Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Essays on the econometrics of program evaluation
(USC Thesis Other)
Essays on the econometrics of program evaluation
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
ESSAYS ON THE ECONOMETRICS OF PROGRAM EVALUATION
by
Kannika Damrongplasit
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
December 2007
Copyright 2007 Kannika Damrongplasit
ii
Dedication
To My Late Grandmother Vipa Wangrattanapranee,
My Parents Khoonavuthi Damrongplasit and Kularb Damrongplasit,
and My Brother Nattapol Damrongplasit
iii
Acknowledgements
I would like to thank my advisor and the chair of my dissertation committee,
Professor Cheng Hsiao, for his guidance, support and encouragement throughout the
preparation of this dissertation. I am indebted to his generosity in answering my
endless questions at every stage of this research. His mentoring and advices during
the past four years have been truly rewarding. They have not only shaped my
academic development at USC, but will also have a life-long impact on my career as
an economist.
I am also thankful to my dissertation committees, Professor Jeffrey Nugent and
Professor Michael Nichol, and other professors including Professor Robert Dekle,
Professor John Ham, Professor John Strauss, and Professor Xueyan Zhao for their
insightful suggestions and guidance, and for challenging and stimulating my
thoughts about many of the ideas presented in this dissertation.
Finally, I want to give a special thank to my family who have always been and will
always continue to be the greatest source of support in my life. I want to thank my
parents for giving me their unconditional love, endless caring, and continuous
encouragement. My deep gratitude goes to my brother who is always willing to help
iv
me with any computer-related problem. Lastly, I am indebted to Nithirut Kampanya
for his patience and support throughout my doctoral study.
v
Table of Contents
Dedication ....................................................................................................................ii
Acknowledgements.....................................................................................................iii
List of Tables .............................................................................................................vii
List of Figures .............................................................................................................ix
Abstract ........................................................................................................................x
1 Decriminalization and Marijuana Smoking Prevalence: Evidence from
Australia...........................................................................................................1
1.1 Introduction..........................................................................................1
1.2 Literature Review.................................................................................6
1.2.1 Demand for Marijuana...............................................................6
1.2.2 Location Choice.........................................................................9
1.3 Parametric Specifications...................................................................13
1.3.1 Models .....................................................................................13
1.3.2 Selection of Explanatory Variables .........................................16
1.3.3 Data..........................................................................................18
1.3.4 Estimation Results ...................................................................22
1.4 Nonparametric Specification: Propensity Score Matching................33
1.4.1 Description of the Model .........................................................33
1.4.2 Empirical Findings..................................................................35
1.5 Specification Analyses.......................................................................39
1.5.1 Parametric Model Specification Test.......................................39
1.5.2 Nonparametric Kernel Consistent Model Specification Test
with Mixed Discrete and Continuous Data..............................42
1.6 Conclusion .........................................................................................57
2 Evaluation of Malaysian Capital Controls in the Short, Medium, and Long
Runs ...............................................................................................................61
2.1 Introduction........................................................................................61
2.2 Malaysian Capital Controls................................................................65
2.3 Data and Explanation for using the Two Approaches .......................67
2.4 Econometrics Models.........................................................................76
2.4.1 Models for Time-Shifted Difference-in-Difference
Estimation ................................................................................76
vi
2.4.2 Models for Conventional Difference-in-Difference
Estimation ................................................................................82
2.5 Empirical Results ...............................................................................85
2.6 Conclusion .......................................................................................105
Bibliography.............................................................................................................109
Appendices...............................................................................................................114
Appendix 1...................................................................................................114
Appendix 2...................................................................................................115
Appendix 3...................................................................................................117
Appendix 4...................................................................................................118
vii
List of Tables
Table 1 Summary Statistics of Dependent and Independent Variables .....................21
Table 2 Coefficient Estimates for Marijuana Smoking Equation ..............................24
Table 3 Coefficient Estimates for Location Choice Equation ...................................26
Table 4 Marginal Effects for Marijuana Smoking Equation......................................27
Table 5 ATE and ATET using Propensity Score Stratification Method....................37
Table 6 Parametric Model Specification Test............................................................41
Table 7 Nonparametric Kernel-Based Test................................................................56
Table 8 Important Dates.............................................................................................72
Table 9 Treatment Periods under the Time-Shifted Approach ..................................78
Table 10 Time-Shifted Difference-in-Difference Estimation....................................80
Table 11 Treatment Periods under the Conventional Approach................................83
Table 12 Conventional Difference-in-Difference Estimation....................................84
Table 13 Estimates from the Time-Shifted Approach using Indonesia, Korea, and
Thailand as Comparators............................................................................................86
Table 14 Estimates from the Time-Shifted Approach using Korea as Comparator ..88
Table 15 Estimates from the Conventional Approach using Indonesia, Korea, and
Thailand as Comparators............................................................................................91
Table 16 Estimates from the Conventional Approach using Korea as Comparator ..93
Table 17 Effectiveness of Malaysian Capital Controls Relative to the IMF
Program(s)................................................................................................................104
Table 18 Definitions of Variables for the First Chapter ..........................................115
viii
Table 19 Definitions of Variables for the Second Chapter......................................117
Table 20 FDI Inflows and Outflows ........................................................................118
ix
List of Figures
Figure 1 Histograms of Estimated Propensity Scores in the Overlapping Region ....36
Figure 2 Foreign Reserves (in log scale) ...................................................................69
Figure 3 Exchange Rates (National currency / US$).................................................70
Figure 4 Map of Australia........................................................................................114
x
Abstract
Through two empirical studies, we explore various methods that can be used to
estimate the average treatment effect of a dichotomous policy variable. The first
essay uses many cross-sectional program evaluation techniques and applies them to a
micro-level data set. The second essay assesses the effect of a macro-level
intervention on outcome variables with the use of time-series approaches.
In the first study, we use the 2001 wave of National Drug Strategy Household
Survey (NDSHS) to assess the impact of marijuana decriminalization policy on
marijuana smoking prevalence in Australia. Both parametric and nonparametric
methods are used. The parametric approach postulates an endogenous probit
switching model and its nested binary probit, endogenous bivariate probit and two-
part models. The nonparametric approach uses the propensity score stratification
matching to compute alternative measures of the treatment effect. Specification
analyses are also conducted. We use likelihood ratio test to choose among nested
parametric models, then construct a nonparametric kernel-based test to select
between a parametric null model and its nonparametric alternative. Both normal and
bootstrap empirical distributions are used to approximate the null distribution of the
test statistic. Our specification analyses demonstrate that the endogenous probit
xi
switching model is superior to other models, which suggests that decriminalization
raises the probability of smoking marijuana by 16.3%.
The second study uses monthly data from January 1993 to June 2005 to assess the
effectiveness of capital controls in restoring the Malaysian economy as compared to
the IMF programs used in Indonesia, Korea, and Thailand. We analyze various
aspects of the economy by employing both conventional and time-shifted difference-
in-difference estimations. To evaluate the effects in the short, medium, and long
runs, this study includes one to six years of treatment periods. Our estimation results
from the time-shifted approach favor capital controls as a more effective crisis’
solution than the IMF programs in the short and medium runs. However, when the
conventional method is used, the performance of capital controls relative to the IMF
programs is found to vary depending on the economic measure of interest, the
treatment period considered and the set of IMF countries used for comparison.
1
1 Decriminalization and Marijuana Smoking
Prevalence: Evidence from Australia
1.1 Introduction
Illicit drug usage is widespread around the world, posing significant social and
economic costs to the health care, justice and social welfare systems in both
developed and developing countries. According to the Economist (2001)
1
, the global
retail sale of illegal drugs is estimated to be US$150 billion a year, which is in the
same league as worldwide sale of tobacco and alcohol and about half the size of the
pharmaceutical industry. Significant amounts of public funds have been spent by
governments worldwide on dealing with the consequences of substance abuse and on
educational programs. For example, the United States’ drugs policy costs
approximately US$35 – US$40 billion a year according to the Economist (2001),
while the Australian government has committed more than AUS$1 billion towards
its National Illicit Drugs Strategy since 1997
2
.
Among illicit drugs, marijuana is by far the most widely used. It is commonly
considered a “softer” drug compared to other “harder” drugs such as cocaine, heroin,
or amphetamines. The prevalence of hydroponic cultivation in recent years has also
1
See The Economist Survey: the case for legalizing drugs, July 28, 2001.
2
See Australian Department of Health and Ageing, National Illicit Drug Strategy,
http://www.health.gov.au/internet/wcms/Publishing.nsf/Content/health-pubhlth-strateg-drugs-illicit-
index.htm.
2
significantly improved the productivity of covert production. While there is more
support for using marijuana for medical purposes in treating patients with nausea,
glaucoma, spasm, and pain, much controversy has surrounded the detrimental health
effects of recreational use of marijuana. Some suggest that marijuana use is linked to
lung cancer, deteriorated immune systems, harmful effect on blood circulation, and
short-term memory loss. For heavy users, there is also the problem of drug
dependency and the related withdrawal symptoms such as anxiety and loss of
appetite
3
.
At the center of the controversy is whether legal sanctions are the best approach to
reduce the use and the associated harm of the drug. The ongoing debates of
marijuana decriminalization concentrate on potential benefits and costs of such
policies. The first supporting argument for decriminalization is that a criminal
charge is too severe a penalty relative to the crime itself. A criminal record can have
many negative consequences on the subsequent life of an otherwise law-abiding
person. For example, an offender may lose out in future employment opportunity or
face problems in international travel. Second, decriminalization would allow the
government to separate the market of marijuana from the market of other harder
drugs, thereby permitting the authorities to redirect their resources used in law
enforcement and criminal justice system from the “softer” cannabis to “harder” drugs
like cocaine, heroin, and amphetamines. For instance, a 2005 report by Jeffrey
3
See Time magazine, Is pot good for you?, November 4, 2002.
3
Miron entitled “The Budgetary Implications of Marijuana Prohibition” finds that
legalizing marijuana can reduce the cost of enforcement in the United States by
US$7.7 billion per year
4
. For those who argue against decriminalization, their first
claim is that decriminalization inevitably lowers both the legal and social costs
associated with the use of marijuana, thus sending a signal that it is acceptable to
smoke marijuana. This may, in turn, encourage higher consumption of the drug as a
result. Another reason against decriminalization is the gateway theory; that is, there
is a growing concern that exposure to marijuana by youths may lead to their
subsequent consumption of other harder drugs. Given the above debates, empirical
evidence of the impact of marijuana decriminalization on marijuana usage is crucial.
In particular, if decriminalization has little or no impact on smoking prevalence, then
there is a strong argument in favor of the policy. On the other hand, if there is
evidence that decriminalization significantly stimulates more marijuana smoking,
then a liberal approach towards marijuana may not be as beneficial as advocated by
its supporters.
This paper uses the 2001 Australian National Drug Strategy Household Surveys
(NDSHS) to study the impact of marijuana decriminalization on marijuana usage.
Australia consists of six states and two territories. The map of Australia is presented
in Appendix 1. As of 2001, South Australia, Australia Capital Territory, and
Northern Territory had already decriminalized the possession and cultivation of
4
See http://www.ProhibitionCosts.org.
4
marijuana for personal consumption. Under this policy, while it is still illegal to
consume or grow marijuana, the criminal status of the offence is changed to a non-
criminal one. If one is caught of using or growing marijuana, one must pay a fine
within a specified period of time
5
, usually within sixty days, in order to be eligible
for the above reduced penalty of receiving no criminal record or imprisonment. If
one fails to pay the fine however, a criminal proceeding may follow possibly leading
to a jail sentence. The policy of decriminalization is commonly known in Australia
as “Cannabis Expiation Notice” system (CEN). Finally, for those states which have
not decriminalized marijuana, criminal offence for possessing, consuming or
cultivating the drug is still retained.
This paper aims to assess the impact of decriminalization on marijuana smoking
prevalence. There are four major differences between this paper and earlier studies.
First, we employ both parametric and nonparametric methods in evaluating the
impact. Second, we test the reliabilities of these competing methods when each is
applied to our data by using a parametric model specification test and by
constructing a nonparametric kernel-based test. Third, we provide a more flexible
marijuana smoking behavior equation by allowing individuals to respond differently
when the legal and institutional arrangement changes. Fourth, we attempt to address
the potential endogeneity of marijuana smoking and the individual’s decision to
5
It must be noted that the maximum amount of marijuana considered a minor offence and the
corresponding fine do differ across jurisdictions, usually between AUS$150-$200.
5
reside in decriminalized versus non-decriminalized states by allowing individual’s
marijuana smoking behavior equation to be correlated with his/her location choice.
Both parametric and nonparametric methods are used in this study because each
method is associated with certain advantages and disadvantages. The advantage of
nonparametric approach is that no functional form or distributional assumption needs
to be imposed. One disadvantage is that the conditional independence assumption
(i.e. ignorable treatment assignment assumption) is a maintained hypothesis.
Another is that the impacts of other socio-demographic effects on marijuana smoking
cannot be estimated. For parametric specification, the main advantage is that a full
model can be estimated. Hence, selection on both observables and unobservables
can be taken account of and the impact of each variable on the outcomes can be
assessed. The disadvantage is that functional form and distributional assumptions
need to be imposed.
This paper is organized as follows. Section 1.2 gives a review of existing literature.
Section 1.3 presents parametric models, which include both the endogenous probit
switching model and its nested two-part model, endogenous bivariate probit and
binary probit models. A description of our data, estimation results as well as the
comparison of our findings to existing literature are also provided. In Section 1.4,
we present alternative measures of treatment effect from the propensity score
stratification matching method and discuss the empirical findings. Specification
6
analyses are conducted in Section 1.5. We first employ a parametric model
specification test and then a nonparametric kernel-based test. Finally, Section 1.6
summarizes our research findings.
1.2 Literature Review
1.2.1 Demand for Marijuana
There have been three empirical studies from Australia. First, Cameron and
Williams (2001) utilize 1988, 1991, 1993, and 1995 NDSHS data to estimate own
price responsiveness, and cross price responsiveness for cannabis, alcohol, and
cigarettes. They employ a binary probit model to estimate participation for each
drug. The impact of decriminalization is also evaluated; however, Northern
Territory and Australia Capital Territory are excluded from their analysis due to
incomplete data. This narrows the impact of decriminalization to only South
Australia. Their results suggest that participation is responsive to both own price and
cross price terms. Furthermore, decriminalization contributes positively and
significantly to marijuana prevalence.
Second, Williams (2004) uses two-part model to estimate demand for marijuana. In
particular, he models participation decision by binary probit. Then, conditional on
using marijuana, he uses an ordered probit model to estimate the frequency of use.
7
The data is taken from 1988, 1991, 1993, 1995, and 1998 waves of NDSHS. The set
of decriminalized states is expanded in this paper to cover South Australia, Australia
Capital Territory, and Northern Territory. He finds that decriminalization has no
significant impact on either prevalence or frequency of smoking when using the
entire sample. However, when focusing on the sub-sample of males aged 25 and
over, the effect of decriminalization is found to be positive and significant.
Finally, through the use of NDSHS data, Zhao and Harris (2004) estimate a
multivariate probit model of marijuana, alcohol, and tobacco participations. Their
model takes into account possible correlations across participations of different
drugs. Their results show positive correlations across the three drugs through
unobservable characteristics. With respect to decriminalization policy, they find a
positive and significant marginal effect of decriminalization on prevalence of about
2%.
