisting during subsequent days. The authors also suggest a reduction in mortality rates

of the elderly, especially when comparing days that follow these announcements.

Similar in spirit to Mullins and Bharadwaj (2015), the present study uses a quasi-

experimental approach to derive causal effects of environmental episodes on air pollution and

urban traffic flows. As an important departure from their work, however, I use instruments

for announcement of episodes, which allows me to obtain unbiased estimators in presence

of omitted variables.11 In addition, I disentangle the effects of temporary driving bans from

other abatement actions also triggered by these episodes, which is not previously addressed in

literature. As an extra feature, I evaluate the effects of these bans using additional evidence

on traffic counts, and on the use of alternative modes of transportation. To my knowledge,

this is the first quasi-experimental study evaluating temporary driving bans using these

datasets. Finally, the use of high-frequency data on air pollution allows me to deal with the

presence of atmospheric persistence, which could affect the concentration of some pollutants

(Gibson and Carnovale, 2015). The inclusion of dynamics in the concentration of pollutants

also fills in an existing gap in the related literature.

11An important consideration deserves the construction of a control group in Mullins and Bharadwaj
(2015). The authors assume perfect compliance with the policy during days with poor air quality, although
this assumption may not hold for all days. There could be some days for which an episode was issued
even under relatively good levels of air quality, and days with some critical airborne contamination but for
which no episode was declared. This is because there is some discretion embedded in the episodes’ issuance
(see Sections 3.3 and 3.4 for more details).
In this non-compliance scenario, estimations in Mullins and
Bharadwaj (2015) are likely biased.

108

3.3. Background Information

The severity of air contamination in Chile has motivated the implementation of mul-

tiple actions to mitigate air The severity of Chile’s air contamination motivated implemen-

tation of multiple actions to mitigate air pollution, especially in the metropolitan area of

Santiago, home of almost half of the country’s population. Santiago is located in the central

part of Chile, in a basin surrounded by mountains with an altitude that ranges between

1,500 and 4,500m. These mountains prevent strong wind formation, which in combination

with thermal inversions decrease the dispersion of pollutants, especially during colder months

(World Health Organization, 2006).12 Figure 3.1 displays hourly average ambient pollutant

concentrations in Santiago for heavy particulate matter (PM10 - panel A), fine particulate

matter (PM2.5 - panel B), carbon monoxide (CO - panel c), and nitrogen oxide (NOX - panel

d). The second axis displays the hourly average number of car trips during both summer and

winter months. During summer, all pollutants peak daily at 7 am, although with a different

concentration pattern throughout the day. Concentrations of particulate matter dissipate at

a lower rate than CO and NOX, the two main tailpipe pollutants from car emissions. These

two last pollutants, however, peak again later in the day in a fashion that is consistent with

the daily pattern observed for vehicle trips. Average car trips peak daily at 7 am as well, and

then later in the day at 1 pm and 6 pm. Even in presence of these three daily peaks, mobile

source pollution is relatively higher during mornings, which reveals the effect that not only

cars have on these concentrations but also radiation levels hitting the ground. Sun radiation

controls the presence of cold and warmer layers of air in the atmosphere. An inversion in

these layers due to low radiation, known as thermal inversion, can reduce the diffusion of

pollutants, and deteriorating Santiago’s air quality (Perez, 2008; Saide et al., 2011). The

role that thermal inversions have in the dispersion of pollutants is clear when looking at am-

bient concentrations during winter months.13 Pollutant concentrations currently peak twice

12Winter goes from April 1st through August 31st.
13Thermal inversion generally occurs during winter nights, where the amount of sun radiation is low
compared to the radiation from the ground, making the air near the Earth’s surface cooler. This phenomenon

109

a day despite the fact that daily patterns in car trips remain similar. This situation reflects

the importance of controlling for inversion in the estimation process since colder hours can

exacerbate the concentration of pollutants.

Figure 3.1. Hourly Average Pollution and Vehicle Trips by Season

(a) PM10

(b) PM2.5

(c) CO

(d) NOX

Notes: Particulate matter is in micrograms per cubic meter, CO is in parts per millions, and NOX is in
parts per billions. Vehicle trips are averaged across stations.

generates a layer of warm air that settles over the layer of cooler air near the ground, and prevents pollutants
from rising.

110

3.3.1. Air Pollution Prevention and Clean-Up Plan

During the late 80s and early 90s, the Special Decontamination Commission of the

Metropolitan Region and the Air Pollution Prevention and Clean-up Plan (PPDA) intro-

duced several abatement actions in Santiago aimed at targeting emissions from both station-

ary and mobile sources (OECD, 2005). Emissions from stationary sources are tackled with

emission standards, tradable permits, and emissions taxes, while mobile source pollution is

addressed with the Critical Episodes Management (Gesti´on de Episodios Cr´ıticos -GEC)

program, which establishes license-plate based driving restrictions on light-duty private cars

every year from April 1st to August 31st. The GEC particularly employs a permanent (sea-

sonal) license plate-based restriction that, fully in operation since 1990, prohibits the driving

dirty vehicles in Santiago of from 7:30 am to 9 pm during weekdays.14 When meteorological

conditions in the city prevent the dispersion of pollutants and lead to particulate matter

concentrations that exceed the PPDA tolerable ceilings, the GEC complements the perma-

nent prohibition by setting short-term driving bans on both dirty and clean vehicles.15 In

these cases, the PPDA considers the issuance a 24-hour environmental episode, or air quality

warning, which activates a set of short-term actions aimed at improving air quality along

with recommendations to avoid outdoor exposure to pollution. Since the introduction of

this clean-up plan, Santiago’s air quality has shown progress in terms of daily PM10 and CO

concentrations (see Figure A3.2 in the Appendix A).16 Yet, it is still difficult to anticipate

the effectiveness of the PPDA when it comes to other air pollutants such as PM2.5 or NOX,

potentially due to the rapid urban growth the city experienced during the last 30 years

14Santiago’s metropolitan area is administratively divided in 6 provinces and 52 communes. Specifically,
driving bans affect cars circulating in the province of Santiago plus cars driving in the communes of Puente
Alto and San Bernardo. In total, 34 communes are affected by these bans. Figure A3.1 (Appendix A) shows
this administratively division. Shaded areas represent communes affected by driving restrictions.

15Despite these short-term actions were established by the PPDA during the early 90s, their full enforce-

ment did not start until 1997 (Mullins and Bharadwaj, 2015).

16Even when there is evidence of an improvement in terms of PM10 concentrations since 1990, annual
average concentrations for this pollutant in Santiago are still well above the WHO guideline levels. These
specific guidelines for 24-hour mean concentrations are set in 50 µg/m3 and 25 µg/m3 for heavy and fine
particulate matter respectively.

111

(Romero et al., 1999).

Figure 3.2 plots the hourly station-average concentrations for these pollutants during

winter days without environmental episodes, before episodes, and with episodes. Consistently

with Figure 3.1 overview, hourly pollutant concentrations peak twice a day: early in the

morning, and late in the evening. These hourly records, however, increase consistently

during the days that precede the issuance of an episode. The middle panels show how these

readings increase throughout the day, especially during evening and mostly for CO (panel c)

and NOX (panel d). As shown by the right-hand panels, these concentrations remain high

for days with episodes although there is visible evidence that they decrease during evenings

with episodes relative to their preceding days. This pattern is a first suggestion that air

quality episodes could be curbing mobile source pollution late at night.

3.3.2. Air Quality Warnings

The issuance of air quality warnings is based on a daily pollution forecasting system

of Santiago’s particulate matter concentrations used by the Ministry of Environment (Min-

isterio de Medio Ambiente -MMA) since 2000. This daily forecast takes place every winter

day at 8pm, and is based on the prediction model called the Cassmassi model -a set of linear

equations (one per station) that incorporates particulate matter concentrations and 24-hour

forecasts of meteorological conditions and atmospheric stability in the city (Saide et al.,

2011).17 These equations generate a 24-hour moving average prediction for PM10 and PM2.5

concentrations for the following day (Perez, 2008; Salini, 2009).18 These expected concen-

trations correlate with two indexes of Air Quality from Particulates (´Indice de Calidad del

Aire Referido a Part´ıculas -ICAP) which, inspired by the former US EPA Air Quality Index,

were created to easily correlate different levels of both PM10 and PM2.5 on the same scale.

17Atmospheric stability is measured by the Meteorological Potential of Atmospheric Pollution Index -

PMCA.

18See Appendix B for an example of the weights used by the Cassmassi model in the prediction of 24-
hour PM10 concentrations for the Pudahuel station, which generally records the highest levels of particle
concentrations in Santiago.

112

Figure 3.2. Hourly Average Pollution by Type of Day -Winter-

(a) PM10

(b) PM2.5

113

Figure 3.2. (cont’d)

(c) CO

Notes: Observations are station-hour. Line represents the hourly average across stations. Days with episodes
include all three different types of episode. Winter goes from April 1st through August 31st.

(d) NOX

114

These indexes (hereafter called ICAP10 and ICAP2.5 respectively), transform measures of

particulate matter concentrations into a convenient and comparable scale, so that the higher

the indexes values, the greater the particulate concentrations and potential health effect.

Different thresholds in these air quality indexes lead to a three-tier label system of

environmental episodes (alert, pre-emergency, and emergency) issued whenever the indexes

exceed the thresholds in at least one of the monitoring stations.19 When this is true, the

Environment Superintendent (Superintendencia del Medio Ambiente -SMA) recommends

that the city’s mayor issue an episode for the next day (starting at midnight), which is

publicly announced through either evening newscast TV and radio shows, the SMA’s official

website, newspapers, and smartphone apps. This announcement continues until next day

early in the morning.

Table 3.1 displays the thresholds considered by the SMA for the PM10- and PM2.5-

based ICAP values, their 24-hour average particulate matter correlations, and the corre-

sponding episode. Forecasts of good air quality conditions are equivalent to ICAP values

below 100, while predictions of regular air conditions relate to values between 100 and 199

for this index. Air quality episodes are announced whenever ICAP values reach the 200-

threshold. The mildest episode, an alert, is announced whenever ICAPs are between 200

and 299, equivalent to 195-239µg/m3 PM10 concentrations, and to 80-109µg/m3 PM2.5 con-

centrations. The 300-threshold leads to an environmental pre-emergency that takes place

whenever ICAPs are between 300 and 499, equivalent to 240-329µg/m3 PM10 concentrations

and to 110-169µg/m3 PM2.5 concentrations. Finally, the 500-threshold leads to an environ-
mental emergency with ICAPs values equivalent to PM10 concentrations ≥ 330µg/m3, or
PM2.5 concentrations ≥ 170µg/m3.

Official documentation highlights ICAP10 and ICAP2.5 as the key variables triggering

19Since 1997, the Cassmassi model, and the announcement of episodes, were both uniquely based on
expected concentrations of heavy particulate matter (PM10). In 2011, the ICAP index based on fine par-
ticulate matter (PM2.5) concentration was introduced so that air quality episodes could also be announced
based on PM2.5 concentrations. Currently, episodes can be issued whenever one of these two indexes exceeds
the tolerable ceilings.

115

Table 3.1. Indexes of Air Quality from Particulates (ICAPs)

24-hour PM10

24-hour PM2.5

Air Quality Environmental

Concentration (µg/m3) Concentration (µg/m3) Condition

Values

0 - 99
100 - 199
200 - 299
300 - 399
400 - 499
+500

≤ 149
150 - 194
195 - 239
240 - 284
285 - 329
≥ 330

≤ 49
50 - 79
80 - 109
110 - 139
140 - 169
≥ 170

Good
Regular

Bad

Episode

-
-

Alert

Critical

Pre-emergency
Dangerous Pre-emergency
Exceeding

Emergency

Notes: Adapted from Morales (2006)

episode issuance. Nevertheless, there is evidence that some other factors play a role in this

process. Saide et al. (2011) argue that this decision might sometimes involve experienced

air quality forecasters, while in some other cases it might be a mere political decision. The

fact that these air quality indexes are not the only variables considered for triggering episode

announcements becomes particularly relevant for the identification strategy; these indexes

are discretely influencing the probability of declaring air quality episodes instead of triggering

their issuance. More details on how this distinction affects the empirical design are provided

in Section 3.5.

3.3.3. Temporary Driving Restrictions

Air quality episodes trigger several abatement actions that affect emissions not only

from mobile sources but also from stationary sources. Emissions from stationary sources

are addressed with shutdowns of industrial facilities and prohibition of wood-burning stoves,

while mobile source emissions are targeted with temporary driving bans aimed at reinforcing

the seasonal (permanent) restriction. In general, the more severe the episode, the stricter

their actions. Table A3.1 (Appendix A) contains a detailed description of the actions trig-

gered by these episodes.

Temporary restrictions match the permanent driving ban by prohibiting driving be-

116

tween 7:30 am and 9 pm. Yet, unlike their long-term counterpart, these short-term re-

strictions can be implemented during any day of the week, and can affect both dirty and

clean cars, that is, without and with green stickers, respectively). Table A3.2 (Appendix

A) summarizes the different versions of this policy over 1999-2015, along with the number

of license plate digits and the type of cars affected. Since its implementation in 1990, the

permanent restriction is set in a 2-digit policy on dirty cars, affecting thus 20% of them.

This policy was tightened in 2008 by banning two additional digits, leading to a 40% of dirty

cars daily banned from driving.20 Clean cars, instead, are curtailed only by temporary bans.

In particular, air quality alerts extend the 2-digit permanent restriction on dirty vehicles to

Saturdays and Sundays; air quality pre-emergencies add two more digits to the ban on dirty

cars, and impose a 2-digit restriction on cleans cars, while air quality emergencies trigger an

8-digit restriction on dirty cars, and impose a 4-digit restriction on clean cars.

The implementation of these restrictions requires planning ahead for a full year the

license plate numbers to be banned from driving (see Table A3.3 in Appendix A for a 2016

example). Public support with this policy reached 89% in Chile, while compliance is also

believed to be high.21 The enforcement of these bans is a role for the national police force

task, who are visible and increase vehicle inspections during days with episodes. Penalties for
violations include fines between US$70 and US$150, and driving license suspensions (Grange

and Troncoso, 2011).22

20Clean cars are not affected by the permanent restriction during the time of analysis. Yet, these restric-

tions started on clean cars in April 2018.

21Information from the 2018 National Survey of Environment, available at http://portal.mma.gob.cl.

Retrieved on March 27, 2018.

