ESSAYS ON GREEN ENERGY POLICIES IN THAILAND
By
Sukampon Chongwilaikasaem

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Economics – Doctor of Philosophy
2015

ABSTRACT
ESSAYS ON GREEN ENERGY POLICIES IN THAILAND
By
Sukampon Chongwilaikasaem
Rural electrification is a major tool that developing countries use to foster economic
development. However, grid-line electrification in remote rural areas is often excessively costly,
or faces environmental challenges. For these reasons, off-grid electrification, such as solar or
wind power, may be the only feasible alternative in many isolated rural areas. In Chapter 1, I
examine the impact of the solar power initiative for rural electrification on economic
development by studying the effect of the Solar Home System Project that Thailand’s
government implemented during the time period 2004-2006. I exploit variation in the year of the
solar installations across households to identify the program’s causal effects using householdlevel panel data, which allows me to control for household fixed effects, trends, and provinceyear effects. I find that solar units increase household income by 6.9%. This result is robust to
other identification strategies, such as lagged dependent variable, propensity score reweighting,
and differences-in-differences with propensity score matching. I also apply a back-of-theenvelope calculation to estimate the costs and benefits of the program. I find that the net present
value of this project is about 1 billion baht or $25 million.
Since the growing concern for climate change in the 1990s, policymakers around the
world have been enthusiastically supporting a wide range of incentive mechanisms for renewable
energy use, including in the electric power sector. In chapter 2, I use a technology-specific
subsidy program for small electricity producers in Thailand to understand the response of
producers to subsidy incentives. In 2009, the Thai government adjusted the subsidy policy to

favor a relatively small power plant. The production subsidy for electricity generators that have
less than 1MW of capacity is higher than the subsidy for generators that have just over 1MW of
capacity. Since the total subsidy is a step function with a “notch” at 1MW, power producers who
plan to build a plant near 1 MW will have an incentive to respond strategically by building a
plant on the side of the notch with a higher subsidy. I develop a structural model of bunching to
understand the linkage between the subsidy policy and power plant producers’ behavior. Then, I
use a test developed by McCrary (2008) to examine the distribution of power plant size for
electricity producers subject to the subsidy policy in 2009. I find evidence of bunching around
the notch. A falsification test shows that similar bunching does not exist in the distribution of
plant size built before the introduction of the notched subsidy payment.
In the final chapter, I examine the impact of the first car buyer tax rebate program in
Thailand on the share of eligible cars and eco-car sales. Province level data of monthly new
vehicle registrations by vehicle model during 2007-2012 are employed in the analysis in this
study. This panel data set allows me to control for vehicle model and province-time fixed effects
when examining the relationship between the program and share of vehicle sales. The estimation
results indicate an 18% increase in the share of eligible cars sales after the introduction of the
program in September 2011. However, the share of eco-car sales is unaffected by the program.

Go Green!

iv

ACKNOWLEDGMENTS

During my journey to complete this dissertation, I have received support and
encouragement from a great number of people. I would like to express the deepest
appreciation to my advisor, Soren Anderson, for his invaluable guidance, caring, patience,
support and encouragement. His comments broadened and deepened my view. His advice
guided me in the right direction and he perpetually encouraged me to improve my work.
He helped me immensely at every stage and was always available whenever I had
questions. I would also like to thank my other committee members: Todd Elder, Christian
Ahlin, and Songqing Jin for their useful comments and great interest in my work.
I am very grateful to the Royal Thai Government for providing me the scholarship
and the opportunity to pursue my doctoral degree. Many thanks to all my friends at MSU,
especially Nakarin Amares, Luke Chu, Seunghwa Rho and Yu Sun for their academic
and emotional support. I appreciate their invaluable friendship.
My heartfelt appreciation goes to my parents for their understanding,
encouragement and endurance during the years. Finally, thanks to Jane for her persistent
patience and love.

v

TABLE OF CONTENTS

LIST OF TABLES……………………………………………………………………...viii
LIST OF FIGURES……………………………………………………………………….x
CHAPTER 1 ........................................................................................................................1
Rural Electrification and Development: Evaluating Thailand’s Solar Power Initiative ......1
1. Introduction .................................................................................................................1
2. Thailand’s rural solar-electricity initiative ..................................................................4
3. Theoretical effect of solar electricity on household income .......................................8
4. Data availability ........................................................................................................11
5. Empirical strategies and results .................................................................................12
5.1. Summary statistics ..............................................................................................14
5.2. Differences in differences (DD) ..........................................................................19
5.2.1. Robustness tests ..........................................................................................21
5.2.2. Heterogeneity of the treatment effect .........................................................22
5.3. Propensity score reweighted regression .............................................................26
6. Sustainability of the solar home system ....................................................................29
7. Comparing costs and benefits of the program ...........................................................29
8. Conclusions ...............................................................................................................30
APPENDIX ...................................................................................................................32
REFERENCES ..............................................................................................................37
CHAPTER 2 ......................................................................................................................40
The Effects of Subsidies on Small-Scale Renewable Electricity Generation in Thailand
............................................................................................................................................40
1. Introduction ...............................................................................................................40
2. Thailand’s very small-scale electric power producers program and subsidies policy
........................................................................................................................................43
3. A structural model of bunching .................................................................................47
4. Data availability ........................................................................................................56
5. Reduced-form test for bunching ................................................................................56
5.1. Graphical evidence .............................................................................................57
5.2. A formal test for bunching .................................................................................60
6. Conclusions ...............................................................................................................65
APPENDIX ...................................................................................................................66
REFERENCES ..............................................................................................................71
CHAPTER 3 ......................................................................................................................73
The Effects of the First-car Buyer Excise Tax Rebate on Small Car Sales in Thailand ...73
1. Introduction ...............................................................................................................73
2. Eco-car and first-car buyer excise tax rebate program ..............................................75
2.1. Eco-car policy .....................................................................................................76
2.2. First-car tax rebate policy ...................................................................................77
vi

3. Data ...........................................................................................................................78
4. Empirical methodology and estimation results .........................................................82
4.1. Effect of first-car excise tax rebate policy on eco-car sales ...............................82
4.2. Empirical methodology ......................................................................................84
5. Conclusions ...............................................................................................................87
APPENDIX ...................................................................................................................89
REFERENCES ..............................................................................................................92

vii

LIST OF TABLES

Table 1.1: Summary Statistics ..........................................................................................17
Table 1.2: Test for pre-trend .............................................................................................18
Table 1.3: Baseline results ................................................................................................20
Table 1.4: Robustness test: Lagged dependent variable ...................................................22
Table 1.5: Heterogeneous treatment effects by quantile ...................................................24
Table 1.6: Heterogeneous treatment effects by region .....................................................25
Table 1.7: Propensity score reweighted regression ...........................................................27
Table 1.8: DD with propensity score matching ................................................................28
Table A.1: 2001-2005 data ...............................................................................................34
Table B.1: Estimation of propensity score ........................................................................36
Table 2.1: The subsidy rates for VSPPs program .............................................................45
Table 2.2: The distribution of VSPPs by energy source in June 2014 ..............................47
Table 2.3: Test for discontinuity in the estimated density of power plant size from the
small data set ..........................................................................................................62
Table 2.4: Tests for discontinuity in the estimated density of power plant size from the
big data set .............................................................................................................64
Table 3.1: List of eco-car models .....................................................................................77
Table 3.2: Incremental subsidy for eligible models ..........................................................79
Table 3.3: Vehicle price before and after the rebate in January 2012 ...............................83
Table 3.4: Baseline results ................................................................................................86
Table 3.5: Heterogeneous effects by exposure .................................................................88
Table C.1: The information of make, nameplate, and cc for all 23 models used in this
study ......................................................................................................................90

viii

Table D.1: Estimation of log (kpl) .....................................................................................91

ix

LIST OF FIGURES

Figure 1.1: The model of SHS in Thailand .........................................................................6
Figure 1.2: Map compares density and distribution of HH04, HH05, and HH06 at the
sub-district level ......................................................................................................7
Figure 1.3a: The effect of a solar electricity system on household income: Working-time
constraint is not binding .........................................................................................10
Figure 1.3b: The effect of a solar electricity system on household income: Working-time
constraint is binding ..............................................................................................10
Figure 1.4: Distribution of log income for treatment and control group: Comparison of
2003 vs. 2005 .........................................................................................................16
Figure 1.5: Comparing the trend of log real income between treatment and control groups
before the program ................................................................................................18
Figure A.1: Comparing pre-treatment trends of other variables between treatment and
control groups .......................................................................................................33
Figure 2.1: Total production subsidy as a function of capacity ........................................45
Figure 2.2: Number of VSPPs and total annual capacity (in MW) for VSPPs 2003-2014
................................................................................................................................46
Figure 2.3: The optimal size for the power plants ............................................................49
Figure 2.4: The unique εc ..................................................................................................54

Figure 2.5: Behavior of power producers in response to a notch in the subsidy ..............55
Figure 2.6: Evidence of bunching from the small data set (binsize = 0.02 MW)……..… 59
Figure 2.7: Evidence of bunching from the big data set (binsize = 0.02 MW) .................59
Figure 2.8: Test for bunching from the small data set ......................................................62
Figure 2.9: Test for bunching from the big data set ..........................................................63
Figure 2.10: Falsification test for bunching from the big data set ....................................64
Figure B.1: Evidence of bunching from the small data set (binsize = 0.01 MW) .............67

x

Figure C.1: Evidence of bunching from the small data set (binsize = 0.03 MW) .............68
Figure D.1: Evidence of bunching from the big data set (binsize = 0.01 MW).................69
Figure E.1: Evidence of bunching from the big data set (binsize = 0.03 MW) .................70
Figure 3.1: Eco-car purchases overtime.............................................................................77
Figure 3.2: Eligible and non-eligible car sales over time .................................................80
Figure 3.3: Share of eligible car sales to all 23 models, and share of eco car sales on
eligible models over time .......................................................................................81
Figure 3.4: The variation of gas price over time ...............................................................82

xi

CHAPTER 1
Rural Electrification and Development: Evaluating Thailand’s Solar Power Initiative
1.

Introduction
Electricity is widely considered to be essential to economic development, and areas

without electricity have far lower living standards (World Bank 2002). However, the 2009 world
energy outlook reports that the electrification rate is only 60.2% in rural areas of developing
countries. In response, many development programs focus on rural electrification. Electric
lighting can improve quality of life by facilitating household production and leisure activity in
the evening and by reducing time spent searching for firewood or other fuel sources. Lighting
can similarly improve educational outcomes by allowing children to study at night after school.
Electric lighting can improve health by reducing indoor pollution and burn injuries, especially
among children, from kerosene lamps and candles (Meier et al. 2010). In addition, when used to
power radios and televisions, electricity provides access to valuable entertainment and
information sources. Finally, electricity can be used to charge cell phone and computer batteries,
thereby empowering the rural poor by increasing access to knowledge and communication
technologies. Thus, many developing countries include rural electrification in their development
plans, hoping to break the cycle of poverty and spur economic growth.
Few studies have attempted to quantify the benefits of rural electrification (Bernard 2010),
and most of these studies have only investigated the effects of expanding the electrical grid to
new areas (Dinkelman 2011; Khandker, Barnes, Samad, and Minh 2009; Lipscomb, Mobarak
and Barham 2013). However, grid-line electrification in remote rural areas is often constrained
by difficult terrain, low population density, and distance to existing electric power plants. In
addition, new coal, natural gas, or nuclear power plants in rural areas can bring pollution

1

problems themselves, and many proposals to build new power plants are rejected by local
residents. For these reasons, off-grid electrification, such as solar or wind power, may be the
only feasible alternative in many isolated rural areas.
Several previous studies attempt to estimate the effects of rural electrification in
developing countries. Dinkelman (2011) finds that female employment increases in the
KwaZulu-Natal province of South Africa, while Lipscomb, Mobarak, and Barham (2013) finds
similar results in Brazil. Lighting allows women to perform household tasks in the evening and
work outside the home during the day. In addition, several studies find that electrification
decreases fertility (Peters and Vance 2011, Grogan and Sadanand 2009). Lighting extends
waking hours, meaning less time in bed, and radio and television provide an alternative to sex for
recreation (World Bank 2008). In the case of solar electricity, however, individual household
systems generate only a limited amount of electricity each day. Thus, previous empirical studies,
which focus on rural electrification projects that typically provide households with unconstrained
electricity access at a relatively low price, are not suitable for assessing the benefits of rural
electrification through solar electricity.
I address this important research need by estimating the causal effect of providing solar
electricity units to households in rural Thailand on household income. Between 2004 and 2006,
the government of Thailand donated approximately 200,000 solar electricity units to individual
rural households throughout the country in an unprecedented attempt to provide electricity to
households living in remote areas too costly or too difficult to connect to the existing grid. I
exploit the variation in the year of the solar unit installations across households to identify the
program’s causal effects using household-level panel data, which allows me to control flexibly
for household fixed effects, trends, and province-year effects.

2

Using a conventional difference-in-difference approach, I find that these solar units
increased average household incomes by 6.9%. The result is robust to other identification
strategies, such as lagged dependent variable, propensity score reweighting, and differences-indifferences with propensity score matching. This result is in line with previous findings in
literature, although the magnitudes of the impacts are somewhat smaller, which is consistent with
the fact that solar electric units generate a limited daily amount of electricity. A simple back of
the envelope calculation suggests that the net present value of the program is $1 billion baht or
$25 million, approximately $125 per treated household.
This paper contributes to the previous literatures by using a credible and transparent
difference-in-difference identification strategy. Most of previous literatures use instrumental
variables. For example, Dinkelman (2011) employs sloped terrain in south Africa as an
instrument for the electricity grid. Lipscomb, Mobarak, and Barham (2013) use simulated
engineering costs for hydroelectric dams in Brazil as an instrument for electricity access. In
contrast, my estimation is based on transparent difference-in-different strategy due to variation in
timing of treatment for eligible households.
The results of this paper also contribute to a recent policy debate about how best to
expand electricity access in rural areas: grid extensions versus off-grid solutions. As noted above,
extending the grid is often excessively costly, or faces environmental challenges. In contrast, offgrid electrification is easier to install, less costly, and generates less pollution. However, the
sustainability of off-grid electrification is also questionable, since many of the solar units in
Thailand’s program or the Soccket project 1 in other developing countries were found to be

1

Soccket is a soccer ball that is able to covert the kinetic energy from playing soccer to power a reading lamp at
night. It provides enough electricity power to an LED lamp for three hours after playing soccer with the ball for
thirty minutes.

3

damaged several years after implementation. Despite skepticism about sustainability, this study
shows a positive net present value from Thailand’s solar project. Therefore, off-grid
electrification may be a good solution, although its cost effectiveness and sustainability needs to
be verified case-by-case.
The rest of the paper is organized as follows. Section 2 discusses Thailand’s solar
electricity initiative. Section 3 presents a simple model to explain how solar electricity system
may alter income of the household. Section 4 describes the availability of data. Section 5
discusses my identification strategy, provides descriptive statistics for the key variables, presents
the econometric model, and presents my estimation results. Section 6 discusses the sustainability
of the program. Section 7 provides a back-of-the-envelope calculation of the program’s net
present value. Section 8 concludes with a summary of my major findings.

2.