There are both theoretical and empirical studies for the United States. Theoretically,
a representative consumer is assumed to maximize the following single-period utility
function
) ; , , (
D M, C,
Z D M C U
Max
subject to I D M C
P P P D M C
≤ + +
where C, M and D stand for composite good, marijuana, and other drugs while P
C
,
P
M
, and P
D
are their corresponding prices, Z represents other variables that may
8
affect the utility function, and I denotes income. This problem is solved by forming
a Lagrangian function and maximizing it with respect to C, M, D, and the Lagrange
multiplier. It turns out that demand for marijuana is a function of prices, income,
and other variables such as age, gender, marital status, educational attainment,
working status, ethnicity, and place of residence. On the empirical side, Saffer and
Chaloupka (1995), Saffer and Chaloupka (1998), and Pacula, Chriqui, and King
(2003) find the impact of decriminalization on marijuana smoking prevalence to be
positive and significant. These three papers all use binary probit model to estimate
the participation decision but employ different data sets. The first two articles use a
nationally representative survey, namely, the National Household Survey on Drug
Abuse (NHSDA), whereas the third paper concentrates on a sample of high school
students from the National Educational Longitudinal Survey (NELS). Furthermore,
Pacula, Chriqui, and King (2003) also control for different dimensions of the
marijuana laws like maximum fines and minimum jail sentences in addition to the
decriminalization dummy variable. In contrast to the above findings, DiNardo and
Lemieux (2001), Pacula (1998), and Thies and Register (1993) discover insignificant
effects of the marijuana policy reform on individual smoking decisions. DiNardo
and Lemieux (2001) use Monitoring the Future data (MTF) of high school seniors to
run binary probit estimation for marijuana and alcohol participation, both separately
and jointly. Because the price of marijuana is unavailable, they use
decriminalization dummy variable as its proxy. Pacula (1998), on the other hand,
utilizes National Longitudinal Survey of Youth (NLSY) to estimate the participation
9
decision and the unconditional demand for marijuana and alcohol by employing
binary probit and ordinary least square methods. Finally, Thies and Register (1993)
use the same NLSY data to estimate binary logit models for marijuana, alcohol and
cocaine participation decisions.
Previous studies give mixed conclusions for the impact of decriminalization on
marijuana smoking prevalence. We believe that there are a few missing points that
have not been addressed by others. First, to our knowledge, there appears to be no
study that attempts to evaluate this impact using nonparametric methods. Second,
even with the parametric approach, most papers typically use just binary probit
models and treat the decriminalization dummy variable as an exogenous explanatory
variable, instead of recognizing the possibility of endogeneity or selection. Our
study tries to bridge these gaps.
1.2.2 Location Choice
Since our study employs cross-section data, we focus on static rather than dynamic
location choice. In theory, individual’s static location choice can be derived from
utility maximization. The utility of individual i as a result of living in location j can
be expressed by where stands for personal attributes and
denotes the locational characteristics of choice j. Individual i then chooses a location
that yields the highest utility level. In empirical studies, static location choice is
usually modeled by using qualitative response framework. When there are only two
) , (
L A U j i ij
F =
Ai L j
10
outcomes, the appropriate model is either binary probit or binary logit. In the case of
multiple outcomes, multinomial logit, conditional logit, or multinomial conditional
logit is employed.
Following the above theoretical and econometric framework, we now discuss the
existing literature on residential and industrial location choice. Toussaint-Comeau
and Rhine (2004), Gabriel and Rosenthal (1989), and Frenkel (2001) include only
individual characteristics, , as explanatory variables in their estimations of
locational choice. In particular, Toussaint-Comeau and Rhine (2004) investigate the
factors that influence Hispanic population in Chicago to reside in ethnic or non-
ethnic enclave areas. They enter socioeconomic, demographic, and life-cycle
characteristics as explanatory variables. Because there are only two locational
choices, probit estimation is employed in their study. Gabriel and Rosenthal (1989),
on the other hand, use multinomial logit model to empirically estimate how white
and black households choose among five mutually exclusive neighborhoods within
the Washington D.C. metropolitan area. The explanatory variables in their paper
comprise household socioeconomic, demographic, and racial characteristics. Frenkel
(2001) studies firms’ locational choices by employing binary logit model to estimate
how high-technology firms in Israel make decision to locate between metropolitan
and non-metropolitan areas. Various plant characteristics are included as
explanatory variables.
Ai
11
We review two papers that take into account only location characteristics, , when
attempting to explain residential location choice. Bartel (1989) and Duncombe,
Robbins, and Wolf (2001) use only place attributes like tax rate, unemployment rate,
and average wage rate as the independent variables of their studies. The former
examines residential location choice of the US immigrants while the latter analyzes
location decision among the US retirees. Both papers use conditional logit as their
methods of estimation.
L j
Finally, Feridhanusetyawan and Kilkenny (1996) and Kittiprapas and McCann
(1999) enter both individual and location characteristics as their independent
variables. Feridhanusetyawan and Kilkenny (1996) attempt to model residential
location choice between rural and urban areas. They employ binary logit model
when there are two choices (i.e. rural or urban), and multinomial logit model when
dealing with three-way typologies (i.e. metro, non-metro adjacent, or non-metro non-
adjacent). The explanatory variables included in their studies are personal as well as
location characteristics. In particular, the location characteristics are measured at the
county level. Similarly, Kittiprapas and McCann (1999) use binary logit model to
investigate how electronic companies in Thailand choose where to locate. The
choices of location consist of Bangkok metropolitan area and non-Bangkok
metropolitan area. They include both firm and regional characteristics in their
estimations.
12
Overall, previous literature suggests that static location choices should be modeled
by using a discrete choice framework. The explanatory variables vary across
different studies. In general, both individual and location characteristics are crucial
determinants of locational decisions. In this research, we only face two location
alternatives, decriminalized and non-decriminalized states, which leads us to choose
binary probit as the method of estimation. Our setup is fairly similar to
Feridhanusetyawan and Kilkenny (1996). The utility of individual i as a result of
living in state that is classified into location type j (i.e. decriminalized or non-
decriminalized states) can be expressed by
s
) , (
L A U sj i isj
F = where stands for
personal attributes, and denotes location characteristics of state s which belongs
to location type j. We employ state level attributes ( ) to explain the choice
between decriminalized and non-decriminalized states in the same manner as
Feridhanusetyawan and Kilkenny (1996) that use county level attributes to explain
the choice between rural and urban areas. Detailed discussion of the chosen
variables is provided in Section 1.3.2.
Ai
Lsj
Lsj
13
1.3 Parametric Specifications
1.3.1 Models
Our observed data are in the form of , where ) , (
d
y
i
i
1 = y
i
denotes that the
i
individual is a marijuana smoker and 0 otherwise, and
th
1 =
d i
if the individual
resides in a decriminalized state and 0 if not. There are four possible states, (1,1),
(1,0), (0,1), and (0,0). To allow for the possibility that an individual’s behavior may
differ when living in a decriminalized state relative to living in a non-decriminalized
state, and the possible joint determination of , we assume a latent marijuana
smoking equation
i
th
) , (
d
y
i
i
(1.3.1)
ε
i i
i
x
g y
1
1
*
1
) ( + =
if an individual resides in a decriminalized state, and
(1.3.2)
ε
i i
i
x
g y
0
0
*
0
) ( + =
if an individual resides in a non-decriminalized state, as well as a latent locational
choice equation,
(1.3.3)
υ i i i z
h
d
+ = ) (
*
where and denote the vectors of observable factors which could have
overlapping elements, and represent the effects of unobservables. The
latent variables are mapped to the observables via:
xi zi
) , , (
0 1 υ ε ε i i i
) , (
d
y
i
i
14
(1.3.4)
⎩
⎨
⎧
≤
>
=
, 0 if , 0
, 0 if , 1
*
*
d
d
d
i
i
i
(1.3.5)
⎪
⎩
⎪
⎨
⎧
≤
> − + =
=
0. y if , 0
, 0 ) 1 ( if , 1
*
i
*
0
*
1
*
y
d
y
d
y
y
i
i
i
i
i
i
We shall take (1.3.1)-(1.3.5) as the maintained hypothesis for evaluating the effect of
decriminalization on marijuana smoking behavior. Both parametric and
nonparametric analyses will be conducted. For parametric specifications, we assume
, (1.3.6)
ε
β
α
i i
i
x
y
1
1
1
*
1
+ + =
, (1.3.7)
ε
β
α
i i
i
x
y
0
0
0
*
0
+ + =
, (1.3.8)
υ
γ
i i i z d
+ =
*
where (
ε i 1
,
ε i 0
,
υ i
) are assumed jointly normally distributed with mean zeros and the
following covariance matrix
. (1.3.9)
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎝
⎛
=
1
1
1
) , , cov(
0 1
0 10
1 10
0 1
ρ ρ
ρ ρ
ρ ρ
υ ε ε
υ υ
υ
υ
i i i
Model (1.3.4)-(1.3.9) leads to an endogenous probit switching model, with equations
(1.3.6)-(1.3.8) being the three latent equations to be estimated. The conventional
dummy variable approach, sample selection model, and two-part model are all nested
in this endogenous switching model. (a) When β
1
= β
0
, and ρ
υ 1
= ρ
υ 0
≠ 0, model
15
(1.3.4)-(1.3.9) becomes a bivariate probit model with the smoking equation being the
same for decriminalized and non-decriminalized states other than the intercept. It is
a generalization of the sample selection model as presented by Amemiya (1985). (b)
When ρ
υ 1
= ρ
υ 0
= , there is no correlation via the unobservable error terms between
the location equation and either one of the two smoking equations. In this case,
model (1.3.4)-(1.3.9) is a generalization of the frequently used two-part model (e.g.
Duan et al. (1983), (1984)). (c) When
0
ρ
υ 1
= ρ
υ 0
= and 0 β
1
= β
0
, model (1.3.4)-
(1.3.9) reduces to a dummy variable approach to evaluate the effect of
decriminalization, i.e.
(1.3.10) ) ( ) (
0 1 0 1
*
0
*
1
ε ε α α
i i
i i
y y − + − = −
or
. (1.3.11)
ε
δ β α
ε ε α α
β α
i i i i i i i
i
d x d x
y + + + = − + − + + = ) ( ) (
0 1 0 1
*
Based on the estimates of model (1.3.4)-(1.3.9) or its restricted versions, we can
estimate the average treatment effect (ATE) under counterfactual condition. In our
model, ATE is the average effect of decriminalization on marijuana smoking
participation for a randomly selected individual among the Australian population.
This ATE can be computed by
∫
+ Φ − + Φ dx x f x x ) ( )] ( ) ( [
0
0
1
1
β
α
β
α
. (1.3.12)
If the sample is randomly drawn, the ATE may be approximated by
16
)]
0 0
( )
1 1
( [
1
^ ^ ^
1
^
x x
n
i i
n
i
β
α
β
α
+ Φ − + ∑ Φ
=
. (1.3.13)
On the other hand, if the conventional model (1.3.11) is the maintained hypothesis,
we may use
∑ + Φ − + + Φ
=
n
i
i i
x x
n 1
^
^ ^
^
^
)] ( ) ( [
1
β
α δ
β
α
(1.3.14)
to estimate ATE.
1.3.2 Selection of Explanatory Variables
The model requires two sets of explanatory variables: (i) determinants of marijuana
smoking decision, , and (ii) factors affecting individual’s choice of living in a
decriminalized or non-decriminalized state, .
xi
zi
Marijuana is an addictive recreational drug, and studies on recreational drugs have
arisen from many disciplines such as psychology, medicine, epidemiology and
sociology, as well as economics. Economists treat the consumption of addictive
commodities as ‘rational’ behaviour that follows the basic law of economics but with
unique characteristics (Becker and Murphy (1988)). Empirical studies on marijuana
smoking have often considered the effects of economic variables such as price and
income. To capture the full price of using an illicit drug, public policies and the
legal status of consumption are often included to represent the legal costs and risks
of use (e.g. Pacula (1998)). Individual socioeconomic and demographic
17
characteristics are also considered to capture heterogeneity in consumer demand and
the effects of health knowledge proxied by education and socioeconomic status (e.g.
Williams (2004), Zhao and Harris (2004)). Following this economic literature, we
choose variables to include the price of marijuana, household income, as well as
some standard socioeconomic and demographic variables that capture heterogeneity
in demand such as age, gender, marital status, educational attainment, work status
and ethnic background for the Australian indigenous population. The impact of the
legal risk of smoking via the decriminalization policy is captured by estimating
separate smoking equations for the two types of states using the endogenous probit
switching model described above.
xi
For , the existing literature (e.g. Feridhanusetyawan and Kilkenny (1996),
Kittiprapas and McCann (1999)) suggests that both personal and location
characteristics are key determinants of individual’s location choice. We include
socioeconomic and demographic variables as personal attributes. They comprise
similar set of variables as those in the marijuana smoking equation with two main
exceptions. First, the age below 20 category is now dropped because individuals in
this age group are usually dependents whose decisions on where to live mostly rely
on their guardians. Second, we add number of dependent children as an additional
explanatory variable because it might influence one’s choice of location. In
particular, individuals with dependent children could display higher tendency to live
in a place with good environment in order to better raise their families. Finally, we
zi
18
include the unemployment rate in each individual’s state of residence as a location
characteristic.
1.3.3 Data
The data used in this study come from three different sources: Australia National
Drug Strategy Household Survey (NDSHS), Australia Bureau of Statistics (ABS),
and Australian Illicit Drug Report. The main data set of this paper is the unit-record
data from the 2001 wave of NDSHS (2001). The NDSHS is a nationally
representative household survey of non-institutionalized civilian Australian
population with age 14 and older. It provides information on individual drug usage,
and many socioeconomic and demographic variables. Three different survey
methods were implemented: a drop-and-collect questionnaire, a face-to-face personal
interview, and a computer assisted telephone interview. For more sensitive questions
like individual drug usage, measures were put in place so that the information is kept
confidential from the interviewer in order to minimize potential underreporting of
drug use. There are altogether 26744 observations available in the 2001 wave of
NDSHS. After deleting observations with missing data, the resulting sample is
14008. South Australia, Australia Capital Territory, and Northern Territory had
already adopted decriminalization of marijuana by 2001, so observations from these
three states are classified as the treatment group.
19
Definitions of all variables are listed in Appendix 2. The main dependent variable
for individual’s marijuana smoking behavior is a discrete binary choice variable ,
which assumes a value 1 if the respondent used marijuana in the past twelve months,
and a value of 0 otherwise. In addition, the 2001 NDSHS also provides information
on explanatory variables such as household income (Income), age (Age1419,
Age2024, Age2529, Age3034, Age3539, Age4069, Age70), gender (Male), marital
status (Married, Divorce, Widow, Never Married), number of dependent children (#
Depchild), educational attainment (Degree), employment status (Working status),
and ethnicity (Aboriginal). A dichotomous variable (Decrim) is also defined
indicating whether a person resides in a decriminalized state.
y
i
The price of marijuana is obtained from the Australia Bureau of Criminal
Intelligence (ABCI, 2002) and the Australian Crime Commissions (ACC, 2003).
Four different prices by state are available: (i) price of head per ounce, (ii) price of
head per gram, (iii) price of leaf per ounce, and (iv) price of leaf per gram. These
prices are first converted to the same unit, price per ounce. Then, for each state, a
weighted average of (i), (ii), (iii), and (iv) is computed by using proportions of the
respondents’ form of purchase as weights. We also deflate each state’s weighted
average price of marijuana by the state’s CPI, and apply logarithmic function. The
final price of marijuana is denoted P
MAR
. A thorough discussion on Australia’s
marijuana price can be found in Clements (2004). State-level CPIs and state-level
unemployment rates are drawn from the Australia Bureau of Statistics (ABS, 2003b),
20
where the latter has its unit expressed in term of percentage (%). Table 1 provides
summary statistics of dependent and independent variables for all observations,
treatment observations and control observations, respectively.
21
Table 1 Summary Statistics of Dependent and Independent Variables
All Data (N = 14008)
Treatment (N = 2968)
Control (N = 11040)
Variable
Mean
S.D.
Mean
S.D.
Mean
S.D.
y
Decrim
P
MAR
Income
Age1419
Age2024
Age2529
Age3034
Age3539
Age4069
Male
Married
Divorce
Widow
# Depchild
Degree
Working
Status
Aboriginal
Unemploy-
ment rate
0.157
0.212
5.876
10.445
0.054
0.077
0.100
0.122
0.126
0.455
0.476
0.619
0.115
0.038
0.594
0.260
0.028
0.013
6.986
0.364
0.409
0.237
0.757
0.227
0.267
0.300
0.327
0.331
0.498
0.499
0.486
0.319
0.191
0.895
0.438
0.164
0.111
1.375
0.181
1
5.929
10.554
0.047
0.075
0.103
0.121
0.130
0.466
0.488
0.621
0.124
0.033
0.583
0.292
0.019
0.017
6.331
0.385
0
0.036
0.709
0.211
0.263
0.304
0.326
0.337
0.499
0.500
0.485
0.330
0.178
0.880
0.455
0.137
0.127
1.386
0.151
0
5.862
10.416
0.056
0.078
0.099
0.122
0.124
0.452
0.472
0.618
0.112
0.039
0.597
0.251
0.030
0.012
7.162
0.358
0
0.264
0.767
0.231
0.268
0.299
0.327
0.330
0.498
0.499
0.486
0.315
0.195
0.899
0.434
0.170
0.107
1.318
Note: S.D. stands for standard deviation.