22The 2016 GEC report from the MMA indicates that 5,578 vehicles were ticketed for driving restriction
violations during morning rush hours in 2016. Most of these violations occurred during June and July, time
during which 16 air quality pre-emergencies and 1 emergency were issued. The same report indicates that
most of these violations correspond to infringements of the permanent restriction. According to Atal (2009),
vehicles inspections are conspicuous in the city during air quality pre-emergencies.

117

3.4. Data

3.4.1. Air Quality Warnings

Historical information on air quality episodes comes from the Operative Unit of Traffic

Control (Unidad Operativa de Control de Tr´ansito -UOCT) from 1997 until 2015. Over this

period, there were 473 episodes issued in Santiago. Due to data availability, however, I only

consider 397 of these occurrences, issued between 2000-2015 (see Table A3.4 in Appendix

A). Air quality alerts are the most common occurrence during this period with a total of

318 issuances, followed by pre-emergencies with 78 issuances. Air quality emergencies are a

rare event with just a single episode declared since 2000.

3.4.2. Air Pollution and Weather Variables

Hourly records on mobile source pollution come from the National Information System

of Air Quality (Sistema de Informaci´on de Calidad del Aire -SINCA)’s network of monitoring

stations located in Santiago’s metropolitan area. The SINCA network in Santiago consists

of eleven stations spread across the city to monitor air quality and weather conditions in the

area. In the empirical analysis, however, I only use records from ten of these stations, which

are located in the communes affected by driving restrictions.23 The eleventh station is used

later in a robustness check analysis.

From the selected monitoring stations, I obtain ambient concentrations of four major

pollutants from vehicle emissions: PM10, PM2.5, CO, and NOX. Particulate matter con-

centrations are important to study, as their concentrations drive the calling of episodes.

Nonetheless, some episodes such as pre-emergencies and emergencies can also affect resi-

dential and stationary source pollution, which threatens the identification of causal effects

23Daily commuting flows in the area move from suburban belts towards the region’s center allowing these
stations to read pollution records from most of the daily economic activity taking place in the region. See
Figure A3.1 (Appendix A) for the spatial location of these stations (red dots), and the communes affected
(shaded area).

118

of driving restrictions on pollution. The examination of two main tailpipe pollutants from

vehicle emissions, CO and NOX, constitutes therefore a better indicator of mobile source

pollution in Santiago.24 Descriptive statistics for these variables, as well as the years with

available information, are both depicted in Table 3.2 (panel A).

Table 3.2. Descriptive Statistics on Pollution and Weather Variables.

Variables

Obs.

Mean

Std. Dev. Min

Max

Years

Panel A. Pollutants
PM10
PM2.5
CO
NOX

1,257,005
956,161
1,385,954
816,772

71.065
28.975
0.891
59.762

Panel B. Weather Variables
Humidity (%)
Wind Speed (m/s)
Temperature ()
Thermal Oscillation ()
Precipitation (mm)

980,343
972,641
1,002,864

4,523
5,844

59.592
1.426
15.783
12.983
0.883

54.864
24.578
1.309
80.378

21.715
1.148
7.805
4.447
4.672

0
0
0
0

0
0

-33.22
0.683

0

998.0
901.9
49.0
1374

1997-2015
2000-2015
1997-2015
2000-2015

105.0
17.3
39.96
24.854
118.738

2003-2015
2003-2015
2003-2015
2003-2015
2000-2015

Notes: Panel A- Observations are station-hour. Particulate matter is in micrograms per cubic meter, CO is
in parts per millions, and NOX is in parts per billions. Panel B- Observations are station-hour for
humidity, wind speed, and temperature; and days for thermal oscillation and precipitation.

Data on weather comes from several sources. Hourly data on humidity, temperature,

and wind speed come from the SINCA stations, while information on daily precipitation

comes from the National System Information on Water (Sistema Nacional de Informaci´on

del Agua -SNIA). Combinations of humidity and temperature affect the dispersion of pol-

lutants in a similar fashion to winds and precipitations, which carry air contaminants away

from the atmosphere. Despite the relevance of accounting for these factors, temperature

inversion is known to be a crucial factor preventing the pollution dispersion in Santiago,

particularly during winter (Perez, 2008; Saide et al., 2011). To account for this effect, I

24The transportation sector is also a primary source of volatile organic compounds (VOCs) emissions.

However, data availability prevents the use of VOCs in this work.

119

consider the difference between the city’s daily maximum and minimum temperature as a

proxy for inversion. This thermal oscillation is positively related to daily average temper-

atures. As warmer days experience less temperature inversion, higher thermal oscillation

is thus expected to reduce pollution. Descriptive statistics for the weather covariates are

displayed in Table 3.2 (panel B).

Table 3.3 displays daily average and maximum pollutant concentrations during sum-

mer and winter, and for days with and without episodes. Average and maximum 24-hour

concentrations are substantially lower for all four pollutants during summer and winter days

without episodes. These concentrations show a steady increase during days with episodes

except for CO, which show a decrease in both average and maximum daily concentrations

during the most critical warning. Yet, the sample only includes one air quality emergency.

Notwithstanding this, values in Table 3.3 mostly indicate how thermal inversion adversely

affects the dispersion of particulates in Santiago during colder months, as pollution concen-

trations are substantially higher during winter.

3.4.3.

Indexes of Air Quality from Particulates (ICAPs)

Daily values for the air quality indexes, ICAP10 and ICAP2.5, are obtained from the

MMA for each of the SINCA stations for the years between 2000 and 2015. However, as the

episodes’ issuance takes place whenever one of these indexes exceeds the norm in at least

one of the reading stations, I use daily maximums across stations to reflect that the issuance

of an episode is driven by the station with the highest daily pollution records.25

The relationship between daily max ICAPs and the announcement of episodes is

displayed in Figure A3.3 (Appendix A) for winter 2015. Dashed lines are drawn to depict

25The reader should remember that air quality warnings are issued based on ICAP levels constructed
from the forecasts of particulate matter concentrations. When getting the access to ICAP levels, the MMA
was not able to clarify whether these values were constructed from PM forecasts, or their actual realizations.
Following a conservative approach, I assume that ICAP values used here are a proxy of the actual variable
used in the episodes’ issuance. This assumption by no means undermines the validity of the identification
strategy, as values for these indexes are used as instruments of the issuance. For more details, see Section
??.

120

Table 3.3. Pollutant Concentrations During Episodes

Pollutant Type of Episode

Season

24-Hour Average

24-Hour Maximum

PM10

PM2.5

CO

NOX

Without Episodes

Summer
Winter

Alerts

Winter
Pre-emergencies Winter
Winter

Emergencies

Without Episodes

Summer
Winter

Alerts

Winter
Pre-emergencies Winter
Winter

Emergencies

Without Episodes

Summer
Winter

Alerts

Winter
Pre-emergencies Winter
Winter

Emergencies

Without Episodes

Summer
Winter

Alerts

Winter
Pre-emergencies Winter
Winter

Emergencies

56.9
78.7

120.3
129.9
148.1

21.8
38.1

57.3
62.4
79.9

0.45
1.29

2.08
2.34
2.15

29.1
90.5

151.2
159.1
224.1

98.9
149.5

216.1
228.3
238.1

37.5
68.7

98.0
108.6
92.8

1.05
3.03

4.61
5.33
4.59

82.3
210.5

327.4
347.7
356.5

Notes: Data from the UOCT and SINCA. Particulate matter is in micrograms per cubic meter, CO is in
parts per millions, and NOX is in parts per billions.

the different ICAP-thresholds considered in issuing an episode (see Table 3.1), while days

with an air quality episode are in either a diamond, cross, and triangle shape, and colored.

Based exclusively on the policy design, days preceding an episode are expected to have daily

max ICAP values above the 200-threshold in at least one of the two indexes (ICAP10 or

ICAP2.5). Yet, a quick examination of Figure A3.3 indicates this is not always true. There

are 2015-days with episodes that were preceded by days with max ICAP values below the 200-

threshold.26 This mismatch reflects the discretion embedded in the calling of these warnings,

as factors other than the ICAPs can affecting the authorities’ decision of announcing one

26The same is true for other years. Similar graphs for previous years are available upon request.

121

of these occurrences as well. This imperfect correlation between daily max ICAPS, and the

episodes’ issuance, drives the decision on the empirical strategy used in this study.

3.4.4. Urban Flows

This study includes three datasets on urban transit flows. Hourly data on vehicle

trips come from the Operative Unit of Traffic Control (UOCT) from 2004 to 2015, obtained

from the readings of traffic-control counting stations located inside Santiago’s main urban

area. These counting stations are placed underground the city’s main roads, and are con-

nected to Santiago’s road network. Despite their inability to distinguish between public and

private transportation, or between light-, medium-, and heavy-duty vehicles, information

from these stations represents a useful proxy for light-duty car use whenever unobservable

factors affecting this counting occur at random.

A second dataset come from Metro S.A. and contains hourly data on metro trips taken

during 2000 and 2015. Santiago’s 103 km subway consists of 5 lines that continuously operate

from 5:30 am to midnight during weekdays, from 6:30 am to midnight during Saturdays, and

from 8:00 am to midnight during Sundays and holidays. Daily ridership in Santiago’s metro

was 640,000 trips before 2007. After this, however, daily ridership increased drastically to

more than 1,750,000 trips due to the ineffective 2007 introduction of Santiago’s new bus rapid

transit called Transantiago. The Transantiago program was aimed at persuading drivers to

get off the roads by improving the quality of the city’s public transportation. Yet, its results

were the opposite due to inefficient route designs and bus schedules, which translated into

increased commuting times pushing public bus users towards the metro instead. After more

than 10 years of improvements, Transantiago is still in place and is used daily by several

riders. To account for the role of Transantiago in Santiago’s daily transportation schemes,

I also include data on daily ridership between 2007 and 2015, from the Ministry of Trans-

portation and Telecommunications (MTT). Descriptive statistics for these three variables,

122

along with the years with available information, are all depicted in Table 3.4.27

Table 3.4. Descriptive Statistics on Urban Traffic Flows

Trips

Obs.

Mean

Std. Dev. Min

Max

Years

Vehicles
Metro
Transantiago

3,297,682
116,880
3,247

678

63,294

2,921,347

689

56,420
967,754

0
0

10,453

11,269
282,664
4,468,663

2004-2015
2000-2015
2007-2015

Notes: Observations are station-hours for vehicle flows; hours for metro ridership; and days for Transantiago
trips. Statistics for Transantiago consider daily totals.

Driving restrictions that work get drivers off the roads, and push them towards cleaner

forms of transportation. Therefore, mass-transit system ridership is at first expected to in-

crease during days in which driving bans are in place.28 For a better understanding of the

traffic patterns during days with and without episodes, Table 3.5 exhibits 24-hour average

and maximum trips taken during summer and winter by each transportation mode. Con-

sistent with an effective policy, average vehicle trips decrease during days with air quality

episodes. Yet, metro ridership increases only during days with air quality alerts, and de-

creases during days with pre-emergencies. A similar pattern is exhibited by Transantiago

ridership. A plausible explanation for a lower mass-transit system ridership during air qual-

ity pre-emergencies is that individuals are positively responding to official recommendations

of avoiding outdoor exposure during days with episodes. The fact that air quality alerts

constitute the most common episode might also suggest that individuals are less sensitive

to this mild episode, and decide to follow these recommendations only during days with

critical or dangerous air pollution records. Under this scenario, daily mass-transit system

ridership would be expected to decrease during severe air quality episodes such as during

27Similar datasets are used in Grange and Troncoso (2011) but only for 2008.
28Drivers can also substitute towards bicycles or motorcycles, where the last one also contributes to mobile
source pollution. Reports on Santiago’s vehicle composition, however, allow to rule out the possibility
that driving restrictions are increasing the use of motorcycles in the city. During 2001, there were two
motorcycles per a hundred of light-duty private vehicles driving in Santiago. In 2015, this number increased
to six motorcycles per a hundred of light-duty cars. These increments, however, are identical to the ratios
exhibited by other cities in the country, for which no driving restrictions are in place. For more details, see
INE (2001) and INE (2015).

123

pre-emergencies. Average ridership during air quality emergencies, however, is hard to in-

terpret as the sample of episodes includes only one emergency.29

Table 3.5. Urban Flows During Episodes

Urban Flow

Type of Episode

Season

24-Hour Average

24-Hour Maximum

Vehicle Trips

Metro Trips

Transantiago Trips

Without Episodes

Summer
Winter

Alerts

Winter
Pre-emergencies Winter
Winter

Emergencies

Without Episodes

Summer
Winter

Alerts

Winter
Pre-emergencies Winter
Winter

Emergencies

Without Episodes

Summer
Winter

Alerts

Winter
Pre-emergencies Winter
Winter

Emergencies

661.38
681.50

680.41
650.61
617.88

61,334.93
65,661.96

69,113.05
61,845.75
120,602.90

120,098.70
123,919.32

125,261.55
117,647.58
137,549.42

2,826.84
2,945.78

2,863.37
2,674.28
2,516.00

119,941.97
133,098.92

139,633.24
124,842.69
265,590.00

-
-

-
-
-

Notes: 24-hour average vehicle trips are obtained by averaging across stations and hours. 24-hour averages
for buses are obtained by dividing daily totals in 24. Winter is defined as April 1st through August 31st.

3.5. Empirical Strategy

In this paper, I address the effectiveness of air quality warnings, and their temporary

driving bans in curbing Santiago’s mobile source pollution. The issuance of these 24-hour

episodes is conditional on authorities foreseeing poor air quality conditions, measured by

air quality indexes (ICAP) values. Due to this structure, a quasi-experimental framework

that exploits the discontinuity embedded in the different ICAP cutoffs would seem to be

29In the remainder of this work, emergency episodes will be merged into pre-emergencies.

124

the natural approach to evaluate Santiago’s policy. A sharp regression discontinuity (SRD)

design that considers an ICAP variable (e.g. ICAP10) as the variable forcing the treatment

assignment (i.e. the issuance of an episode), constitutes a common strategy for identification

of causal effects under the identifying assumption that in absence of an episode, air pollu-

tion and urban flows would have changed smoothly around the ICAP thresholds. Yet, the

authorities’ discretion embedded in this process prevents this announcement from being a

deterministic function of ICAPs, inhibiting a perfect compliance with the policy. When the

probability of treatment jumps not necessarily from 0 to 1 at the running variable cutoffs,

a sharp design becomes an impractical option to evaluate policy impacts, as factors other

than the running variable may also affect the treatment assignment.