Thailand’s rural solar-electricity initiative
In 2004-2005, the Thai government donated approximately 200,000 solar home systems

(SHSs) to individual rural households throughout Thailand. The recipients of these systems were
households that could not be connected to the electric power grid due to the high cost of
connection or because the households lived in nationally protected land. Every household who
did not already have electricity was eligible to receive a solar panel eventually.
By providing solar electricity, the government hoped to increase access to radio and
television news, thereby allowing rural households to participate more fully in the political and
administrative affairs of the country, with the ultimate goal being to provide electricity for every
household in the country by the year 2005. Each system provided by the Thai government
consists of one 120 Watt-Peak (maximum power under ideal conditions) solar module, an

4

inverter of at least 150 Watts, a 125 Ampere-hours battery, two 10-Watt fluorescent lamps, and a
set of electrical outlets. Each complete system costs approximately 25,000 Baht or about $800.
Figure 1.1 presents a schematic of the solar electricity systems provided by the Thai
government. Solar panels were installed on a separately constructed pole because the roof
structures of most houses in remote areas are not strong enough to support the system, and to
ensure better exposure to sunlight throughout the day. The solar panels are photovoltaic cells that
produce a direct current from the radiant energy of the sun that reaches the panel. The panel is
connected to a solar charger and controller, which regulates the voltage and current coming from
solar panels to a standard household electrical voltage (220 volt, a standard voltage for electrical
appliances in Thailand). Then, fluorescent lamps and a battery are connected to the system to
serve for a basic usage and as storage devices respectively. Batteries store electricity produced
during the day so that it can be used as needed at nighttime or during overcast weather. Batteries
also serve to power the solar array, so that it functions at a stable voltage, since the amount of
solar radiation being absorbed by the array varies throughout the day (Florida Solar Energy
Center, 2006).
These systems generate about 350 to 450 Watt-hours per day. Maximum continuous
power output is limited to about 150 Watts by the inverter’s capacity. Thus, the systems generate
electricity sufficient to provide, for example, several hours of lighting (using two 10-Watt light
bulbs) and about one hour of television per night.
The government’s goal was to provide electricity to the majority of Thailand’s 203,000
off-grid households. The program planned to install solar electricity systems in two phases over a
period of two years in 2004–2005. About 153,000 systems were installed in the first phase, while
about 50,000 systems were installed in the second phase.

5

Figure 1.1: The model of SHS in Thailand
Source: Kruanpradit et al. (2004)

The actual installation was delayed and some installations were finished later than
initially planned; i.e. installations for some households in Phase 1 were completed at the end of
2005, while installations for some households in Phase 2 were finished at the end of 2006.
Therefore, households that obtained the solar electricity system can be categorized into 3 groups
based on installation date;
1.

Households that received systems in 2004 (HH04),

2.

Households that received systems in 2005 (HH05), and

3.

Households that received systems in 2006 (HH06).

Figure 1.2 compares the geographic distribution of HH04, HH05, and HH06.

6

Province border
HH04
No. of households

HH05
No. of households

HH06
No. of households

Figure 1.2: Map compares density and distribution of HH04, HH05, and HH06 at the subdistrict level
Note: The shaded area is the central region for which the installation year is unavailable.

7

3.

Theoretical effect of solar electricity on household income
This section introduces a simple model to explain how solar electricity system may alter

the income of the household. For simplicity, in this model, the benefits of solar units were
limited to the provision of lighting only. I do not include the benefits from using the solar
electricity system to power radio and television, or to charge a cell phone battery, which the
household could use to learn about new technology, ideas, commodity prices in a distant market,
or the weather forecast.
Suppose household i’s welfare is represented by the following utility function:
𝑢𝑢𝑖𝑖 (𝑙𝑙𝑖𝑖 , 𝑌𝑌𝑖𝑖 ) = 𝑣𝑣𝑖𝑖 (𝑙𝑙𝑖𝑖 ) + 𝑌𝑌𝑖𝑖 ,

(1)

where 𝑣𝑣𝑖𝑖 (𝑙𝑙𝑖𝑖 ) is utility from leisure, 𝑙𝑙𝑖𝑖 , with 𝑣𝑣𝑖𝑖′ > 0 and 𝑣𝑣𝑖𝑖′′ < 0, and 𝑌𝑌𝑖𝑖 is the household’s income.
The household’s budget constraint is

𝑌𝑌𝑖𝑖 = 𝑤𝑤𝑖𝑖 𝐿𝐿𝑖𝑖 , (2)

where 𝐿𝐿𝑖𝑖 is hours worked and 𝑤𝑤𝑖𝑖 is wage. The household’s time constraints can be written as
24 = 𝐿𝐿𝑖𝑖 + 𝑙𝑙𝑖𝑖 ,

(3)

and 𝐿𝐿𝑖𝑖 ≤ 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 or 𝑙𝑙𝑖𝑖 ≥ 24 − 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 ,

(4)

Equation (4) shows the working-time constraint due to the limitation of daylight when
there is no electricity in the household, assuming 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 hours of light per day. For instance, a
household that weaves or produces handicrafts for a living can work only under the sunlight.
Thus, the maximum hours of work per day for this household is 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 .

8

If the household allocates time between work and leisure to maximize utility, under the
Kuhn-Tucker conditions, there are two possible maxima. If 𝑙𝑙𝑖𝑖∗ ≥ 24 − 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 , then 𝑣𝑣𝑖𝑖′ (𝑙𝑙𝑖𝑖∗ ) = 𝑤𝑤𝑖𝑖 .
The other possible maximum is the corner solution in which 𝑙𝑙𝑖𝑖∗ = 24 − 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 .

After the household receives the solar home system, the working-time constraint changes
to
𝐿𝐿 ≤ 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 + 𝑇𝑇𝑠𝑠ℎ𝑠𝑠 or ≥ 24 − 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 − 𝑇𝑇𝑠𝑠ℎ𝑠𝑠 ,

(5)

where 𝑇𝑇𝑠𝑠ℎ𝑠𝑠 is hours of lighting from light bulbs using electricity generated by the solar panel.
With lights during the night, the length of the effective day is increased. It allows the households
2

to prepare or fix their equipment and to make a plan for tomorrow’s job . This change will not
alter the solution of the household whose working time constraint is not binding before getting
solar panel. However, the household in corner solution before getting solar panel will decrease
their optimal leisure to satisfied 𝑣𝑣𝑖𝑖′ (𝑙𝑙𝑖𝑖∗ ) = 𝑤𝑤𝑖𝑖 condition or choose 𝑙𝑙𝑖𝑖∗ = 24 − 𝑇𝑇𝑑𝑑𝑑𝑑𝑑𝑑 − 𝑇𝑇𝑠𝑠ℎ𝑠𝑠 .
Graphically, Figure 1.3a represents the tangency condition between an indifference curve
and the household budget constraint when the working-time constraint is not binding. In this
case, after the household receives the solar panel, the income of household will not change
(staying at 𝑌𝑌 0 ). Figure 1.3b illustrates the case when the working-time constraint is binding
before the household receives the solar panel. The solar home system changes the budget

2

Faculty of Engineering and Architecture (2007) questioned sample households who received the solar unit (3,350
households all over the country) about the benefits of the program. They found that households use the lights at
night to increase their household income by separating paddy, making handicrafts, preserving fruit, fertilizing,
collecting insects, and repairing their farm or fishing equipment.

9

constraint by increasing possible working time. Thus, the income of household will increase
from 𝑌𝑌1 to 𝑌𝑌 2 .

Figure 1.3a: The effect of a solar electricity system on household income:
Working-time constraint is not binding

Figure 1.3b: The effect of a solar electricity system on household income:
Working-time constraint is binding
10

4.

Data availability
The data employed in this study come from two main sources: a list of the households

that received a Solar Home System (SHS) from the Provincial Electricity Authority (PEA), and
the socioeconomics data in the Basic Minimum Need (BMN) survey. PEA information consists
of the name and address of households that acquired SHS (202,998 households), installation date,
name of the solar system installer, and the hiring contract number. To compare the living
standards of households before and after receiving a solar electricity system, this study matches
household and installation date in the data from PEA with their socioeconomics data in the BMN
survey.
The BMN survey measures the life quality of household members in different dimensions
during a specific period. The BMN data are collected at the household level every year in every
village all over the country (approximately 8 million households per year). Information covers
every household member who lived in the house for at least 6 months of the last year.
Information is generally collected in January to March about living standards during the previous
year. BMN data in this study cover the period of 2001-2005. Information in BMN consists of
household income, the number of pregnant women, the number and ages of children, child
education, and various demographic variables and other measures of well-being.
Household addresses were used to link households across BMN surveys to generate
household-level panel data on well-being for 2001-2005. These data were then matched to SHS
installation dates by household address. Names were ignored in the matching process due to
errors or incomplete information, such as misspelling and other typos, repeated names, and
changes in household members interviewed for BMN. In the matching process, data for 43,944
households was found in at least one year.

11

There are several factors that influence the matching. First, all information in the central
region from PEA is missing because it does not contain village name or installation date. These
missing data account for about 8.44% of the total solar installations. Second, some solar units
were installed in a public area, such as a religious site, temple, or school, which has no BMN
data. Finally, the data contain some errors, such as strange letters in the village identification
number, misspelling in a village name, missing village name, strange letters in the home number,
or a home number format that was inadvertently set to a date-month format in Excel in a way
that makes it impossible to recover the original house number. Most of the errors in the data
come from the administrative PEA data, since PEA is not proficient in data collection and
management.
Data on annual rainfall (millimeters) in each province was obtained from the National
Statistics Office of Thailand. However, the information of some provinces is missing since there
is no weather station in those areas. To cope with this problem, the missing data were estimated
using the average annual rainfall of surrounding provinces.

5.

Empirical strategies and results
To avoid selection bias problems due to the policy’s non-random implementation, this

study employs difference-in-differences techniques. I compare the before (2003) and after (2005)
income of households that received solar panels in 2004 (HH04) and in 2006 (HH06).
Since the program was started in 2004, data in 2003 were chosen to reflect the status of
households prior to when the program was implemented. There was no household that had a
solar unit installed in this year. This study uses 2005 as an after program year. The reason that
data in 2004 are not employed for this task is because the solar unit installations in 2004 are

12

distributed throughout the year. The income in 2004 of HH04 incompletely reflects the impact
of the solar unit. If a solar unit was installed at the beginning of the year, the income would
reflect the impact of the program. However, if households received their unit late in the year,
annual income of that year will not reflect the impact of the solar unit. Besides the obvious fact
that households without a solar unit cannot generate electricity, upon receiving their unit,
households may adapt slowly or may delay in acquiring electronic devices, such as a radio, TV,
or cell phone, that would increase their productivity. Moreover, households may only learn
gradually about the benefits of electricity-using devices, such as using a cell phone to sell
agricultural products in a distant market for a higher price than in the local market. In addition,
household production may only adjust slowly in response to lighting, perhaps because it takes
time to reschedule household activities from the daytime to nighttime.
The HH04 and HH06 were selected as a treatment and control group, respectively. This
study does not include HH05 in either the control or treatment group. The HH05 should not be
included in the control group, since HH05 already received solar unit in 2005. I do not include
HH05 in the treatment group because some households in HH05 had received the solar unit at
the end of the year. As discussed above, there was an inadequate amount of time for them to
3

learn to use the solar panel . Since I used HH06, the group of households who are also eligible in
the program but still did not get the installation in 2005, as a control group, the treatment and
control households should share similar preprogram characteristics.
Another potential concern is the effect of other policies. If the timing of implementation
of the solar units was correlated with timing of other programs, it would be hard to identify the
impact of electrification program. Although there were many populist policies launched in
3

Table A.1 in appendix shows the results of the estimation of equation (6) in section 5.2 using the full set of data
during 2001-2005, including HH05.

13

Thailand in the past decade, there was none starting in 2004. The “one million baht one village”
4

5

fund program , “one tambon one product” program (OTOP) , and micro-credit program were all
6

started in 2001, while the “thirty-baht health care” program was started in 2002. Therefore, the
effect of solar unit should not be contaminated by these other programs.

5.1

Summary statistics
In order to do the empirical analysis, I use data of 14,255 households for which I have

household income data in both 2003 and 2005. Table 1.1 presents the summary statistics for
several key variables in the data set, classified by installation year.

Panel A shows the

geographic distribution of households who received a solar panel. Most of HH04 (71%) are in
north region, while most of HH06 are distributed in the north and northeast regions (40% and 49%
respectively).
In 2003 (Panel B), average logged real income was 11.03 for HH04, compared with an
average of 11.22 for HH06. Annual rainfall (Rain), number of household members age 18-60
year old members (Age 18-60), and number of member older than 60 years (Elder) are higher for
HH06 than in HH04. On the other hand, the number of household members (HHmem), number
of children age 3-5 years (Age3-5), number of children age 6-11 years (Age6-11), and number of
children age 12-14 years (Age12-14) are lower in HH06 than in HH04. The number of disabled
in the household (Disabled) is not significantly different between HH04 and HH06. In 2005
(Panel C), average logged real income was 11.41 for HH04, compared with an average of 11.44
for HH06. The Rain and Elder variables are higher in HH06 than in HH04. On the other hand,
4
5
6

This program was allocated roughly 1 million baht (around $33,000) to each village.
This program stimulates the development of small and medium-size rural enterprises at the sub-district level.
This program limits a hospital’s charge to 30 baht (approximately $1).

14

HHmem, Age3-5, Age6-11, and Age12-14 variables are lower in HH06 than in HH04. Age 1860 and Disable variables are not significantly different between HH04 and HH06.
The average logged real income of treated households is lower than non-treated
households in both 2003 and 2005. In both type of households, average logged real income
increased over time. However, it increased more in HH04 relative to their HH06 counterparts
after the solar panels were installed. The relative gain (the “difference in differences”) of the
changes in logged real income is 0.147, implying an approximately 15% relative increase for
HH04.
Further insight into this change is provided in Figure 1.4, which shows the distribution of
logged income for the treated and non-treated households before and after the program. For the
treatment group (HH04), there was a notable rightward shift in the distribution from 2003 to
2005. On the other hand, the density looks similar for the control group (HH06) between 2003
and 2005.
Figure 1.5 compares the trend of logged real income for treatment households (HH04)
and non-treated households (HH06) before the first phase of the program arrived in 2004. When
comparing 2001 and 2003, the trend of HH04 vs. HH06 looks similar for logged real income.
Trends of other variables are also compared and shown in Figure A.1 in the appendix to this
chapter.
I test formally for a difference in pre-treatment trends in logged real income conditional
on controls by estimating the following equation:
𝑌𝑌𝑖𝑖𝑖𝑖 = 𝛼𝛼0 + 𝛼𝛼1 𝑇𝑇𝑖𝑖𝑖𝑖+2 + 𝑋𝑋𝑖𝑖𝑖𝑖′ 𝛼𝛼 + 𝛿𝛿𝑡𝑡 + 𝜇𝜇𝑖𝑖 + 𝜀𝜀𝑖𝑖𝑖𝑖 ,

15

where t is 2001 and 2003, Y it is logged real income for household i in year t, 𝑇𝑇𝑖𝑖𝑖𝑖+2 is the lead of

the treatment variable, X it is a vector of covariates that I will discuss in more detail in the next

1

section, µi is a household fixed effect, and δ t is a year effect.

Density

.6

.8

Treatment group (HH04)

.4

2005

0

.2

2003

10

11
logged income

12

13

12

13

.8

1

9

.4

Density

.6

Control group (HH06)

.2

2003

0

2005
9

10

11
logged income

Figure 1.4: Distribution of log income for treatment and control group:
Comparison of 2003 vs. 2005

16

Table 1.1: Summary Statistics
Variables

HH04

HH06

70.84
7.76
21.39

40.19
48.59
11.21

11.03
(0.008)
1088.84
(2.289)
4.31
(0.022)
2.31
(0.014)
0.32
(0.007)
0.02
(0.002)
0.26
(0.006)
0.56
(0.009)
0.29
(0.006)

11.22
(0.011)
1270.13
(4.731)
4.23
(0.023)
2.42
(0.017)
0.40
(0.009)
0.02
(0.002)
0.19
(0.006)
0.45
(0.009)
0.24
(0.006)

11.41
(0.006)
1249.41
(2.690)
4.26
(0.022)
2.22
(0.015)
0.34
(0.007)
0.02
(0.002)
0.23
(0.005)
0.54
(0.009)
0.29
(0.006)

11.45
(0.008)
1476.03
(4.935)
4.12
(0.023)
2.21
(0.019)
0.42
(0.009)
0.02
(0.003)
0.17
(0.006)
0.41
(0.009)
0.24
(0.006)

A. Geographic distribution of household
(percentages):
North
Northeast
South
B. Means in 2003:
log real income
Rain (millimeter)
HHmem
Age 18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14
C. Means in 2005:
log real income
Rain (millimeter)
HHmem
Age 18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14

Note: Standard deviations are presented in parentheses.