22
1.3.4 Estimation Results
The maximum likelihood method is used to derive our parameter estimates. Because
we do not simultaneously observe and , the joint distribution of or y
*
1
y
*
0
) , (
0 1 ε ε
ρ
10
is not identified. Log-likelihood function can be constructed from four mutually
exclusive outcomes as follows:
∑ + ∑ +
∑ + ∑ =
= = = =
= = = =
0 , 0 :
00
0 , 1 :
10
1 , 0 :
01
1 , 1 :
11
1 0 1 0
1 0
log log
log log ) , , , , , , (
d
y
P
d
y
P
d
y
P
d
y
P
L
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
ρ ρ γ β β
α α
υ υ
where
) , , ( ) 1 , 1 Pr(
1 1
1
11
ρ γ β
α
υ
z x d
y
P
i i i
i i
+ Φ = = = =
6
) , , ) ( ( ) 1 , 0 Pr(
1 1
1
01
ρ γ β
α
υ
− + − Φ = = = =
z x d
y
P i i i
i
i
) , , ( ) 0 , 1 Pr(
0 0
0
10
ρ γ β
α
υ
− − + Φ = = = =
z x d
y
P i i i
i
i
) , , ) ( ( ) 0 , 0 Pr(
0 0
0
00
ρ γ β
α
υ
z x d
y
P i i i
i
i
− + − Φ = = = = .
6
ε υ
ρ
υ
i i
)
i i
υ
i
i
(
i i
υ
υ
i i i
d d
e
υ ε
ρ
υ ε
x
β
α z
γ
ρ Π
) ρ ,
z
, γ
x
β
α
Φ(
P 1
)
2
1
1 ( 2
1
1
2 2
1 1
1
2
1
1 1
1
11
2
1 2
1
−
− + −
+
∫
∞ −
∫
∞ − −
= + = . Similar
expressions can be obtained for the other three terms.
23
As discussed earlier, the proposed endogenous probit switching model nests three
other commonly used models in the empirical literature. In this section, we present
results for all four different models. In decreasing order of restrictiveness in
specification, the four models are: (i) a binary probit as in equation (1.3.11), (ii) a
bivariate probit model with endogenous treatment variable, (iii) a two-part model
without endogeneity of treatment, and (iv) an endogenous probit switching model
that allows for both the flexibility in marijuana participation behavior and
endogeneity of treatment.
Table 2 reports estimated coefficients for the marijuana equation, and the average
treatment effects for all four models. Table 3 reports the estimated location choice
model. Estimates of the marginal effects on marijuana participation probability are
provided in Table 4. They are computed at the mean of marijuana price, the mean of
household income, and for a reference person whom we define to be male, aged 14-
19 years old, never married, having less than university education, not unemployed
7
,
not of Aboriginal origin, and living in a non-decriminalized state.
7
A person who is not unemployed may do one of the following activities: employed part-time,
employed full-time, a student, a retiree, or someone who performs home duties.
24
Table 2 Coefficient Estimates for Marijuana Smoking Equation
Variable Binary Probit Bivariate
Probit
Two-Part:
Treatment
Two-Part:
Control
Switching:
Treatment
Switching:
Control
Decrim
P
MAR
Income
Age1419
Age2024
Age2529
Age3034
Age3539
Age4069
Male
Married
Divorce
0.173***
(0.033)
-0.338***
(0.057)
-0.008
(0.021)
1.952***
(0.205)
2.105***
(0.202)
2.113***
(0.200)
1.908***
(0.200)
1.848***
(0.200)
1.212***
(0.197)
0.238***
(0.028)
-0.470***
(0.037)
-0.001
(0.053)
0.187
(0.130)
-0.341***
(0.061)
-0.009
(0.022)
1.952***
(0.205)
2.105***
(0.202)
2.112***
(0.200)
1.908***
(0.200)
1.848***
(0.200)
1.211***
(0.197)
0.238***
(0.028)
-0.470***
(0.037)
-0.001
(0.053)
-5.305***
(0.810)
-0.016
(0.049)
1.962***
(0.418)
2.084***
(0.409)
2.023***
(0.404)
1.854***
(0.404)
1.894***
(0.402)
1.113***
(0.398)
0.201***
(0.060)
-0.429***
(0.080)
-0.051
(0.112)
-0.315***
(0.057)
-0.007
(0.024)
1.961***
(0.240)
2.109***
(0.237)
2.133***
(0.235)
1.920***
(0.234)
1.822***
(0.234)
1.237***
(0.232)
0.250***
(0.032)
-0.493***
(0.043)
0.008
(0.061)
-5.324***
(0.829)
-0.018
(0.052)
1.963***
(0.418)
2.084***
(0.409)
2.022***
(0.405)
1.853***
(0.404)
1.893***
(0.402)
1.113***
(0.398)
0.201***
(0.060)
-0.428***
(0.080)
-0.053
(0.112)
-0.409***
(0.060)
-0.022
(0.024)
1.915***
(0.237)
2.048***
(0.235)
2.068***
(0.233)
1.869***
(0.232)
1.774***
(0.231)
1.206***
(0.228)
0.236***
(0.032)
-0.470***
(0.044)
-0.005
(0.059)
25
Table 2 Coefficient Estimates for Marijuana Smoking Equation, Continued
Variable Binary Probit Bivariate
Probit
Two-Part:
Treatment
Two-Part:
Control
Switching:
Treatment
Switching:
Control
Widow
Degree
Working Status
Aboriginal
Constant
ρ
1 υ =
ρ
0 υ
ρ
1 υ
ρ
0 υ
Log Likelihood
Average Treatment
Effect (ATE)
-0.441***
(0.127)
-0.034
(0.034)
0.256***
(0.074)
0.282**
(0.107)
-0.468
(0.429)
-5218.030
0.037***
(0.0002)
-0.441***
(0.127)
-0.034
(0.034)
0.257***
(0.075)
0.281**
(0.107)
-0.451
(0.455)
-0.008
(0.077)
-11969.827
0.040***
(0.0002)
-0.561†
(0.310)
-0.262***
(0.073)
0.306
(0.187)
-0.304
(0.218)
29.338***
(4.832)
-1183.790
0.137***
(0.002)
-0.435***
(0.140)
0.047
(0.039)
0.249***
(0.081)
0.438***
(0.124)
-0.650
(0.457)
-3997.998
-0.561†
(0.310)
-0.262***
(0.073)
0.307
(0.188)
-0.308
(0.222)
29.486***
(5.021)
-0.012
(0.109)
-11929.312
0.163***
(0.002)
-0.422***
(0.136)
0.042
(0.038)
0.266***
(0.080)
0.377***
(0.123)
-0.011
(0.475)
-0.469**
(0.181)
Note: (1) Standard errors are in parentheses, (2) *** significant at 1% (two-tailed test), (3) ** significant at 5% (two-tailed test),
(4) * significant at 10% (two-tailed test), (5) † significant at 10% (one-tailed test)
26
Table 3 Coefficient Estimates for Location Choice Equation
Variable
Binary Probit Bivariate
Probit
Two-Part Switching
Income
Age2024
Age2529
Age3034
Age3539
Age4069
Male
Married
Divorce
Widow
# Depchild
Degree
Working Status
Aboriginal
Unemployment rate
Constant
N/A
0.065***
(0.019)
0.052
(0.059)
0.080
(0.055)
0.055
(0.054)
0.084
(0.055)
0.073
(0.044)
0.024
(0.025)
0.015
(0.035)
0.137**
(0.050)
-0.001
(0.074)
-0.026†
(0.016)
-0.013
(0.029)
-0.160†
(0.083)
0.316***
(0.107)
-0.286***
(0.011)
0.378†
(0.220)
0.065***
(0.019)
0.052
(0.059)
0.080
(0.055)
0.055
(0.054)
0.084
(0.055)
0.073
(0.044)
0.024
(0.025)
0.015
(0.035)
0.137**
(0.050)
-0.001
(0.074)
-0.026†
(0.016)
-0.013
(0.029)
-0.161†
(0.083)
0.316***
(0.107)
-0.286***
(0.011)
0.378†
(0.220)
0.065***
(0.019)
0.055
(0.059)
0.080
(0.055)
0.052
(0.054)
0.083
(0.055)
0.073†
(0.044)
0.027
(0.025)
0.019
(0.035)
0.140***
(0.050)
0.003
(0.074)
-0.025
(0.015)
-0.014
(0.029)
-0.155†
(0.083)
0.317***
(0.107)
-0.287***
(0.011)
0.381†
(0.220)
Note: (1) Standard errors are in parentheses.
(2) *** significant at 1% (two-tailed test)
(3) ** significant at 5% (two-tailed test)
(4) * significant at 10% (two-tailed test)
(5) † significant at 10% (one-tailed test)
27
Table 4 Marginal Effects for Marijuana Smoking Equation
Variable Binary Probit Bivariate
Probit
Two-Part:
Treatment
Two-Part:
Control
Switching:
Treatment
Switching:
Control
Decrim
P
MAR
Income
Age1419
Age2024
Age2529
Age3034
Age3539
Age4069
Male
Married
Divorce
0.067***
(0.013)
-0.127***
(0.021)
-0.003
(0.008)
0.352***
(0.020)
0.598***
(0.022)
0.598***
(0.022)
0.578***
(0.027)
0.570***
(0.029)
0.443***
(0.055)
0.085***
(0.010)
-0.157***
(0.012)
-0.0004
(0.020)
0.072
(0.051)
-0.128***
(0.022)
-0.003
(0.008)
0.351***
(0.021)
0.599***
(0.024)
0.599***
(0.023)
0.578***
(0.028)
0.571***
(0.029)
0.443***
(0.055)
0.085***
(0.010)
-0.157***
(0.012)
-0.001
(0.020)
-2.101***
(0.324)
-0.006
(0.019)
0.433***
(0.048)
0.523***
(0.044)
0.520***
(0.045)
0.507***
(0.049)
0.510***
(0.048)
0.388***
(0.098)
0.078***
(0.023)
-0.161***
(0.028)
-0.020
(0.044)
-0.118***
(0.021)
-0.003
(0.009)
0.351***
(0.022)
0.599***
(0.026)
0.601***
(0.025)
0.580***
(0.031)
0.567***
(0.034)
0.450***
(0.063)
0.089***
(0.012)
-0.163***
(0.013)
0.003
(0.023)
-2.112***
(0.339)
-0.007
(0.021)
0.439***
(0.169)
0.518***
(0.163)
0.514***
(0.161)
0.502***
(0.161)
0.505***
(0.160)
0.385**
(0.158)
0.078***
(0.024)
-0.161***
(0.032)
-0.021
(0.045)
-0.145***
(0.020)
-0.008
(0.008)
0.304***
(0.092)
0.628***
(0.089)
0.631***
(0.088)
0.604***
(0.087)
0.588***
(0.086)
0.451***
(0.082)
0.078***
(0.013)
-0.144***
(0.018)
-0.002
(0.021)
28
Table 4 Marginal Effects for Marijuana Smoking Equation, Continued
Variable Binary Probit Bivariate
Probit
Two-Part:
Treatment
Two-Part:
Control
Switching:
Treatment
Switching:
Control
Widow
Degree
Working Status
Aboriginal
-0.149***
(0.037)
-0.013
(0.013)
0.099***
(0.029)
0.110**
(0.042)
-0.148***
(0.037)
-0.013
(0.013)
0.100***
(0.030)
0.109**
(0.043)
-0.204†
(0.098)
-0.101***
(0.027)
0.121
(0.074)
-0.116
(0.080)
-0.147***
(0.041)
0.018
(0.015)
0.097***
(0.032)
0.172***
(0.049)
-0.206
(0.123)
-0.101***
(0.029)
0.122
(0.074)
-0.118
(0.088)
-0.131***
(0.049)
0.015
(0.013)
0.099***
(0.028)
0.143***
(0.044)
Note: Marginal effect is evaluated at the mean of marijuana price, the mean of household income, and for a reference person whom
we define to be male, aged 14-19 years old, never married, having less than university education, not unemployed, not of Aboriginal origin,
and living in a non-decriminalized state.
29
Our results show that decriminalization has positive and generally significant
impacts on marijuana smoking behavior although their magnitudes differ across
different models. We present the average treatment effect (ATE) of
decriminalization for all models, as well as the marginal effect of the
decriminalization dummy for the probit and the endogenous bivariate probit models.
However, we believe that ATE should provide a more accurate measurement
because it represents the effect of decriminalization for a randomly selected
individual from the population as opposed to just focusing on a reference person. As
a result, we will only discuss each model’s estimated ATE here. When using a
simple binary probit model without accounting for endogeneity of treatment and
flexibility in behavior, ATE is estimated to be 3.7%. This is very similar to those
estimated by Cameron and Williams (2001), and Zhao and Harris (2004). When
accounting for endogeneity of treatment as in model (ii), ATE rises to 4%. However,
when allowing for behavioral differences between the treatment and the control
groups but ignoring endogenous treatment as in the two-part model, we obtain ATE
of 13.7%. Finally, in the most general model (iv), our estimated ATE is found to be
16.3%. All these figures are statistically significant at 1% level. It is clear that the
two-part and the endogenous probit switching models provide stronger support to the
opponents of marijuana decriminalization policy because they yield substantially
larger ATEs than the binary probit and the bivariate probit models.
30
For other explanatory variables, models (i) and (ii) yield results that are similar while
models (iii) and (iv) generate comparable outcomes. The coefficients of P
MAR
are
negative and significant for all models, implying that there is negative own price
responsiveness. For a reference person, a 10% increase in the price of marijuana
decreases the probability of using it by 1.27% and 1.28% in the binary probit and the
bivariate probit models, respectively. When allowing for behavioral differences
between the treatment and the control groups alone or together with taking into
account possible endogeneity, we find negative own price effect to be much larger in
decriminalized states. In particular, for a reference person, as marijuana price
increases by 10%, the probability of using the drug is estimated to fall by
approximately 21% in decriminalized states but only drop by 1.18% to 1.45% in
non-decriminalized states. The higher price responsiveness in the decriminalized
states is expected, as we would anticipate that price plays a much smaller role in the
non-decriminalized states where the risk premium of being caught is expected to
play a bigger role as a proxy of price in the smoking decision. This negative price
effect is also found in other studies. For example, Cameron and Williams (2001),
and Williams (2004) both discover that increase in the price of marijuana lessens the
likelihood of using it.
The coefficient of household income is negative but insignificant for all models. It
indicates that income effect is absent in this study. Existing literature also reports
mixed results on income. For example, the full-sample estimation by Saffer and
31
Chaloupka (1998) finds that income has insignificant effect on the probability of
marijuana use while Pacula (1998), and Thies and Register (1993) both report
significant negative income effect.
Age is an important determinant of marijuana smoking. Cameron and Williams
(2001) and Williams (2004) both report that the probability of participating in
marijuana peaks for people in the 20-24 years old age-group, and then monotonically
declines for subsequent ages. Our maximum likelihood coefficients in Table 2
obtain this same finding for the treatment group of models (iii) and (iv). However,
for the rest of the models (i.e. model (i), model (ii), the control group of model (iii),
and the control group of model (iv)), the peak of smoking prevalence is found to
occur among those aged 25-29 years old. Our results indicate that young adults have
the highest risk of becoming marijuana smokers in decriminalized states while adults
have the greatest exposure in non-decriminalized states.
Shifting our discussion towards gender, it is commonly presented in existing
literature that males are more likely to be marijuana smokers than females. Our
positive and significant coefficient on the gender dummy variable supports this
point. We find that the impact of gender on marijuana smoking decision is quite
uniform across models as well as across the treatment and the control groups.
32
Marital status also influences marijuana usage. Married individuals are less likely to
use marijuana compared to their never married counterparts. For model (i), model
(ii), the control group of model (iii), and the control group of model (iv), we find that
widowed individuals are less inclined to be marijuana smokers. Finally, we find no
difference in cannabis usage between those who are divorced and those who are
single across all models.