Values for the air quality indexes, however, still can discretely impact the conditional

expectation of the treatment. For instance, ICAP10 forecasts above the PPDA thresholds

may not uniquely determine the announcement of an episode for the following day, but

instead, they can discontinuously affect the probability of their issuance. Whether this is

the case, identification of causal effects can be tackled by exploiting the discontinuity in the

conditional probability of treatment, in which case a fuzzy regression discontinuity (FRD)

design constitutes a valid identification strategy (Imbens and Lemieux, 2008). In an FRD,

the discontinuities in the running variable serve as instruments for issuing an episode instead

of deterministically defining it (Angrist and Pischke, 2008).

In this study, I explore two

FRD designs: a pooled FRD approach aimed at combining the treatment effects of different

episodes, and a multi-cutoff FRD setup that exploits the treatment heterogeneity (Cattaneo

et al., 2016) using the distinct thresholds in the PPDA program.

3.5.1. Pooled Fuzzy Regression Discontinuity Design

Different ICAP thresholds determine three types of air quality warnings, expected

to affect both pollution concentrations and urban transit flows differently. When dealing

with threshold heterogeneities, the standard in the literature is the score normalization, so

125

that the cutoff becomes the same for all units (i.e. days) in the analysis (Cattaneo et al.,

2016). When pooling the episodes, a treatment is defined as having an air quality episode

during day t, regardless of its type. Formally, let Tt be the random variable representing the

treatment during day t. When the episodes are pooled, the probability of issuance is defined

to jump as follows:

 g1(¯xt−1)

if ¯xt−1 ≥ 0,
g0(¯xt−1) otherwise

P (Tt = 1|¯xt−1) =

(3.1)

where g(·) is any function with g1(0) > g0(0), and ¯xt−1 is the running variable defined

as the daily max ICAP10,t−1 during day t − 1 normalized as follows:30

¯xt−1 = ICAP10,t−1 − 200.

(3.2)

From the normalized running variable, days with ¯xt−1 values equal or higher than 0

are more likely to have an episode.31 The first-stage equation relating episode announcement

and the running variable, ¯xt−1, is:

Tt = α0 + α11[¯xt−1 ≥ 0]t +

p(cid:88)

j=1

p(cid:88)

j=1

γlj ¯xj

t−1 +

γrj ¯xj

t−1 × 1[¯xt−1 ≥ 0]t + Xtδ + νt,

(3.3)

where Tt is equal to 1 if an episode was announced on day t (=0 otherwise); 1[¯xt−1 ≥ 0]t

is an indicator variable taking 1 for normalized ICAP values equal and above 0 during day t

(=0 otherwise); Xt is a vector of covariates that include current and 24-hour lags of quartics

in humidity, temperature, precipitation, wind speed, and thermal oscillation, and seasonal

fixed effects on year, month of the year (month), day of the week (dow), hour of the day

30During 1997 till 2011, ICAP10 was the unique index considered in the episodes’ issuance. The empirical
analysis, therefore, considers this index as the running variable throughout all the estimations. Nonetheless,
and due to the potential influence of ICAP2.5 in the calling of more recent episodes, I consider this variable
as an additional instrument as well all along, unless otherwise is noted.

31As the episodes are called only during winter, equation (3.1) can also be complemented with a season
dummy variable. In this case, the conditional probability of having an episode would be defined as g1(¯xt−1)
if ¯xt−1 ≥ 0 and day t − 1 is a winter day.

126

(hour), and interactions between weekends and hour of the day; and νt is an error term.

The default specification of equation (3.3) allows for a flexible global polynomial fit below

and above the discontinuity. The global polynomial fit in equation (3.3) is introduced with

subscripts l and r that identify coefficients predicting the relationship between the ICAP10

and the treatment below (l) and above (r) the 0-threshold, respectively, where p indicates the

order of the polynomial fit. The discontinuity at the threshold in the episodes’ announcement
is indicated by the coefficient α1 next to 1[¯xt−1 ≥ 0]t.

A similar structure is fitted on the outcome variables where the second-stage estima-

tion equations take the following form:

yt = β0 + β1Tt +

ηlj ¯xj

t−1 +

p(cid:88)

j=1

p(cid:88)

j=1

ηrj[¯xj

t−1 × Tt] + Xtζ + 
t,

(3.4)

where yt is the outcome of interest (e.g. log of air pollution, log of urban transit flows)

during day t, 
t is an error term, and coefficients ηl and ηr identify the global polynomial

fit on the outcome variables before (l) and after (r) the treatment, respectively. In subse-

quent specifications, this global polynomial fit is replaced with a non-parametric approach

through a local linear estimation.32 The episode impact on the outcome variables is given

by the coefficient β1, which under the assumption of the discontinuity in 3.3 inducing the

discontinuity in 3.4, indicates the episodes’ weighted average treatment effect, weighted by

the number of occurrences in each threshold.

3.5.2. Multi-Cutoff Fuzzy Regression Discontinuity Design

Unlike a pooled FRD, a multi-cutoff FRD allows the running variable to affect the

probability of one or more treatments, through several discontinuities. When these treat-

32Local linear designs in RD approaches could be challenging; controls are hard to include when estima-
tions consider only few observations around a cutoff. As a solution, I follow Hausman and Rapson (2017)
and estimate an augmented local linear specification, where a local linear approach is fitted on the residuals
from the (full) regression of the response variables on weather and seasonality controls. For more details,
see Hausman and Rapson (2017).

127

ments have heterogeneous effects on the outcome variable, a multiple cutoff FRD fully ex-

ploits this feature by using all the information available on the data (Cattaneo et al., 2016),

although at the expense of some power. As pre-emergencies and emergencies episodes are

grouped together, the multi-cutoff approach here uses only two discontinuities on the running

variable: at the 200-threshold, and at the 300-threshold of ICAP10. The first-stage equation

in this case is defined as follows:

t = δ0+δ11[Alerts]t + δ21[P re − em]t +
T A

θljICAP j

10,t−1

p(cid:88)

p(cid:88)

j=1

p(cid:88)

+

θmj1[Alerts]t × ICAP j

10,t−1 +

j=1

j=1

θrj1[P re − em]t × ICAP j

10,t−1 + Xtϕ + υt

(3.5)

where T A
t

is the response variable taking the value of 1 if an alert episode was

announced during day t (=0 otherwise); 1[Alerts]t
is an indicator variable taking 1 if
ICAP10,t−1 ∈ [200, 299] (=0 otherwise); 1[P re − em]t is an indicator variable taking 1 if
ICAPt−1 ≥ 300 (=0 otherwise); Xt is a vector of of weather covariates and fixed effects as

explained before; and υt is an error term. The default specification of equation (3.5) includes

a flexible polynomial fit at each side of the discontinuities, where the subscripts l, m, and

r identify the coefficients predicting the relationship between ICAP10,t−1 and the outcome

variables below, in the middle, and above the two discontinuities. An expression similar to

3.5 is also specified for pre-emergencies, with the response variable being now T P

t . For the

second-stage equation:

p(cid:88)

j=1

p(cid:88)

yt = λ0+λ1T A

t + λ2T P

t +

ωljICAP j

10,t−1

p(cid:88)

+

ωmjICAP j

10,t−1 × T A

t +

ωrjICAP j

10,t−1 × T P

t + Xtµ + ξt

(3.6)

j=1

j=1

Following Gelman and Imbens (2017), default specifications of the estimation equa-

128

tions consider low-order (up to third) polynomials fitted at each side of the cutoffs, while

alternative specifications consider local estimations around these thresholds but using a small

neighborhood of observations around them. To test for an intertemporal shift in driving, all

the different specifications consider the outcome variables measured during peak (restricted)

and off-peak (unrestricted) hours. If drivers substitute their driving towards unrestricted

hours, estimation results should indicate significant impacts of episodes on these hours, es-

pecially for vehicle flows. In terms of pollution, however, changes in car emissions between

these hours may not be as sharply reflected in ambient pollution concentrations due to dif-

ferent lifetimes of pollutants in the atmosphere (Gibson and Carnovale, 2015; Zhang, Lawell,

and Umanskaya, 2017).33 Potential heterogeneities in how driving bans affect pollutants

due to factors such as their atmospheric chemistry, their formation and lifetime in the at-

mosphere, and how several emission quantities at different points in time affect their total

damage, might all impact the pace whereby eventual reductions in traffic flows are reflected

in ambient pollution concentrations. Failing to control for this factor could lead to inaccu-

rate conclusions regarding the policy effects on pollution. Therefore, to account for some of

this pollution persistence, all estimation equations on pollution include the outcome variable

lagged by some hours.34

Due to the time series characteristic of all the outcome variables, the estimation pro-

cess is carried out using a Generalized Method of Moments (GMM) procedure that accounts

for serial correlation in the data. This is performed with the Newest-West heteroskedastic-

and autocorrelation-consistent HAC weighting matrix (Barlett kernel), which based on the

analysis of partial autocorrelations, allows for 24-hour lags in air pollution, a 2-hour lag in

vehicle trips, and 1-day lag in mass-transit system use. As episode issuance is based not

33Persistence has been shown to affect the concentration of several pollutants in Santiago, mostly PM2.5

(P´erez, Trier, and Reyes, 2000).

34The analysis of partial autocorrelations performed on the time series on pollution indicates 24-hour
autocorrelations with values equal or higher than 0.5, suggesting the inclusion of these number of lags in the
estimations. Yet, the inclusion as covariates of that number of lagged periods turns the estimations into a
very intensive computational task. For the sake of a parsimonious set of estimations, I only include 6-hour
lags on pollution. However, the inclusion of 24-hour lags in weather covariates should capture any effect left
for pollutants persisting for more than 6 hours in the atmosphere.

129

only on forecasts of PM10, but recently on PM2.5 as well, the ICAP2.5 variable is considered

as an additional instrument in all the estimations.

3.6. The Effectiveness of Air Quality Episodes and their

Short-Term Driving Restrictions

A preliminary examination of the relationship between the outcomes and the forcing

variable is depicted in Figures 3.3 and 3.4, which display discontinuity plots of the residuals

from the augmented local linear regressions on hourly pollution concentrations, and vehicle

flows, respectively. In terms of pollution concentrations, Figure 3.3 suggests that air quality

episodes, through abatement actions, effectively reduce airborne contamination in Santiago.

All the four panels indicate reductions in pollution concentrations after the air quality index

reaches each of the thresholds, with higher impacts for the CO (panel c) and NOX (panel d),

especially during air quality pre-emergencies. Reductions in heavy (panel A) and fine (panel

B) particulate matter concentrations are also observed, mostly during alert episodes. The

lines in Figure 3.3 also indicate the existence of a potential kink in the trend of pollution con-

centrations after the announcement of an episode, suggested by the change in the slopes of

the linear curves fitted at each side of the thresholds. These trend changes are all consistent

at first with an effective policy. Nevertheless, an overview of car trips in Figure 3.4 offers a

different perspective. Figure 3.4 ssuggests no reductions in car trips during air quality alerts.

There is also a slight suggestion of an increase in driving during air quality pre-emergencies,

which would undermine the effectiveness of the policy. In spite of these preliminary obser-

vations, is important to consider that Figures 3.3 and 3.4 assume sharp discontinuities at

the selected thresholds, despite the fact that these cutoffs are not deterministically affecting

130

calling of episodes. This situation makes these figures merely informative.35,36

Figure 3.3. Discontinuity Plot on Hourly Average Pollution Concentrations

(a) PM10

(b) PM25

(c) CO

(d) NOX

Notes: All hours. Using a linear fit on winter days and with ICAP10,t−1 as the running variable. Outcome
variables are residuals from the regressions of hourly average pollution concentrations (in logs) on weather
variables (current and 24-hour lags of quartics in precipitation, humidity, temperature, wind speed, and
thermal oscillation), and year, month, dow, hour, and hour × weekend fixed effects. The linear fit uses bins
at each side of the cutoffs following the mimicking variance evenly-spaced method (see Calonico, Cattaneo,
and Titiunik (2015b) for more details).

35See Figure A3.4 in the Appendix A for the kernel density estimates of the daily air quality indexes
(ICAP10 and ICAP2.5) during days that precede (a) alert and (b) pre-emergency episodes. Panel (a) shows
that days with alert episodes have ICAP10 values that are in general below the 200-threshold. This situation
is slightly similar for days with pre-emergencies, as shown in panel (b). The fact that most of the episodes
have been issued with ICAP10 values below their corresponding threshold, evidences the fuzzy feature of the
discontinuities in this policy.

36Linear fits in Figures 3.3 and 3.4 at each side of the cutoffs use sample averages of residuals within opti-
mal disjoint bins based on Calonico, Cattaneo, and Titiunik (2015a). Estimates of reduced form equations,
however, consider the full information available.

131

Figure 3.4. Discontinuity Plot on Hourly Average Traffic Flow

(a) All Hours

(b) Peak Hours

(b) Off-Peak Hours

Notes: Using a linear fit on winter days and with ICAP10,t−1 as the running variable. Outcome variables
are residuals from the regressions of hourly average vehicle trips (in logs) on weather variables (current and
24-hour lags of quartics in precipitation, humidity, temperature, wind speed, and thermal oscillation), and
year, month, dow, hour, and hour × weekend fixed effects. The linear fit uses bins at each side of the cutoffs
following the mimicking variance evenly-spaced method (see Calonico, Cattaneo, and Titiunik (2015b) for
more details).

3.6.1. The Overall Effect of Episodes and their Driving Bans -

Pooled FRD

Tables 3.6, 3.7, and 3.8 depict the estimated episode impact on hourly average pol-

lution, vehicle trips, and daily average mass-transit system use, respectively, from a pooled

FRD estimation for peak and off-peak hours. Tables C3.1, C3.2, and C3.3 in Appendix C

display similar results for maximums instead. Columns (1)-(4) show estimation results from

132

a polynomial fit at each side of the (normalized) cutoff, while columns (5)-(7) display results

from local estimations.37,38 To facilitate comparisons, results from an ordinary least-squares

(OLS) estimation are always in column (1), while results from the GMM-instrumental vari-

able (IV) estimations are presented from columns (2) to (7).