17

11.4
11.2
11
10.8

HH04

10.6

HH06

10.4
10.2
10

2001

2003

Figure 1.5: Comparing the trend of log real income between treatment and control groups
before the program
Table 1.2 shows the estimates from the above equation with and without controlling for
the vector of covariates. The coefficients on the leads of the treat variable are insignificant in
both specifications. Thus, there is no statistically significant difference in the trend of logged
income for treatment and control households before the program, conditional on controls.
Table 1.2: Test for pre-trend

VARIABLES
𝑇𝑇𝑖𝑖𝑖𝑖+2

Constant
Controls
Year Effect
Household Effect

(1)
log real inc

(2)
log real inc

-0.100
(0.099)
10.509***
(0.032)
No
Yes
Yes

-0.058
(0.111)
10.409***
(0.489)
Yes
Yes
Yes

23,700
0.115

20,109
0.128

Observations
R-squared
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

18

5.2

Differences in differences (DD)
The difference-in-differences estimate can be calculated within a regression framework.

The estimating equation is specified as follows:
Yit = β 0 + β1Tit + X ' it α + µi + δ t + Pit + ∈it , (6)

where Y it is an outcome variable of interest (for example, logged real income) for household i in
year t, T it is an indicator variable for whether a household has received a solar unit by time
period t, X it is a vector of time-varying household controls, µi is a household effect that is
constant over time, δ t is a year effect that is constant across households, and 𝑃𝑃𝑖𝑖𝑖𝑖 is the provinceyear effects. Province-year effects control for unobserved differences between provinces over
time, such as natural disasters or recessions. Moreover, it might be the case that the solar panel
installation time is correlated by province. Excluding province-year effects could therefore
introduce a selection bias.
The vector of covariates (X it ) controls for predetermined factors affecting household
outcomes that vary over time. Covariates include annual rainfall (millimeter), number of
household members, number of members age 18-60; number of elder members (older than 60
years), number of disabled in the household; number of children age 3-5, number of children age
6-11, and number of children age 12-14. I do not include number of infant children because this
variable is more plausibly an endogenous outcome of the treatment than the above controls.
Table 1.3 shows the estimates from equation (6) using the log of real household income
as the outcome variable.

The table provides estimated coefficients for the electrification

indicators, control variables, time trends, and province-year dummy variables. I cluster standard
errors at the province level to account for annual shocks to household income that are correlated
across households in the same province, such as weather patterns that affect agricultural
19

productivity. Thus, standard errors (in parentheses) are fully robust to heteroskedasticity, as well
as serial correlation and spatial correlation both within and across households living in the same
province.
Table 1.3: Baseline results

VARIABLES
T

(1)
log real inc

(2)
log real inc

(3)
log real inc

(4)
log real inc

0.147*
(0.0761)

0.064***
(0.021)

11.110***
(0.0362)
Yes
Yes

0.130*
(0.0711)
0.001***
(0.0004)
-2.88e-07***
(7.84e-08)
0.157***
(0.0086)
0.019
(0.0124)
-0.009
(0.0149)
-0.032
(0.0296)
-0.018
(0.0114)
-0.047**
(0.0179)
-0.022
(0.0202)
9.448***
(0.3730)
Yes
Yes

11.300***
(0.0030)
Yes
Yes

0.069**
(0.0307)
-0.001***
(3.50e-05)
8.60e-08***
(3.67e-09)
0.158***
(0.0083)
0.013
(0.0097)
-0.010
(0.0131)
-0.028
(0.0288)
-0.021**
(0.0077)
-0.040**
(0.0186)
-0.019
(0.0209)
11.290***
(0.0669)
Yes
Yes

No

No

Yes

Yes

28,494
0.166

26,846
0.257

28,494
0.142

26,846
0.343

rain
rainsq
HHmem
Age18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14
Constant
Year Effect
Household Effect
Province-year
Effect
Observations
R-squared

Note: Robust standard errors in parentheses, clustered at province level.
*** p<0.01, ** p<0.05, * p<0.1

20

Column (1) shows the estimate from equation (6) without controlling for any covariates.
This estimation is directly comparable to the simple difference-in-differences of logged income
change in section 5.1. Column (2) adds eight control variables. Column (3) adds province-year
effects without control variables. Column (4) adds control variables and the province-year effects.
The coefficient of 0.069 on the solar electricity dummy in column (4) implies that solar
electricity increases real household income by approximately 6.9%. This result aligns with the
work of Khandker et al. (2009), which investigates the impacts of the World Bank financed
Rural Electrification project (grid electrification) in Vietnam on household welfare using panel
survey fielded in 2002 and 2005. They found that electrification increases household income by
36%. The magnitude of the impact in this study is smaller, however, which is consistent with the
fact that solar electric units generate a limited daily amount of electricity.

5.2.1

Robustness tests
I check the robustness of the results in Table 1.3 using the lagged dependent variable

model suggested by Angrist and Pischke (2009). Also the lagged dependent variable model
7

accounts for the possibility of the preprogram dip problem . Table 1.4 shows the estimate from
the following equation:
Yit = β 0 + β1TTit + X ' it α + γYit −2 + Pi + ∈it , (7)

where t is 2005, 𝑇𝑇𝑇𝑇𝑖𝑖𝑖𝑖 is an indicator for whether a household is treatment or control. 𝑌𝑌𝑖𝑖𝑖𝑖−2 is the
lag of logged real income (logged real income in 2003), and 𝑃𝑃𝑖𝑖 is the province effect.

From Table 1.4, the coefficient on the policy dummy variable is 0.062, which is about the

same as the estimation of the DD model in column (4) of Table 1.3.

7

See Ashenfelter and Card (1985).

21

Table 1.4: Robustness test: Lagged dependent variable
VARIABLES

log real inc

TT

0.062**
(0.0278)
-3.55e-05
(2.60e-05)
4.80e-09
(5.14e-09)
0.151***
(0.0110)
0.011
(0.0144)
-0.053***
(0.0104)
-0.075***
(0.0208)
-0.039**
(0.0179)
-0.029**
(0.0129)
-0.024*
(0.0130)
0.242***
(0.0306)
8.169***
(0.3580)
Yes

rain
rainsq
HHmem
Age18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14
Lagged log real inc
Constant
Province Effect
Observations
R-squared

13,685
0.4040

Note: Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

5.2.2

Heterogeneity of the treatment effect
It is interesting to know whether the solar home system is similarly effective for the

households in different quantiles of income, or in different regions. To examine the quantile
heterogeneous effect, I create 4 sub-group samples by income in 2003 from the DD-data, and
then estimate equation (6) separately for each group. I find that there is no significant impact on
22

income for households in the bottom or top income quartiles, although the point estimate is large
and positive for high income group. I find positively significant effects on middle-income
households (see Table 1.5).
One possible way to explain these results is through the model in section 3. It might be
the case that, before the policy, treated household in the lowest and highest income sub-groups
are at an interior solution, while middle-income households are at a corner solution. If
households relatively preferred leisure to income, they are probably not going to use electricity
from a solar unit for working during nighttime. Then, their household income would not be
affected.
To allow for heterogeneous treatment effects by region, I interact the policy variable (T)
in equation (6) with regional dummy variables. In particular, the estimating equation is:
Yit = β 0 + β1Tit + β 2Tit ⋅ N i + β 3Tit ⋅ NEi + X 'it α + µi + δ t + Pit + ∈it ,

(8)

where N i and NE i are dummy variables indicating whether household i lives in the north or
northeast region, respectively.
Table 1.6 provides the effect of solar units on logged real income from equation (8). The
effect on household income in the north is about 10.4% higher than in the south. However, the
effect in the northeast is not statistically different than in the south. These results are consistent
with the fact that rural households in the north region weave or make bamboo basketry
handicrafts for a living, which with lights, they can work in the night. On the other hand, rural
households in the south and northeast regions mostly work in the fishing industry, on rubber
plantation, or on farms, for which the small amount of extra light is not particularly helpful. Thus,
households in the north gain more benefits from the change in the budget constraint due to the
solar units than households in other regions.

23

Table 1.5: Heterogeneous treatment effects by quantile

Percentile
VARIABLES
T
rain
rainsq
HHmem
Age18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14
Constant
Year Effect
Household Effect
Province-year
Effect
Observations
R-squared

(1)
less than 25
log real inc

(2)
[25-50)
log real inc

(3)
[50-75)
log real inc

(4)
more than 75
log real inc

-0.034
(0.0483)
-0.002***
(7.30e-05)
5.32e-07***
(2.25e-08)
0.122***
(0.0217)
0.045***
(0.0121)
0.028
(0.0235)
0.003
(0.0644)
-0.018
(0.0280)
-0.050***
(0.0180)
-0.010
(0.0295)
11.440***
(0.0601)
Yes
Yes

0.119*
(0.0692)
-0.0003***
(0.000107)
4.53e-08***
(1.28e-08)
0.117***
(0.0108)
-0.005
(0.0155)
-0.009
(0.0326)
-0.040
(0.0380)
-0.041*
(0.0203)
-0.043***
(0.0152)
-0.039**
(0.0191)
10.820***
(0.126)
Yes
Yes

0.054*
(0.0299)
-0.0007***
(1.55e-05)
1.33e-07***
(2.35e-09)
0.128***
(0.0181)
0.007
(0.0096)
-0.018
(0.0280)
-0.007
(0.0293)
-0.011
(0.0268)
-0.038**
(0.0169)
-0.023
(0.0185)
11.380***
(0.0792)
Yes
Yes

0.108
(0.0703)
0.0003***
(2.31e-05)
-4.42e-08***
(3.73e-09)
0.117***
(0.0129)
0.004
(0.0081)
0.009
(0.0258)
-0.012
(0.0433)
-0.029*
(0.0152)
-0.028
(0.0173)
-0.037*
(0.0189)
10.780***
(0.0636)
Yes
Yes

Yes

Yes

Yes

Yes

6,770
0.6960

6,494
0.3800

6,801
0.2130

6,781
0.2410

Robust standard errors in parentheses, clustered at province level
*** p<0.01, ** p<0.05, * p<0.1

24

Table 1.6: Heterogeneous treatment effects by region
VARIABLES

log real inc

T

-0.001
(0.0166)
0.105***
(0.0375)
-0.028
(0.0405)
-0.0005***
(5.23e-05)
7.10e-08***
(6.11e-09)
0.157***
(0.0082)
0.013
(0.0096)
-0.009
(0.0130)
-0.028
(0.0290)
-0.020**
(0.0078)
-0.039**
(0.0184)
-0.019
(0.0208)
11.080***
(0.0879)
Yes
Yes
Yes

T*N
T*NE
Rain
rainsq
HHmem
Age18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14
Constant
Year Effect
Household Effect
Province-year Effect
Observations
R-squared

26,846
0.3430

Note: Robust standard errors in parentheses,
clustered at province level
*** p<0.01, ** p<0.05, * p<0.1

25

5.3

Propensity score reweighted regression
To balance the observable covariates between treatment and control groups, I use the

propensity score reweighted regression. In particular, the estimating equation is the following
regression:
∆𝑌𝑌𝑖𝑖 = 𝛽𝛽0 + 𝛽𝛽1 ∆𝑇𝑇𝑖𝑖 + ∆𝑋𝑋′𝑖𝑖 𝛼𝛼 + 𝜖𝜖𝑖𝑖 , (9)

with weights of 1/𝑝𝑝̂ (𝑥𝑥) for treated households and 1/(1 − 𝑝𝑝̂ (𝑥𝑥)) for controls, where 𝑝𝑝̂ (𝑥𝑥) is

an estimator of propensity score. ∆𝑌𝑌𝑖𝑖 , ∆𝑇𝑇𝑖𝑖 , and ∆𝑋𝑋𝑖𝑖 are the differences of logged real income, the
policy variable, and the control variables between 2005 and 2003 of household i.

The validity of using propensity score in impact evaluation analysis depends on the
following two assumptions: unconfoundedness and overlap. Although there is no formal test for
these two assumptions, the balancing test is always used as a diagnostic tool for the propensity
score specification. To estimate propensity score, I used the data in 2003 to run the logit model of
treatment variable on exogenous variables. The result is shown in Table B.1 in the appendix to
this chapter. Note that I adjust the model specification in the logit model until the balance test
under the pscore command in the Stata is satisfied. Table 1.7 shows that the solar unit increases
household income by 4.08%
As a further robustness check, I estimate a DD with propensity score matching model.
First, I matched household in 2003 using propensity score from the model in Table B.1. After
that I merge them with DD data to keep only the matched households in the panel sample. Then,
I use this new data set to estimate equation (6). I find that the solar units statistically increase the
household income by 6.94% (see Table 1.8).

26

Table 1.7: Propensity score reweighted regression
∆log real inc

VARIABLES
∆T

0.041***
(0.0138)
0.001
(0.0008)
-9.65e-08
(3.73e-07)
0.163***
(0.0129)
0.032**
(0.0130)
-0.041
(0.0367)
-0.050
(0.0411)
-0.013
(0.0136)
-0.064**
(0.0259)
-0.051**
(0.0244)
0.113***
(0.0341)

∆rain

∆rainsq

∆HHmem

∆Age18-60
∆Elder

∆Disabled
∆Age3-5

∆Age6-11

∆Age12-14
Constant

Observations
R-squared

10,666
0.111

Note: Robust standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

27

Table 1.8: DD with propensity score matching
VARIABLES

log real inc

T

0.069**
(0.0309)
-0.0001***
(2.72e-06)
-5.37e-08***
(9.24e-10)
0.157***
(0.0100)
0.013
(0.0117)
-0.010
(0.0149)
-0.041
(0.0300)
-0.020**
(0.0086)
-0.044*
(0.022)
-0.019
(0.0239)
10.840***
(0.0403)

Rain
Rainsq
HHmem
Age18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14

Constant
Year Effect
Household Effect
Province-year Effect

Yes
Yes
Yes

Observations
R-squared

21,548
0.355

Note: Robust standard errors in parentheses, clustered at province level
*** p<0.01, ** p<0.05, * p<0.1

28

6.

Sustainability of the solar home system
Even though the solar home system gives a positive impact on household income, many

of the systems malfunctioned several years after installation. In summer 2011, I had an in-depth
interview with the PEA officer who was in charge of this program and a member of a nonprofit
organization that volunteers to repair the systems in Tak province. They stated that malfunction
is the most important problem of the program, and more than 80% of the systems have already
failed.
In 2007, the Faculty of Engineering and Architecture of Rajamangala University of
Technology in Thailand asked 3,350 treated households in 24 provinces all over the country
about the experiences and satisfaction level from using the solar home system. Faculty of
Engineering and Architecture (2007) found that 90% of the sample does not know how to repair
the systems properly when the system is broken. Moreover, 74% of the treated households have
never been educated about maintenance. Based on the data of this survey, I found that 3.89% of
the treated households had a problem with the battery, solar panel, or inverter during the first
year after they received the system. The failure rate for systems surviving past the first year
increases to 6.34% in the second year, and the failure rate for systems surviving past the second
year is 10.54% in the third year. Thus, on average, the failure rate is about 7% per year.

7.

Comparing costs and benefits of the program
The main result of this study indicates that solar unit increases the household income by

6.9%. In this section, I conduct a back-of-the-envelope calculation to compare the costs and
8

benefits of the program.