Educational attainment does not seem to play a role in the decision to consume
marijuana when using models (i) and (ii). However, when allowing for behavioral
differences due to policy’s change as in model (iii) and possible endogeneity of
treatment as in model (iv), we find that the effects of education do differ across the
treatment and the control groups. For those who live in decriminalized states, having
a university degree substantially reduces their likelihood of becoming marijuana
smokers. On the other hand, in non-decriminalized states, there is no difference in
marijuana smoking prevalence between those with and without tertiary education.
In this paper, Working Status is assigned a value of 1 if that person is unemployed
and a value 0 otherwise. The coefficient of Working Status is found to be positive
and significant for binary probit, bivariate probit, the control group of two-part, and
the control group of endogenous probit switching models. The estimated marginal
effect on Table 4 suggests that a reference person can experience up to 10% higher
chance of becoming marijuana smoker when he is unemployed. On the contrary, we
33
find no evidence that being unemployed leads to higher prevalence of the drug for
the treatment group of model (iii) and (iv).
Finally, we turn to the ethnic variable. With models (i) and (ii), we find that being
an Aboriginal or Torres Strait Islander has positive and significant effect on
participation in marijuana use. However, when using models (iii) and (iv) to do the
estimation, we find that in decriminalized states respondents with this ethnic origin
have more or less the same probability of becoming marijuana smokers as
individuals from other ethnic backgrounds.
1.4 Nonparametric Specification: Propensity Score
Matching
1.4.1 Description of the Model
When marijuana smoking equations, and , are unspecified, the ATE may still
be identifiable and estimable under the assumption that conditional on a set of
confounding variables,
y
i
*
1
y
i
*
0
{ }
z x w i i i
∪ = , (ignorable treatment selection,
see Heckman and Robb (1985), Rosenbaum and Rubin (1983)). In this section, we
use Rosenbaum and Rubin (1983)’s propensity score method to correct for selection
on observables, where the propensity score is defined as the conditional probability
of being assigned into treatment given the covariates. In our context, this is simply
the conditional probability of residing in decriminalized states given observable
d
y y
i
i i
C ) , (
*
0
*
1
34
} variables. Let . We denote the propensity score by
. Under the assumptions
{
z x w i i i
∪ = ) | 1 Pr(
w d i i
=
) (
w
p
i
1 ) | 1 Pr( ) ( 0 < = = <
w d w
p
i i i
, (1.4.1)
and
, (1.4.2)
w d
y y
i i
i i
| ) , (
*
0
*
1
C
we have
, (1.4.3) ) ( | ) , (
*
0
*
1
w
p
d
y y
i i
i i
C
and
. (1.4.4) ) ( |
w
p
d w i i i
C
Equation (1.4.4) establishes that conditioning on the propensity score, the
distribution of covariates must be the same across the treatment and the control
groups. In other words, given the propensity score, the assignment into treatment is
random. We compute ATE under the assumptions (1.4.1) and (1.4.2).
wi
Propensity score stratification matching can be implemented by following these
steps: (i) estimating the propensity score either parametrically or nonparametrically,
(ii) dividing the propensity score into different intervals such that for each interval
there is a presence of both treated and untreated units, (iii) within each stratum,
calculating means difference of the treatment and the control outcomes, and finally
(iv) computing ATET and ATE by simply taking the weighted average of these
35
differences with the weight being the frequency of treated observations or the
frequency of both treated and untreated observations in each interval respectively
8
.
1.4.2 Empirical Findings
We follow Dehejia and Wahba (1999) and Becker and Ichino (2002) in doing the
empirical estimation. First, we estimate the propensity score by running a binary
probit estimation given
9
wi
. This step provides us with the estimated propensity
score, , which we can use to plot histograms for both the treatment and the
control groups. We draw histograms in Figure 1 by focusing on a range of
propensity score between 0.05 and 0.45 because both the treated and the control units
are presented in this region.
^
) (
w
p
i
We proceed to calculate ATE, ATET, and their associated standard errors in Table
5
10
. Four different ranges of overlapping region (i.e. 0.05-0.45, 0.075-0.425, 0.05-
0.4, and 0.1-0.35) and two different ways of partitioning the propensity scores (i.e.
length of interval 0.025 and 0.05) are being considered in this study. In addition to
our manual partition, we also use STATA’s pscore command to divide propensity
scores into smaller stratums. The results in Table 5 demonstrate that ATE varies
between 0.059 and 0.112, while ATET fluctuates from -0.069 to -0.021.
8
See Cameron and Trivedi (2005), p.875-876, and Becker and Ichino (2002), p.7, for the exact
mathematical expression.
9
A study conducted by Newey, Powell, and Walker (1990) actually demonstrates that there is not
much difference in predicting the outcomes using parametric or nonparametric methods.
10
See Becker and Ichino (2002), p.7, for the computation of analytical standard error.
36
Figure 1 Histograms of Estimated Propensity Scores in the Overlapping Region
Treatment group
2
9
15
74
185
341
363
125
38
19
33
66
12
15
31
55
124
158
181
104
43
22
19 17
21
33
44
80
133
174 175
178
2
0 10 0 20 0 30 0 40 0
F req ue n c y
.05 .075 .1 .125 .15 .175 .2 .225 .25 .275 .3 .325 .35 .375 .4 .425 .45
p(w)
Control group
99
300
725
672
202
74
30 27
72
96
183
304
417
344
271
147
167
244
389
674
890
967
1118
1010
437
219
73
52
36
17 12
4 6 5 7 6 4 1
0 50 0 10 00
F req uen c y
.05 .075 .1 .125 .15 .175 .2 .225 .25 .275 .3 .325 .35 .375 .4 .425 .45
p(w)
37
Table 5 ATE and ATET using Propensity Score Stratification Method
Range of
Estimated
Propensity
Score
Number of
Treatment
observations
Number of
Control
observations
ATE ATET
0.05 – 0.45
Length of
interval 0.025
2810
10301
0.092***
(0.026)
-0.032
(0.053)
0.05 – 0.45
With STATA
interval
2810
10301
0.096***
(0.025)
-0.069*
(0.038)
0.075 – 0.425
Length of
interval 0.025
2432
9509
0.074***
(0.018)
-0.056**
(0.025)
0.075 – 0.425
With STATA
interval
2432
9509
0.086***
(0.020)
-0.056**
(0.025)
0.05 – 0.4
Length of
interval 0.05
2116
10295
0.098***
(0.017)
-0.024†
(0.014)
0.05 – 0.4
With STATA
interval
2116
10295
0.112***
(0.026)
-0.026**
(0.012)
38
Table 5 ATE and ATET using Propensity Score Stratification Method,
Continued
Range of
Estimated
Propensity
Score
Number of
Treatment
observations
Number of
Control
observations
ATE ATET
0.1 – 0.35
Length of
interval 0.05
1909
8263
0.067***
(0.016)
-0.021†
(0.015)
0.1 – 0.35
With STATA
interval
1909
8263
0.059***
(0.020)
-0.022†
(0.015)
Note: (i) *** significant at 1%
(ii) ** significant at 5%
(iii) * significant at 10%
(iv) † significant at 10% one-tailed test
(v) standard errors are in the parentheses
Since both ATE and ATET are highly sensitive to the way in which the propensity
score is stratified, we check whether condition (1.4.4) is met by testing whether there
is any difference in the first moment between the treatment and the control groups
for each interval. To do so, we employ both two-sample t-test and two-sample t-
squared generalized means test
11
. Regardless of how we or STATA partition the
propensity score, the t-test and the t-squared test always rejects the null hypothesis of
means’ equality between the treatment and the control groups. This result suggests
that conditioning on the propensity score the distribution of is different between
the treated and the control units, which implies that balancing condition is violated.
wi
11
The details of these tests are available per request from the authors.
39
The violation could be because our sample size is not large enough to perform
reliable nonparametric estimates or because conditional independence assumption
does not hold. Thus, the propensity score matching estimates of the ATE and ATET
could capture not only the treatment effect but also the impact of differences in
observable and unobservable covariates on smoking outcome.
1.5 Specification Analyses
1.5.1 Parametric Model Specification Test
One prominent feature of our analysis is that different models yield very different
inferences about the impact of decriminalization on marijuana smoking prevalence.
In this section, we conduct specification analysis to select a most appropriate
representation among different parametric models. Taking the endogenous
switching model (model (1.3.4)-(1.3.9)) as the maintained hypothesis that other
models are nested in, if
0 :
0 1
0
= = ρ ρ
υ υ
H
,
the model has the form of a generalized two-part model. If
ρ ρ β β
υ υ 0 1 0 1
*
0
and : = =
H
,
the endogenous switching model is reduced to the sample selection model (Amemiya
(1985)). If
0 and :
0 1 0 1
* *
0
= = = ρ ρ β β
υ υ
H
,
40
the endogenous switching model is reduced to the conventional dummy variable
approach (model (1.3.11)). To conduct specification tests for , , and , we
compute likelihood ratio statistics LR
1
, LR
2
, and LR
3
, respectively, where each of
them is associated with different degree of freedom.
H 0 H
*
0 H
* *
0
Let ( , , , , , , ) be the maximum likelihood estimates of model
(1.3.4)-(1.3.9). To test against model (1.3.4)-(1.3.9), the likelihood ratio statistic
is
^
*
1
α
^
*
0
α
^
*
1
β
^
*
0
β
^
*
γ
^
*
1
ρ
υ
^
*
0
ρ
υ
H 0
LR
1
= -2[ln L( , , , , , )
^
1
α
^
0
α
^
1
β
^
0
β
^
γ 0
^
0
^
1
= = ρ ρ
υ υ
– ln L( , , , , , , )]
^
*
1
α
^
*
0
α
^
*
1
β
^
*
0
β
^
*
γ
^
*
1
ρ
υ
^
*
0
ρ
υ
LR
1
~ . (1.5.1) χ
2
2
To test , the likelihood ratio statistic is
H
*
0
LR
2
= -2[ln L( , , = = , , )
^
1
α
^
0
α
^
1
β
^
0
β
^
β
^
γ
^ ^
0
^
1
ρ ρ ρ
υ υ
= =
– ln L( , , , , , , )]
^
*
1
α
^
*
0
α
^
*
1
β
^
*
0
β
^
*
γ
^
*
1
ρ
υ
^
*
0
ρ
υ
LR
2
~ . (1.5.2) χ
2
1 + K
To test , the likelihood ratio statistic is
H
* *
0
LR
3
= -2[ln L( , , = = , , )
^
1
α
^
0
α
^
1
β
^
0
β
^
β
^
γ 0
^
0
^
1
= = ρ ρ
υ υ
– ln L( , , , , , , )]
^
*
1
α
^
*
0
α
^
*
1
β
^
*
0
β
^
*
γ
^
*
1
ρ
υ
^
*
0
ρ
υ
41
LR
3
~ . (1.5.3) χ
2
2 + K
Table 6 provides the results of this specification analysis. The LR
1
and LR
3
firmly
reject the independence assumption between the errors of marijuana smoking
equation and the errors of location choice equation. The LR
2
and LR
3
also show that
individuals do behave differently if marijuana smoking is decriminalized. In other
words, our specification analysis appears to favor the endogenous probit switching
model over other parametric models namely the two-part model, the endogenous
bivariate probit model, and the binary probit model.
Table 6 Parametric Model Specification Test
Model under
Null Hypothesis
Two-Part
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
= =
≠
0
,
0 1
0 1
ρ ρ
β β
υ υ
Bivariate Probit
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
≠ =
=
0
,
0 1
0 1
ρ ρ
β β
υ υ
Binary Probit
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
= =
=
0
,
0 1
0 1
ρ ρ
β β
υ υ
Model under
Alternative
Hypothesis
Endogenous
Switching
Endogenous
Switching
Endogenous
Switching
Degree of Freedom
of the Chi-Square
Dist.
2
16
17
Likelihood Ratio
Test Statistics
LR
1
= 8.559**
LR
2
= 81.03***
LR
3
= 81.042***
Note: (1) *** significant at 1% (two-tailed test)
(2) ** significant at 5% (two-tailed test)
42
1.5.2 Nonparametric Kernel Consistent Model Specification Test
with Mixed Discrete and Continuous Data
In the previous section, we conduct parametric model specification test by using
likelihood ratio test to choose among nested parametric models. Although it is
computationally attractive, the parametric model specification test requires the user
to specify both the null hypothesis and the set of parametric alternatives. This set of
parametric alternatives is crucial because it is assumed that, when the null model is
rejected, the data generating process indeed follows this specified alternative.
However, it is possible that the data may in fact follow other alternative
specifications that have not been identified by the user and therefore are not captured
by the test. If the latter occurs, this parametric test is said to be inconsistent. As a
result, we can say that parametric model specification test does not have power in
every direction that departs from the null model.
Unlike its parametric counterpart, nonparametric specification test has power in
every direction that departs from the null model. This is because for this type of test
the alternative hypothesis is a nonparametric model, which neither places restriction
on the functional form nor makes any distributional assumption, allowing it to
capture all data generating processes. Nonparametric specification test is desirable
because it is always consistent as opposed to the test that is based on parametric
approach. We follow Hsiao, Li, and Racine (2006) and Li and Racine (2006) in
43
constructing a nonparametric kernel-based test for our application on marijuana
decriminalization policy and marijuana smoking prevalence.
According to previous studies, the test is constructed by first assuming that the
parametric model is
u x
m y
i i
i
+ = ) , ( β (1.5.4)
where ) , ( β
x
m
i
is a known function of unknown parameters β , β is a p-
dimensional vector, is a compact subset in Β
R
p
, and consists of both
continuous and discrete variables. To test parametric model (1.5.4) against
nonparametric alternative, the null hypothesis is
xi
) , ( ) | ( :
0
β
x
m
x
y E
H i i
i
= , for some Β ∈ β (1.5.5)
and the alternative hypothesis is
) , ( ) | ( :
1
β
x
m
x
y E
H i i
i
≠ , for all Β ∈ β .