Table 3.6. Episodes Impact on Hourly Average Pollution

Hours

Pollutant

Polynomial Approach

Local Approach

Peak

Off-Peak

PM10

PM2.5

CO

NOX

PM10

PM2.5

CO

NOX

Weather
Time FE
Estimation
Functional Form
Bandwidth k

(1)

-0.017∗∗
(0.006)
-0.011∗
(0.006)
-0.018∗∗
(0.007)
-0.010∗
(0.006)

0.013
(0.009)
0.003
(0.008)
0.012∗
(0.007)
0.004
(0.010)

×
×
OLS
Linear

-

(2)

-0.138∗∗
(0.045)
-0.079∗
(0.042)
-0.095∗∗
(0.039)
-0.056
(0.043)

0.004
(0.056)
-0.006
(0.052)
0.011
(0.045)
0.191∗∗
(0.080)

×
×
IV

(3)

-0.162∗∗
(0.075)
0.003
(0.061)
-0.081
(0.056)
-0.039
(0.066)

0.116
(0.094)
0.197∗∗
(0.098)
0.067
(0.070)
0.295∗∗
(0.138)

×
×
IV

(4)

-0.126∗
(0.068)
-0.007
(0.062)
-0.095∗
(0.057)
-0.064
(0.066)

0.133
(0.090)
0.165∗
(0.091)
0.087
(0.068)
0.256∗∗
(0.126)

×
×
IV

(5)

-0.025∗∗∗
(0.007)
-0.010
(0.007)
-0.024∗∗∗
(0.007)
-0.026∗∗
(0.009)

0.007
(0.008)
-0.002
(0.008)
0.021∗∗
(0.007)
-0.002
(0.011)

×
×
IV

(6)

-0.016∗∗
(0.007)
-0.013∗∗
(0.007)
-0.022∗∗
(0.007)
-0.028∗∗∗
(0.007)

0.006
(0.007)
0.018∗∗
(0.006)
0.027∗∗∗
(0.006)
-0.002
(0.010)

×
×
IV

(7)

-0.028∗∗∗
(0.005)
-0.020∗∗∗
(0.005)
0.00004
(0.006)
-0.020∗∗∗
(0.006)
0.022∗∗∗
(0.005)
0.020∗∗∗
(0.004)
0.030∗∗∗
(0.004)
0.019∗∗
(0.006)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as an
additional instrument. All estimations include 6-hour lags on pollution. Weather variables include current
and 24-hour lags of quartics in humidity, temperature, wind speed, precipitation, and thermal oscillation.
Time fixed effects include dummies for year, month, dow, hour, and hour × weekend. Linear local approach
uses as outcome variables the residuals from the (full) regressions of hourly average pollution concentrations
(in logs) on covariates. Peak hours: N full sample = 34,796; N local = 3,056 (5), 1,816 (6), 960 (7). Off-Peak
hours: N full sample = 17,366; N local = 1,530 (5), 910 (6), 482 (7). Standard errors robust to a 24-hour
serial correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

37I estimate local approaches using a bandwidth of h = 25, and allowing for K0 = 3 and K1 = 3 number

of bins at each side of the cutoff. See Imbens and Lemieux (2008) for more details.

38Pooling estimations are performed after normalizing the running variable to 0. Given the size of bins
used in the local approaches, estimations from these schemes are merely occurring in the alerts-neighborhood,
and therefore, their results are reflecting the impacts of only this type of episodes.

133

Results in Table 3.6 (Table C3.1) suggest that air quality episodes, through their

temporary driving restrictions, reduce hourly average (hourly maximum) air pollution con-

centrations from vehicle emissions during peak hours. The OLS estimation results in column

(1) of Table 3.6 indicate a 1-2 percentage point reduction in all the four pollutants, with

higher reductions for both PM10 and CO. When the GMM-IV estimation is introduced, find-

ings for a polynomial fit in columns (2)-(4) indicate stronger reductions for all four cases.

In particular, findings for peak hours indicate that, on average, episodes reduce PM10 con-

centrations by 13-16 percentage points, PM2.5 concentrations by 8 percentage points, CO

concentrations around 10 percentage points, and NOX around 4-6 points, although these

last estimations are not statistically different from zero. These results are consistent with

findings in Table C3.1 for hourly maximum concentrations across stations. Estimation re-

sults from a local approach in columns (5)-(7) of Table 3.6 are coherent with effective driving

restrictions as well. Local estimates indicate an episode impact of around 2-3 percentage

point reductions in PM10, around 1-2 percentage point reduction in PM2.5, a 2 percentage

point reduction in CO, and a 2-3 percentage point reduction in NOX. Results in Table

C3.1 also indicate an episode impact on hourly maximum concentrations of 4-6 percentage

point reductions in PM10, around 3-5 percentage point reductions in PM2.5, and around a

2-3 percentage point reduction in both CO and NOX, the two main tailpipe pollutants from

car emissions. All the local estimates are statistically different from zero.

Findings for air pollution reverse sign when off-peak hours are considered. Results

in columns (2)-(4) of Table 3.6 indicate that environmental episodes increase hourly average

PM2.5 and NOX concentrations during unrestricted hours. Whether these results are ex-

plained either by drivers intertemporally substituting their driving towards hours unaffected

by the policy, or merely by a longer lifetime of these pollutants in the atmosphere during

days that precede an episode, should be clarified after looking at the results on vehicle trips

in Table 3.7. If temporary driving restrictions push drivers to switch their driving towards

off-peak hours, then car trips during these hours should increase on days with episodes.

134

Results for the GMM-IV estimations in columns (2) to (7) of Table 3.7, however, all consis-

tently reject this idea. Findings for car trips in Table 3.7 all suggests that temporary driving

restrictions decrease hourly average vehicle trips during both peak and off-peak hours. These

results are statistically different from zero in all the cases, and pass the Stock-Wright-Yogo

rule of thumb on weak instruments.39

Table 3.7. Episodes Impact on Hourly Average Vehicle Trips

Hours

Peak

Off-Peak

×
Weather
×
Time FE
Estimation
OLS
Functional Form Linear
Bandwidth k

-

Polynomial Approach

Local Approach

(1)

(2)

(3)

(4)

(5)

(6)

(7)

-0.004
(0.005)
-0.010
(0.008)

-0.145∗∗
(0.048)
-0.191∗∗
(0.078)

×
×
IV

-0.261∗∗∗
(0.068)
-0.314∗∗
(0.112)

×
×
IV

-0.271∗∗∗
(0.071)
-0.292∗∗
(0.107)

×
×
IV

-0.055∗∗∗
(0.008)
-0.077∗∗∗
(0.015)

×
×
IV

-0.052∗∗∗
(0.009)
-0.068∗∗∗
(0.017)

×
×
IV

-0.038∗∗∗
(0.010)
-0.065∗∗∗
(0.018)

×
×
IV

Linear Quadratic

Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as
an additional instrument. Weather variables include current and 24-hour lags of quartics in humidity,
temperature, wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for
year, month, dow, hour, and hour × weekend. Linear local approach uses as outcome variables the residuals
from the (full) regressions of hourly average vehicle trips (in logs) on covariates. Peak hours: N full sample
= 58,425; N local = 5,240 (5), 3,113 (6), 1,654 (7). Off-peak hours: N full sample = 37,488; N local =
3,367 (5), 2,006 (6), 1,074 (7). Standard errors robust to heteroskedasticity and 2-hour serial correlation in
parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

In particular, results for a GMM-IV polynomial fit in columns (2)-(4) of Table 3.7

indicate a large effect of temporary driving restrictions on car use. Accordingly, episodes

reduce car flows by around 15-27 percentage points during peak hours, and around 19-31

percentage points during off-peak hours. These estimation results are also consistent with

reductions in hourly maximum car trips displayed in Table C3.2 (Appendix C). When a

39Results for hourly average vehicle trips all pass the Stock-Wright-Yogo rule of thumb on weak instru-
ments. In other words, the HAC-F tests are in all the cases higher than 10, which allows to discard the
possibility of bias in these results. In terms of pollution, results in Table 3.6-column (2) also pass test. For
more details, see Stock, Wright, and Yogo (2002).

135

local linear approach is fitted to the data, results in columns (5)-(7) of Table 3.7 suggests

conservative, but more realistic reductions in car use. These reductions, on average, are of

around 4-6 percentage points during peak hours, and around 7-8 percentage points during

off-peak hours. Smaller reductions in car use are expected when episodes are pooled together,

as air quality alerts trigger driving bans that only affect dirty vehicles, which constitute a tiny

fraction of light-duty vehicles in the city.40 Additionally, reductions in car trips taken during

off-peak hours reject the idea of an intertemporal switching in driving towards unrestricted

hours, and suggest as well that increments in PM2.5 and NOX found in Table 3.6 are likely

due to the longer persistence of these pollutants in the atmosphere during days of critical

air pollution. This situation reflects the importance of evaluating causal impacts of driving

restrictions using indicators of automobile use in conjunction with indicators of pollution, as

factors other than the policy could be driving the results for this last variable.

The overall episode impact on daily average (daily maximum) use of mass-transit

system ridership is displayed in Table 3.8 (Table C3.3 in Appendix C). Results from a poly-

nomial fit in columns (2)-(4) of Table 3.8 weakly indicate a reduction in mass-transit system

use, although these estimates are not significantly different from zero. When a local lin-

ear approach is fitted to the data, estimation results in columns (5)-(7) suggest instead an

episode impact of a 3 percentage points increase in metro ridership during peak hours, and

a 9-15 percentage points increase during off-peak hours, time at which this mass-transit

system runs with some capacity. Estimates on the use of Transantiago suggests no effects

of environmental episodes on the use of this transportation mode. In general, Transantiago

runs at full capacity and serves a demographic segment of Santiago’s population character-

ized by lower average incomes, and not necessarily car owners. Results in Table C3.3 on

daily maximum metro trips consistently indicate increases in subway ridership, again mostly

40Table A3.5 (Appendix A) contains the expected number of light-duty private vehicles in the Santiago’s
Metropolitan Region affected by driving bans during the period 2001-2000. Temporary driving bans during
air quality alerts are expected to affect a 5.7% of cars, a number that is close to the magnitude found in Table
3.7. Nevertheless, it’s important to highlight that estimations in Table 3.7 are on the impacts of episodes on
car trips, which is not necessarily equal to the number of light-duty private vehicles registered the city.

136

during non-rush hours.41

Table 3.8. Episodes Impact on Daily Average Mass-Transit System Trips

Polynomial Approach

Local Approach

(1)

(2)

(3)

(4)

(5)

(6)

-0.161
(0.150)
-0.265
(0.267)

-0.252
(0.210)
-0.398
(0.400)

-0.245
(0.202)
-0.396
(0.378)

0.009
(0.015)
0.033
(0.029)

0.008
(0.016)
0.088∗∗
(0.031)

-0.208
(0.139)

×
×
IV

-0.197
(0.161)

×
×
IV

-0.199
(0.161)

×
×
IV

-0.019
(0.013)

×
×
IV

-0.015
(0.012)

×
×
IV

(7)
0.031∗
(0.018)
0.151∗∗∗
(0.030)

0.012
(0.014)

×
×
IV

System

Hours

Metro

Peak

Off-Peak

Transantiago All Hours

Weather
Time FE
Estimation
Functional Form
Bandwidth k

0.016
(0.023)
0.031
(0.039)

0.002
(0.019)

×
×
OLS
Linear

-

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as an
additional instrument. Weather variables include current and 1-day lags of quartics in humidity, temperature,
wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for year, month, and
dow. Linear local approach uses as outcome variables the residuals from the (full) regressions of daily mass-
transit system trips (in logs) on covariates. Metro: N full sample = 4,373; N local = 384 (5), 229 (6), 122
(7). Transantiago: N full sample = 3,222; N local = 290 (5), 175 (6), 98 (7). Standard errors robust to
heteroskedasticity and 1-day serial correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05,
∗∗∗p < 0.001

3.6.2. Heterogeneous Effectiveness of Episodes - Multi-Cutoff FRD

Heterogeneous effects of alerts and pre-emergency episodes on hourly average (hourly

maximum) pollution concentrations are depicted in Table 3.9 (Table C3.4 in Appendix C)

for both peak and off-peak hours. The GMM-IV estimation results from a polynomial fit

in columns (2)-(4) of Table 3.9 indicate that air quality alerts and pre-emergencies both

reduce average pollution concentrations during peak hours, although these results are only

statistically different from zero for PM10 concentrations. Specifically, environmental episodes

reduce PM10 concentrations in around 16-22 percentage points during air quality alerts, and

41Estimates using daily maximum or peak and off-peak trips in Transantiago are not feasible due to

frequency of the data available.

137

in around 26 percentage points during pre-emergencies. The fact these impacts are smaller

during air quality alerts than during pre-emergencies (and emergencies) is consistent with

the implementation of more drastic mitigation actions during the latter. When switching

to a local linear estimation approach, results in columns (5)-(7) indicate that air quality

alerts reduce hourly average pollutant concentrations only during peak hours. Air quality

alerts reduce both PM10 and PM2.5 concentrations in around 2-5 percentage points, while

both CO and NOX concentrations are reduced in around 2-3 percentage points. For the

case of air quality pre-emergencies, however, results weakly suggest reductions in PM2.5 and

in CO concentrations. For all the other cases, however, estimation results for air quality

pre-emergencies suggest increments in ambient pollution concentrations of around 2-8 per-

centage points (local approach) during both peak and off-peak hours. Again, whether this is

explained by an ineffective policy, or by a misspecification of the estimation equations that

fail capturing the pollution trend, could be answered by examining the results for car use.

Results on the effects of air quality alerts and pre-emergencies on car use are dis-

played in Table 3.10 (Table C3.5) for hourly average (hourly maximum) use. Starting with

the impacts of air quality alerts, GMM-IV polynomial estimations in columns (2)-(5) of

Table 3.10 suggest this type of episode generates a reduction in car trips of around 13-21

percentage points during peak hours, and of around 20-33 percentage points during off-peak

hours. These estimates are consistent with results from a local linear approach.

Indeed,

GMM-IV local estimation results in columns (5)-(7) indicate that air quality alerts reduce

hourly average car trips in around 4-6 percentage points during peak hours, and in around

8-10 percentage points during off-peak hours. Car trips reductions of this magnitude during

restricted hours are coherent with expected reductions for days with air quality alerts, as

shown in Table A3.5. The fact that higher reductions are found for unrestricted hours during

days with alerts could be explained by some affected and unaffected drivers staying at home

to avoid outdoor exposure to pollution during these days. The temporary characteristic

of environmental episodes, and its driving restrictions, enables this possibility given that

138

Table 3.9. Alerts and Pre-emergencies Impact on Hourly Average Pollution
Concentrations.