8

In this study, I measure the benefits of the program by measuring the increase in household income only.

29

From the data, the mean of the control household (HH06) annual income equals
93,901.35 baht in 2005 and my preferred estimates imply a 6.9 percent increase in income. Thus,
the solar electricity system increased household income by about 6,479.19 baht
(93,901.35*0.069). Since the Thai government provided approximately 200,000 solar electricity
9

systems, the total benefits of the program were 1.295 million baht per year . Assume that this
total benefit is steady over time, that the probability that the solar unit breaks in a given year is
constant at 0.07, and that all the solar units will be broken after 2010 (a conservatively low
estimate of lifetime). Then, the present value (in 2005) of the total benefits is
PV = ∑𝑡𝑡=5
𝑡𝑡=0

1,295,838,000(1−0.07)𝑡𝑡
(1+𝑟𝑟𝑡𝑡 )𝑡𝑡

where 𝑟𝑟𝑡𝑡 is the real interest rate in time t. I use the annual average of real overnight interbank rate

as a proxy. The data of real interbank rates from 2006 to 2010 are from Bank of Thailand. Thus,
the present value of the total benefits is 6 billion baht. The cost of each solar electricity systems
is approximately 25,000 baht. Thus, the total costs of the program is about 5 billion baht. Thus,
10

the net present value of the program is about 1 billion baht or $25 million . Note that this
calculation does not include the installation costs and maintenance costs of the systems.
Moreover, I also ignore the tradeoff between income and leisure. Nevertheless, this result
indicates that the effects of the program on household income are meaningful.

8.

Conclusions
This paper examines the impact of solar power initiatives for rural electrification on

economic development by studying the effect of the Solar Home System project that Thailand’s
9

Assuming that all 200,000 solar units were installed in 2005.

10

The exchange rate is 40.22 baht/$ in 2005.

30

government implemented during 2004-2006. I exploit variation in the year of the solar
installations across households to identify the program’s causal effects using household-level
panel data, which allows me to control for household fixed effects, trends, and province-year
effects.
Despite the fact that the solar home system generates a small service of electricity, this
research found significant impacts on income from the solar units. Solar units increased average
household incomes by 6.9%. The result is robust to other identification strategies such as lagged
dependent variable, propensity score reweighting, and differences-in-differences with propensity
score matching. Even with skepticism about the sustainability of the program, a simple back-ofthe-envelope calculation suggests that the net present value of the program (if all the value is
income) is about 1 billion baht. Thus, it is not an exaggeration to conclude that rural
electrification brings development and improved living standards to rural areas.

31

APPENDIX

32

Figure A.1: Comparing pre-treatment trends of other variables between treatment and
control groups

33

Table A.1: 2001-2005 data

VARIABLES
T
𝑇𝑇 ′

Rain
rainsq
HHmem
Age18-60
Elder
Disabled
Age3-5
Age6-11
Age12-14
Constant
Year Effect
Household Effect
Province-year Effect
Observations
R-squared

(1)
log real inc

(2)
log real inc

0.083
(0.0514)

0.010
(0.0359)

(3)
log real inc

(4)
log real inc

0.017
(0.0332)
0.0001***
(2.38e-06)
-4.63e-08***
(3.69e-10)
0.111***
(0.0141)
0.053***
(0.0137)
0.009
(0.0200)
-0.054*
(0.0278)
-0.012
(0.0136)
-0.024*
(0.0124)
-0.012
(0.0118)
9.972***
(0.0238)
Yes
Yes
Yes
50,374
0.431

0.0002
(0.0002)
-4.54e-08
(2.95e-08)
0.113***
(0.0125)
0.053***
(0.0156)
-0.007
(0.0189)
-0.057*
(0.0325)
-0.024
(0.0164)
-0.025*
(0.0131)
-0.023**
(0.0093)
9.656***
(0.2740)
Yes
Yes
No

4.78e-05***
(2.31e-06)
-1.35e-08***
(5.19e-10)
0.111***
(0.0141)
0.053***
(0.0138)
0.010
(0.0200)
-0.054*
(0.0278)
-0.012
(0.0137)
-0.024*
(0.0124)
-0.012
(0.0118)
9.984***
(0.0251)
Yes
Yes
Yes

0.105*
(0.0570)
0.0002
(0.0002)
-4.67e-08
(2.88e-08)
0.113***
(0.0125)
0.053***
(0.0155)
-0.007
(0.0190)
-0.056*
(0.0324)
-0.024
(0.0164)
-0.025*
(0.0130)
-0.023**
(0.0093)
9.646***
(0.2680)
Yes
Yes
No

50,374
0.332

50,374
0.431

50,374
0.332

Note: Robust standard errors in parentheses, clustered at province level
*** p<0.01, ** p<0.05, * p<0.1

I employ data of 10,075 households that have the complete panel data set during 20012005 to estimate equation (6). Table A.1, columns (1) and (2) show that there is no significant
impact on household income in this model specification. This might be due to the timing that
solar unit was installed. Since the installations are distributed throughout the year, if households
34

Table A.1 (cont’d)
received their unit at the end of the year, annual income of that year will not reflect the impact of
the solar unit. Moreover, households may take time to reschedule household activities from the
daytime to nighttime in response to lighting. Thus, I allow treatment households to adjust their
working behavior for at least 1 month by defining the treatment policy dummy (𝑇𝑇 ′ ) equals to 1 if
the household owns the solar unit at least 1 month in year t. I find that the solar units statistically
increase the household income by 10.5% when the province-year effect is not included (See
column (3) of the table).

35

Table B.1: Estimation of propensity score
VARIABLES

TT

HHmem

0.060***
(0.0173)
-0.076*
(0.0409)
0.007
(0.158)
0.023
(0.0583)
0.049
(0.0388)
0.031
(0.0534)
1.670***
(0.1030)

Elder
Disabled
Age3-5
Age6-11
Age12-14
Constant
Province Effect

Yes

Observations
Pseudo R-squared

11,053
0.245

Note: Standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

36

REFERENCES

37

REFERENCES

Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An
Empiricists Companion. Princeton, NJ: Princeton University Press.
Ashenfelter, Orley, and David Card. 1985. “Using the Longitudinal Structure of Earnings to
Estimate the Effect of Training Programs.” Review of Economics and Statistics 67(4): 648–
660.
Bernard, Tanguy. 2010. “Impact Analysis of Rural Electrification Projects in Sub-Saharan
Africa.” The World Bank Research Observer 2010: 1-19.
Dinkelman, Taryn. 2011. “The Effects of Rural Electrification on Employment: New Evidence
from South Africa.” American Economic Review 101(7): 3078–3108.
Faculty of Engineering and Architecture. 2007. “Evaluation of Electrical Service Acceleration
Project by Solar Home System.” Rajamangala University of Technology (in Thai).
Florida Solar Energy Center.
(http://www.fsec.ucf.edu.).

2006.

“Photovoltaics.”

Retrieved

April

21,

2011

Grogan, Louise, and Asha Sadanand. 2009. “Electrification and the Household.” Retrieved Jan.
15, 2014 (http://www.isid.ac.in/~pu/conference/dec_09_conf/Papers/AshaSadanand.pdf).
Gruber, Jonathan. 1994. “The Incidence of Mandated Maternity Benefits.” American Economic
Review 84(3): 622-41.
Khandker, Shahidur, Douglas Barnes, Hassain Samad, and Nguyen Minh. 2009. “Welfare
Impacts of Rural Electrification: Evidence from Vietnam.” Policy Research Working Paper
5057.
Khandker, Shahidur, Gayatri Koolwal and Hassain Samad. 2009. Handbook on
Impact Evaluation – Quantitative Methods and Practices. Washington, DC: World Bank.
Kruanpradit, P., S. Saengsrithorn, and S. Sriphirom. 2004. “Electricity Supply by Solar Home
System in Thailand.” Paper presented at Technical Digest of the International PVSEC-14,
Bangkok, Thailand.
Lipscomb, Molly, A. Mushfiq Mobarak, and Tania Barham. 2013. “Development Effects of
Electrification: Evidence from the Topographic Placement of Hydropower Plants in Brazil.”
American Economic Journal: Applied Economics 5(2): 200–231.
Meier, Peter, Voravate Tuntivate, Douglas Barnes, Susan Bogach, and Daniel Farchy. 2010.
“Peru: National Survey of Rural Household Energy Use.” Retrieved Jan. 15, 2014
(http://www.esmap.org/sites/esmap.org/files/ESMAP_PeruNationalSurvey_Web_0.pdf).
Peters, Jörg, and Colin Vance. 2011. “Rural Electrification and Fertility – Evidence from Côte
d’Ivoire.” Journal of Development Studies 47(5): 753-766.
38

World Bank. 2002. “Rural Electrification and Development in the Philippines: Measuring the
Social
and
Economic
Benefits.”
Retrieved
Jan.
15,
2014
(http://siteresources.worldbank.org/INTPSIA/Resources/4900231120845825946/philippines_
rural_electrification.pdf).
World Bank. 2008. “The Welfare Impact of Rural Electrification: A Reassessment of the Costs
and
Benefits.”
Retrieved
Jan.
15,
2014
(http://siteresources.worldbank.org/EXTRURELECT/Resources/full_doc.pdf).

39

CHAPTER 2
The Effects of Subsidies on Small-Scale Renewable Electricity Generation in Thailand
1.

Introduction
The growing concern about climate change in the 1990s, policymakers around the world

have been enthusiastically supporting a wide range of incentive mechanisms for renewable
energy use, including in the electric power sector. Renewable energy is also supported to
stimulate economies, enhance energy security, and diversify energy supply. As of early 2013, the
Renewable Energy Policy Network for the 21st Century (REN21) reported that renewable energy
support policies were implemented in 127 countries, more than two-thirds of which were
developing countries or emerging economies (REN21 2013). While these renewable energy
support policies are popular, less attention has been paid to how well the supporting policies
work.
This paper analyzes the effects of a production subsidy targeting small-scale electric
power producers in Thailand. In 2007, the Thai government introduced a technology-specific
subsidy program for electricity producers with less than 10 MW of capacity. The program gives
a subsidy for renewable energy generators on top of the normal prices that electricity producers
would receive when selling electricity to the power utilities. For example, the subsidy is 0.30
baht

11

per kilowatt-hour (kWh) of energy produced using biomass, 3.5 baht per kWh of wind

power, and 8 baht per kWh produced through solar energy. The subsidy was adjusted in 2009 to
favor a relatively small power plant. For example, the production subsidy for electricity
generators using biomass or biogas that have less than 1MW of capacity is 0.50 baht per kWh,
while the subsidy for generators using biomass or biogas that have over 1MW of capacity is 0.30

11

1 baht is approximately 0.03 U.S. dollars in 2007.

40

baht per kWh. Thus, a biomass or biogas electricity producer generating at capacity of 1MW
would earn a subsidy of 500 baht per MWh (approximately 1,350,000 baht or $40,500 per year,
assuming the plant runs at 75% capacity for 10 hours per day), while a producer of
infinitesimally higher capacity would earn a subsidy of just 300 baht per MWh (about 810,000
baht or $24,300 per year).
Since the total production subsidy in 2009 is a step function with a large “notch” at 1MW
based on the power plant’s capacity, power producers who plan to build a plant near 1 MW will
have an incentive to respond strategically by building a plant on the side of the notch with a
12

higher subsidy . I develop a structural model of bunching to understand the linkage between
subsidy policy and power plant producers’ behavior. Then, I use a test developed my McCrary
(2008) to examine the distribution of power plant size for electricity producers subject to the
subsidy rate in 2009. I find the evidence of bunching around the notch. A falsification test shows
that similar bunching does not exist in the distribution of plant size built before the introduction
of the notched subsidy payment. Thus, power producers respond to the subsidy strategically,
leading to bunching in the distribution of power plant size below the notch. The subsidy is
intended to increase total electricity produced, but the notch structure gives producers an
incentive to build a small power plant in order to receive a higher subsidy, thereby potentially
reducing capacity. In addition, bunching implies variation in the long-run variable cost of
generating electricity at notch, and therefore allocative inefficiencies in production.
While an existing literature studies the effect of subsidies on renewable energy
production (Astranda and Neijb 2006; Carley 2009; Dong 2012; Lipp 2007; Palmer et al. 2011;
12

This is not a “tiered” tariff scheme in which a producer would earn the high price on the first units of production
and then a lower price on subsequent units of production. Rather, if a producer’s capacity is “epsilon” above the
threshold, then it earns the lower price on all production.

41

Palmer and Burtraw 2005; Rio Gonzalez 2008; Haito and Powers 2010), few provide a
structural model to explain the linkage between the subsidy policy on producer behavior. This
paper fills the gap by developing a model to explain that mechanism.
Although subsidy policies encourage electricity producers to invest in renewable energy,
the step function (notch) character of this policy may discourage them from investing in a largescale generator, introducing economic distortions. While the notched subsidy scheme affects a
large number of small-scale power producers, it does not affect a large quantity of installed
capacity (since the producers are by definition very small). However, the Thai government
recently began considering a new subsidy policy to initiate more electricity from renewable
energy. They plan to increase amount of renewable energy from 10% of total electricity
generation in 2012 to 20% in 2021 (Ministry of Energy, Thailand). Thus, the lessons learned
from the small-scale subsidy program in this paper could help inform the expansion to larger
producers, for which the potential welfare consequences could be quite large.
The rest of the paper is organized as follows. Next section discusses Thailand’s very
small-scale electric power producers program and subsidies policy. Then a simple model for
bunching is introduced in section 3. Section 4 describes the data availability. Section 5 presents
the reduced-form test for bunching and estimation results. Section 6 concludes with a summary
of major findings.

42

2.

Thailand’s very small-scale electric power producers program and subsidies policy
In 2002, the Thai government began allowing very small power producers (VSPPs) with

a maximum of 1 MW capacity that generated electricity from renewable energy sources
cogeneration

14

13

or

to connect to the grid and sell excess electricity to the country’s two power

distribution utilities, i.e. the Provincial Electricity Authority (PEA) and the Metropolitan
Electricity Authority (MEA). The objectives of this program are to reduce electricity generation
using commercial fuels, to decrease expenditures on fuel imports from foreign countries, to
distribute power generation to remote areas, and encourage public participation in the power
generation, and to decrease the environmental impacts (Piyasvasti Amranand, Thailand’s energy
minister). Later, in 2006, the government raised the maximum purchase capacity from each
VSPP from 1 MW to 10 MW.
To induce more power plants to produce electricity from renewable energy, in February
2007, the government began providing a subsidy on top of the normal prices that power
producers receive when selling electricity to the Power Utilities. The amounts of the subsidies
vary depending on the energy sources used. These subsidies will be provided for 7 years.
On 19 August 2009, the government adjusted the subsidy rates and the duration of the
subsidy provision to encourage more investment in the VSPPs sector. For example, the subsidy
for electricity producers using biomass or biogas with less than 1MW of capacity is 0.50 baht per
kWh, while the subsidy for electricity producers using biomass or biogas with a capacity of
1MW or greater is 0.30 baht per kWh. To alleviate the investment risks for power generation

13

The renewable energy sources under VSPP regulations include power plants using wind, solar, hydro, wastes,
biogas, and biomass.

14

Cogeneration, or combined heat and power, uses waste heat or by-product heat from industrial production or
electricity generation to produce electricity.