(1.5.6)
There is a different way to write both the null and alternative hypotheses. The null
hypothesis can be expressed as
0 ) | ( :
0
=
x u
E
H i i
(1.5.7)
while the alternative hypothesis is
0 ) | ( :
1
≠
x u
E
H i i
. (1.5.8)
Hsiao, Li, and Racine (2006) and Li and Racine (2006) develop a consistent test
statistic based on the following conditional moment expression
44
{ } ) ( ) | (
x
f
x u
E
u
E
i i i i
(1.5.9)
where is a joint probability density function of . Because testing
is equivalent to testing
) (
x
f
i xi
0 ) | ( =
x u
E
i i
{ } 0 ) ( ) | ( =
x
f
x u
E
u
E
i i i i
and the test based on
the latter is easier to develop, our focus will be on (1.5.9). The sample analogue of
(1.5.9) is
∑
=
n
i
i i i i x
f
x u
E
u
n 1
) ( ) | (
1
. (1.5.10)
The test statistic based on (1.5.10) can be computed by replacing
u
with where
is a residual from parametric null model (i.e. ), and by using a
leave-one-out kernel estimator to estimate . The replacement of the
conditional mean function, , by the leave-one-out kernel estimator is
the main reason for naming this test a “nonparametric kernel-based” test. The leave-
one-out kernel estimator is simply equal to
i
^
ui
^
ui
) , (
^ ^
β
x
m y
u i
i
i
− =
) ( ) | (
x
f
x u
E
i i i
) ( ) | (
x
f
x u
E
i i i
∑
≠
n
i j
ij ij h j L K u
n
, ,
^
1
λ
(1.5.11)
where is a multivariate kernel function for continuous variables and is a
multivariate kernel function for discrete variables. If one further assumes product
kernel functions, (1.5.11) becomes
K ij h, L ij , λ
⎥
⎦
⎤
⎢
⎣
⎡
∏
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛ −
∏ ∑
= = ≠
) , , (
) (
1 1
1 1
^
λ s
d
js
d
is
r
s
s
c
js
c
is
s
q
s
n
i j
j x x
l
h
x x
k
h
u
n
(1.5.12)
45
where and q r is a dimension of continuous and discrete variables, and is used
to denote continuous and discrete variables, and
c d
hs λ s
is a smoothing parameter for
continuous and discrete variables s , and finally, ) ( ⋅ k and ) ( ⋅ l is a univariate kernel
function for continuous and discrete variables, respectively. The gaussian or normal
kernel function is used for continuous variables while, for discrete variables, two
different kernel functions are available depending on the type of regressor. When
dealing with unordered discrete regressors, the kernel function is
(1.5.13)
⎩
⎨
⎧ =
=
otherwise, ,
if , 1
) , , (
λ
λ
s
d
js
d
is
s
d
js
d
is
x x
x x
l
where . And, if the variable is an ordered discrete regressor, the kernel
function is given by
] 1 , 0 [ ∈
λ s
(1.5.14)
⎪
⎩
⎪
⎨
⎧ =
=
−
otherwise. ,
if , 1
) , , (
| |
λ
λ
d
js
d
is
x x
s
d
js
d
is
s
d
js
d
is
x x
x x
l
In our study of marijuana decriminalization policy and marijuana smoking
prevalence, all of the discrete regressors are unordered. Although age is
conventionally treated as an ordered discrete regressor, here we treat different age
categories as different regressors to capture the nonlinear effect of age on marijuana
consumption. Thus, (1.5.14) is irrelevant in our case. In term of smoothing
parameters ( )
λ λ r q h h
,..., , ,...,
1 1
, they can be obtained from a least square cross
validation method or an ad hoc plug-in method as long as they satisfy the condition
46
that and 0 →
hs
∞ →
h h
n
q
K
1
when ∞ → n . Combining the expressions of
residual and leave-one-out kernel estimator, (1.5.10) can be written as
. ) , , (
) (
1 1
) , , (
) (
1 1 1
1 1
^
1
^
2
1 1 1
^ ^
⎥
⎦
⎤
⎢
⎣
⎡
∏
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛−
∏ ∑∑ =
∑
⎪
⎭
⎪
⎬
⎫
⎪
⎩
⎪
⎨
⎧
⎥
⎦
⎤
⎢
⎣
⎡
∏
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛ −
∏ ∑ =
= = =≠
= = = ≠
λ
λ
s
d
js
d
is
r
s
s
c
js
c
is
s
q
s
j
n
i
n
i j
i
n
i
s
d
js
d
is
r
s
s
c
js
c
is
s
q
s
n
i j
j i n
x x
l
h
x x
k
h
u u
n
x x
l
h
x x
k
h
u
n
u
n
I
(1.5.15)
Hsiao, Li, and Racine (2006) and Li and Racine (2006) have proven the asymptotic
null distribution of test statistic . It is shown to be normally distributed with mean
zero and variance one,
I n
() ) 1 , 0 (
^
2 / 1
1
N
I h h n
J n n q
→ Ω = K (1.5.16)
where
∑∑ =
Ω
≠ i
ij ij h
i j
j i
q
L K u u
n
h h
2
,
2
,
^
2
^
2
2
1
^ 2
λ
K
. (1.5.17)
After replacing multivariate kernel functions by product kernels,
Ω
becomes
^
⎥
⎦
⎤
⎢
⎣
⎡
∏
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛ −
∏ ∑∑ =
Ω
= =
≠
) , , (
) (
1
2
1 1
2
2
^
2
^
2
2
1
^
λ
s
d
js
d
is
r
s
s
c
js
c
is
s
q
s
x x
l
h
x x
k
h
u u
n
h h
j
ij i
i
q
K
. (1.5.18)
We can say that test statistic is test statistic that has already been normalized.
Expressions (1.5.15) through (1.5.18) basically tells us that we can conduct the
model specification test by first computing test statistic , then , and finally .
Once is obtained, we can compare it to the critical values from the standard
J n I n
I n Ω
^
J n
J n
47
normal distribution. Previous studies show that the test based on (1.5.16) is
consistent because approaches
J n
∞ whenever is false, resulting in the definite
rejection of the null hypothesis in this case. There is, however, one major drawback
for the test based on (1.5.16). Under , converges to the standard normal
distribution at an extremely slow rate of
H 0
H 0 J n
( ) ) (
1
2 / 1
h h O q p
K . To overcome this slow
convergence problem, bootstrapping procedure is proposed as an alternative way to
approximate the finite-sample null distribution. In particular, a wild bootstrapping
method is strongly suggested by other studies
12
.
So far, we discuss a nonparametric kernel-based test that has previously been
proposed by others. This test has been designed for models with continuous
outcomes. However, for our application on marijuana decriminalization policy and
marijuana smoking prevalence, our dependent variable is not continuous rather it has
a binary choice structure. Therefore, we need to modify the above nonparametric
kernel-based test in order to make it workable for our binary outcome model.
In the preceding section, we use the likelihood ratio test to choose among nested
parametric models and it turns out that the endogenous probit switching model is the
most favorable one. In this section, we want to confirm the robustness of this result
by testing the endogenous probit switching model against the much wider range of
alternative models that are captured under the umbrella of nonparametric alternative.
12
See Li and Racine (2006), p.357 for detailed discussion of a wild bootstrapping procedure.
48
This can be done by conducting a nonparametric kernel-based test. If the test turns
out to favor the endogenous probit switching model, then we can make an even
stronger case that our proposed endogenous probit switching model is the most
appropriate one for the study of marijuana decriminalization policy and its impact on
smoking. On the other hand, if the endogenous probit switching model is rejected
and the test inclines towards nonparametric alternative, then more attention should
be given to nonparametric specification when studying marijuana decriminalization
policy. We consider both normal null distribution and bootstrap empirical
distribution when conducting our test. In what follows, we discuss the steps
involved in conducting the nonparametric kernel-based test. First, we show how to
compute , , and when the parametric null model is the endogenous probit
switching model. We then describe our proposed bootstrapping procedure in order
to obtain the bootstrap empirical distribution.
I n Ω
^
J n
Our goal is to test the null hypothesis of endogenous probit switching model against
the alternative hypothesis of nonparametric model. Model (1.3.4)-(1.3.9) represents
the endogenous probit switching model. We follow the existing literature in using
(1.5.15) to construct the test statistic . However, unlike previous studies that
replace
u
in (1.5.15) by , we need to replace by
I n
i
) , (
^ ^
β
x
m y
u i
i
i
− =
ui
49
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎧
=
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
⎟
⎠
⎞
⎜
⎝
⎛
Φ −
− +
Φ − = = −
=
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
Φ
+ +
Φ − = = −
=
0 if
1
) ( 1
) (
) 0 | (
and , 1 if
1
) (
) (
) 1 | (
^
2
0
2 / 1
^
^
^
0
^
0
^
0
^
2
1
2 / 1
^
^
^
1
^
1
^
1
^
d
z
z
x
y
d
y E y
d
z
z
x
y
d
y E y
u
i
i
i
i
i
i
i i
i
i
i
i
i
i
i i
i
ρ
γ
γ φ
ρ β
α
ρ
γ
γ φ
ρ β
α
υ
υ
υ
υ
(1.5.19)
because our parametric null model has a discrete choice structure. For kernel
functions, we use normal distribution for continuous explanatory variables while
only include (1.5.13) as a kernel function for discrete explanatory variables because
all of them are unordered regressors. In term of smoothing parameters
13
( )
λ λ r q h h
, , , , ,
1 1
K K , we can either use a least square cross validation method or an
ad hoc plug-in method to choose them. We decide to use an ad hoc plug-in method
because there are over 14,000 observations in our sample, making it very
computationally intensive to use a least square cross validation approach.
Furthermore, smoothing parameters obtained from these two approaches should be
quite similar when the number of observations is sufficiently large. A commonly
used ad hoc plug-in bandwidth for continuous explanatory variable is
13
The words “smoothing parameter” and “bandwidth” are used interchangeably in this paper.
50
) (
) 2 (
1
n
c h
l p
s s s
+
−
=
σ
where is the resulting ad hoc plug-in bandwidth for continuous explanatory
variable , is the scale factor that is set to 1.06 following a so-called “normal
reference rule of thumb”,
hs
s
cs
σ s
is the sample standard deviation for variable ,
stands for the number of observation which in our case is equal to 14008, is the
order of the kernel function that is defaulting to 2 for the normal kernel function, and
is the number of continuous explanatory variables. Our endogenous probit
switching model consists of 16 explanatory variables, of which 2 are continuous
variables and 14 are unordered discrete regressors. Thus, we set l to 2. Ultimately,
we use the following formula to compute ,
s n
p
l
hs
) 14008 ( 06 . 1 ) 14008 ( 06 . 1
6
1
) 2 2 * 2 (
1 −
+
−
= =
σ σ s s s h
. (1.5.20)
For discrete regressors, the ad hoc plug-in bandwidths are simply set at . We
plug from (1.5.19), from (1.5.20),
0 =
λ s
^
u
i h s
0 =
λ s
, the normal kernel function for
continuous variables, and the kernel function (1.5.13) for discrete variables into
(1.5.15) to compute . From , we can get by calculating (i.e. (1.5.17)-
(1.5.18)) and plugging it into (1.5.16). It must be noted that the same , ad hoc
plug-in bandwidths, and kernel functions are used when computing , , and
I n I n J n Ω
^
^
u
i
I n J n Ω
^
.
Once is determined, we can compare this test statistic to the critical values from
the standard normal distribution with mean zero and variance one.
J n
51
Even though approaches in distribution to the standard normal distribution, it is
done so only at a very slow convergence rate. Bootstrapping procedure can be used
to approximate the empirical null distribution, which in turn can substantially
improve the finite-sample performance of the test. Because our parametric null
model has discrete choice structure, our bootstrapping procedure will be different
from the wild bootstrapping procedure proposed by Hsiao, Li, and Racine (2006) and
Li and Racine (2006). We discuss the steps involved in conducting our
bootstrapping procedure below.
J n
(i) From maximum likelihood estimates
^
1
β ,
^
0
β ,
^
1 α
,
^
0 α
, γ
^
,
^
1
ρ
υ
,
^
0
ρ
υ
of the
endogenous probit switching model (1.3.4)-(1.3.9), we can compute
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
Φ
+ +
Φ = = = =
^
2
1
2 / 1
^
^
^
1
^
1
^
1
^
11
1
) (
) (
) 1 | 1 Pr(
ρ
γ
γ φ
ρ β
α
υ
υ
z
z
x
d
y p
i
i
i
i
i i
,
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
Φ
+ +
Φ − = = = =
^
2
1
2 / 1
^
^
^
1
^
1
^
1
^
01
1
) (
) (
1 ) 1 | 0 Pr(
ρ
γ
γ φ
ρ β
α
υ
υ
z
z
x
d
y p
i
i
i
i
i i
,
52
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
⎟
⎠
⎞
⎜
⎝
⎛
Φ −
− +
Φ = = = =
^
2
0
2 / 1
^
^
^
0
^
0
^
0
^
10
1
) ( 1
) (
) 0 | 1 Pr(
ρ
γ
γ φ
ρ β
α
υ
υ
z
z
x
d
y p
i
i
i
i
i i
, and
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
⎟
⎠
⎞
⎜
⎝
⎛
Φ −
− +
Φ − = = = =
^
2
0
2 / 1
^
^
^
0
^
0
^
0
^
00
1
) ( 1
) (
1 ) 0 | 0 Pr(
ρ
γ
γ φ
ρ β
α
υ
υ
z
z
x
d
y p
i
i
i
i
i i
.
(ii) This step involves finding bootstrapping samples { }
x
y
i
b
i
n
i
,
1 =
where y
b
i
is
the dependent variable from bootstrapping,
xi
is the explanatory
variables from the original data, and n is the number of observations that
is equal to 14008 in our case. We obtain y
b
i
by randomly drawing it from
a binomial distribution. In particular,
if , , and 1 =
d i
⎪
⎩
⎪
⎨
⎧
=
^
01
^
11
0
1
p y probabilit with
p y probabilit with
y
i
i b
i
if , . 0 =
d i
⎪
⎩
⎪
⎨
⎧
=
^
00
^
10
0
1
p y probabilit with
p y probabilit with
y
i
i b
i
53
(iii) Using the bootstrapping samples { }
x
y
i
b
i
n
i
,
1 =
from (ii), we estimate the
endogenous probit switching model (i.e. model (1.3.4)-(1.3.9)) by
maximum likelihood method. We obtain a new set of maximum
likelihood estimators, which we denote
1
β
b
, , , , , ,
and .
^ ^
0
β
b
^
1
α
b
^
0
α
b
^
γ
b
^
1
ρ
υ
b
^
0
ρ
υ
b
(iv) We use the maximum likelihood estimators , , , , , ,
and from step (iii) and the bootstrapping samples {}
x
y
i
b
i
n
i
,
1 =
from step
(ii) to compute bootstrapping residual as follows. If 1 =
d i
,
^
1
β
b
^
0
β
b
^
1
α
b
^
0
α
b
^
γ
b
^
1
ρ
υ
b
^
0
ρ
υ
b
^
u
b
i
.
1
) (
) (
) 1 | 1 Pr(
) 1 | (
2
^
1
2 / 1
^
^
^
1
^
1
^
1
^
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
Φ
+ +
Φ − =
= = − =
= − =
ρ
γ
γ φ
ρ β
α
υ
υ
b
i
b
i
b
b
i
b
b
b
i
i
b
i
b
i
i
b
i
b
i
b
i
z
z
x
y
d
y y
d
y E y
u
54
If , 0 =
d i
.
1
) ( 1
) (
) 0 | 1 Pr(
) 0 | (
2
^
0
2 / 1
^
^
^
0
^
0
^
0
^
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎟
⎠
⎞
⎜
⎜
⎜
⎝
⎛
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
−
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
Φ −
− +
Φ − =
= = − =
= − =
ρ
γ
γ φ
ρ β
α
υ
υ
b
i
b
i
b
b
i
b
b
b
i
i
b
i
b
i
i
b
i
b
i
b
i
z
z
x
y
d
y y
d
y E y
u
(v) Once bootstrapping residual is obtained, we can compute
I
b
n
,
J
b
n
, and
in (1.5.15)-(1.5.17) by simply replacing
^
u
i
by . It must be noted
that the same ad hoc plug-in bandwidths are used here.
^
u
b
i
^
Ω
b
^
u
b
i
(vi) We repeat steps (ii) through (v) a large number of times. In particular,
we repeat these steps 399 times as recommended by Hsiao, Li, and
Racine (2006) and Li and Racine (2006). This step should provide us
with test statistics { }
J
b
nj
399
1 j =
, which we can later sort them in sequential
order.
55
(vii) We use these sorted test statistics { }
J
b
nj
399
1 j =
to construct a bootstrap
empirical distribution. We conduct a two-sided test by comparing
J n
to
the 1%, 5%, and 10% critical values from the bootstrap empirical
distribution. The null hypothesis of endogenous probit switching model
is rejected if
J n
is greater than the critical values. On the other hand, if
J n
is smaller than the critical values, our specification test indicates that
the endogenous probit switching model is more favorable than its
nonparametric alternative.
We follow the above steps in conducting our nonparametric kernel-based test. Both
the computation of test statistics and the bootstrapping procedure were undertaken
using R software. Table 7 provides the results of our nonparametric kernel-based
test
14
. Our test statistic is equal to 1.190. When the asymptotic normal
distribution is used as the null distribution, we obtain a p-value of 0.117. Because
the p-value of 0.117 is greater than 0.005, 0.025, and 0.05, we cannot reject the null
hypothesis of endogenous probit switching model at 1%, 5%, and 10% significant
levels (using two-sided test).
J n
14
The details of these specification tests and the R code are available per request from the authors.
56
Table 7 Nonparametric Kernel-Based Test
Bandwidths (i.e. from (1.5.20) and
h s
0 =
λ s
)
P
MAR
h 0.051
h Income
0.163
Statistics
I n
7.402*10
-6
Ω
^
6.342*10
-5
J n
1.190
P-value when normal distribution is the null distribution
P-value 0.117
Critical values and P-value when using bootstrap empirical distribution
Two-sided critical value at 1% 2.627
Two-sided critical value at 5% 2.219
Two-sided critical value at 10% 1.754
P-value 0.213
Notes: (1) When normal distribution is the null distribution, P-value = |) (| 1
J n
Φ − .
(2) When using bootstrap empirical distribution as null distribution, P-value is
the proportion of bootstrap test statistic, , that is greater than .