Hours

Pollutant

Episode

Polynomial Approach

Local Approach

Peak

Off-Peak

1[Alerts]

PM10

1[Pre-emergencies]

1[Alerts]

PM2.5

1[Pre-emergencies]

1[Alerts]

CO

1[Pre-emergencies]

1[Alerts]

NOX

1[Pre-emergencies]

1[Alerts]

PM10

1[Pre-emergencies]

1[Alerts]

PM2.5

1[Pre-emergencies]

1[Alerts]

CO

1[Pre-emergencies]

1[Alerts]

NOX

1[Pre-emergencies]

Weather
Time FE
Estimation
Functional Form
Bandwidth k

(1)

0.027∗∗
(0.012)
-0.024
(0.025)
0.036∗∗
(0.011)
0.035
(0.026)
0.023∗
(0.014)
-0.008
(0.037)

0.019
(0.013)
-0.026
(0.032)
0.062∗∗∗
(0.014)
0.088∗∗
(0.042)
0.060∗∗∗
(0.012)
0.044
(0.032)
0.040∗∗
(0.013)
0.061
(0.041)

0.022
(0.016)
0.084
(0.055)

×
×
OLS
Linear

-

(2)

-0.155∗∗
(0.059)
-0.255∗∗
(0.126)

-0.025
(0.043)
0.012
(0.094)

-0.0483
(0.042)
-0.049
(0.091)

-0.052
(0.053)
0.014
(0.107)

0.089
(0.068)
-0.026
(0.145)
0.150∗∗
(0.064)
0.239
(0.148)
0.092∗
(0.054)
0.075
(0.113)
0.272∗∗
(0.095)
0.346
(0.233)

×
×
IV

(3)

-0.169∗∗
(0.066)
-0.119
(0.136)

(4)

-0.221∗∗
(0.107)
-0.260∗∗
(0.128)

-0.026
(0.055)
0.128
(0.119)

-0.062
(0.051)
-0.113
(0.112)

-0.056
(0.060)
-0.011
(0.129)

0.120
(0.082)
0.241
(0.183)
0.194∗∗
(0.091)
0.400∗∗
(0.202)

0.077
(0.071)
0.251
(0.161)
0.292∗∗
(0.129)
0.465∗
(0.282)

×
×
IV

0.005
(0.070)
0.028
(0.088)

-0.049
(0.065)
-0.069
(0.088)

0.070
(0.085)
-0.091
(0.119)

0.144
(0.113)
0.022
(0.143)
0.323∗∗
(0.158)
0.234
(0.167)

0.109
(0.092)
0.047
(0.107)
0.436∗∗
(0.219)
0.217
(0.236)

×
×
IV

(5)

-0.042∗∗∗
(0.012)
-0.021
(0.014)
-0.020∗
(0.011)
-0.016
(0.013)
-0.032∗∗
(0.011)
-0.051∗∗∗
(0.012)
-0.026∗∗
(0.012)
0.003
(0.016)

-0.0003
(0.012)
0.049∗∗∗
(0.010)

0.008
(0.011)
0.016∗
(0.009)
0.021∗∗
(0.010)
0.034∗∗∗
(0.010)

-0.012
(0.016)
0.050∗∗∗
(0.014)

×
×
IV

(6)

-0.020∗
(0.011)
-0.012
(0.012)

-0.016
(0.011)
-0.016∗
(0.010)
-0.034∗∗
(0.011)
0.010
(0.009)

-0.013
(0.011)
0.015
(0.013)

-0.005
(0.010)
0.062∗∗∗
(0.006)
0.021∗∗
(0.010)
0.018∗∗
(0.006)
0.036∗∗∗
(0.010)
0.027∗∗∗
(0.006)

-0.002
(0.014)
0.057∗∗∗
(0.008)

×
×
IV

(7)

-0.052∗∗∗
(0.008)
-0.0001
(0.005)
-0.051∗∗∗
(0.009)
0.005
(0.005)
-0.021∗∗
(0.009)
0.042∗∗∗
(0.004)
-0.020∗∗
(0.009)
0.064∗∗∗
(0.006)
0.024∗∗∗
(0.007)
0.020∗∗∗
(0.001)
0.020∗∗∗
(0.006)
0.013∗∗∗
(0.001)
0.033∗∗∗
(0.006)
0.002
(0.001)

0.004
(0.011)
0.081∗∗∗
(0.001)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as an
additional instrument. All estimations include 6-hour lags on pollution. Weather variables include current
and 24-hour lags of quartics in humidity, temperature, wind speed, precipitation, and thermal oscillation.
Time fixed effects include dummies for year, month, dow, hour, and hour × weekend. Linear local approach
uses as outcome variables the residuals from the (full) regressions of hourly average pollution (in logs) on
covariates. Peak hours: N full sample = 34,796; N local alerts = 3,056 (5), 1,816 (6), 960 (7); N local
pre-emergencies = 1,040 (5), 704 (6), 344 (7). Off-peak hours: N full sample = 17,366; N local alerts =
1,530 (5), 910 (6), 482 (7); N local pre-emergencies = 516 (5), 348 (6), 172 (7). Standard errors robust to a
24-hour serial correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

139

episodes occur only on a sporadic base.

Table 3.10. Alerts and Pre-emergencies Impact on Hourly Average Vehicle Trips

Hours

Episode

1[Alerts]

Peak

1[Pre-emergencies]

1[Alerts]

Off-Peak

1[Pre-emergencies]

Weather
Time FE
Estimation
Functional Form
Bandwidth k

Polynomial Approach

Local Approach

(1)

(2)

(3)

(4)

(5)

(6)

(7)

0.012
(0.008)
-0.112∗∗∗
(0.031)

0.005
(0.014)
-0.040
(0.048)

×
×
OLS
Linear

-

-0.129∗∗
(0.048)
-0.303∗∗
(0.113)
-0.195∗∗
(0.083)
-0.385∗∗
(0.184)

×
×
IV

-0.146∗∗
(0.062)
-0.585∗∗∗
(0.145)
-0.225∗∗
(0.097)
-0.515∗∗
(0.204)

×
×
IV

-0.206∗∗
(0.077)
-0.197∗∗
(0.094)
-0.334∗∗
(0.142)
-0.408∗∗
(0.163)

×
×
IV

-0.061∗∗∗
(0.010)
-0.059∗∗∗
(0.014)
-0.097∗∗∗
(0.018)
0.018
(0.022)

×
×
IV

-0.059∗∗∗
(0.010)
-0.052∗∗∗
(0.013)
-0.094∗∗∗
(0.020)
0.020
(0.022)

×
×
IV

-0.037∗∗∗
(0.011)
-0.048∗∗∗
(0.011)
-0.078∗∗∗
(0.020)
0.010
(0.017)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as
an additional instrument. Weather variables include current and 24-hour lags of quartics in humidity,
temperature, wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for
year, month, dow, hour, and hour × weekend. Linear local approach uses as outcome variables the residuals
from the (full) regressions of hourly average vehicle trips (in logs) on covariates. Peak hours: N full sample
= 58,425; N local alerts = 5,240 (5), 3,113 (6), 1,654 (7); N local pre-emergencies = 1,800 (5), 1,225 (6),
601 (7). Off-peak hours: N full sample = 37,488; N local alerts = 3,367 (5), 2,006 (6), 1,074 (7); N local
pre-emergencies = 1,149 (5), 778 (6), 387 (7). Standard errors robust to heteroskedasticity and 2-hour serial
correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

Likewise, GMM-IV results for a polynomial fit of car use during air quality pre-

emergencies, indicate that hourly average vehicle trips are curtailed by around 20-59 per-

centage points during peak hours, and by around 39-52 percentage points during off-peak

hours (columns 2-4). These estimates are all statistically different from zero. They are also

consistent with findings for hourly maximum trips depicted in Table C3.5, mostly during

peak hours. Findings for a local linear approach in columns (5)-(7) of Table 3.10 indicate a

more conservative reduction of 5-6 percentage points during peak hours, and no effects dur-

ing off-peak hours. All the results in Table 3.10 pass the HAC F-test on weak instruments,

except those in column (5).

At first, higher causal impacts on car trips are expected for days with air quality

140

pre-emergencies relative to those with alerts, as more severe driving bans are triggered by

the former episode relative to the latter. Results for a polynomial fit in columns (2)-(4)

of Table 3.10 are consistent with this intuition; stricter environmental episodes reduce car

use in a higher proportion. Notwithstanding, estimation results from a local linear ap-

proach suggest smaller reductions in car use during pre-emergencies. Whether the local

linear approach is preferred over low-order polynomial fits is still an outstanding issue in the

literature. Nevertheless, a possible explanation for eventual smaller reductions in car use

during more stringent driving bans comes from a behavioral aspect. Drivers that are not

banned from driving can take advantage of reduced road congestion due to the activation of

severe driving restrictions. In this case, potential reductions in mobile source pollution dur-

ing pre-emergencies and emergencies from driving bans would be outweighed by an increase

in driving due to reduced road congestion. Unfortunately, whether this behavior is being

triggered by severe driving restrictions in Santiago is hard to test with the data available.

Table 3.11 exhibits estimation results for heterogeneous impacts of episodes on daily

average mass-transit system ridership. Results for the impacts on daily maximum are dis-

played in Table C3.6 in the Appendix C. Like results from pooled estimations in Table 3.8,

findings for a polynomial fit in columns (2)-(4) reveal no effects of driving bans on mass-

transit system ridership, except for a slight suggestion of reductions when a cubic polynomial

is fitted to the data. When a local linear fit is introduced, findings in columns (5)-(7) suggest

an increase in metro ridership during the two episodes and during both peak and off-peak

hours. According to findings in column (7), air quality alerts increase daily average metro

ridership by around 8 percentage points during peak hours, and results in columns (6) and

(7) indicate increments of around 13-24 percentage points during off-peak hours. Higher

increments in the use of alternative modes of transportations during non-rush hours can

be explained first, by regular metro users substituting their riding towards non-rush hours

under the threat of a massive use during peak hours of days with episodes, and second, by

banned drivers substituting their driving towards alternative modes of transportation during

141

times at which they run with some extra capacity. In the case of air quality pre-emergencies,

estimates in column (7) of Table 3.11 also suggest an increase in metro ridership, namely of

5 percentage points during peak hours, and of 16 percentage points during off-peak hours.

Regarding the use of Transantiago, estimates in column (7) also indicate an increase in this

transportation mode of 8 percentage points during alerts, and of 3 percentage points during

pre-emergencies.

Table 3.11. Alerts and Pre-emergencies Impact on Daily Average Mass-Transit Systems

System

Hours

Episode

Metro

1[Alerts]

Peak

1[Pre-emergencies]

1[Alerts]

Off-Peak

1[Pre-emergencies]

Transantiago All Hours

1[Pre-emergencies]

1[Alerts]

Weather
Time FE
Estimation
Functional Form
Bandwidth k

Polynomial Approach

Local Approach

(2)

(3)

(4)

(5)

(6)

(7)

-0.163
(0.123)
-0.214
(0.334)

-0.285
(0.227)
-0.512
(0.641)

-0.138
(0.090)
-0.175
(0.178)

×
×
IV

-0.155
(0.166)
-0.377
(0.370)

-0.320
(0.327)
-0.885
(0.717)

-0.134
(0.103)
-0.083
(0.202)

×
×
IV

-0.231
(0.266)
-0.442
(0.314)

-0.399
(0.505)
-1.010∗
(0.608)

-0.096
(0.099)
-0.334∗
(0.193)

×
×
IV

0.012
(0.016)
-0.008
(0.019)

0.047
(0.032)
-0.002
(0.040)

-0.021
(0.015)
-0.045∗∗
(0.016)

×
×
IV

0.021
(0.019)
0.006
(0.017)
0.132∗∗∗
(0.038)
-0.018
(0.030)

-0.006
(0.016)
-0.016
(0.013)

×
×
IV

0.075∗∗
(0.023)
0.053∗∗
(0.022)
0.244∗∗∗
(0.038)
0.161∗∗
(0.051)
0.079∗∗∗
(0.019)
0.033∗
(0.019)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

(1)
0.050∗
(0.029)
0.010
(0.117)

0.079
(0.054)
0.022
(0.210)

0.039
(0.024)
-0.044
(0.112)

×
×
OLS
Linear

-

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as an
additional instrument. Weather variables include current and 1-day lags of quartics in humidity, temperature,
wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for year, month, and
dow. Linear local approach uses as outcome variables the residuals from the (full) regressions of daily mass-
transit system trips (in logs) on covariates. Metro: N full sample = 4,373; N local alerts = 384 (5), 229 (6),
122 (7). N local pre-emergencies = 130 (5), 88 (6), 43 (7). Transantiago: N full sample = 3,222; N local
alerts = 290 (5), 175 (6), 98 (7); N local pre-emergencies = 104 (5), 72 (6), 36 (7). Standard errors robust
to heteroskedasticity and 1-day serial correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05,
∗∗∗p < 0.001

A plausible explanation for mass-transit system ridership increasing to a lesser degree

during more severe driving restrictions comes from the idea that drivers follow the author-

ities’ recommendations of avoiding outdoor exposure to pollution, especially during days

where particulate matter concentrations exceed safe exposure levels.

Is important to re-

142

call that environmental episodes in Santiago trigger temporary driving restrictions, but also

recommend the population the avoidance of outdoor exposure to ambient contamination, es-

pecially for those more vulnerable to pollution. These recommendations are highlighted even

more during the more severe air quality episodes. Thus, the possibility that pre-emergency

and emergency episodes will induce an avoidance behavior not only on affected drivers but on

the entire population is highly likely. Some individuals other than drivers, may stay at home

during the more severe environmental episodes, which is consistent with lesser reductions

observed in car trips during environmental pre-emergencies.

3.6.3. Robustness Checks

3.6.3.1 Monitoring Station Outside the Restricted Area as a Control Station

To check the robustness of the previous findings on pollution, I use information on

airborne concentrations originated in one of the SINCA network stations but omitted from

the previous empirical analysis. This unit, labeled before as the eleventh station (see Section

??), corresponds to Talagante station, which is located approximately at 38 km. (23 miles)

SE Santiago’s downtown. Talagante is a commune that belongs to the Talagante province,

however, is not part of the GEC plan that issues environmental episodes for Santiago, despite

being part of Santiago’s Metropolitan Area. The existence of a station that, due to its

location will likely follow a similar pollution trend than in the rest of Santiago, but for which

no policy is in effect, allows the specification of a difference-in-difference (DID) robustness

analysis. The DID design considers all SINCA’s monitoring stations located in Santiago as

“treated” stations during days with environmental episodes, except the Talagante station,

which is defined as the “control” unit. The identifying assumption in this case is that,

conditional on weather covariates and several fixed effects, station-specific unobservables

affecting pollution trends are uncorrelated with the issuance of episodes.