43

from renewable energy in the three southernmost provinces (i.e., Yala, Pattani and
15

Narathivath) , the government gave the special subsidy rates to VSPPs located in these three
provinces

16

(see Table 2.1). In order to qualify for the new subsidy rates, VSPPs must submit

their proposal to sell electricity to PEA or MEA on or after Aug 19, 2009.
The subsidy policy in 2009 creates a notch in the total production subsidy for electricity
producers using biomass, biogas, and wind energy. To illustrate, assume that a biomass power
plant produces electricity at capacity for 4,000 hours per year. If the producer chooses to build
17

the power plant at a size of 1 MW, he will receive a subsidy of 2 million baht per year .
However, if he chooses to build the power plant just bigger than 1 MW, say 1.01 MW, he will
18

receive a subsidy of around 1.2 million baht per year . Figure 2.1 shows the total production
subsidy as a function of capacity based on this example, with a large “notch” at 1 MW.
In order to prevent fraudulent reporting of size to gain the subsidy, the PEA’s and MEA’s
authority officers will cross-check the power plant to confirm its size. Also, an electricity meter
will be installed to measure the amount of electric energy produced by each plant.
Presently, in June 2014, there are 460 VSPPs registered, which have total capacity equal
to 1,621MW—an increase from only 3 VSPPs in 2003 which had 0.06 MW total capacity (see
Figure 2.2). Table 2.1 shows the distribution of VSPPs by energy source. About 56% of VSPPs
use solar power as a source of energy and produce half of total capacity. It is noteworthy that
more than 90% of VSPPs generate electricity using solar power, biomass, and biogas.

15
16
17
18

In the last decade, these areas have been faced several terrorism incidents.
Unfortunately, there is only one VSPP operating in these three provinces.
1 MW*0.5 baht per kWh *4,000 hours per year.
1.01 MW*0.3 baht per kWh *4,000 hours per year.

44

Table 2.1: The subsidy rates for VSPPs program
Energy sources &
capacity

Subsidies in
2007
(Baht/kWh)

New
subsidies
(Aug 2009)
(Baht/kWh)

Special subsidies
for 3 Southern
provinces
(Baht/kWh)

Duration
(years)

Biomass ≤ 1MW
Biomass >1 MW
Biogas ≤ 1MW
Biogas > 1MW
Wastes-Landfill/
anaerobic digestion
Wastes-Thermal Process
Wind ≤ 50 kW
Wind > 50 kW
Hydro 50 - 200 kW
Hydro < 50 kW
Solar

0.30
0.30
0.30
0.30
2.50

0.50
0.30
0.50
0.30
2.50

1.00
1.00
1.00
1.00
1.00

7
7
7
7
7

2.50
3.50
3.50
0.40
0.80
8.00

3.50
4.50
3.50
0.80
1.50
8.00

1.00
1.50
1.50
1.00
1.00
1.50

7
10
10
7
7
10

Figure 2.1: Total production subsidy as a function of capacity
Note: The subsidy policy in 2009 creates a notch in the total production subsidy for electricity producers using
biomass and biogas. The subsidy for electricity producers using biomass or biogas with less than 1MW of capacity
is 0.50 baht per kWh, while the subsidy for electricity producers using biomass or biogas with a capacity of 1MW or
greater is 0.30 baht per kWh. If the producer produces electricity at capacity for 4,000 hours a year, the total subsidy

45

Figure 2.1 (cont’d)
for a 1 MW power plant is 2 million baht per year, while the total subsidy for a slightly larger power plant, say 1.01

500

1800

450

1600

400

1400

350

1200

300

1000

250

800

200
600

150
100

400

50

200

Total capacity (MW)

Number of projects

MW, is about 1.2 million baht per year.

0

2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014*
Year
# projects

Total Capacity (MW)

Figure 2.2: Number of VSPPs and total annual capacity (in MW) for VSPPs 2003-2014*
Note: * The 2014 data include through June 2014.

46

Table 2.2: The distribution of VSPPs by energy source in June 2014
Energy Sources
Biomass
Biogas
Hydro
Solar
Wind
Cogeneration
Wastes-Landfill digestion
Wastes-Thermal process
Total

3.

Projects
#
Percent
70
15.22
101
21.96
6
1.30
256
55.65
5
1.09
2
0.43
15
3.26
5
1.09
460
100

Capacity (MW)
#
Percent
501.41
30.94
214.87
13.26
1.13
0.07
828.52
51.12
8.78
0.54
13.60
0.84
32.75
2.02
19.60
1.21
1621
100

A structural model of bunching
Since the total subsidy

19

for biomass and biogas generation in 2009 is a step function

(notch) at 1MW of capacity, the power producers who plan to build a plant near 1 MW will
strategically build a plant at the incentive-preferred side of notch. The simple explanation for
bunching is modeled as follows.
Suppose that a potential renewable electricity producer (i.e., with access to renewable
energy supply) maximizes his/her long run profits by choosing size of the power plant. For
simplicity, assume that the producer plans to generate electricity at capacity. Let the long run
variable costs 𝐶𝐶𝑖𝑖 with respect to size 𝑆𝑆 be given by 𝐶𝐶𝑖𝑖 (𝑆𝑆𝑖𝑖 ) = 𝑒𝑒 𝛼𝛼+𝛽𝛽𝑆𝑆𝑖𝑖 +𝜀𝜀𝑖𝑖 , where 𝛼𝛼 is the variable

cost for an average plant (𝜀𝜀𝑖𝑖 = 0) with 𝑆𝑆𝑖𝑖 = 0 , 𝛽𝛽 is percentage rate at which variable cost
increases with size, and 𝜀𝜀𝑖𝑖 is an unobserved producer-specific cost shifter, such as the ability of
the producer or the cost and availability of local renewable energy sources. Assume for

19

It is important to note that the total subsidy payment is notched, not just the marginal subsidy.

47

illustrative purposes that 𝜀𝜀𝑖𝑖 has a normal distribution with mean 0 and standard deviation 𝜎𝜎. The
variable revenue with respect to the size of the power plant is 𝑃𝑃(𝑆𝑆𝑖𝑖 ) is given by:
𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ if 𝑆𝑆𝑖𝑖 ≤ 1 MW
𝑃𝑃(𝑆𝑆𝑖𝑖 ) = �
,
𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 if 𝑆𝑆𝑖𝑖 > 1 MW

where 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ > 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙

Figure 2.3 shows the optimal size for power plant i for a range of possible values of 𝜀𝜀𝑖𝑖 .

The optimal power plant size (𝑆𝑆𝑖𝑖∗ ) can be described in 5 possible cases.

Case (I): The producer does not build the power plant, or 𝐶𝐶𝑖𝑖 (0) > 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ (see point A in

Figure 2.3 for illustration). In this case, 𝛼𝛼 + 𝜀𝜀𝑖𝑖 > 𝑙𝑙𝑙𝑙 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ . In other words, the cost of building

the power plant is too high, even under the highest subsidy from the government. The probability
that 𝑆𝑆𝑖𝑖∗ ≤ 0 equals

𝑙𝑙𝑙𝑙 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼

1 − Φ�

𝜎𝜎

where Φ(. ) is the cumulative distribution function of 𝜀𝜀𝑖𝑖 .

�,

Case (II): The marginal cost of power plant size equals 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ or 𝐶𝐶𝑖𝑖 (𝑆𝑆𝑖𝑖∗ ) = 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ for

𝑆𝑆𝑖𝑖∗ ∈ (0,1). In this case, 𝛼𝛼 + 𝛽𝛽𝑆𝑆𝑖𝑖∗ + 𝜀𝜀𝑖𝑖 = 𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ , or 𝑆𝑆𝑖𝑖∗ =

𝑙𝑙𝑙𝑙𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

. The producer will build

a power plant size between 0 MW and 1 MW (see point B in Figure 2.3 for illustration). The
probability that 0 < 𝑆𝑆𝑖𝑖∗ < 1 is

𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼

Φ�

𝜎𝜎

while the probability density function is
∅�

𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼−𝛽𝛽

�−Φ�

𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼

𝑆𝑆𝑖𝑖∗ −�

𝜎𝜎
𝛽𝛽

𝛽𝛽

�

𝜎𝜎

�,

� for 𝑆𝑆𝑖𝑖∗ ∈ (0,1),

where ∅(. ) is the probability density function of 𝜀𝜀𝑖𝑖 .

48

Figure 2.3: The optimal size for the power plants
Note: Let the long run variable costs 𝐶𝐶𝑖𝑖 with respect to size 𝑆𝑆 be given by 𝐶𝐶𝑖𝑖 (𝑆𝑆𝑖𝑖 ) = 𝑒𝑒 𝛼𝛼+𝛽𝛽𝑆𝑆𝑖𝑖 +𝜀𝜀𝑖𝑖 , where ∝ is a constant
term, 𝛽𝛽 is a scalar and 𝜀𝜀𝑖𝑖 is unobserved producer-specific cost shifter. The variable revenue with respect to size of

the power plant is 𝑃𝑃(𝑆𝑆𝑖𝑖 ), which is given by

𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ if 𝑆𝑆𝑖𝑖 ≤ 1 MW
,
𝑃𝑃(𝑆𝑆𝑖𝑖 ) = �
𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 if 𝑆𝑆𝑖𝑖 > 1 MW

where 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ > 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 .

Figure 2.3 shows the optimal size for power plant i for a range of possible values for 𝜀𝜀𝑖𝑖 . Consider producers that

have 𝐶𝐶𝑠𝑠4. Those producers will strategically change their plant size from their optimal size,𝑆𝑆 4∗ , to 1 MW if the

benefits from moving to the 1 MW notch (area GHIJ) is greater than the benefits from staying at 𝑆𝑆 4∗ (area IDK).
This strategic behavior leads to the bunching in the distribution of the power plant size at the notch.

49

Case (III): The marginal cost of power plant size at 𝑆𝑆 = 1 is located between 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 and

𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ , or 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 < 𝐶𝐶𝑖𝑖 (1) < 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ . In this case, 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 < 𝛼𝛼 + 𝛽𝛽 + 𝜀𝜀𝑖𝑖 < 𝑙𝑙𝑙𝑙𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ . The producer

will build the power plant at size 1 MW (see point C in Figure 2.3 for example). The probability
that 𝑆𝑆𝑖𝑖∗ = 1 in this case is

𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼−𝛽𝛽

Φ�

𝜎𝜎

� − Φ�

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝛽𝛽
𝜎𝜎

�.

Case (IV): The marginal cost of power plant size equals 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 , or 𝐶𝐶𝑖𝑖 (𝑆𝑆𝑖𝑖∗ ) = 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 for

𝑆𝑆𝑖𝑖∗ ∈ (1, 𝑆𝑆 𝑐𝑐 ]20 , because the loss of the subsidy at the notch causes the producer to revert back to
P

the notch. Suppose firm i chooses to build his/her plant size at 𝑆𝑆𝑖𝑖∗ > 1 with 𝑆𝑆𝑖𝑖∗ =

Then its profit is

𝜋𝜋𝑖𝑖∗ (𝜀𝜀𝑖𝑖 ; 𝑆𝑆𝑖𝑖∗ > 1) = �

𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

� . 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −

𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙
𝛽𝛽

𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

.

.

Suppose instead that a firm chooses to build at 𝑆𝑆𝑖𝑖 = 1. Then its profit is
𝜋𝜋𝑖𝑖 (𝜀𝜀𝑖𝑖 ; 𝑆𝑆𝑖𝑖 = 1) = 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ −

𝑒𝑒 𝛼𝛼+𝛽𝛽+𝜀𝜀𝑖𝑖
𝛽𝛽

.

There is a critical value of 𝜀𝜀𝑖𝑖 , 𝜀𝜀 𝑐𝑐 , above which firms bunch at 𝑆𝑆𝑖𝑖 = 1 and below which firms
optimally choose 𝑆𝑆𝑖𝑖∗ > 1. Note that this 𝜀𝜀 𝑐𝑐 cannot be solved analytically. However, it can be

solved numerically by equating 𝜋𝜋𝑖𝑖∗ (𝜀𝜀𝑖𝑖 ; 𝑆𝑆𝑖𝑖∗ > 1) = 𝜋𝜋𝑖𝑖 (𝜀𝜀𝑖𝑖 ; 𝑆𝑆𝑖𝑖 = 1) (see Figure 2.4 for illustration).
There is a unique critical value of size, 𝑆𝑆 𝑐𝑐 associated with this 𝜀𝜀 𝑐𝑐 :
𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 − 𝛼𝛼 − 𝜀𝜀 𝑐𝑐
𝑆𝑆 =
𝛽𝛽
𝑐𝑐

𝑆𝑆𝑖𝑖𝑐𝑐 is the critical power plant size that make power producers indifferent between moving back to the 1
MW notch, or staying at their optimal size.
20

50

Thus, in this case,

𝛼𝛼 + 𝛽𝛽 + 𝜀𝜀𝑖𝑖 < 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑤𝑤 and 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ −

𝑒𝑒 𝛼𝛼+𝛽𝛽+𝜀𝜀𝑖𝑖

> �

𝛽𝛽

𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

� . 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −

𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙
𝛽𝛽

,

or the benefit from moving to the 1 MW notch (area GHIJ in Figure 2.3) is greater than the
benefit from staying at 𝑆𝑆𝑖𝑖∗ (area IDK if the optimal size of the producer is 𝑆𝑆 4∗ ). The probability
that 1 < 𝑆𝑆𝑖𝑖∗ ≤ 𝑆𝑆 𝑐𝑐 in this case is

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝛽𝛽

Φ�

𝜎𝜎

𝜀𝜀 𝑐𝑐

� − Φ � 𝜎𝜎 �.

This strategic behavior leads to further bunching in the density distribution of the power plant
size at the notch. Combined with case III, the cumulative probability of bunching at the 1 MW
notch is
𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼−𝛽𝛽

Φ�

𝜎𝜎

𝜀𝜀 𝑐𝑐

� − Φ � 𝜎𝜎 �.

𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 , or

Case (V) the marginal cost of power plant size equals
𝑆𝑆 ∈ (𝑆𝑆 𝐶𝐶 , 10). In this case, 𝛼𝛼 + 𝛽𝛽𝑆𝑆𝑖𝑖∗ + 𝜀𝜀𝑖𝑖 = 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 or 𝑆𝑆𝑖𝑖∗ =
�

𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

� . 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −

𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙
𝛽𝛽

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 = 𝐶𝐶𝑖𝑖 (𝑆𝑆𝑖𝑖∗ ) for

and 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ −

𝑒𝑒 𝛼𝛼+𝛽𝛽+𝜀𝜀𝑖𝑖
𝛽𝛽

<

. Power plant producers do not have any incentive to move from their

𝑆𝑆𝑖𝑖∗ since the benefit from moving to the 1 MW, now, is less than the benefit from staying at 𝑆𝑆𝑖𝑖∗
(see point F in Figure 2.3 for illustration). The probability that 𝑆𝑆𝑖𝑖𝑐𝑐 < 𝑆𝑆𝑖𝑖∗ < 10 equals
𝜀𝜀 𝑐𝑐

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−10𝛽𝛽

Φ � 𝜎𝜎 � − Φ �

while the probability density function is
∅�

𝑆𝑆𝑖𝑖∗ −�

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼
�
𝛽𝛽
𝜎𝜎
𝛽𝛽

𝜎𝜎

�,

� for 𝑆𝑆𝑖𝑖∗ ∈ (𝑆𝑆𝑖𝑖𝑐𝑐 , 10)

Suppose that 𝛼𝛼, 𝛽𝛽, and the distribution of 𝜀𝜀𝑖𝑖 are the same before and after the notched

policy is introduced. Panels A and C of Figure 2.5 show the relationship between the unobserved
producer-specific cost shifter (𝜀𝜀𝑖𝑖 ) and optimal power plant size (𝑆𝑆𝑖𝑖∗), and the distribution of 𝑆𝑆𝑖𝑖∗
51

both before and after the notched policy, respectively. Panel B shows the distribution of 𝜀𝜀𝑖𝑖 ,

which has a normal distribution with mean 0 and standard deviation 𝜎𝜎.