J
b
n
| |
J n
When the bootstrap empirical distribution is used, the two-sided critical values at
1%, 5%, and 10% are found to be 2.627, 2.219, and 1.754, respectively. We can
compare these critical values to our test statistic , which is equal to 1.190. It turns
out that is always smaller than the above critical values, leading to an unanimous
J n
J n
57
In this study, we use the 2001 wave of National Drug Strategy Household Survey
(NDSHS) to empirically examine the impact of marijuana decriminalization policy
on marijuana smoking prevalence in Australia. Our main econometric focuses are on
program evaluation and model specification test for mixed data. With regards to
program evaluation, we adopt both parametric and nonparametric approaches in
evaluating the impact because there are both advantages and disadvantages
associated with each of them. In particular, the benefits of the parametric approach
are the shortcomings of the nonparametric approach and vice versa.
acceptance of the parametric null model at 1%, 5%, and 10% significant levels.
Furthermore, a p-value of 0.213 strongly confirms that our endogenous probit
switching model is more favorable than its nonparametric alternative.
The tests based on normal null distribution and bootstrap empirical distribution
suggest exactly the same story. That is, for our study of marijuana decriminalization
policy and marijuana smoking prevalence, the endogenous probit switching model is
superior to its nonparametric alternative. Furthermore, when we combine the results
of nonparametric kernel-based test in this section to those of the parametric model
specification test in section 1.5.1, it is strongly evident that our proposed endogenous
probit switching model is the most appropriate model for the study of marijuana
decriminalization policy and its impact on smoking.
1.6 Conclusion
58
Postulating a parametric model to take account selection on observables and
unobservables, we can obtain consistent estimates of the parameters to allow the
computation of the ATE conditional on observable covariates and unconditional
ATE. Thus, whether or not we have overlapping x between the treatment and the
control group is not a problem. Furthermore, parametric estimation allows us to
investigate the behavioral impacts of other explanatory variables on smoking
outcome in addition to evaluating the effect of decriminalization as shown in Table
2. However, the main drawback is that both functional form and distributional
assumptions are imposed. Four different parametric models have been proposed.
They include binary probit, endogenous bivariate probit, two-part, and endogenous
probit switching models. For nonparametric model, the key advantage is that neither
functional form nor distributional assumption is needed. The disadvantages are that
we need to make an assumption about conditional independence, the impacts of other
socio-demographic variables on marijuana smoking cannot be estimated, and finally,
it cannot take account of selection on unobservables. Our nonparametric model is
the propensity score stratification matching.
Another main focus of our study is to examine the reliabilities of these different
models through specification tests. First, we use the likelihood ratio test to choose
among nested parametric models. Then, we construct a nonparametric kernel-based
test to select between a parametric null model and its nonparametric alternative. Our
nonparametric kernel-based test is different from the existing literature because it is
59
able to accommodate a model with discrete choice structure. Both asymptotic
normal distribution and bootstrap empirical distribution are used as the null
distribution of the test statistic. To obtain the bootstrap empirical distribution, we
propose a bootstrapping procedure that also takes account of our model’s discrete
choice structure.
Our specification analyses unanimously favor the endogenous probit switching
model over other parametric and nonparametric models. The endogenous probit
switching model, which takes into account selection on observables and
unobservables, suggests that decriminalization policy leads to higher marijuana
smoking participation. It indicates that on average living in decriminalized states
significantly increases the probability of smoking marijuana by 16.3%. This
estimate is larger than those obtained from the binary probit model, the bivariate
probit model, the two-part model, the propensity score matching and other existing
studies. For example, using the same dataset from Australia, the estimated marginal
effects of decriminalization from Cameron and Williams (2001) and Zhao and Harris
(2004) are less than 3%, which is similar to our binary probit estimate. In the United
States, Saffer and Chaloupka (1995) estimate the average treatment effect of
decriminalization to be about 6-7% when using annual data and 4-5% when
employing monthly data.
60
Our study also gives interesting finding about the effect of marijuana price on
marijuana smoking decision. The endogenous probit switching model, which allows
for behavioral differences between the treatment and the control groups, finds the
negative own price responsiveness to be much larger in decriminalized states. For a
reference person, a 10% increase in the price of marijuana decreases the probability
of using it by 21% in decriminalized states and only by 1.45% in non-decriminalized
states. This result points to the fact that in non-decriminalized states price plays a
much smaller role where the risk premium of getting caught plays a much bigger role
as a proxy for price in the smoking decision. Unlike our paper, previous studies
cannot separately identify these own price effects for the treatment and the control
groups because they do not allow for the flexibility in marijuana participation
behavior.
In conclusion, the discrepancies of our findings to the existing literature, both in
terms of decriminalization impact and own price effect, could be due to the use of
different data sets. It could also be due to different model specifications as
demonstrated in this paper. However, our specification analyses appear to favor the
model that allows both the endogeneity of the decriminalization dummy and the
flexibility of different behavioral patterns due to changes in legal or institutional
environment.
61
2 Evaluation of Malaysian Capital Controls in the
Short, Medium, and Long Runs
2.1 Introduction
The Asian financial crisis, which began in 1997, negatively influenced many
economies in Asia. At that time, some analysts thought that it would take a
generation for these countries to recover. Now, it is 10 years after the crisis and
these countries have shown a strong sign of recovery. For the four most-affected
Asian nations including Indonesia, Korea, Malaysia, and Thailand, their real GDP
growth in 2006 were at 5.5%, 5%, 5.9%, and 5%, respectively. All of them have
experienced current account surplus, strong private sector consumption, large inflow
of foreign capital, and the appreciation of their currencies against the US dollar.
Although these economic indicators demonstrate that the region has bounced back
strongly post-crisis, the UN’s Economic and Social Commission for Asia and the
Pacific has issued a report warning that these crisis-affected countries, except
Malaysia, are showing sign of renewed vulnerability in 2006
15
. In particular,
Indonesia, Korea, and Thailand are now facing the risk due to a decline in the ratio
of foreign reserves to short-term debt and an increase in the short-term capital
inflows that can possibly be reversed in the future. In addition to good economic
performances, Malaysia’s stock market also appears to be more favorable than its
15
See Economic and Social Survey of Asia and the Pacific 2007, United Nations Economic and
Social Commission for Asia and the Pacific.
62
neighboring countries. According to the International Herald Tribute
16
, Malaysian
stocks are the only ones in Asia that have been rising for 11 straight months. In sum,
even though those countries that were at the front-line of the financial crisis have
shown a strong sign of recovery, Malaysia seems to outperform the others after a
decade of economic downfall in Asia.
One reason for different countries’ varying economic performances may be the way
in which each of them handled the crisis a decade ago. Indonesia, Korea, and
Thailand sought IMF assistance where the IMF’s policy prescription called for tight
monetary and fiscal policies, floating exchange rate, openness of the financial
market, and various institutional and financial reforms. Malaysia took an alternative
path towards the recovery. Initially, it adopted somewhat similar policy to the IMF
prescription. However, in September 1998, Malaysia decided to fix its exchange rate
at RM 3.80 per US$ and impose capital controls.
There is no uniform consensus among economists regarding the effect of capital
controls. From the opponents’ point of view, a country is worse off from the
implementation of capital controls because these policies can reduce investor’s
confidence in a country, generate more capital flight, induce higher domestic
inflation, and give a country the incentive to postpone the restructuring of its
fundamentals. The supporters of these policies, on the other hand, believe that
16
See International Herald Tribute, Around The Markets: Malaysia stocks look set to continue
winning streak, June 1, 2007.
63
capital controls actually have positive impact on the economy. In particular, a
country which adopts these policies can regain internal balance without having to
worry about exchange rate volatility. Detailed discussion of capital controls, their
associated benefits and costs can be found in Neeley (1999) and Hartwell (2001).
This paper aims to evaluate the effectiveness of capital controls in restoring
Malaysian economy as compared to the IMF programs used in Indonesia, Korea, and
Thailand. Our study closely follows Kaplan and Rodrik (2000). In particular, we
employ the same program evaluation technique, namely “time-shifted” difference-in-
difference estimation, in studying the above issue. There are two main contributions
to the paper. First, we include one to six years of treatment periods. Unlike Kaplan
and Rodrik (2000) that considers just one year of treatment period, our paper can
assess the effectiveness of capital controls in the short, medium, and long runs.
Second, in addition to the “time-shifted” difference-in-difference estimation, we also
employ a “conventional” difference-in-difference estimation in our study. Because
the time-shifted approach compares Malaysia post-September 1998 to Indonesia
post-October 1997, Korea post-December 1997, and Thailand post-August 1997, its
point estimates can be considered an upper bound for the impact of Malaysian capital
controls relative to the IMF programs. On the contrary, the conventional approach
uses post-September 1998 as a common benchmark to compare the four Asian
countries. The estimated results from this approach can be regarded as a lower
bound for the effect of the controls policy. Hence, one important feature of our study
64
is that it is able to capture the real effect of Malaysian capital controls which lies
between these two bounds.
As mentioned earlier, the main focus of this paper is to check whether the two ways
of handling the crisis (i.e. capital controls versus the IMF programs) lead to varying
economic performances in the short, medium, and long runs. The point estimates
from the time-shifted approach provide strong evidence in favor of capital controls in
the short and medium runs. In particular, Malaysia was able to stabilize both the
exchange rate and interest rate, and control inflation more effectively than Indonesia,
Korea, and Thailand. Malaysia also yielded higher level of export, import, industrial
production, and share price compared to its neighboring countries. When using the
conventional approach to do the estimation, the performance of capital controls
relative to the IMF programs is found to be mixed depending on the economic
measure of interest, the treatment period considered as well as the set of IMF
countries used for comparison.
The remainder of this paper is organized as follows. Section 2.2 gives a brief review
of Malaysian capital controls. Section 2.3 discusses our data and explains the
reasons for including both the time-shifted and conventional difference-in-difference
estimations in this study. Section 2.4 presents the econometrics models. Estimation
results and interpretations are reported in Section 2.5. Section 2.6 provides a
conclusion.
65
2.2 Malaysian Capital Controls
After the Thai crisis in 1997, Malaysia initially followed a conventional path, which
is similar to the IMF’s prescription for Indonesia, Korea, and Thailand. Interest rates
were raised in order to prevent the Malaysian Ringgit from further depreciating. In
addition, the government also ran a tight fiscal policy. This orthodox policy was
promoted by Anwar Ibrahim, a Deputy Prime Minister at that time, who supported
exchange rate flexibility and rejected capital controls. This policy, however, proved
unsuccessful. By the end of June 1998, Prime Minister Mahathir Mohammed, an
opponent to Anwar’s policy, formulated an alternative plan. His goal was to
implement an expansionary monetary policy via cutting interest rates. Nonetheless,
when interest rates were decreased, Malaysia experienced more speculative attacks
from the offshore markets. These speculative attacks caused more capital flight.
Thus, the crisis intensified in Malaysia while Thailand and Korea were already on
the path towards recovery. In response, Malaysia imposed capital controls in
September 1998.
The main purpose of these controls was to stop speculative attacks against the
Ringgit and at the same time leave foreign direct investment unaffected. The
Malaysian currency was fixed at RM 3.80 per $US. All sales of Ringgit were
required to go through authorized domestic intermediaries. This made offshore
trading illegal. All Ringgit assets abroad also needed to be repatriated back home.
Moreover, to prevent capital outflows, short-term investors had to keep their capitals
66
in Malaysia for a minimum period of one year. It, however, allowed for the
repatriation of profits earned from foreign direct investment (FDI). Many believe
that this feature of the controls leaves Malaysian FDI unaffected. Nevertheless, it is
still an open question whether foreign investors perceive Malaysia as a less desirable
destination for FDI. We will briefly examine FDI data later in the paper. In
February 1999, there was a change in the implementation of capital controls. In
particular, the one-year ban on repatriation of investment was replaced by a
graduated tax.
Most initial reactions from the international communities were against the policy.
For example, according to Kaplan and Rodrik (2000), Morgan Stanley dropped
Malaysia from its international index while Moody’s downgraded Malaysian
securities. The IMF mentioned that any restriction on capital movement is not
conducive for building investor confidence. On the supporting side, a prominent
economist like Paul Krugman claimed to highly advocate temporary restriction on
the ability of investors to pull money out of crisis economies – a curfew on capital
flight – as part of a recovery strategy (see Krugman 1998). After many signs of
recovery in Malaysia, the attitudes towards capital controls were not as hostile as
before.
67
2.3 Data and Explanation for using the Two Approaches
This paper employs monthly data from Indonesia, Korea, Malaysia, and Thailand,
which spans from January 1993 to June 2005. The data come from the IMF’s
International Financial Statistics. We compare Malaysian capital controls to the IMF
programs by focusing on eight different measures of economic performance. These
measures include (i) exchange rate, (ii) foreign reserves, (iii) interest rate, (iv)
inflation, (v) merchandise export, (vi) merchandise import, (vii) industrial production
index, and (viii) share price index. For variables (vii) and (viii), the data is only
available for Korea and Malaysia
17
. Thus, for these two aspects of the economy, the
comparison will only be made between Korea and Malaysia. The definitions of all
variables are listed in Appendix 3.
Next, we want to explain the reasons for using both the time-shifted and
conventional difference-in-difference estimations to perform our empirical analysis.
There are both supporting and opposing arguments associated with each approach.
The shortcoming of the time-shifted approach is the strength of the conventional
approach and vice versa. Thus, for comparison and completeness purposes, we
include both of them in our study. In what follows, we begin our discussion with the
time-shifted approach then move on to the conventional approach.
17
Both Central Bank of Indonesia and Central Bank of Thailand provide no information on industrial
production index. Jakarta Stock Exchange and Stock Exchange of Thailand provide stock price
indexes for these two countries respectively; however, their formulas for computing monthly
weighted average are different from the International Financial Statistics. As a result, for consistency,
we only include the information from the IFS in this paper.
68
The main justification for using the time-shifted difference-in-difference estimation
is the argument that each country entered the financial crisis at a different time.
Kaplan and Rodrik (2000) argue that Malaysia’s financial crisis occurred at the time
when Korea and Thailand had begun to recover. We present some important
empirical observations to support this view. In 1997, when the crisis hit Asia, central
banks in the affected countries initially responded by using their foreign reserves to
defend fixed exchange rates. However, this approach was proven unsuccessful. A
large amount of foreign reserves was depleted during a short period of time as
demonstrated in Figure 2. From this figure, one may view the timing of the financial
crisis to be different across the four countries. Thailand’s foreign reserves fell
dramatically in July 1997 but started to increase again at the beginning of 1998.
Korea, on the other hand, experienced significant drop of foreign reserves at the end
of 1997 with another round of accumulation in the early 1998. While Thailand and
Korea already set off for their recovery in the middle of 1998, Malaysia still suffered
from sizeable loss of foreign reserves. In fact, the loss of foreign reserves in
Malaysia intensified during the period prior to the implementation of capital
controls.
69
Figure 2 Foreign Reserves (in log scale)
9 10 11 12
ln(Foreign Reserve)
1993m1 1995m1 1997m1 1999m1 2001m1 2003m1 2005m1
Thailand Malaysia
Korea Indonesia
Note: The vertical line represents August 1998, one month before the implementation of Malaysian
capital controls
As more foreign reserves were depleted, central banks allowed their exchange rates
to depreciate and float freely. Figure 3 displays these countries’ exchange rates.
There is a variation in the timing at which each country’s exchange rate was most
depreciated. Thailand and Korea experienced the greatest depreciation at the
beginning of 1998 thereafter their exchange rates started to appreciate gradually. For
Malaysia, its currency was still under intense speculative attacks in August 1998, a
month prior to the imposition of capital controls.