Table 3.12 displays the results for the DID setup using a panel fixed effects estimator

with Newey-West standard errors. Panel A contains the average treatment effect on the

143

Table 3.12. Robustness Check - Using Talagante as a Control Station

PM10

PM2.5

CO

NOX

(1)

(2)

(1)

(2)

(1)

(2)

(1)

(2)

Panel A. Pooling Episodes
1[Episode]

-0.084∗∗∗
(0.011)

-0.084∗∗∗
(0.011)

-0.083∗∗∗
(0.010)

-0.083∗∗∗
(0.010)

-0.032∗∗∗
(0.008)

-0.032∗∗∗
(0.008)

-0.102∗∗∗
(0.012)

-0.103∗∗∗
(0.012)

Panel B. Heterogeneous Effects
1[Alerts]

1[Pre-emergencies]

Time Trend
N panel

-0.074∗∗∗
(0.013)
-0.107∗∗∗
(0.024)

-0.075∗∗∗
(0.012)
-0.100∗∗∗
(0.021)

-0.073∗∗∗
(0.013)
-0.107∗∗∗
(0.024)

×

-0.075∗∗∗
(0.012)
-0.098∗∗∗
(0.021)

×

-0.023∗∗
(0.009)
-0.046∗∗
(0.019)

-0.023∗∗
(0.009)
-0.045∗∗
(0.019)

×

-0.100∗∗∗
(0.013)
-0.110∗∗∗
(0.029)

-0.100∗∗∗
(0.013)
-0.110∗∗∗
(0.029)

×

394,879

394,879

426,041

426,041

426,176

426,176

388,943

388,943

Notes: Including all hours. Estimations include 6-hour lags on pollution; current and 24-hour lags of quartics
in humidity, temperature, wind speed, precipitation, and thermal oscillation; fixed effects for year, month,
dow, hour, hour × weekend; and stations fixed effects. Using Newey-West standard errors robust to 6-hour
correlation. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

treated (ATT) stations after pooling the episodes, while panel B displays the ATT consid-

ering heterogeneous impacts of these occurrences. Results in panel A consistently suggest

that overall, environmental episodes reduce mobile source pollution in the treated area by

around 3-10 percentage points. Particularly, episodes on average reduce particulate matter

concentrations by 8 percentage points, CO by 3 percentage points, and NOX concentrations

by 10 percentage points. When heterogeneous effects are allowed, results indicate that air

quality alerts and pre-emergencies both curb mobile source pollution, with air quality pre-

emergencies (and emergencies) reducing pollution, as expected, in a higher proportion. Air

quality alerts, on average, reduce particulate matter concentrations by 7-8 percentage points,

CO by 2 percentage points, and NOX by 10 percentage points. Air quality pre-emergencies,

instead, curb PM concentrations by 10-11 percentage points, CO by 5 percentage points,

and NOX by 11 percentage points. Results in panel B of Table 3.12 also represents a simple

test on the role of driving restrictions triggered by episodes, as opposed to other abatement

actions instead targeting stationary sources. The fact that air quality alerts effectively re-

duce ambient concentrations for all the four pollutants considered, suggests the effectiveness

of temporary driving restrictions relative to other mitigation actions, as these episodes only

144

affect pollution from mobile sources.

3.6.3.2 False ICAP

To check the validity of results on car trips, Table 3.13 displays the episodes’ impact on

vehicle flows from both a pooled FRD (panel A) and a multi cutoff FRD (panel B) estimation

using ICAP10,t−7 as the running variable. The probability of issuing an environmental episode

on day t, is affected by the maximum value that the air quality index takes across stations
during day t− 1. Conditional on current and 24-hour lags of quartiles in weather covariates,

there are no reasons to expect significant impacts on this probability of the air quality

index when lagged 7 days. Therefore, the use of ICAP10,t−7 as an instrument of the episodes’

announcement serves as a placebo tests on the treatment effects of episodes, and their driving

restrictions, on car flows. Results for the GMM-IV estimation in Table 3.13 indicate that,

except for the isolated estimation result on pre-emergencies in column (2), there are no

statistically significant effects on vehicle trips when values 1-week before for the air quality

index are used as the forcing variable.

3.7. Conclusions

Driving restrictions are a common regulatory mechanism used by several governments

to reduce mobile source pollution and traffic congestion. Empirical evidence of how these

restrictions affect the number of cars in circulation, and so their effects on pollution is

not yet conclusive. In this study, I provide suggestive indications that driving bans, when

issued on a temporary basis and in conjunction with air quality warnings, represent an

efficient mechanism to curb emissions from mobile sources. I test this idea using evidence for

Santiago’s temporary license-plate based driving restrictions, issued by official authorities

whenever they foresee critical levels of air pollution. Using data on several air pollutants

from mobile sources, and on urban traffic flows, I estimate a fuzzy regression discontinuity

145

Table 3.13. Robustness Check - Using ICAP10,t−7 as the Running Variable

Polynomial Approach

Local Approach

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Panel A. Pooling Episodes
0.001
1[Episode]
(0.005)

0.235
(0.147)

0.088
(0.118)

0.129
(0.105)

-0.017
(0.012)

-0.001
(0.013)

-0.020
(0.016)

Panel B. Heterogeneous Effects
1[Alerts]

1[Pre-emergencies]

-0.012∗
(0.006)
-0.040∗∗
(0.016)

0.277
(0.176)
-0.236∗∗
(0.114)

0.125
(0.170)
-0.546
(0.453)

0.332
(0.207)
-0.511
(0.483)

-0.024
(0.016)
-0.035
(0.027)

-0.005
(0.017)
-0.020
(0.024)

-0.027
(0.019)
-0.001
(0.018)

Estimation
Functional Form
Bandwidth k

OLS
Linear

-

IV

IV

IV

IV

IV

IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: Including all hours. GMM estimations in columns (2)-(7) use ICAP10,t−7 as the running variable
and ICAP2.5,t−7 as an additional instrument. Estimations include current and 24-hour lags of quartics in
humidity, temperature, wind speed, precipitation, and thermal oscillation; and fixed effects for year, month,
dow, hour, hour × weekend. Linear local approach uses as outcome variables the residuals from the (full)
regressions of hourly average vehicle trips (in logs) on covariates. N full sample = 95,913; N local = 8,713
(5), 5,216 (6), 2,767 (7); N local alerts = 8,713 (5), 5,216 (6), 2,767 (7); N local pre-emergencies = 2,923 (5),
1,997 (6), 1,001 (7). Standard errors robust to heteroskedasticity and 2-hour serial correlation in parentheses.
Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

design with air quality index used as an instrument for episode announcements.

Results on the overall impact of environmental episodes show strong evidence of a

reduction in mobile source pollution of around 2-3 percentage points during peak hours.

These results are consistent with hourly vehicle trips decreasing by 4-6 percentage points on

average, and with metro ridership increasing by 3 percent during peak hours. Results for

car trips also indicate reductions during hours in which driving restrictions are not in place,

which rejects the idea of an intertemporal substitution of driving towards hours unaffected

by the policy.

When I allow heterogeneous effects of episodes on the outcome variables, results for

peak hours suggest that airborne contamination is curtailed by 2-5 percentage points on

average during air quality alerts and pre-emergencies. Results for the impacts of episodes on

pollution during off-peak hours indicate an increase in airborne contamination during non-

rush hours, although these increments are presumably due to the persistence of pollutants in

146

the atmosphere, which is likely worsened during the more critical air quality episodes. Results

for car trips support this conclusion as they all suggests reductions in vehicle flows during

days with episodes. Average cars in circulation during restricted hours are curtailed by 4-6

percentage points during air quality alerts, and by 5-6 percentage points during air quality

pre-emergencies. The fact that these reductions are not necessarily higher when more severe

driving bans are in place suggests that unaffected drivers take advantage of a reduced road

congestion during restricted days and eventually increase their driving. Findings for mass-

transit system ridership ratify this conclusion. The use of alternative modes of transportation

increases during days with temporary driving restrictions, mostly during non-rush hours at

which the system is not running at full capacity. Unexpected results in urban traffic flows

during pre-emergencies relative to alerts suggest first, that some drivers might decide to

drive more due to a potential reduction in road congestion during days in which more severe

restrictions are in place, and second, that some affected drivers might decide to stay at home

and avoid outdoor exposure during days with exacerbated air pollution.

The empirical exercise in this paper suggests that Santiago’s mechanism of tempo-

rary banning cars from driving is an effective policy to curb pollution from mobile sources.

Multiple factors can make this policy effective in Santiago relative to similar designs applied

to other cities. Short-term driving bans, as opposed to long-term restrictions, might induce

different behavioral responses on drivers. This policy restricts drivers with a less-than-one

probability, which might imply that potentially affected drivers are not facing strong incen-

tives to purchase a second car. Moreover, the fact that these restrictions are triggered by

air quality warnings increases awareness about air quality conditions, pushing the marginal

driver to stay at home and avoid exposure to critical air pollution. The existence of mass-

transit systems with over a wide extent of the city may also incentivize the switching of

driving towards alternative forms of transportation instead of switching towards days or

hours unrestricted by this policy. Finally, it is also possible that Santiago’s driving re-

strictions success may only be reflecting the role of enforcement and cultural habits in the

147

effectiveness of a policy mechanism to reduce airborne contamination.

148

APPENDICES

149

APPENDIX A: SUPPLEMENTAL MATERIAL

Figure A3.1. Spatial Location of Santiago’s Monitoring Stations

Notes: Areas affected by driving restrictions in color. Only one monitoring station is outside this range
-Talagante station. Borders represent communes’ limits.

150

Figure A3.2. Daily Average Pollutant Concentrations

(a) PM10

(b) PM2.5

151

Figure A3.2. (cont’d)

(c) CO

(d) NOX

152

Table A3.1. 2015 Protocols in Environmental Episodes

Episode

Alert

Pre-emergency

Emergency

Protocols

• Temporary driving restriction on 20% (weekends) of dirty vehicles
• Prohibition in the use of wood and other biomass for residential heat-

ing

• Temporary driving restriction on 60% (all days) of dirty vehicles
• Temporary driving restriction on 20% (all days) of clean vehicles
• Temporary suspension of stationary emissions sources responsable
for 30% of total stationary emissions of particulate matter. This is
equivalent to the shutdown of 750 facilities.

• Elementary and high-school physical education classes and commu-

nity sports are suspended by the Ministry of Education

• Prohibition in the use of wood and other biomass for residential heat-

ing

• Temporary driving restriction on 80% (all days) of dirty vehicles
• Temporary driving restriction on 40% (all days) of clean vehicles
• Temporary suspension of stationary emissions sources responsable
for 50% of total stationary emissions of particulate matter. This is
equivalent to the shutdown of 2,461 facilities.

• Recommendation by the Ministry of Education of classes cancella-

tions in elementary schools and high-schools

• Prohibition in the use of wood and other biomass for residential heat-

ing

Notes: Adapted from Mullins and Bharadwaj (2015).

153

Table A3.2. Number of License Plate Digits Restricted by Permanent and Temporary
Driving Restrictions. 1990-2015

Stage

Type of

Day of

Permanent

Temporary Restrictions

Light Vehicle

the Week Restriction Alert Pre-emergency Emergency

1990-1996

Dirty

Clean

1997-2000

Dirty

Clean

2001-2007

Dirty

Clean

2008-2015

Dirty

Clean

Weekdays
Weekends

Weekdays
Weekends

Weekdays
Weekends

Weekdays
Weekends

Weekdays
Weekends

Weekdays
Weekends

Weekdays
Weekends

Weekdays
Weekends

2
0

0
0

2
0

0
0

2
0

0
0

4
0

0
0

-
-

-
-

4
2

0
0

4
2

0
0

4
2

0
0

-
-

-
-

6
4

0
0

6
4

2
2

6
6

2
2

-
-

-
-

8
6

0
0

8
6

4
4

8
8

4
4

Notes: Permanent and temporary driving restrictions are in place only between 7:30am to 9:00pm and from
April 1st to August 31st. Temporary driving restrictions show the total number of digits restricted during
days with environmental episodes (i.e. number of digits restricted under the permanent restriction plus
additional digits). Dirty = Vehicles without a green sticker. Clean = Vehicles with a green sticker.

154

Table A3.3. 2016 Calendar of Driving Restrictions for Santiago

Type of Restriction

Day/Episode

Digits Affected

Permanent Restriction

Temporary Restrictions

Monday
Tuesday

Wednesday
Thursday

Friday

First Episode
Second Episode
Third Episode
Fourth Episode
Fifth Episode

3-4-5-6
7-8-9-0
1-2-3-4
5-6-7-8
9-0-1-2

0-1
2-3
4-5
6-7
8-9

Notes: Permanent restrictions on dirty vehicles only.

155

Table A3.4. Historical Issuance of Environmental Episodes. 1997-2015

Environmental Episodes

Alerts Pre-emergencies Emergencies Total

Year

1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015

7
19
11
27
21
22
21
13
7
21
27
21
23
7
19
23
6
22
38

12
11
14
11
4
11
5
2
2
3
4
8
0
2
7
2
0
1
16

0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1

3
1

19
31
26
38
25
33
26
15
9
24
31
29
23
9
26
25
6
23
55

473
397

Episodes Since 1997
Episodes Since 2000

355
318

115
78

Notes: Data from the Unidad Operativa de Control de Tr´ansito (UOCT). Currently, environmental episodes
are announced during winter days defined as April 1st through August 31st. Yet, five episodes were an-
nounced on March and September during 1997 and 1998. These episodes are not listed in this table.

156

Figure A3.3. Daily Maximum ICAP10 and ICAP2.5 during Winter 2015

(a) ICAP based on PM10

Notes: Using daily maximum ICAP. Winter is defined as April 1st through August 31st.

(b) ICAP based on PM2.5

157

Figure A3.4. Air Quality Index Densities During Episodes

(a) Alerts

(b) Pre-emergencies

Notes: Densities of the daily max air quality indexes during day t − 1. Using an Epanechnikov kernel, with
a bandwidth of 22.88 (alerts) and of 26.52 (pre-emergencies).