Panel A illustrates the relationship between 𝜀𝜀𝑖𝑖 and 𝑆𝑆𝑖𝑖∗ . Before the notched policy, every

power producer received the same amount of subsidy, which is equal to 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 . The producers

who have “high” marginal costs (𝛼𝛼 + 𝜀𝜀𝑖𝑖 > 𝑙𝑙𝑙𝑙 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 ) will not build the power plant. In other

words, the producers who have 𝜀𝜀𝑖𝑖 > 𝑙𝑙𝑙𝑙 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 − 𝛼𝛼

will have 𝑆𝑆𝑖𝑖∗ =0. The producers who have

𝜀𝜀𝑖𝑖 < 𝑙𝑙𝑙𝑙 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 − 𝛼𝛼 will choose the size of their power plant according to the following equation:
𝑆𝑆𝑖𝑖∗ =

𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

. The relationship between 𝜀𝜀𝑖𝑖 and 𝑆𝑆𝑖𝑖∗ before the notch policy is shown as a blue-

dashed line with slope =

−1
𝛽𝛽

.

After the notched policy is implemented, power producers who have power plant size less
than 1 MW will get the subsidies equal to 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ per kWh of the electricity that they produce. The
power producers who have a power plant size larger than 1 MW will get the subsidies equal to
𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 , which is less than 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ . As described in cases (I)-(V) above, the relationship between 𝜀𝜀𝑖𝑖

and 𝑆𝑆𝑖𝑖∗ after the notched policy is shown as a kinked red line. The producers who have 𝜀𝜀𝑖𝑖 >

𝑙𝑙𝑙𝑙 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ − 𝛼𝛼 will continue to have 𝑆𝑆𝑖𝑖∗ =0. The producers who have 𝑙𝑙𝑙𝑙 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ − 𝛼𝛼 − 𝛽𝛽 < 𝜀𝜀𝑖𝑖 <

𝑙𝑙𝑙𝑙 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ − 𝛼𝛼 will continue to choose the size of their power plant according to the following

equation: 𝑆𝑆𝑖𝑖∗ =

𝑙𝑙𝑙𝑙𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

. The producers who have 𝑙𝑙𝑙𝑙 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ − 𝛼𝛼 − 𝛽𝛽 < 𝜀𝜀𝑖𝑖 < 𝑙𝑙𝑙𝑙 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 − 𝛼𝛼 ,

however, will bunch at 1 MW. In addition, producers who have 𝜀𝜀 𝑐𝑐 < 𝜀𝜀𝑖𝑖 < 𝑙𝑙𝑙𝑙 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 − 𝛼𝛼 will

strategically change their plant size from their optimal size to 1 MW (see case IV above for
detail). The producers who have 𝜀𝜀𝑖𝑖 < 𝜀𝜀 𝑐𝑐 will continue to choose the size of their power plant
according to the following equation: 𝑆𝑆𝑖𝑖∗ =

𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖
𝛽𝛽

52

.

If the number of the potential producers is normalized to 1, under the assumptions of the
model, this study can conclude that the notched policy will induce the entry of new power
producers Φ �

𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ −𝛼𝛼
𝜎𝜎

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼

� − Φ�

𝜎𝜎

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑤𝑤 −𝛼𝛼

�. A fraction Φ �

𝜎𝜎

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝛽𝛽

�−Φ�

𝜎𝜎

� of potential

producers will increase their power plant sizes up to 1 MW. Some producers,
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝛽𝛽

fraction Φ �

𝜎𝜎

𝜀𝜀 𝑐𝑐

� − Φ � 𝜎𝜎 �, will decrease their power plant sizes to the 1 MW notch. The

rest of the producers will not change their power plant sizes (see panel B for illustration).

Panel C depicts the effects of the change in the subsidy on the density of power plant
size. Before the change in the subsidy, the density is smooth around the 1 MW notch (the bluedashed density line). After the change in the subsidy, all power producers with 𝜀𝜀 between

𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑖𝑖𝑖𝑖ℎ − 𝛼𝛼 − 𝛽𝛽and

𝜀𝜀 𝑐𝑐 bunch at 1 MW, creating a spike in the density distribution (the red

density line). So the density is no longer smooth at 1 MW. There also is a hole in the postdensity distribution since producers do not want to build their power plants between 1 MW and
𝑆𝑆 𝑐𝑐 MW.

53

Figure 2.4: The unique 𝜺𝜺𝒄𝒄

Note: The concave red line shows 𝜋𝜋𝑖𝑖 (𝜀𝜀𝑖𝑖 ; 𝑆𝑆𝑖𝑖 = 1) = 𝑃𝑃ℎ𝑖𝑖𝑖𝑖ℎ −
𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀𝑖𝑖

�

𝛽𝛽

� . 𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −

𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙
𝛽𝛽

𝑒𝑒 𝛼𝛼+𝛽𝛽+𝜀𝜀𝑖𝑖
𝛽𝛽

. The linear dashed line is 𝜋𝜋𝑖𝑖∗ (𝜀𝜀𝑖𝑖 ; 𝑆𝑆𝑖𝑖∗ > 1) =

𝑙𝑙𝑙𝑙𝑙𝑙
−𝛼𝛼−𝜀𝜀�
. Let 𝜀𝜀̃ = 𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 − 𝛼𝛼 − 𝛽𝛽 corresponds to 𝑆𝑆̃ = 𝑙𝑙𝑙𝑙𝑙𝑙
= 1 . If a producer optimally

chooses power plant size at 𝑆𝑆̂ =

𝑙𝑙𝑙𝑙𝑃𝑃𝑙𝑙𝑙𝑙𝑙𝑙 −𝛼𝛼−𝜀𝜀�
𝛽𝛽

𝛽𝛽

> 1 (note: 𝜀𝜀̂ < 𝜀𝜀̃), the profit at 𝑆𝑆̂ or 𝜋𝜋(𝜀𝜀̂; 𝑆𝑆 ∗ > 1) is less than the profit

if the firm instead chooses to produce at 𝑆𝑆𝑖𝑖 = 1 or 𝜋𝜋(𝜀𝜀̂; 𝑆𝑆 = 1). Thus, as 𝜀𝜀 declines from 𝜀𝜀̃, 𝜋𝜋(𝜀𝜀; 𝑆𝑆 = 1) > 𝜋𝜋(𝜀𝜀; 𝑆𝑆 ∗ >
1) until 𝜀𝜀 reaches 𝜀𝜀 𝑐𝑐 . If 𝜀𝜀 < 𝜀𝜀 𝑐𝑐 , then it is optimal for firms to go above 1 MW.

54

Figure 2.5: Behavior of power producers in response to a notch in the subsidy
Note: Panel A shows the relationship between the unobserved producer-specific cost shifter (𝜀𝜀𝑖𝑖 ) and optimal power

plant size (𝑆𝑆𝑖𝑖∗ ). The distribution of 𝜀𝜀𝑖𝑖 is shown in panel B. Panel C shows the density distribution of power plant size

before and after notch policy. From panel A, the relationship between 𝜀𝜀𝑖𝑖 and 𝑆𝑆𝑖𝑖∗ before the notch policy is shown as
a blue dashed line with slope =

−1
𝛽𝛽

, and the relationship between 𝜀𝜀𝑖𝑖 and 𝑆𝑆𝑖𝑖∗ after the notch policy is shown as a kinked

red line. From panel C, the blue dashed line shows the normal density distribution of the power plant size without
the notch policy. The solid red line shows the bunching in the density at the notch, corresponding to the analysis in
Figure 2.4 when the notch policy is implemented.

55

4.

Data availability
This study focuses on the sizes of biomass and biogas power plants, since the subsidy for

solar energy does not change during the period of study and since there is very little electricity
generated from other renewable energy sources in Thailand (see Table 2.2). Thus, this study uses
cross sectional data for 171 biomass and biogas power plants from the Energy Regulatory
Commission in Thailand. The data include generator name (household, firm, or non-profit
organization), address (province), types of fuel used, power plant capacity (MW), project
approval date (D/M/Y), and start date for electricity sales (D/M/Y) of the power plants.
To improve precision of the density estimates in the next section, I combine the data
above with the data for power plants that have applied to the subsidy program but have not yet
delivered electricity. Some of these applications are still waiting for approval or are under
construction, while others have cancelled their applications. This yields a bigger data set (634
power plants). I will call this data set the “big” data set and the earlier data set the “small” data
set throughout the paper.

5.

Reduced-form test for bunching
The subsidy for producers using biomass or biogas was adjusted in 2009 to favor a

relatively small power plant. The subsidy rate is 0.50 Baht per kWh for a generator below 1MW
capacity, and 0.30 Baht per kWh for a generator above 1MW capacity. As a result, the total
subsidy for biomass and biogas in 2009 is a step function (notch) at 1MW. Thus, power
producers who plan to build a plant near 1 MW will have a strong incentive to build a plant at the
side of the notch with the larger subsidy.

56

The identifying assumption in bunching estimation is that the probability density function
(pdf) of power plant size would be smooth at 1 MW in the absence of the notched subsidy. I use
the data of power plants established before the change in the policy in August 2009 to verify that
the bunching is in fact due to the notch in the subsidy function and not some other feature of
renewable electricity technology or the economic environment. The difference in the distribution
of power plants founded before and after August 2009 near the threshold provides strong
evidence of bunching.

5.1

Graphical evidence
I employ the data of 171 existing power plants that use biomass and biogas as energy

sources to draw the histogram before and after the change in the policy. Figure 2.6 presents
histograms of power plant generating capacity. The dark-shaded histogram shows the
distribution for plants built between August 2009 and June 2014 (post-policy), during which time
the notch was present at 1MW. The light-shaded histogram shows the distribution for plants built
before August 2009, before the introduction of the notched subsidy (pre-policy). The post-policy
distribution shows a noticeable deviation from the pre-policy, distribution with a mass of
observations at 1 MW. It is worth noting that there is no obvious gap in the density immediately
above 1 MW, as shown in Figure 2.5 (Panel C). Some producers might face optimization
frictions such as inattention and inertia (Kleven and Waseem 2013). The subsidy does appear,
however, to have increased the mass of small producers, in part by shifting some of the mass
above 1 MW backwards to the notch, which is expected. Figure B.1 and C.1 in appendix show
the same information for narrower and wider bin sizes.

57

These histograms are based on a relatively small number of power producers, which can
lead to considerable noise in the density estimates. To reduce this noise, I combined the data of
the power plants that are already producing with the data of the plants that have applied to the
subsidy program but have not yet started producing. Some of these plants are still in the waiting
process, while some of them have cancelled their applications. Figure 2.7 shows the similar
evidence of bunching at 1 MW. Figure D.1 and E.1 in appendix show the same information for
narrower and wider bin sizes.
It is worth noting that there is also a bunching at 10 MW, which can be explained as
follows. To gain the subsidy, electricity generators who have capacity between 10 MW-90 MW
(small power producers, or SPPs) have to satisfy additional permitting requirements, including
the provision of an environmental impact assessment. Also, the subsidy rate for such plants is
determined by competitive bidding, with a maximum subsidy rate of 0.3 baht per kWh.
Therefore, electricity producers who plan to build a power plant near 10 MW have a strategic
incentive to be just below 10 MW, to receive a guaranteed of subsidy rate of 0.3 baht per kWh,
and to avoid extra permitting requirements. However, the magnitude of bunching at 10 MW
decreases in 2009 because the 0.3 baht per kWh was in the fact promised to all electricity
generators who had capacity between 10 MW-90 MW. In any case, this study will be dedicated
to the bunching at 1MW.

58

.2
.15
Fraction
.1
.05
0
0

2

6

4

10

8

size
After

Before

0

.2

Fraction

.4

.6

Figure 2.6: Evidence of bunching from the small data set (binsize = 0.02 MW)

0

2

4

6

8

10

cap
After

Before

Figure 2.7: Evidence of bunching from the big data set (binsize = 0.02 MW)
59

5.2.

A formal test for bunching
McCrary (2008) develops a formal test for bunching by estimating the discontinuity in

the estimated density of a variable potentially subject to manipulation in response to a notched
incentive.

21

The McCrary test has 2 steps. The first step is to create a histogram of the variable

in which there is no bin overlapping the notch. For the second step, local linear regression with
kernel-weighting is used to smooth the histogram on either side of the threshold and test for a
discontinuity.
The test statistic for bunching is derived by taking the log difference in the height of the
density of the running variable when the density is estimated separately with points to the left
and to the right of the notch:
𝜃𝜃� = 𝑙𝑙𝑙𝑙𝑓𝑓�+ − 𝑙𝑙𝑙𝑙𝑓𝑓�− ,
where 𝑓𝑓�+ and 𝑓𝑓�− are the estimated density function just above and just below the cutoff point,

respectively. The null hypothesis is that 𝜃𝜃� = 0 at the notch, or that there is no bunching.

I use the data of power plants established after the change in the policy in August 2009 to

perform a test of bunching. Figure 2.8 gives graphical results of the test. The figure strongly
suggests that the density function of power plant size is discontinuous at the 1 MW notch. The
estimated parameter 𝜃𝜃� and its standard error are shown in Table 2.3. The t-test rejects the null

hypothesis of continuity at the 95% significant level. The test indicates that the estimated density
to the left of the notch is 2.3 times the estimated density to the right of the notch. As theory
predicts, power producers strategically manipulate their power plant size to be at or just below

21

The test was developed as a diagnostic for manipulation of the running variable in a regression discontinuity
design. Used in this original context, evidence of bunching would imply that I cannot use the notch in the subsidy to
identify intensive-margin responses of electricity supply to prices.

60

the notch to receive the higher subsidy. Unfortunately, the “small” data set is not rich enough to
do a falsification test. Since there is no power plant with a size less than 1 MW before the change
in policy in August 2009, the kernel-weighted linear regression on the left side of the threshold
cannot be estimated.
Thus, I use the “big” data set to perform a bunching test. Figure 2.9 and 2.10 give an
estimate of the density function of the power plant size after and before the change in policy in
August 2009, respectively. Table 2.4 gives the results from the test for a discontinuity and
indicates a significant discontinuity in the distribution of power plant size for the power plants
founded after August 2009 (column 1). Moreover, column 2 shows that the null hypothesis of
continuity cannot be rejected for the power plants established before August 2009.
These results are in line with previous findings. Saez (2010) and Kleven and Waseem
(2013) observe large bunching in the distribution of income at kink points in the income tax
schedule in the U.S. and Pakistan, respectively. Sallee and Slemrod (2012) find that the Gas
Guzzler Tax creates bunching in the distribution of fuel economy ratings on the tax-preferred
side of notches. They also show that welfare losses may occur when firms strategically exploit
notched policies. Ito and Salee (2014) observe bunching in the distribution of vehicle weight in
response to fuel economy targets in Japan, which are a step function (notches) of vehicle weight.

61

Figure 2.8: Test for bunching from the small data set
Note: Dashed line represents 95% confidence intervals. Circles represent undersmoothed histogram of data.

Table 2.3: Test for discontinuity in the estimated density of power plant size
from the small data set
After notch policy
-0.8343*
(0.4135)
1.8836
102

difference in log density at notch
Standard error
Bandwidth
N

Notes : 1) I used the bandwidth guided by an automatic procedure suggested by McCrary (2008).
2) Standard error is in parentheses.
3)*Significant at the 95% level.

62

Figure 2.9: Test for bunching from the big data set
Note: Dashed line represents 95% confidence intervals. Circles represent undersmoothed histogram of data.

63

Figure 2.10: Falsification test for bunching from the big data set
Note: Dashed line represents 95% confidence intervals. Circles represent undersmoothed histogram of data.

Table 2.4: Tests for discontinuity in the estimated density of power plant size from the big
data set

difference in log density at notch
Standard error
Bandwidth
N

After notch policy
(1)
-1.4849*
(0.20099)
2.5752
387

Before notch policy
(2)
-0.3436
(0.9251)
1.0556
236

Notes : 1) I used the bandwidth guided by an automatic procedure suggested by McCrary (2008).
2) Standard error is in parentheses.
3)*Significant at the 95% level.