70
Figure 3 Exchange Rates (National currency / US$)
0 50 00 10 000 15 0 00
Ru pia h / US $
1993m1 1995m1 1997m1 1999m1 2001m1 2003m1 2005m1
Indonesia Exchange Rate
80 0 10 0 0 12 0 0 14 00 16 0 0 18 0 0
Wo n / U S $
1993m1 1995m1 1997m1 1999m1 2001m1 2003m1 2005m1
Korea Exchange Rate
71
Figure 3 Exchange Rates (National currency / US$), Continued
2.5 3 3.5 4 4.5
R ing git / U S $
1993m1 1995m1 1997m1 1999m1 2001m1 2003m1 2005m1
Malaysia Exchange Rate
20 30 40 50 60
Ba h t / U S $
1993m1 1995m1 1997m1 1999m1 2001m1 2003m1 2005m1
Thailand Exchange Rate
Note: The vertical line represents August 1998, one month before the implementation of Malaysian
capital controls
72
Another piece of evidence in favor of different crisis timing is the variation of the
date that each country either requested the IMF assistance or imposed capital
controls. These two ultimate crisis solutions got implemented at the time when other
initial approaches failed to stabilize the economies (i.e. at the time when the
economies were at the most vulnerable positions). Table 8 summarizes these dates.
Table 8 Important Dates
Country Date that a country seek IMF
assistance
Date of IMF approval of the
program
Indonesia
October 8, 1997
November 5, 1997
Korea
November 21, 1997 December 4, 1997
Thailand
July 28, 1997 August 20, 1997
Date of capital controls
implementation
Malaysia
September 1, 1998
Note: This table comes from Kaplan and Rodrik (2000)
The three pieces of evidence that we present above seem to support the view that
each country entered the financial crisis at a different time, which in turn suggests
that we should compare the economic performances of Malaysia post-September
1998 to Indonesia, Korea, and Thailand post-IMF assistance. We choose post-IMF
assistance to be a period of time starting from the date that a country seeks help from
the IMF. To be more specific, we assess the average economic performances of
Malaysia post-September 1998 to Indonesia post-October 1997, Korea post-
73
December 1997, and Thailand post-August 1997. A unique program evaluation
technique namely time-shifted difference-in-difference estimation is needed for this
purpose.
This time-shifted approach does not come without criticism. Since the crisis hit
Thailand, Malaysia did not just sit still and take no action until the imposition of
capital controls. It initially tried to implement other policies that are quite similar to
the IMF programs such as tightening its fiscal policy, raising interest rates, and
allowing its currency to depreciate. Subsequently, it conducted an expansionary
monetary policy via cutting interest rates. The fact that Malaysia had taken these
actions right after the Thai crisis and prior to the imposition of capital controls
weakens the claim that the timing of the crisis is different across different countries.
Another crucial argument against this method is the possibility that it may lead to an
unfair comparison between Malaysia and the three IMF countries. The time-shifted
approach compares Malaysia post-September 1998 to Indonesia post-October 1997,
Korea post-December 1997, and Thailand post-August 1997. Because capital
controls were imposed later than the IMF programs, it is possible that both internal
and external economic environments might have improved by September 1998. This
could lead to a much faster economic recovery in Malaysia when the time-shifted
difference-in-difference estimation is used to conduct the analysis. Thus, we expect
the point estimates from the time-shifted approach to be more favorable towards
Malaysian capital controls and less favorable towards the IMF programs. The point
74
estimates from this approach can be viewed as the upper bound for the impact of
Malaysian capital controls relative to the IMF programs.
The second method of estimation is the conventional difference-in-difference
estimation. Under this approach, we compare Malaysia to the three IMF countries
by using the same starting date. We choose a common benchmark date to be post-
September 1998. This is the period immediately after the imposition of capital
controls. There are a few reasons for choosing this date over post-August 1997 (i.e.
the period following the request made by Thailand for the IMF assistance). First, if
post-August 1997 is used, we will not be able to separate the impact of capital
controls from other prior policies that Malaysia had undertaken since the advent of
the Thai crisis. In such case, we will not be able to get a clear effect of the capital
controls policy, which undermines our goal. Second, it was not until October 1997
and December 1997 that Indonesia and Korea asked for the IMF bail-out. Thus, no
meaningful comparison can be made between these two countries and Malaysia if
post-August 1997 is used as a benchmark date.
The conventional approach is associated with certain pros and cons. On the
supporting side, the point estimates from the conventional approach do not provide
any favorable outcomes for Malaysia. This is because the same starting date is used
to compare every country. This conventional technique manages to overcome the
criticism underlying the time-shifted approach; however, it is still subject to certain
75
negative remarks. By using a common benchmark date to compare these countries,
the conventional approach does not take into account any possible variation in the
timing of the financial crisis across nations. This is the exact opposite of the time-
shifted approach, where it relies on the assumption that different countries entered
the financial crisis at a different time. If one seriously takes the view that the timing
of the crisis differs across countries, one can argue that in September 1998 Malaysia
was at its most vulnerable position while Korea and Thailand were already on the
path towards their recovery. The conventional approach that uses post-September
1998 as a common benchmark date to compare these four Asian countries can put
Malaysia at a disadvantage position while place Korea and Thailand at a more
advantage point. As a result, we can regard the point estimates from the
conventional approach as the lower bound for the impact of Malaysian capital
controls relative to the IMF programs.
One of the main highlights of this paper is the fact that we employ both the time-
shifted and conventional difference-in-difference estimations in our empirical
analysis. These two approaches provide us with the upper and lower bounds of the
impact, in which the real effect should fall between them. Thus, unlike previous
studies, this paper can accurately capture the real effect of Malaysian capital controls
in the short, medium, and long runs.
76
2.4 Econometrics Models
2.4.1 Models for Time-Shifted Difference-in-Difference Estimation
To evaluate the effect of Malaysian capital controls under the time-shifted approach,
we employ the following models:
ε
ϕ
δ
φ
γ β
α α α α
τ τ
it
kj
j
t j
k
t
k
i
t M
i
t M M T T K K
it
time
X Z
d d d d d d
y
+ + ∑∑ + +
+ + + + + =
> >
0
(2.4.1)
and
ε
ϕ
δ
φ γ β
α α
τ τ
it
kj
j
t j
k
t
k
i
t M
i
t M M
it
time
X Z d d d d
y + + ∑∑ + + + + + =
> >
0
, (2.4.2)
where
y
it
is a measure of economic performance (for example, exchange rate);
d K
, , and are country-specific dummies for Korea, Malaysia, and Thailand
with Indonesia being treated as an omitted category in (2.4.1);
d M d T
d
i
t τ >
is a time-varying dummy that equals 1 during the treatment period and 0
otherwise;
d d
i
t M τ >
is the interaction term between and ;
d M d
i
t τ >
Z
k
t
is a set of time-varying variables that captures external economic environments
(for example, interest rates in the US and Japan);
X
j
t
is a set of monthly dummies with January being treated as an omitted category;
77
time is a time trend; and
ε it
d
i
t τ >
Note that model (2.4.1) and (2.4.2) are similar to the specification proposed by
Kaplan and Rodrik (2000); however, there are some slight modifications here. We
use to represent time-varying post-treatment dummy. Because the notation
used is
is the error term.
τ i
instead of τ , it implies that the treatment was implemented in different
countries at different times. For Malaysia, the treatment period starts right after the
implementation of capital controls. For Indonesia, Korea, and Thailand, their
treatment periods begin after they start seeking help from the IMF. This paper also
includes one to six years of treatment periods in order to evaluate the short, medium,
and long runs’ impacts of the controls. In particular, we consider one year of
treatment period as short run, two to four years as medium run, and five to six years
as long run. For clarification, Table 9 lists the treatment periods for all four
countries under the time-shifted approach.
78
Table 9 Treatment Periods under the Time-Shifted Approach
Treatment Period (i.e. 1 =
>
d
i
t
τ
)
Country
Date that a
country seek
IMF assistance
1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Indonesia
Oct 8, 1997
10/97-9/98
10/97-9/99
10/97-9/00
10/97-9/01
10/97-9/02
10/97-9/03
Korea
Nov 21, 1997
12/97-11/98
12/97-11/99
12/97-11/00
12/97-11/01
12/97-11/02
12/97-11/03
Thailand
Jul 28, 1997
8/97-7/98
8/97-7/99
8/97-7/00
8/97-7/01
8/97-7/02
8/97-7/03
Date of capital
controls
implementation
1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Malaysia
Sep 1, 1998
9/98-8/99
9/98-8/00
9/98-8/01
9/98-8/02
9/98-8/03
9/98-8/04
79
For any period prior to or after the treatment, is equal to 0. On the other hand,
for any period within the treatment window,
d
takes the value of 1. The
coefficient of (i.e.
d
i
t τ >
i
t τ >
d
i
t τ >
β ) captures the effect of undergoing the IMF policy during
the financial crisis relative to the outcome in more normal time. In other words, β
establishes the baseline for post-treatment response. The coefficient of (i.e.
d d
i
t M τ >
γ ) is what we most interested in this paper. It captures the differential effect of
capital controls in Malaysia compared to the IMF programs in its three neighboring
countries. This coefficient tells us whether or not capital controls are more effective
than the IMF programs at stabilizing the economy. Because this coefficient is
obtained from the time-shifted difference-in-difference estimation, it should be
viewed as the upper bound for the effect of Malaysian capital controls. Table 10
clearly demonstrates the time-shifted difference-in-difference specification as well as
provides the meaning of each coefficient.
80
Table 10 Time-Shifted Difference-in-Difference Estimation
Measure of economic performance ( ) y
it
Indonesia Korea Thailand Malaysia
Before seeking
IMF assistance
or imposing
capital controls
α 0
α α K
+
0
α α T
+
0
α α M
+
0
After seeking
IMF assistance
or imposing
capital controls
β
α
+
0
β
α α
+ +
K 0
β
α α
+ +
T 0
γ β
α α
+ + +
M 0
After – Before
β
β
β
γ β +
Note: β captures the average impact of undergoing the IMF treatment while γ (i.e. ( γ β + ) - β =
γ ) measures the differential effect of capital controls in Malaysia compared to the IMF programs in
Indonesia, Korea, and Thailand.
We also control for external economic environments, . These variables include
the US and Japan’s interest rates, inflations, and industrial production indexes. The
inclusion of these variables is essential under the time-shifted difference-in-
difference estimation because many argue that Malaysia’s faster recovery may come
from the fact that it imposed capital controls at a more favorable time than Indonesia,
Korea, and Thailand when they first went to the IMF. Under this argument,
Malaysia’s recovery is induced by better external economic environments, and not
by capital controls. To overcome this contention, we control for external economic
factors. We choose economic indicators from the US and Japan to serve this purpose
Z
k
t
81
because they are the two largest economies in the world and also major trading
partners of Korea, Malaysia, and Thailand
18
.
We further include monthly dummies, , in order to control for any possible
seasonality effect. Time trend is also used to proxy any variable that may affect the
dependent variable but is not directly observable yet highly correlated with time.
Finally, we propose two econometrics models under the time-shifted approach.
Model (2.4.1) compares Malaysian capital controls to the IMF programs in
Indonesia, Korea, and Thailand. In this case, we must interpret
X
j
t
β and γ in term of
the average of the IMF programs in the three comparator countries. For Model
(2.4.2), we compare Malaysia to only Korea. Because there are more data available,
two additional dependent variables including industrial production index and share
price index can also be investigated here. Furthermore, we can infer from this
comparison whether Malaysia performs better or worse than Korea, a country that is
viewed as the IMF’s most successful case in the region.
18
China is also a major trading partner of Korea, Malaysia, and Thailand; however, no data is
available. According to the Central Bank of Indonesia, Indonesia’s major trading partners in non-oil
and gas are Philippines, Japan, Singapore, and Brunei. Thus, the US does not seem to be a major
trading partner of Indonesia.
82
2.4.2 Models for Conventional Difference-in-Difference Estimation
When the conventional approach is used as a method of estimation, we employ the
following models:
ε
ϕ
δ
φ
γ β
α α α α
τ τ
it
kj
j
t j
k
t
k
t M t M M T T K K
it
time
X Z
d d d d d d
y
+ + ∑∑ + +
+ + + + + =
> >
0
(2.4.3)
and
ε
ϕ
δ
φ γ β
α α
τ τ
it
kj
j
t j
k
t
k
t M t M M
it
time
X Z d d d d
y + + ∑∑ + + + + + =
> >
0
. (2.4.4)
The main difference between (2.4.3) and (2.4.4) in this section and (2.4.1) and
(2.4.2) in the previous section is the use of a time-varying dummy variable
instead of
d
. Because in this case we use the notation
d t τ >
i
t
τ
>
τ instead of
τ i
, it implies
that treatment periods are exactly the same for all four countries. Under the
conventional difference-in-difference estimation, we propose using post-September
1998 as a common benchmark date to compare the four Asian nations, which
explains why their treatment periods coincide. We list these mutual treatment
periods on Table 11. Again, one to six years of treatment periods are included here
to allow for the short, medium, and long runs’ assessment of Malaysian capital
controls.
83
Table 11 Treatment Periods under the Conventional Approach
Treatment Period (i.e. ) 1 =
>
d
t
τ
1 Year (Short run)
September 1998 – August 1999
2 Years (Medium run)
September 1998 – August 2000
3 Years (Medium run)
September 1998 – August 2001
4 Years (Medium run)
September 1998 – August 2002
5 Years (Long run)
September 1998 – August 2003
6 Years (Long run)
September 1998 – August 2004
For any time period within the treatment window, takes a value 1, otherwise it
assumes a value 0. The coefficient
d t τ >
β and γ from model (2.4.3) and (2.4.4) have a
slightly different interpretation from the description given for them in model (2.4.1)
and (2.4.2). The coefficient β now describes the average impact of the IMF
programs post-September 1998 relative to their pre-September 1998 performance.
The coefficient γ captures the differential effect of the controls policy in Malaysia
when comparing to the post-September 1998 IMF programs in its three comparator
countries. These estimated coefficients are treated as the lower bound for the impact
of Malaysian capital controls because they are derived from the conventional
method. Table 12 summarizes the conventional difference-in-difference estimation.
It closely resembles Table 10. The only difference is the interpretation of the
coefficients.
84
Table 12 Conventional Difference-in-Difference Estimation
Measure of economic performance ( ) y
it
Indonesia Korea Thailand Malaysia
Before
September 1998
(i.e. before the
imposition of
Malaysian capital
controls)
α 0
α α K
+
0
α α T
+
0
α α M
+
0
After September
1998
(i.e. after the
imposition of
Malaysian capital
controls
β
α
+
0
β
α α
+ +
K 0
β
α α
+ +
T 0
γ β
α α
+ + +
M 0
After – Before
β
β
β
γ β +
Note: β captures the average impact of the IMF programs post-September 1998 while γ (i.e.
( γ β + ) - β = γ ) measures the differential effect of capital controls in Malaysia compared to the
post-September 1998 IMF programs in Indonesia, Korea, and Thailand.
With respect to other explanatory variables, we include exactly the same set of
variables as those in the time-shifted approach. The detailed discussion of each
variable can be found in the previous section. Again, we propose two econometrics
models here. Model (2.4.3) compares Malaysia to the three IMF countries; thus, one
must be careful when interpreting the result. The interpretation of the coefficients
should always be in term of average of the IMF programs. For Model (2.4.4), it only
compares Malaysia to Korea. It allows us to assess whether Malaysian capital
controls perform better or worse than the IMF program in Korea. This narrow set of
Table 13 and 15 provide the point estimates when Indonesia, Korea, and Thailand
are used as comparator countries. Table 14 and 16 give the results when the
comparator country only consists of Korea. In what follows, we start by explaining
Table 13 and 15 then move on to compare Table 14 and 16. We believe that by
concurrently discussing the results from both the time-shifted and conventional
approaches a clear picture of the upper and lower bounds for the impact of capital
controls can then be obtained.
This paper estimates altogether one hundred sixty-eight monthly regressions, of
which half of them comes from the time-shifted approach and the other half is from
the conventional approach. For each dependent variable, we estimate model (2.4.1)-
(2.4.4) with one to six years of treatment periods. Table 13 and 14 provide the
estimation results from the time-shifted difference-in-difference estimation. On the
other hand, Table 15 and 16 present the point estimates from the conventional
approach. For all these tables, we report only the coefficient
85
comparison is a tough test for Malaysia because Korea is usually considered to be
the most successful IMF recipient in this region.
2.5 Empirical Results
β and γ because these
two estimated coefficients are the main focus of this study.