Table A3.5. Expected Percentage of Light-Duty Private Cars Affected by Restrictions Over
2001-2015.

Years

Permanent Restriction

Temporary Restrictions

Alerts

Pre-emergencies

Emergencies

Clean Cars Dirty Cars Clean Cars Dirty Cars Clean Cars Dirty Cars Clean Cars Dirty Cars

2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015

-
-
-
-
-
-
-
-
-
-
-
-
-
-
-

8.83
8.22
7.68
6.89
5.94
7.03
3.10
4.82
3.81
2.51
1.99
1.60
1.32
1.07
0.87

-
-
-
-
-
-
-
-
-
-
-
-
-
-
-

26.49
24.65
23.03
20.68
17.82
21.08
9.30
7.24
5.72
3.76
2.99
2.40
1.98
1.60
1.30

22.34
23.57
24.64
26.21
28.12
25.94
33.80
35.18
36.19
37.49
38.01
38.40
38.68
38.93
39.13

44.15
41.08
38.39
34.47
29.71
35.14
15.51
14.47
11.43
7.52
5.97
4.80
3.96
3.20
2.61

44.68
47.14
49.29
52.43
56.24
51.89
67.59
70.35
72.38
74.99
76.02
76.80
77.36
77.87
78.26

61.82
57.51
53.75
48.25
41.59
49.19
21.71
19.30
15.24
10.02
7.96
6.40
5.28
4.26
3.48

Average

2.19

5.67

25.97

45.97

Notes: Considering estrictions placed during both weekdays and weekends. Calculations are based on the
number of clean and dirty light-duty private cars registered in Santiago’s Metropolitan Area, retrieved from
the Annual Reports on Road Vehicles (Anuario del Parque Vehicular de Veh´ıculos en Circulaci´on) available
at http://ine.cl.

158

APPENDIX B: EXAMPLE OF THE CASSMASSI FORECAST MODEL

The following equation describes the weights used by the Cassmassi model to forecast

PM10 concentrations in the Pudahuel station:

yt+1 = 39.4νt + 0.33yt + 2.06xt + 0.21ht − 21.7

(3.7)

where yt+1 is the expected 24-hour moving average of PM10 for day t + 1; νt is the

forecasted atmospheric stability for the following day taking discrete values from 1 to 5; yt

is the 24-hour moving average of PM10 measured during day t at 10:00am (local time); xt

is the temperature () of the 925 hP a level registered by monitoring station Santo Domingo

(located at 80km west from Santiago) at 12:00pm UTC of day t, and ht is 24-hour change in

height measured at 500 level registered by monitoring station Santo Domingo during day t

at 12:00pm UTC. The inclusion of xt and ht as part of the equations in the Cassmassi model

are intended to control for the expected strength of thermal inversions in Santiago (Perez,

2008).

159

APPENDIX C: ADDITIONAL ESTIMATIONS

Table C3.1. Episodes Impact on Hourly Maximum Pollution

Hours

Pollutant

Polynomial Approach

Local Approach

Peak

Off-Peak

PM10

PM2.5

CO

NOX

PM10

PM2.5

CO

NOX

Weather
Time FE
Estimation
Functional Form
Bandwidth k

(1)

-0.024∗∗
(0.008)
-0.017∗∗
(0.007)
-0.026∗∗
(0.010)
-0.017∗∗
(0.008)
0.023∗∗
(0.010)
0.025∗∗
(0.010)
0.025∗∗
(0.010)
0.013
(0.012)

×
×
OLS
Linear

-

(2)

-0.235∗∗∗
(0.065)
-0.093∗
(0.052)
-0.134∗∗
(0.058)
0.013
(0.012)

-0.041
(0.068)
0.025
(0.065)
-0.001
(0.062)
0.184∗∗
(0.086)

×
×
IV

(3)

-0.263∗∗
(0.104)
0.024
(0.079)
-0.124
(0.078)
0.013
(0.012)

0.060
(0.097)
0.290∗∗
(0.124)
0.032
(0.089)
0.308∗∗
(0.147)

×
×
IV

(4)

-0.226∗∗
(0.097)
0.006
(0.079)
-0.148∗
(0.081)
0.013
(0.012)

0.108
(0.094)
0.280∗∗
(0.120)
0.059
(0.087)
0.271∗∗
(0.137)

×
×
IV

(5)

(6)

(7)

-0.057∗∗∗
(0.010)
-0.031∗∗
(0.010)
-0.021∗∗
(0.010)
-0.032∗∗
(0.011)
0.028∗∗
(0.011)
0.041∗∗∗
(0.009)
0.040∗∗∗
(0.010)
0.024∗
(0.014)

×
×
IV

-0.037∗∗∗
(0.009)
-0.027∗∗
(0.009)
-0.027∗∗
(0.009)
-0.033∗∗∗
(0.009)
0.016∗
(0.009)
0.043∗∗∗
(0.007)
0.050∗∗∗
(0.009)
0.014
(0.011)

×
×
IV

-0.048∗∗∗
(0.007)
-0.052∗∗∗
(0.008)
-0.007
(0.007)
-0.026∗∗∗
(0.007)
0.016∗∗
(0.006)
0.034∗∗∗
(0.005)
0.045∗∗∗
(0.006)
0.017∗∗
(0.007)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as an
additional instrument. All estimations include 6-hour lags. Weather variables include current and 24-hour
lags of quartics in humidity, temperature, wind speed, precipitation, and thermal oscillation. Time fixed
effects include dummies for year, month, dow, hour, and hour × weekend. Linear local approach uses as
outcome variables the residuals from the (full) regressions of hourly maximum pollution concentrations (in
logs) on covariates. Peak hours: N full sample = 34,796; N local = 3,056 (5), 1,816 (6), 960 (7). Off-Peak
hours: N full sample = 17,366; N local = 1,530 (5), 910 (6), 482 (7). Standard errors robust to a 3-week
serial correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

160

Table C3.2. Episode Impact on Hourly Max Vehicle Trips

Hours

Peak

Off-Peak

Weather Variables
Time Fixed Effects
Estimation
Functional Form
Bandwidth k

Polynomial Approach

Local Approach

(1)

(2)

(3)

(4)

(5)

(6)

(7)

-0.004
(0.005)
-0.019∗∗
(0.009)

×
×
OLS
Linear

-0.053
(0.047)
-0.161∗∗
(0.080)

×
×
IV

-0.168∗∗
(0.065)
-0.228∗∗
(0.108)

×
×
IV

-0.155∗∗
(0.065)
-0.216∗∗
(0.106)

×
×
IV

-0.0003
(0.009)
-0.057∗∗∗
(0.015)

×
×
IV

-0.007
(0.011)
-0.082∗∗∗
(0.017)

×
×
IV

0.002
(0.011)
-0.095∗∗∗
(0.018)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as
an additional instrument. Weather variables include current and 24-hour lags of quartics in humidity,
temperature, wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for
year, month, dow, hour, and hour × weekend. Linear local approach uses as outcome variables the residuals
from the (full) regressions of hourly maximum vehicle trips (in logs) on covariates. Peak hours: N full
sample = 58,425; N local = 5,240 (5), 3,113 (6), 1,654 (7). Off-peak hours: N full sample = 37,488; N local
= 3,367 (5), 2,006 (6), 1,074 (7). Standard errors robust to heteroskedasticity and 2-hour serial correlation
in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

161

Table C3.3. Episodes Impact on Daily Max Mass-Transit System Trips

Polynomial Approach

Local Approach

(1)

(2)

(3)

(4)

(5)

(6)

-0.227
(0.173)
-0.201
(0.247)

-0.311
(0.250)
-0.292
(0.362)

-0.325
(0.242)
-0.268
(0.338)

0.025
(0.016)
0.046
(0.029)

0.020
(0.017)
0.122∗∗∗
(0.034)

-
-
×
×
IV

-
-
×
×
IV

-
-
×
×
IV

-
-
×
×
IV

-
-
×
×
IV

(7)
0.036∗
(0.021)
0.154∗∗∗
(0.030)

-
-
×
×
IV

System

Hours

Metro

Peak

Off-Peak

Transantiago All Hours

Weather
Time FE
Estimation
Functional Form
Bandwidth k

0.020
(0.025)
0.028
(0.037)

-
-
×
×
OLS
Linear

-

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as
an additional instrument. Weather variables include current and 24-hour lags of quartics in humidity,
temperature, wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for
year, month, dow, hour, and hour × weekend. Linear local approach uses as outcome variables the residuals
from the (full) regressions of daily maximum metro trips (in logs) on covariates. Metro: N full sample =
4,373; N local = 384 (5), 229 (6), 122 (7). Transantiago: Not available. Significance levels: ∗p < 0.10,
∗∗p < 0.05, ∗∗∗p < 0.001

162

Table C3.4. Alerts and Pre-emergencies Impact on Hourly Average Pollution
Concentrations.

Hours

Pollutant

Episode

Polynomial Approach

Local Approach

Peak

Off-Peak

1[Alerts]

PM10

1[Pre-emergencies]

1[Alerts]

PM2.5

1[Pre-emergencies]

1[Alerts]

CO

1[Pre-emergencies]

1[Alerts]

NOX

1[Pre-emergencies]

1[Alerts]

PM10

1[Pre-emergencies]

1[Alerts]

PM2.5

1[Pre-emergencies]

1[Alerts]

CO

1[Pre-emergencies]

1[Alerts]

NOX

1[Pre-emergencies]

Weather
Time FE
Estimation
Functional Form
Bandwidth k

(1)

0.044∗∗
(0.016)
-0.023
(0.036)
0.041∗∗
(0.015)
0.059
(0.037)

0.022
(0.021)
0.030
(0.056)

0.017
(0.017)
-0.026
(0.040)
0.081∗∗∗
(0.017)
0.089
(0.059)
0.091∗∗∗
(0.016)
0.071∗
(0.042)
0.040∗∗
(0.018)
0.004
(0.055)
0.037∗
(0.020)
0.083
(0.066)

×
×
OLS
Linear

-

(2)

-0.228∗∗
(0.082)
-0.440∗∗
(0.178)

(3)

-0.236∗∗
(0.089)
-0.306
(0.189)

(4)

-0.305∗∗
(0.140)
-0.424∗∗
(0.176)

-0.010
(0.056)
0.043
(0.120)

-0.073
(0.059)
-0.145
(0.131)

-0.052
(0.066)
-0.013
(0.132)

0.112
(0.083)
-0.047
(0.167)
0.230∗∗
(0.082)
0.265
(0.186)

0.106
(0.072)
0.049
(0.149)
0.303∗∗
(0.107)
0.375
(0.251)

×
×
IV

-0.001
(0.067)
0.124
(0.147)

-0.102
(0.069)
-0.212
(0.155)

-0.058
(0.074)
-0.063
(0.163)

0.100
(0.098)
0.330
(0.219)
0.291∗∗
(0.119)
0.594∗∗
(0.266)

0.120
(0.086)
0.256
(0.185)
0.363∗∗
(0.144)
0.523∗
(0.314)

×
×
IV

0.043
(0.091)
-0.006
(0.117)

-0.109
(0.089)
-0.097
(0.126)

0.075
(0.108)
-0.169
(0.158)

0.182
(0.137)
0.005
(0.176)
0.462∗∗
(0.211)
0.451∗∗
(0.223)

0.086
(0.117)
0.049
(0.133)
0.493∗∗
(0.241)
0.261
(0.257)

×
×
IV

(5)

-0.075∗∗∗
(0.015)
-0.019
(0.017)
-0.054∗∗∗
(0.015)
-0.026
(0.017)
-0.025∗
(0.014)
-0.090∗∗∗
(0.016)
-0.047∗∗
(0.016)
0.026
(0.020)
0.038∗∗
(0.014)
0.082∗∗∗
(0.013)
0.043∗∗
(0.014)
0.083∗∗∗
(0.013)
0.052∗∗∗
(0.014)
0.120∗∗∗
(0.015)

0.023
(0.020)
0.097∗∗∗
(0.015)

×
×
IV

(6)

-0.060∗∗∗
(0.015)
-0.014
(0.015)
-0.038∗∗
(0.015)
-0.012
(0.014)
-0.047∗∗∗
(0.014)
0.006
(0.013)
-0.037∗∗
(0.015)
0.054∗∗∗
(0.016)
0.036∗∗
(0.013)
0.111∗∗∗
(0.008)
0.043∗∗∗
(0.012)
0.075∗∗∗
(0.007)
0.064∗∗∗
(0.013)
0.090∗∗∗
(0.009)

0.017
(0.017)
0.111∗∗∗
(0.010)

×
×
IV

(7)

-0.078∗∗∗
(0.011)
0.021∗∗
(0.007)
-0.104∗∗∗
(0.012)
0.040∗∗∗
(0.006)
-0.033∗∗
(0.012)
0.064∗∗∗
(0.006)

-0.015
(0.012)
0.084∗∗∗
(0.008)
0.028∗∗
(0.010)
0.014∗∗∗
(0.001)
0.023∗∗
(0.008)
0.019∗∗∗
(0.002)
0.063∗∗∗
(0.009)
0.022∗∗∗
(0.001)

0.002
(0.010)
0.084∗∗∗
(0.003)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as an
additional instrument. All estimations include 6-hour lags. Weather variables include current and 24-hour
lags of quartics in humidity, temperature, wind speed, precipitation, and thermal oscillation. Time fixed
effects include dummies for year, month, dow, hour, and hour × weekend. Linear local approach uses as
outcome variables the residuals from the (full) regressions of hourly average pollution (in logs) on covariates.
Peak hours: N full sample = 34,796; N local alerts = 3,056 (5), 1,816 (6), 960 (7); N local pre-emergencies
= 1,040 (5), 704 (6), 344 (7). Off-peak hours: N full sample = 17,366; N local alerts = 1,530 (5), 910 (6),
482 (7); N local pre-emergencies = 516 (5), 348 (6), 172 (7). Standard errors robust to a 24-hour serial
correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

163

Table C3.5. Alerts and Pre-emergencies Impact on Hourly Max Vehicle Trips

Hours

Episode

1[Alerts]

Peak

1[Pre-emergencies]

1[Alerts]

Off-Peak

1[Pre-emergencies]

Weather
Time FE
Estimation
Functional Form
Bandwidth k

Polynomial Approach

Local Approach

(2)

(3)

(4)

(5)

(6)

(7)

-0.104∗∗
(0.047)
-0.189∗
(0.114)
-0.171∗∗
(0.082)
-0.143
(0.187)