64

6.

Conclusions
This paper examines the response of small-scale electric power producers to a notched

subsidy policy. The Thai government gives a technology-specific subsidy for electricity
producers with less than 10 MW capacity. The program gives a subsidy for renewable energy
generators on top of the normal prices that electricity producers would receive when selling
electricity to the power utilities. The total production subsidy in 2009 is a step function with a
large “notch” at 1MW based on the power plant’s capacity. This notch provides an incentive for
power producers who plan to build a plant near 1 MW to strategically build a plant on the side of
the notch with a higher subsidy. I find evidence of bunching in the distribution of power plant
size around the notch for the power plants subject to the subsidy rate in 2009. A falsification test
shows that similar bunching does not exist in the distribution of plant size in 2006, before the
introduction of the notched subsidy payment. I also use a test developed by McCrary (2008) to
evaluate whether power producers alter their power plant size to avoid a discontinuous decrease
in the subsidy. The McCrary test shows significant bunching, with a statistically significant
discontinuity in the density of plant size that implies 2.3 times as many plants are just small
enough to receive the subsidy than are just barely too big. Even though the subsidy is intended to
increase total electricity produced by small producers, the notch structure creates a perverse
incentive to build a small power plant in order to receive a higher subsidy.

65

APPENDIX

66

.2
.15
Fraction
.1
.05
0
0

2

8

6

4

10

size
After

Before

Figure B.1: Evidence of bunching from the small data set (binsize = 0.01 MW)

67

.25
.2
Fraction
.1
.15
.05
0
0

2

4

6

8

10

cap
After

Before

Figure C.1: Evidence of bunching from the small data set (binsize = 0.03 MW)

68

.5
.4
Fraction
.3
.2
.1
0
0

2

4

6

8

cap
After

Before

Figure D.1: Evidence of bunching from the big data set (binsize = 0.01 MW)

69

10

.6
.4
Fraction
.2
0
0

2

8

6

4

10

cap
After

Before

Figure E.1: Evidence of bunching from the big data set (binsize = 0.03 MW)

70

REFERENCES

71

REFERENCES

Astranda, Kerstin, and Lena Neijb. 2006. “An Assessment of Governmental Wind Power
Programmes in Sweden—Using a Systems Approach.” Energy Policy 34: 277–296.
Carley, Sanya. 2009. “State Renewable Energy Electricity Policies: An Empirical Evaluation of
Effectiveness.” Energy Policy 37: 3071–3081.
Dong, Changgui. 2012. “Feed-in Tariff vs. Renewable Portfolio Standard: An Empirical Test of
Their Relative Effectiveness in Promoting Wind Capacity Development.” Energy Policy 42:
476–485.
Ito, Koichiro, and James M. Sallee. 2014. “The Economics of Attribute-Based Regulation:
Theory and Evidence from Fuel-Economy Standards.” NBER Working Paper 20500.
Kleven, Henrik J., and Mazhar Waseem. 2013. “Using Notches to Uncover Optimization
Frictions and Structural Elasticities: Theory and Evidence from Pakistan.” The Quarterly
Journal of Economics 128(2), 669-723.
Lipp, Judith. 2007. “Lessons for Effective Renewable Electricity Policy from Denmark,
Germany and the United Kingdom.” Energy Policy 35: 5481–5495.
McCrary, Justin. 2008. “Manipulation of the Running Variable in the Regression Discontinuity
Design: A Density Test.” Journal of Econometrics 142 (2), 698-714.
Palmer, Karen, Anthony Paul, Matt Woerman, and Daniel C. Steinberg. 2011. “Federal Policies
for Renewable Electricity: Impacts and Interactions.” Energy Policy 39, 3975–3991.
Palmer, Karen, and Dallas Burtraw. 2005. “Cost-effectiveness of Renewable Electricity Policies.”
Energy Economics 27: 873 – 894.
REN21. 2013. “Renewables 2013 Global Status Report.” Retrieved Jan. 15, 2014
(http://www.ren21.net/portals/0/documents/resources/gsr/2013/gsr2013_lowres.pdf ).
Gonzalez, Pablo del Rio. 2008. “Ten Years of Renewable Electricity Policies in Spain: An
Analysis of Successive Feed-in Tariff Reforms.” Energy Policy 36: 2917– 2929.
Saez, Emmanuel. 2010. “Do Taxpayers Bunch at Kink Points?” American Economic Journal:
Economic Policy 2(3): 180-212.
Sallee, James M., and Joel Slemrod. 2012. “Car Notches: Strategic Automaker Responses to Fuel
Economy Policy.” Journal of Public Economics 96(11): 981-999.
Yin, Haitao, and Nicholas Powers. 2010. “Do State Renewable Portfolio Standards Promote Instate Renewable Generation?” Energy Policy 38: 1140–1149.

72

CHAPTER 3
The Effects of the First-Car Buyer Excise Tax Rebate on Small Car Sales in Thailand
1.

Introduction
The dramatic increase in gasoline prices around the world has encouraged the

governments of many countries to focus on fuel efficiency policy. The government of Thailand
is also compelled to impose a fuel efficiency policy since the price of gasoline in Thailand has
22

more than doubled from 1998 to 2008 . In order to inspire consumers to consider a small
engine car which consumes less fuel, the government introduced the eco-car policy into the
market in October 2009. Under the eco-car program, the Finance Ministry offered to reduce the
excise tax rate from 30% to 17% on cars that have small engine sizes and that are fuel efficient.
Shortly after the policy was put in place, the first eligible eco-car was introduced to the
automobile market at a price of 459,000 baht,

23

which is almost 20% lower than the next

affordable car. Sales of eco-car models grew to 7,000 cars within six months and have continued
to rise to 40,000 cars in August 2011.
In late 2011, the government launched a first-car buyer program in order to stimulate the
growth of the automotive industry, which has become a large share of the Thai economy. Under
the first-car buyer scheme, eligible first-time car buyers

24

will receive an excise tax deduction.

The goal of the policy was to make it easier for low income buyers to purchase a vehicle and
boost domestic automotive manufacturing. The goal of stimulating the economy through the
growth of the automotive industry is employed by several governments around the world, e.g.
22

Price of Benzene 91 was increased from 12.42 baht per liter in January 1998 to 31.88 baht per liter in January
2008.

23
24

1 baht is approximately 0.03 U.S. dollars.
See section 2.2 for detail.

73

the cash-for-clunkers program in the U.S. This idea appears to fit Thailand’s economy because
25

the number of people who own passenger vehicles in Thailand is relatively low . There are
great possibilities for growth in automotive industry.
Several studies have examined the effect of this kind of policy in the U.S. and developed
countries (Li, Linn, and Spiller 2013; Mian and Sufi 2012; Kaul, Pfeifer, and Witte 2012; Nina
and Frank 2013). However, the lessons learned from these studies might not be applicable to the
case of developing countries, such as Thailand. Car in many developing countries is considered
luxury goods, which has a higher price elasticity of demand than in the U.S. Therefore, in
developing countries, a policy that lowers the price of a vehicle would have a larger impact on
the demand.
The introduction of the new car buyer subsidy reduces costs of eco-car models, but by
less than non-eco-car models. Thus, the first car buyer policy might create unintended effects on
the eco-car project that the government also strongly promotes. On the one hand, since the first
car buyer program was announced, potential eco-car buyers may have started to reconsider their
choices, including cars with larger engines that are now eligible for a large tax reduction.
Because eco-cars already have a lower excise tax than other vehicles, the new car buyer tax
rebate is lower. A larger vehicle (1,400cc), such as Honda Jazz, Honda City, Toyota Vios, or
Chevrolet Aveo, offer savings of 100,000 baht, while the eco-car offers savings of just 50,00080,000 baht. On the other hand, the first car buyer program may also induce an increase in ecocar sales, since it allows a lower income household to buy a vehicle. A person who cannot afford,
for example the Nissan March at 450,000 baht, is now able to purchase one for less than 400,000
25

According to data from the World Bank, the number of passenger cars per 1,000 people in Thailand was 62 in
2009 and 67 in 2010 (compared to 439 and 426 respectively in the U.S.). Cars are relatively expensive in Thailand,
with the country standing in 50th place out of 58 countries in the Economist Intelligence Unit's affordability
rankings for compact cars (ranking by car price/ personal income).

74

baht. The theoretical effect of the program on eco-car sales is therefore ambiguous, making it an
important empirical question. A systematic look at the data to understand the impact of the
program is essential.
In this paper, I examine the effects of the first-car buyer program on the share of eligible
cars and the share of eco-car sales using province-level data of monthly new vehicle registrations
by vehicle model from 2007-2012. Due to the availability of the data, this study focuses on
vehicle models less than 1,600cc, which accounts for 50% of all new personal vehicle
registrations. The panel data set allows me to control for vehicle model and province-time fixed
effects when examining the relationship between the program and share of vehicle sales. I find
that the first-car buyer program significantly increases the sales share of eligible cars. I also find
that the program decreases the share of eco-car sales, but not significantly.
While the program was national in scope, its impact on a given province might depend
heterogeneously on the number of qualifying car buyers for the program. I take account of this
issue by including a variable that measures exposure to the first-car buyer program. I find that
the effect of the program on the share of eligible cars is significantly higher in the provinces with
a larger number of qualifying car buyers.
The rest of this paper proceeds as follows. Section 2 describes the programs and potential
effects of the programs in more detail. Section 3 presents the data. Section 4 provides the
empirical methodology and estimation results, and Section 5 concludes.

2.

Eco-car and first-car buyer excise tax rebate program
This section describes the eco-car policy and the first-car buyer excise tax rebate program.

75

2.1

Eco-car policy
To promote small and environmental friendly cars, in October 2009, the Finance Ministry

of Thailand introduced a preferential excise tax rate of 17 per cent for eligible eco-cars. The
normal excise tax rate that is applied for standard passenger cars is between 30% and 50%,
depending on the engine size. An eco-car is defined as a car with an engine size smaller than
1,300cc for petrol engines and 1,400cc for diesel engines, that is fuel efficient (consuming less
than 5 liters per 100 kilometers), and that complies with a minimum pollution standard of
EURO4 or better (emitting no more than 120 grams of CO2 per kilometer). Moreover, the ecocar should satisfy passenger safety standards, both for front and side impact, as specified by the
UN Economic Commission for Europe (UNECE) Regulations 94 and 95 respectively.
As a response to this policy, the first eco-car, the Nissan March, was released into the
Thailand automobile market at a price of 459,000 baht. It was priced 17% lower than the
Chevrolet Aveo 1,400 cc (550,000 baht), which is the next lowest price. Compared to a car at the
median price, the Suzuki Swift and Honda Jazz 1,400 cc (599,000 baht), the March was priced
23.37% lower. The policy received a strong response from the market. The sales of eco-car
models were over 7,000 cars within six months.
As of December 2012, there were 6 eco-car models released into the market (see Table
3.1). All of these eco-cars use a petrol engine. There was no diesel engine eco-car in the market
at that time. Figure 3.1 displays the number of new eco-car registrations on a monthly basis from
2007 to 2012. The vertical dashed line represents the beginning of the eco-car policy. The data
were obtained from Department of Land Transport in Thailand, which I describe in more detail
in the next section. As shown in Figure 3.1, there is a noticeable increase in eco-car purchases
after the program was launched.

76

Table 3.1: List of eco-car models
Make
NISSAN
SUZUKI
HONDA
NISSAN
HONDA
MITSUBISHI

Model
MARCH
SWIFT
BRIO
ALMERA
BRIO AMAZE
MIRAGE

CC
1198
1242
1198
1198
1198
1198

Figure 3.1: Eco-car purchases over time

2.2

First-car tax rebate policy
To stimulate the automotive industry, at the end of 2012, the government endorsed a first-

car policy that returned the auto excise tax to persons that purchased their first car between
September 16, 2011 and December 31, 2012. The policy applied to the purchase of vehicles
under the following terms and conditions:

77

a) The vehicle engine capacity is smaller than 1,500 cc or is a double-cab pick-up truck with
any engine capacity, but whose price does not exceed one million baht per unit.
b) The vehicles must be manufactured in Thailand.
c) Buyers must be 21 years of age or older and can seek a refund of the excise tax deduction
up to one year after the purchase, but the refund must not exceed 100,000 baht.
d) Buyers are required to retain ownership of the new vehicle for at least five years.
Table 3.2 shows eligible vehicle models under the first car tax rebate policy and the amount
of rebate for each vehicle model and style. It is worth noting that all eco-car models are eligible
for the first car policy. Since the excise tax on eco-car models was already reduced to 17% prior
to the first car policy, however, the amount of the excise tax rebate for eco-car models under the
first car policy is lower than the rebate for other eligible vehicle models.
The response to the first car policy is also noticeable. Figure 3.2 displays the number of
new cars registrations of eligible and non-eligible

26

vehicle models during 2007-2012. The

vertical dashed line and the solid line represent the beginning of eco-car policy and the beginning
of the first car policy respectively. It is obvious that sales of cars that were eligible for the first
car policy increased dramatically after the policy. Sales of car that were not eligible for the first
car policy show no deviation from their trend.

3.

Data
This study uses the data of new vehicle registrations as a proxy for vehicle sales. The data

of new vehicle registrations were obtained from Department of Land Transport, using the data
for sedans under 1600 cc (less than 7 passengers). Since the Department of Land Transport only

26

Lists of eligible and non-eligible car model are shown in Table C.1 in appendix.

78

27

has monthly provincial level data for sedans under 1600 cc , other models will be omitted in
this study. The data is collected for each 76 provinces and available from Jan 2007 to Dec 2012.
In 2011, a new province, Bungkan, was established by separating from Nongkai province. Since
this study covers periods both before and after the separation, I put the data of Bungkan back to
Nongkai province. All registered vehicles are broken down by make, broad vehicle model or
nameplate, and the size of the engine (cc). For example, a 1,400 cc Aveo made by Chevrolet is
one model, while a 1,600 cc Aveo is a different model. A total of 23 vehicle models are included
in this study, and each of them has at least 1% sales share in at least one year during 2007 – 2012.
These 23 models account for 96% of total car sales among models with 1,600 cc engines or
smaller. Information on make, nameplate, and cc for all 23 models are presented in Table C.1 in
the appendix to this chapter.

Table 3.2: Incremental subsidy for eligible models
Make
HONDA
MITSUBISHI
NISSAN
NISSAN
SUZUKI
KIA
FORD
KIA
CHEVROLET
CHEVROLET
HONDA
HONDA
TOYOTA
MAZDA

Nameplate
BRIO
MIRAGE
ALMERA
MARCH
SWIFT
PICANTO
FIESTA
RIO
AVEO
SONIC
JAZZ
CITY
YARIS
MAZDA 2

CC
1198
1198
1198
1198
1242
1248
1388
1396
1398
1398
1497
1497
1497
1498

27

Eco-car
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
No
No
No
No
No

Rebate (baht)
63,000-73,000
55,000-77,000
60,000-84,000
53,000-79,000
65,000-79,000
54,000-80,000
100,000
100,000
100,000
100,000
100,000
100,000
100,000
100,000

In 2010, new vehicle registrations for vehicles smaller than 1600cc covered about 50% of total new vehicle
registrations (for vehicles with capacity less than 7 passengers).

79

Figure 3.2: Eligible and non-eligible car sales over time

Among these 23 models, 11 models are eligible for the first car rebate policy, and five of
them are eco-car models. Figure 3.3 shows the share of eligible car sales as a fraction of sales for
all 23 models, as well as the share of eco-car sales as a fraction of sales for all eligible models
over time. Prior to the first car policy, the share of eligible car seems to fluctuate around 0.8.
However, a positive trend in the share of eligible car sale is noticeable after the policy. The share
of eco-car sales seems to have a similar positive trend before and after first car policy.
Information

about

the

attributes

of

each

car

model

was

obtained

from

www.redbookasiapacific.com/th. Redbook (Automated Data Services Pty Ltd) is a known
company for providing vehicle specification and price information in Australia, New Zealand,
Malaysia,

80

Figure 3.3: Share of eligible car sales to all 23 models,
and share of eco car sales on eligible models over time.