×
×
IV

-0.162∗∗
(0.057)
-0.279∗∗
(0.122)
-0.218∗∗
(0.095)
-0.249
(0.186)

×
×
IV

-0.181∗∗
(0.079)
-0.223∗∗
(0.097)
-0.236∗
(0.141)
-0.272∗
(0.152)

×
×
IV

0.003
(0.010)
-0.069∗∗∗
(0.015)
-0.063∗∗∗
(0.017)
-0.009
(0.021)

×
×
IV

-0.002
(0.012)
-0.059∗∗∗
(0.013)
-0.086∗∗∗
(0.020)
-0.049∗∗
(0.021)

×
×
IV

0.014
(0.013)
-0.055∗∗∗
(0.012)
-0.075∗∗∗
(0.021)
-0.047∗∗
(0.017)

×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

(1)
0.015∗
(0.008)
-0.143∗∗∗
(0.029)

-0.002
(0.015)
-0.106∗∗
(0.046)

×
×
OLS
Linear

-

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as
an additional instrument. Weather variables include current and 24-hour lags of quartics in humidity,
temperature, wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for
year, month, dow, hour, and hour × weekend. Linear local approach uses as outcome variables the residuals
from the (full) regressions of hourly maximum vehicle trips (in logs) on covariates. Peak hours: N full sample
= 58,425; N local alerts = 5,240 (5), 3,113 (6), 1,654 (7); N local pre-emergencies = 1,800 (5), 1,225 (6),
601 (7). Off-peak hours: N full sample = 37,488; N local alerts = 3,367 (5), 2,006 (6), 1,074 (7); N local
pre-emergencies = 1,149 (5), 778 (6), 387 (7). Standard errors robust to heteroskedasticity and 2-hour serial
correlation in parentheses. Significance levels: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.001

164

Table C3.6. Alerts and Pre-emergencies Impact on Daily Max Mass-Transit System Trips

Systems

Hours

Episode

Metro

1[Alerts]

Peak

1[Pre-emergencies]

1[Alerts]

Off-Peak

1[Pre-emergencies]

Transantiago All Hours

Weather
Time FE
Estimation
Functional Form
Bandwidth k

Polynomial Approach

Local Approach

(1)
0.061∗
(0.034)
0.024
(0.136)

0.072
(0.051)
-0.023
(0.200)

-
-
×
×
OLS
Linear

-

(2)

(3)

(4)

(5)

(6)

-0.232
(0.143)
-0.242
(0.397)

-0.218
(0.206)
-0.340
(0.584)

-
-
×
×
IV

-0.167
(0.196)
-0.436
(0.431)

-0.228
(0.289)
-0.536
(0.626)

-
-
×
×
IV

-0.238
(0.313)
-0.586
(0.372)

-0.268
(0.458)
-0.831
(0.548)

0.027
(0.017)
0.024
(0.023)
0.065∗∗
(0.032)
0.001
(0.036)

0.032
(0.020)
0.024
(0.022)
0.181∗∗∗
(0.040)
-0.050∗∗
(0.024)

-
-
×
×
IV

-
-
×
×
IV

-
-
×
×
IV

(7)

0.088∗∗
(0.028)
0.121∗∗∗
(0.014)
0.231∗∗∗
(0.037)
0.116∗∗∗
(0.022)

-
-
×
×
IV

Linear Quadratic Cubic

Linear

Linear

Linear

-

-

-

75

50

25

Notes: GMM estimations in columns (2)-(7) use ICAP10,t−1 as the running variable and ICAP2.5,t−1 as an
additional instrument. Weather variables include current and 1-day lags of quartics in humidity, temperature,
wind speed, precipitation, and thermal oscillation. Time fixed effects include dummies for year, month, and
dow. Linear local approach uses as outcome variables the residuals from the (full) regressions of daily mass-
transit system trips (in logs) on covariates. Metro: N full sample = 4,373; N local alerts = 384 (5), 229
(6), 122 (7). N local pre-emergencies = 130 (5), 88 (6), 43 (7). Transantiago: Not available. Standard
errors robust to heteroskedasticity and 1-day serial correlation in parentheses. Significance levels: ∗p < 0.10,
∗∗p < 0.05, ∗∗∗p < 0.001

165

REFERENCES

166

REFERENCES

Angrist, Joshua D and J¨orn-Steffen Pischke (2008). Mostly Harmless Econometrics: An Em-

piricist’s Companion. Princeton University Press.

Atal, R (2009). “Efectos de las Medidas de Gesti´on de Episodios Cr´ıticos en la Calidad del
Aire en Santiago, Chile”. In: Instituto de Econom´ıa, Pontificia Universidad Cat´olica de
Chile, Santiago.

Barahona, Hern´an, Francisco Gallego, and Juan-Pablo Montero (2016). “Adopting a Cleaner
Technology: The Effect of Driving Restrictions on Fleet Turnover”. In: Working Paper.

Bonilla, Jorge A (2016). “The More Stringent, the Better? Rationing Car Use in Bogot´a with
Moderate and Drastic Restrictions”. In: The World Bank Economic Review, lhw053–25.

Calonico, Sebastian, Matias D Cattaneo, and Roc´ıo Titiunik (2015a). “Optimal Data-Driven
Regression Discontinuity Plots”. In: Journal of the American Statistical Association
110.512, pp. 1753–1769.

Calonico, Sebastian, Matias D Cattaneo, and Rocio Titiunik (2015b). “Optimal Data-Driven
Regression Discontinuity Plots”. In: Journal of the American Statistical Association
110.512, pp. 1753–1769.

Cantillo, Victor and Juan De Dios Ort´uzar (2014). “Restricting the Use of Cars by License

Plate Numbers: A Misguided Urban Transport Policy”. In: DYNA 81.188, pp. 75–82.

Carrillo, Paul E, Arun S Malik, and Yiseon Yoo (2016). “Driving Restrictions that Work?
Quito’s Pico y Placa Program”. In: Canadian Journal of Economics/Revue Canadienne
d’ ´Economique 49.4, pp. 1536–1568.

Cattaneo, Matias D et al. (2016). “Interpreting Regression Discontinuity Designs with Mul-

tiple Cutoffs”. In: The Journal of Politics 78.4, pp. 1229–1248.

Chang, Tom et al. (2016). “Particulate Pollution and the Productivity of Pear Packers”. In:

American Economic Journal: Economic Policy 8.3, pp. 141–69.

Chay, Kenneth Y and Michael Greenstone (2003). “The Impact of Air Pollution on Infant
Mortality: Evidence from Geographic Variation in Pollution Shocks Induced by a Reces-
sion”. In: The Quarterly Journal of Economics 118.3, pp. 1121–1167.

Chen, Yuyu et al. (2013). “The Promise of Beijing: Evaluating the Impact of the 2008
Olympic Games on Air Quality”. In: Journal of Environmental Economics and Manage-
ment 66.3, pp. 424–443.

167

Crocker, Thomas D and Robert L Horst (1981). “Hours of Work, Labor Productivity, and
Environmental Conditions: A Case Study”. In: The Review of Economics and Statistics
63.3, p. 361.

Currie, Janet and Matthew Neidell (2005). “Air Pollution and Infant Health: What Can We
Learn from California’s Recent Experience?” In: The Quarterly Journal of Economics
120.3, pp. 1003–1030.

Currie, Janet et al. (2009). “Does Pollution Increase School Absences? ” In: Review of Eco-

nomics and Statistics 91.4, pp. 682–694.

Davis, Lucas W (2008). “The Effect of Driving Restrictions on Air Quality in Mexico City”.

In: The Journal of Political Economy 116.1, pp. 38–81.

Eskeland, G S and T Feyzioglu (1997). “Rationing can Backfire: the “Day without a Car”

in Mexico City”. In: The World Bank Economic Review, pp. 383–408.

Gallego, Francisco, Juan-Pablo Montero, and Christian Salas (2013). “The Effect of Trans-
port Policies on Car Use: Evidence from Latin American Cities”. In: Journal of Public
Economics 107, pp. 47–62.

Gelman, Andrew and Guido Imbens (2017). “Why High-Order Polynomials Should Not Be
Used in Regression Discontinuity Designs”. In: Journal of Business & Economic Statistics
66.4.

Gibson, Matthew and Maria Carnovale (2015). “The Effects of Road Pricing on Driver

Behavior and Air Pollution”. In: Journal of Urban Economics 89, pp. 62–73.

Grange, Louis de and Rodrigo Troncoso (2011). “Impacts of Vehicle Restrictions on Urban
Transport Flows. The Case of Santiago, Chile”. In: Transport Policy 18.6, pp. 862–869.

Greenstone, M and R Hanna (2014). “Environmental Regulations, Air and Water Pollution,
and Infant Mortality in India”. In: The American Economic Review 104.10, pp. 3038–
3072.

Greenstone, Michael and B Kelsey Jack (2015). “Envirodevonomics: A Research Agenda for

an Emerging Field”. In: Journal of Economic Literature 53.1, pp. 5–42.

Hanna, Rema and Paulina Oliva (2015). “The Effect of Pollution on Labor Supply: Evidence
from a Natural Experiment in Mexico City”. In: Journal of Public Economics 122, pp. 68–
79.

Hausman, Catherine and David S Rapson (2017). Regression Discontinuity in Time: Consid-
erations for Empirical Applications. Tech. rep. National Bureau of Economic Research.

Hausman, Jerry A, Bart D Ostro, and David A Wise (1984). Air Pollution and Lost Work.

National Bureau of Economic Research Cambridge.

168

Imbens, Guido W and Thomas Lemieux (2008). “Regression Discontinuity Designs: A Guide

to Practice”. In: Journal of Econometrics 142.2, pp. 615–635.

INE (2001). Parque de Veh´ıculos En Circulaci´on 2001. Santiago: INE.

— (2015). Parque de Veh´ıculos En Circulaci´on 2015. Santiago: INE.

Knittel, Christopher R, Douglas L Miller, and Nicholas J Sanders (2016). “Caution, Drivers!
Children Present: Traffic, Pollution, and Infant Health”. In: Review of Economics and
Statistics 98.2, pp. 350–366.

Lichter, Andreas, Nico Pestel, and Eric Sommer (2017). “Productivity Effects of Air Pollu-

tion: Evidence from Professional Soccer”. In: Labour Economics 48, pp. 54–66.

Morales, Ra´ul GE (2006). Contaminaci´on Atmosf´erica Urbana: Episodios cr´ıticos de Con-

taminaci´on Ambiental en la Ciudad de Santiago. Editorial Universitaria.

Mullins, Jamie and Prashant Bharadwaj (2015). “Effects of Short-Term Measures to Curb
Air Pollution: Evidence from Santiago, Chile”. In: American Journal of Agricultural Eco-
nomics 97.4, pp. 1107–1134.

Neidell, Matthew J (2004). “Air Pollution, Health, and Socio-Economic Status: The Effect
of Outdoor Air Quality on Childhood Asthma”. In: Journal of Health Economics 23.6,
pp. 1209–1236.

OECD (2005). OECD Environmental Performance Reviews: Chile 2005. OECD Publishing.

Ostro, Bart D (1983). “The Effects of Air Pollution on Work Loss and Morbidity”. In: Journal

of Environmental Economics and Management 10.4, pp. 371–382.

Perez, P (2008). “Air Quality Forecasting in a Large City”. In: WIT Transactions on Ecology

and the Environment 116.116, pp. 21–28.

P´erez, Patricio, Alex Trier, and Jorge Reyes (2000). “Prediction of PM2.5 Concentrations
Several Hours in Advance Using Neural Networks in Santiago, Chile”. In: Atmospheric
Environment 34.8, pp. 1189–1196.

Romero, H et al. (1999). “Rapid Urban Growth, Land-Use Changes and Air Pollution in

Santiago, Chile”. In: Atmospheric Environment 33.24-25, pp. 4039–4047.

Saide, Pablo E et al. (2011). “Forecasting Urban PM10 and PM2.5 Pollution Episodes in
Very Stable Nocturnal Conditions and Complex Terrain using WRF–Chem CO Tracer
Model”. In: Atmospheric Environment 45.16, pp. 2769–2780.

Salas, Christian H (2010). “Evaluating Public Policies with High Frequency Data: Evidence
for Driving Restrictions in Mexico City Revisited”. In: Documentos de Trabajo (Instituto
de Econom´ıa PUC) 374, p. 1.

169

Salini, G. A. (2009). “Desarrollo de un Modelo para Pronosticar Concentraciones Extremas

de PM2.5 en Santiago”. In: Universidad de Santiago de Chile, Santiago.

Stafford, Tess M (2015). “Indoor Air Quality and Academic Performance”. In: Journal of

Environmental Economics and Management 70, pp. 34–50.

Stock, James H, Jonathan H Wright, and Motohiro Yogo (2002). “A Survey of Weak In-
struments and Weak Identification in Generalized Method of Moments”. In: Journal of
Business & Economic Statistics 20.4, pp. 518–529.

Troncoso, Rodrigo, Louis de Grange, and Luis A Cifuentes (2012). “Effects of Environmental
Alerts and Pre-Emergencies on Pollutant Concentrations in Santiago, Chile”. In: Atmo-
spheric Environment 61, pp. 550–557.

United Nations Environment Programme (2002). Global Environment Outlook 3. Earthscan

Publications, Ltd.

Viard, V Brian and Shihe Fu (2015). “The Effect of Beijing’s Driving Restrictions on Pollu-

tion and Economic Activity”. In: Journal of Public Economics 125, pp. 98–115.

Wolff, H (2014). “Keep your Clunker in the Suburb: Low-Emissions Zones and Adoption of

Green Vehicles”. In: The Economic Journal 124.578, F481–F512.

World Health Organization (2006). World Health Organization Air Quality Guidelines. Global

Update 2005. World Health Organization.

— (2014). Ambient (Outdoor) Air Pollution in Cities Database. Data retrieved from the
World Health Organization Ambient Air Pollution in Cities Database 2014, http : / /
www.who.int/phe/health_topics/outdoorair/databases/cities-2014/en/.

Zhang, Wei, C Y Cynthia Lin Lawell, and Victoria I Umanskaya (2017). “The Effects of Li-
cense Plate-Based Driving Restrictions on Air Quality: Theory and Empirical Evidence”.
In: Journal of Environmental Economics and Management 82, pp. 181–220.

Zivin, Joshua Graff and Matthew Neidell (2012). “The Impact of Pollution on Worker Pro-

ductivity”. In: The American Economic Review 102.7, pp. 3652–3673.

170