Thailand, and China. In Thailand, information is available for all vehicles in the market since
1992. Since the data for new vehicle registrations from Department of Land Transport does not
include body style, I used the attributes of base style with automatic transmission for each model.
Information on attributes includes vehicle length (mm), width (mm), height (mm), curb weight
(kg), and fuel tank capacity (liters). However, the information obtained from Redbook websites
did not contain vehicle fuel efficiency (liter per kilometer or lpk). Therefore, I use the data of lpk
and other vehicle attributes for a vehicle in 2014 obtained from other websites (e.g.
www.checkraka.com). I use the attributes and fuel efficiency of these 2014 models to predict the
fuel efficiency as a function of vehicle attributes acquired from Redbook in 2007-2012. The
prediction method and results are shown in the next section.
For gasoline prices, I use the national nominal price of Benzene 91 (baht per liter) from
January 2007 to December 2012, which I converted to real prices using the consumer price index

81

(CPI) obtained from Bank of Thailand. Figure 3.4 shows the variation in real gas price over the
period of study.

50
45
40
35
30
Baht/litre 25
20
15
10
5
0

2007m1
2007m5
2007m9
2008m1
2008m5
2008m9
2009m1
2009m5
2009m9
2010m1
2010m5
2010m9
2011m1
2011m5
2011m9
2012m1
2012m5
2012m9

Real Benzene 91 Price

Time

Figure 3.4: The variation of gas price over time

4.

Empirical methodology and estimation results
In this section, I first discuss the channels through which the program could affect the

share of eco-car sales. Then, I describe the empirical methodology and results.

4.1

Effect of first-car excise tax rebate policy on eco-car sales
The first-car excise tax rebate scheme may have unintended effects on eco-car sales. The

excise tax rebate is lower for eco-car models. Thus, prices decreases more for eligible non ecocars models in the presence of the subsidy. People who had previously intended to buy an ecocar might therefore deviate from their decision. Thus, in this case, the first car buyer policy
might decrease the share of eco car sales.

82

However, the lower price of eco car (after the excise tax rebate) may induce a higher
demand due to its affordability. People who had not previously planned to buy a new car might
encouraged to buy a car, and the eco-car seems to be the first choice for low-income people,
since it is very affordable. For example, people who cannot afford a car (at 450,000 baht) are
now able to purchase a Nissan March at less than 400,000 baht. In this case, the first car buyer
policy might increase the share of eco-car sales.
Table 3.3 shows vehicle prices before and after the rebate for eligible car models. The
price and rebate information are for a base style with automatic transmission for each vehicle
model in January 2012.

Table 3.3: Vehicle price before and after the rebate in January 2012
Make
NISSAN
NISSAN
MITSUBISHI
HONDA
SUZUKI
CHEVROLET
HONDA
HONDA
TOYOTA
TOYOTA
MAZDA
FORD
FORD
FORD
CHEVROLET

Nameplate
MARCH
ALMERA
MIRAGE
BRIO
SWIFT
AVEO
CITY
JAZZ
YARIS
SOLUNA
MAZDA 2
FIESTA
FIESTA
FIESTA
AVEO

cc
1198
1198
1198
1198
1242
1399
1497
1497
1497
1497
1498
1388
1596
1499
1598

Price (baht)
459,000
489,000
460,000
473,000
469,000
550,000
599,000
630,000
574,000
564,000
580,000
584,000
644,000
644,000
649,000

83

Eco
Yes
Yes
Yes
Yes
Yes
No
No
No
No
No
No
No
No
No
No

Rebate
65,000
69,000
69,000
72,000
72,000
100,000
100,000
100,000
100,000
100,000
100,000
100,000
100,000
100,000
100,000

Price after rebate
394,000
420,000
391,000
401,000
397,000
450,000
499,000
530,000
474,000
464,000
480,000
484,000
544,000
544,000
549,000

4.2

Empirical methodology
I estimate a reduced-form model for the effects of the first-car tax rebate program on

eligible cars and eco-car sales, controlling for vehicle model and province-time effects. The
province-time dummies can be treated as a market fixed effect, which will capture unobserved
movements in the outside good (Huang and Rojas 2013). The baseline specification is then:
ln 𝑠𝑠ℎ𝑎𝑎𝑎𝑎𝑎𝑎𝑗𝑗𝑗𝑗𝑗𝑗 = 𝛼𝛼 + 𝛽𝛽0 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 + 𝛽𝛽1 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 + 𝛾𝛾𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗 ∙ 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑡𝑡 + 𝛿𝛿𝑗𝑗 + 𝜃𝜃𝑝𝑝𝑝𝑝
+ 𝜀𝜀𝑗𝑗𝑗𝑗𝑗𝑗 , (1)

where 𝑠𝑠ℎ𝑎𝑎𝑎𝑎𝑎𝑎𝑗𝑗𝑗𝑗𝑗𝑗 is the share of vehicle model j sales in province p at time t. 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 is the

policy dummy variable, which equals one for any vehicle model j that is eligible for the first-car

buyer program from September 2009, and zero otherwise. 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 is a dummy variable for eco-

car models. 𝛿𝛿𝑗𝑗 is a vehicle model effect that captures brand or perceptions of an individual
model’s quality. 𝜃𝜃𝑝𝑝𝑝𝑝 is a province-time effect that captures the outside good, and which also

controls for unobservable effects, such as demand shocks at the province level. Lastly, 𝜀𝜀𝑗𝑗𝑗𝑗𝑗𝑗 is the

error term.

I do not include 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 in the equation (except when it interacts with the 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗

dummy) since all eco-car models are introduced to the market after the eco-car policy started in
October 2009. Thus, variation in the 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 dummy comes only from the variation in vehicle
model, which is already captured by the vehicle model fixed effect.

To control for other characteristics that are responsible for the differential purchase
patterns in eligible and non-eligible vehicle models, I include vehicle characteristics that are
likely to affect the car sales. I use 𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗 ∙ 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑡𝑡 as a control variable, where 𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗 is liters
per kilometer, which varies across vehicle models and over time. 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑡𝑡 is the fuel price in
84

real terms. Thus, 𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑡𝑡 measures the energy costs of running vehicle model j for 1

kilometer. Since data for lpk are only available for some vehicle models (from 32 models) in
2014, I use vehicle attributes and lpk information for these models to predict lpk as a function of
vehicle attributes. To estimate the relationship between kpl (or 1/lpk) and other vehicle attributes,
I assume the Cobb-Douglas functional form (Knittel, 2011). Thus, fuel economy is modeled as:
ln 𝑘𝑘𝑘𝑘𝑘𝑘𝑗𝑗 = 𝛽𝛽0 + 𝛽𝛽1 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑗𝑗 + 𝛽𝛽2 𝑙𝑙𝑙𝑙ℎ𝑒𝑒𝑒𝑒𝑒𝑒ℎ𝑡𝑡𝑗𝑗 + 𝛽𝛽3 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙ℎ𝑡𝑡𝑗𝑗 + 𝛽𝛽2 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗 + 𝜀𝜀𝑗𝑗 , (2)

where 𝑘𝑘𝑘𝑘𝑘𝑘𝑗𝑗 is kilometer per liter of vehicle model j. Width, height, weight, and cc are the
information of width, height, weight, and cc of vehicle j, respectively. Estimates for equation (2)
are shown in Table D.1 in appendix to this chapter.
Results from estimating equation (1) are presented in Table 3.4. The first specification
includes the first car policy dummy, an interaction between the first car policy and the eco car
policy, and province-time and vehicle model fixed effects. The second specification adds control
variable. The standard errors in both specifications are clustered at the province level. I find that
the first car tax rebate policy significantly increases the share of eligible vehicles in both
specifications. In addition, the coefficient values are around 0.18-0.21. The coefficient on the
interaction between the first car and eco-car policy is negative, but small and statistically
insignificant in all specifications. In other words, the first car buyer policy does not significantly
detract from eco-car sales, on net.
The impact of the rebate policy on a given province might depend heterogeneously on the
number of qualifying car buyers for the first-car buyer program. The effect in the provinces with
more qualifying buyers should be higher than in provinces with fewer qualifying buyers. I
account for this issue by including a variable that measures exposure to the first-car buyer

85

Table 3.4: Baseline results
(1)
lnshare

(2)
lnshare

0.2106***
(0.0290)
-0.0139
(0.0376)

-2.1954***
(0.0573)

0.1787***
(0.0287)
-0.0014
(0.0373)
-1.2840***
(1.8805)
0.1444
(0.3616)

Yes
Yes

Yes
Yes

52,970
0.6787

52,970
0.6794

VARIABLES
Firstcar
Firstcar*Ecocar
lpk*gas price
Constant

Model effects
Province-time effects
Observations
R-squared

Note: Robust standard errors in parentheses, cluster at province
level
*** p<0.01, ** p<0.05, * p<0.1

program in the analysis. To measure exposure to the program, I construct a new variable called
exposure as:
exposurep =pop20upp - car stock p ,

where pop20upp is the number of people older than 20 years in province p in 2010. Although
eligible buyers for the first-car buyer program must be 21 years or older, there is no data
available on the population older than age 21. Thus, I use the population over age 20 as a proxy
variable. The variable car stock p is the number of all private vehicles

28

that were registered in

province p as of December 2010. The data for the population older than 20 years old were
obtained from the Department of Provincial Administration. The data for the car stock as of
December 2010 come from the Department of Land Transport. I focus on population and car
28

Private vehicles include private sedans (less than 7 passengers), private microbuses and passenger vans, and
private vans and pick-up trucks.

86

stock in 2010, since these numbers were pre-determined at the time of the first-car policy in
2011.
To facilitate interpretation, I rescale the exposure variable to have a mean of zero and a
standard deviation of one. Then, I interact the exposure variable with all policy variables to
equation (1). Thus, I estimate the following specification:
𝑙𝑙𝑙𝑙 𝑠𝑠ℎ𝑎𝑎𝑎𝑎𝑎𝑎𝑗𝑗𝑗𝑗𝑗𝑗 = 𝛼𝛼 + 𝛽𝛽1 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 + 𝛽𝛽2 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 ∙ 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 + 𝛽𝛽3 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑝𝑝 ∙ 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗

+ 𝛽𝛽4 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑝𝑝 ∙ 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 + 𝛽𝛽5 𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑒𝑝𝑝 ∙ 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 ∙ 𝐸𝐸𝐸𝐸𝐸𝐸𝐶𝐶𝐶𝐶𝐶𝐶𝑗𝑗𝑗𝑗 + 𝛾𝛾𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗
∙ 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑡𝑡 + 𝛿𝛿𝑗𝑗 + 𝜃𝜃𝑝𝑝𝑝𝑝 + 𝜀𝜀𝑗𝑗𝑝𝑝𝑡𝑡

(3)

Table 3.5 shows the estimation results for equation (3). The estimate of 𝛽𝛽 implies that

the first-car policy leads to a 15.7% increase in the share of eligible car in a province with mean
exposure. In a province with exposure one standard deviation above the mean, however, the
effect of the policy is almost doubled (i.e., 𝛽𝛽̂1 + 𝛽𝛽̂3 = 0.1570 + 0.1333 = 0.2903). Put simply,

the effect of the policy on the share of eligible vehicles is higher in the provinces which have
more qualifying buyers. The coefficients on the interactions between exposure and the other
policy variables are small and insignificant. These results are consistent with the results in Table
3.4.

5. Conclusions
In this paper, I use monthly data on new car registrations to estimate reduced-form
models for the effects of the first car buyer tax rebate policy on the share of eligible cars and the
share of eco-car sales. The estimation results indicate an 18% relative increase in the share of
eligible cars sales after the introduction of the policy in September 2011. However, I do not find
significant change in the share of eco-car sales after the first car policy was introduced. To

87

address the potentially heterogeneous effect of the program due to different numbers of
qualifying buyers in each province, I model exposure to the first-car buyer program. I find that
the effect of program on the share of eligible cars is significantly higher in the provinces with a
larger number of qualifying car buyers.

Table 3.5: Heterogeneous effects by exposure
(1)
lnshare

VARIABLES
Firstcar
Firstcar*eco
Firstcar*exposure
Eco*exposure
Firstcar*eco*exposure
lpk*gas price
Constant
Model effects
Province-time effects
Observations
R-squared

0.1570***
(0.0348)
0.0005
(0.0382)
0.1333***
(0.0234)
-0.0273
(0.0315)
-0.0204
(0.0268)
-1.3044***
(1.8781)
0.1864
(0.3594)
Yes
Yes
52,970
0.6802

Note: Robust standard errors in parentheses, cluster
at province level
*** p<0.01, ** p<0.05, * p<0.1

88

APPENDIX

89

Table C.1: Information on make, nameplate, and cc for all 23 models used in this study

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

Make
Nameplate
CHEVROLET
AVEO
CHEVROLET
OPTRA
HONDA
CITY
HONDA
JAZZ
MITSUBISHI LANCER
TOYOTA
AVANZA
TOYOTA
YARIS
TOYOTA
SOLUNA
NISSAN
TIIDA
TOYOTA
COROLLA
MAZDA
MAZDA 3
MAZDA
MAZDA 2
HONDA
FREED
NISSAN
MARCH
SUZUKI
SWIFT
FORD
FIESTA
FORD
FIESTA
CHEVROLET
AVEO
HONDA
BRIO
NISSAN
ALMERA
SUZUKI
SWIFT
FORD
FIESTA
MITSUBISHI MIRAGE

cc
1399
1598
1497
1497
1584
1495
1497
1497
1598
1598
1598
1498
1497
1198
1490
1388
1596
1598
1198
1198
1242
1499
1198

90

Eco-car
No
No
No
No
No
No
No
No
No
No
No
No
No
Yes
No
No
No
No
Yes
Yes
Yes
No
Yes

First car eligible
Yes
No
Yes
Yes
No
No
Yes
No
No
No
No
Yes
No
Yes
No
Yes
No
No
Yes
Yes
Yes
No
Yes

Table D.1: Estimation of log (kpl)
VARIABLES
Lnwidth

lnkpl
0.211
(0.735)
-1.476***
(0.370)
-0.0954
(0.245)
-1.102***
(0.123)
20.64***
(5.633)

Lnheight
Lnweight
Lncc
Constant

Observations
R-squared

32
0.665

Note: Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

91

REFERENCES

92

REFERENCES

Huang, Dongling, and Christian Rojas. 2013. “The Outside Good Bias in Logit Models of Demand
with Aggregate Data.” Economics Bulletin 33(1): 198-206.
Kaul, Ashok, Gregor Pfeifer, and Stefan Witte. 2012. “The Incidence of Cash for Clunkers: An
Analysis of the 2009 Car Scrappage Scheme in Germany.” University of Zurich Working
Paper 68.
Knittel, Christopher R. 2011. “Automobiles on Steroids: Product Attribute Trade-Offs and
Technological Progress in the Automobile Sector.” American Economic Review 101(7): 336899.
Leheyda, Nina and Frank Verboven. 2013. “Scrapping Subsidies During the Financial Crisis Evidence from Europe.” ZEW Discussion Paper 13-079.
Mian, Atif, and Amir Sufi. 2012. “The Effects of Fiscal Stimulus: Evidence from the 2009 ‘Cash
for Clunkers’ Program.” The Quarterly Journal of Economics 127(3): 1107-1142.
Li, Shanjun, Joshua Linn, and Elisheba Spiller. 2013. “Evaluating Cash-for-Clunkers: Program
Effects on Auto Sales and the Environment.” Journal of Environmental Economics and
Management 65(2): 165-193.

93