Table 13 Estimates from the Time-Shifted Approach using Indonesia, Korea, and Thailand as Comparators
Variable Treatment period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
.286***
(.044)
.149***
(.042)
.073
(.046)
.201***
(.044)
.215***
(.04)
.214***
(.036)
Difference in
Malaysia ( γ )
-.381***
(.079)
-.247***
(.06)
-.193***
(.052)
-.283***
(.046)
-.309***
(.043)
-.332***
(.042)
Exchange
Rate
(log,
HC/$)
F-Statistics
4427.17
4245.19
4193.65
4407.15
4522.27
4619.64
Baseline Effect
( β )
-.124**
(.055)
.024
(.051)
.083
(.056)
.0961
†
(.054)
.06
(.048)
.087*
(.043)
Difference in
Malaysia ( γ )
.085
(.098)
-.053
(.074)
-.214***
(.062)
-.294***
(.056)
-.356***
(.053)
-.388***
(.051)
Foreign
Reserves
(log)
F-Statistics
97.35
96.38
98.77
101.97
106.28
108.44
Baseline Effect
( β )
17.887***
(1.356)
12.708***
(1.339)
9.117***
(1.531)
5.284***
(1.513)
5.672***
(1.401)
3.885***
(1.272)
Difference in
Malaysia ( γ )
-21.322***
(2.42)
-12.646***
(1.929)
-8.082***
(1.713)
-4.988***
(1.609)
-3.981**
(1.522)
-1.684
(1.509)
Interest
Rate
(%)
F-Statistics
34.53
27.84
23.43
21.32
21.46
20.79
86
Table 13 Estimates from the Time-Shifted Approach using Indonesia, Korea, and Thailand as Comparators, Continued
Variable Treatment period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
1.901***
(.157)
.86***
(.16)
.734***
(.177)
.585***
(.172)
.529***
(.159)
.392**
(.144)
Difference in
Malaysia ( γ )
-1.711***
(.281)
-.634**
(.23)
-.479**
(.198)
-.327
†
(.183)
-.286
†
(.173)
-.185
(.171)
Inflation
Rate
(CPI %
change,
monthly)
F-Statistics
13.68
7.4
6.81
6.47
6.43
6.23
Baseline Effect
( β )
.033*
(.016)
.029*
(.015)
.048***
(.016)
-.028
†
(.016)
-.053***
(.014)
-.053***
(.013)
Difference in
Malaysia ( γ )
.037
(.028)
.051**
(.021)
.004
(.018)
.036*
(.017)
.039**
(.016)
.043***
(.015)
Export
(log)
F-Statistics
902.06
910.56
902.94
894.57
909.6
914.81
Baseline Effect
( β )
-.09***
(.021)
-0.055**
(.02)
-.036
†
(.022)
-.125***
(.021)
-.127***
(.019)
-.132***
(.017)
Difference in
Malaysia ( γ )
.157***
(.038)
.143***
(.029)
.063**
(.025)
.122***
(.022)
.098***
(.021)
.094***
(.02)
Import
(log)
F-Statistics
573.8
575.74
557.02
600.02
600.71
614.26
Note: (1) *** significant at 1% (two-tailed test), (2) ** significant at 5% (two-tailed test), (3) * significant at 10% (two-tailed test),
(4) † significant at 10% (one-tailed test), (5) Standard errors are in the parentheses, (6) F-Statistics for all regressions correspond to
F(23, 576) with a critical value between 1.72 and 1.90 at 1% significant level.
87
Table 14 Estimates from the Time-Shifted Approach using Korea as Comparator
Variable Treatment Period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
.235***
(.02)
.104***
(.021)
.059**
(.022)
.103***
(.019)
.116***
(.018)
.12***
(.017)
Difference in
Malaysia ( γ )
-.27***
(.026)
-.121***
(.023)
-.051**
(.02)
-.085***
(.018)
-.084***
(.016)
-.069***
(.016)
Exchange
Rate
(log,
HC/$)
F-Statistics
33599.29
24235.21
22218.13
24199.66
25114.88
25598.98
Baseline Effect
( β )
-.236**
(.102)
-.017
(.091)
.006
(.092)
.074
(.08)
.082
(.069)
.165***
(.058)
Difference in
Malaysia ( γ )
.269*
(.132)
.014
(.101)
-.238***
(.084)
-.425***
(.073)
-.587***
(.063)
-.747***
(.056)
Foreign
Reserves
(log)
F-Statistics
57.34
55.88
58.52
65.62
81.6
105.57
Baseline Effect
( β )
7.654***
(.709)
2.585***
(.727)
1.166
(.769)
-.142
(.698)
-.086
(.657)
-.535
(.608)
Difference in
Malaysia ( γ )
-6.797***
(.921)
-2.212**
(.805)
-1.185
†
(.696)
.378
(.64)
.967
(.604)
1.924***
(.588)
Interest
Rate
(%)
F-Statistics
44.79
29.6
28.12
27.66
28.08
29.3
88
Table 14 Estimates from the Time-Shifted Approach using Korea as Comparator, Continued
Variable Treatment Period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
.315**
(.125)
.089
(.111)
-.021
(.116)
-.059
(.105)
-.024
(.099)
.011
(.093)
Difference in
Malaysia ( γ )
-.205
(.163)
-.046
(.124)
-.049
(.105)
.006
(.096)
-.029
(.091)
-.055
(.09)
Inflation
Rate
(CPI %
change,
monthly)
F-Statistics
4.38
4.01
4.00
4.00
3.99
4.00
Baseline Effect
( β )
.012
(.024)
.041
†
(.021)
.036
(.022)
-.034
†
(.02)
-.081***
(.019)
-.092***
(.017)
Difference in
Malaysia ( γ )
.052
(.032)
.041
†
(.023)
-.001
(.02)
.026
(.019)
.028
†
(.017)
.016
(.017)
Export
(log)
F-Statistics
446.11
459.82
440.98
440.74
467.01
488.11
Baseline Effect
( β )
-.244***
(.028)
-0.158***
(.025)
-.132***
(.028)
-.138***
(.024)
-.142***
(.023)
-.117***
(.022)
Difference in
Malaysia ( γ )
.252***
(.036)
.174***
(.028)
.069**
(.025)
.09***
(.022)
.052**
(.021)
.016
(.021)
Import
(log)
F-Statistics
348.77
316.52
289.7
300.3
305.51
300.78
89
Table 14 Estimates from the Time-Shifted Approach using Korea as Comparator, Continued
Variable Treatment Period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
-.158***
(.015)
-.112***
(.014)
-.082***
(.015)
-.048***
(.014)
-.006
(.013)
.012
(.012)
Difference in
Malaysia ( γ )
.092***
(.019)
.058***
(.015)
.026
†
(.014)
-.001
(.013)
-.032**
(.012)
-.049***
(.012)
Industrial
Production
Index
(log)
F-Statistics
539.7
459.06
410.06
389.8
383.05
393.79
Baseline Effect
( β )
-.521***
(.061)
-.096
(.06)
-.064
(.063)
-.165***
(.056)
-.117*
(.053)
-.165***
(.049)
Difference in
Malaysia ( γ )
.351***
(.079)
.123
†
(.067)
.013
(.057)
.061
(.051)
-.033
(.049)
-.036
(.047)
Share Price
Index
(log)
F-Statistics
15.61
9.84
9.61
10.24
10.29
11.14
Note: (1) *** significant at 1% (two-tailed test), (2) ** significant at 5% (two-tailed test), (3) * significant at 10% (two-tailed test),
(4) † significant at 10% (one-tailed test), (5) Standard errors are in the parentheses, (6) F-Statistics for all regressions correspond to
F(21, 278) with a critical value between 1.72 and 1.90 at 1% significant level.
90
Table 15 Estimates from the Conventional Approach using Indonesia, Korea, and Thailand as Comparators
Variable Treatment period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
-.092
†
(.049)
-.193***
(.046)
-.028
(.042)
-.005
(.039)
.02
(.035)
.054
(.036)
Difference in
Malaysia ( γ )
-.093
(.080)
-.085
(.058)
-.153***
(.051)
-.202***
(.046)
-.232***
(.043)
-.259***
(.042)
Exchange
Rate
(log,
HC/$)
F-Statistics
4151.31
4280.59
4178.57
4249.04
4306.72
4362.01
Baseline Effect
( β )
.163**
(.059)
.262***
(.056)
.206***
(.05)
.163***
(.046)
.125***
(.042)
.133***
(.043)
Difference in
Malaysia ( γ )
-.123
(.096)
-.144*
(.07)
-.254***
(.06)
-.313***
(.054)
-.379***
(.051)
-.396***
(.05)
Foreign
Reserves
(log)
F-Statistics
97.89
101.02
101.91
103.97
107.96
109.75
Baseline Effect
( β )
-2.149
(1.656)
-9.521***
(1.54)
-11.188***
(1.361)
-8.992***
(1.288)
-8.294***
(1.181)
-8.865***
(1.215)
Difference in
Malaysia ( γ )
-5.703*
(2.706)
-1.589
(1.942)
-.369
(1.630)
.21
(1.519)
1.39
(1.449)
3.008*
(1.418)
Interest
Rate
(%)
F-Statistics
20.85
23.73
25.97
24.20
24.05
24.23
91
Table 15 Estimates from the Conventional Approach using Indonesia, Korea, and Thailand as Comparators, Continued
Variable Treatment period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
-.896***
(.185)
-1.136***
(.175)
-1.084***
(.157)
-.933***
(.147)
-.83***
(.135)
-.761***
(.14)
Difference in
Malaysia ( γ )
0.394
(.303)
0.31
(.22)
.208
(.188)
.227
(.174)
.237
(.166)
.243
(.164)
Inflation
Rate
(CPI %
change,
monthly)
F-Statistics
7.12
8.12
8.43
8.00
7.87
7.41
Baseline Effect
( β )
-.007
(.017)
0.038**
(.016)
-.038**
(.015)
-.059***
(.013)
-.063***
(.012)
-.055***
(.013)
Difference in
Malaysia ( γ )
.068**
(.028)
.052**
(.02)
.037**
(.018)
.045***
(.016)
.042**
(.015)
.039**
(.015)
Export
(log)
F-Statistics
895.33
913.55
899.08
919.83
929.60
918.35
Baseline Effect
( β )
-.007
(.023)
0.069***
(.022)
-.051**
(.02)
-.035
†
(.019)
-.053***
(.017)
-.057***
(.018)
Difference in
Malaysia ( γ )
.09**
(.038)
.084***
(.028)
.066**
(.024)
.085***
(.022)
.068***
(.021)
.057**
(.02)
Import
(log)
F-Statistics
556.08
577.93
560.77
565.72
565.34
563.97
Note: (1) *** significant at 1% (two-tailed test), (2) ** significant at 5% (two-tailed test), (3) * significant at 10% (two-tailed test),
(4) † significant at 10% (one-tailed test), (5) Standard errors are in the parentheses, (6) F-Statistics for all regressions correspond to
F(23, 576) with a critical value between 1.72 and 1.90 at 1% significant level.
92
Table 16 Estimates from the Conventional Approach using Korea as Comparator
Variable Treatment Period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
-.075***
(.025)
-.16***
(.021)
-.056**
(.021)
-.038*
(.019)
-.023
(.018)
-.004
(.018)
Difference in
Malaysia ( γ )
-.022
(.031)
.021
(.021)
.001
(.02)
-.02
(.018)
-.019
(.017)
-.008
(.017)
Exchange
Rate
(log,
HC/$)
F-Statistics
23031.71
26841.79
22231.13
22254.73
21948.84
21564.82
Baseline Effect
( β )
0.243**
(.106)
.439***
(.094)
.393***
(.083)
.38***
(.073)
.351***
(.06)
.44***
(.053)
Difference in
Malaysia ( γ )
-.107
(.130)
-.216**
(.094)
-.414***
(.079)
-.56***
(.068)
-.715***
(.059)
-.859***
(.049)
Foreign
Reserves
(log)
F-Statistics
57.31
61.32
64.38
73.13
92.59
130.90
Baseline Effect
( β )
-4.322***
(.841)
-6.006***
(.713)
-6.667***
(.594)
-5.832***
(.566)
-5.281***
(.521)
-5.792***
(.508)
Difference in
Malaysia ( γ )
2.696**
(1.03)
2.337***
(.711)
2.393***
(.567)
2.903***
(.533)
3.434***
(.508)
4.191***
(.475)
Interest
Rate
(%)
F-Statistics
31.55
38.19
46.35
43.27
43.36
49.02
93
Table 16 Estimates from the Conventional Approach using Korea as Comparator, Continued
Variable Treatment Period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
-.261*
(.131)
-.277**
(.119)
-.228**
(.107)
-.207*
(.099)
-.182*
(.091)
-.102
(.094)
Difference in
Malaysia ( γ )
.248
(.161)
.146
(.119)
.045
(.102)
.07
(.093)
.047
(.089)
-.005
(.088)
Inflation
Rate
(CPI %
change,
monthly)
F-Statistics
4.23
4.31
4.28
4.25
4.24
4.07
Baseline Effect
( β )
.02
(.025)
.078***
(.022)
-.021
(.021)
-.057***
(.019)
-.078***
(.017)
-.031
†
(.018)
Difference in
Malaysia ( γ )
.047
(.031)
.027
(.022)
.026
(.02)
.036*
(.018)
.027
(.017)
-.015
(.017)
Export
(log)
F-Statistics
446.68
473.66
438.51
450.55
469.67
445.85
Baseline Effect
( β )
-.052
(.032)
.09***
(.029)
-.018
(.027)
.017
(.025)
.013
(.023)
.041
†
(.023)
Difference in
Malaysia ( γ )
.092**
(.040)
.034
(.029)
.016
(.025)
.019
(.023)
-.021
(.022)
-.059**
(.021)
Import
(log)
F-Statistics
271.86
286.35
266.97
268.56
267.31
274.33
94
95
Table 16 Estimates from the Conventional Approach using Korea as Comparator, Continued
Variable Treatment Period 1 Year
(Short run)
2 Years
(Med run)
3 Years
(Med run)
4 Years
(Med run)
5 Years
(Long run)
6 Years
(Long run)
Baseline Effect
( β )
-.039*
(.018)
.03
†
(.016)
.038**
(.015)
.078***
(.013)
.082***
(.011)
.095***
(.011)
Difference in
Malaysia ( γ )
-.007
(.022)
-.024
(.016)
-.029**
(.014)
-.057***
(.012)
-.073***
(.011)
-.085***
(.01)
Industrial
Production
Index
(log)
F-Statistics
381.17
373.43
378.84
424.47
455.69
497.35
Baseline Effect
( β )
.086
(.071)
.417***
(.06)
.132**
(.058)
.201***
(.053)
.067
(.049)
.058
(.05)
Difference in
Malaysia ( γ )
-.139
(.087)
-.143**
(.06)
-.076
(.055)
-.104*
(.05)
-.12**
(.048)
-.142***
(.047)
Share Price
Index
(log)
F-Statistics
9.74
13.64
9.96
10.73
10.03
10.28
Note: (1) *** significant at 1% (two-tailed test), (2) ** significant at 5% (two-tailed test), (3) * significant at 10% (two-tailed test),
(4) † significant at 10% (one-tailed test), (5) Standard errors are in the parentheses, (6) F-Statistics for all regressions correspond to
F(21, 278) with a critical value between 1.72 and 1.90 at 1% significant level.
96
Table 13 and 15 can be interpreted as follows. When the dependent variable is the
log of exchange rate, the figures from the time-shifted approach show that after the
implementation of the IMF programs in Indonesia, Korea, and Thailand, these three
countries experience on average 20% (i.e. β ≈ 0.2) deprecation of their currencies in
comparison to their exchange rates during more normal time. However, when we
use the conventional method to conduct the analysis, the post-September 1998
exchange rates in the three IMF countries are found to appreciate in the short run and
return to the pre-September 1998 levels in the medium and long runs. Our main
interest is on the coefficient γ . Regardless of the estimation method, our results
demonstrate that Malaysian Ringgit following the imposition of capital controls
always appreciates relative to its neighbors’ currencies. The results from the time-
shifted approach consistently show a larger appreciation of the Malaysian currency
than the conventional approach, which allow us to identify the upper and lower
bounds of the effect. In particular, the bounds of the Malaysian currency’s
appreciation are found to be 0% to 38% in the short run, 18% to 24% in the medium
run, and 25% to 32% in the long run. These findings suggest that Malaysian capital
controls are more effective at stabilizing the Ringgit than the average of the IMF
programs during all treatment periods. Furthermore, these results match our prior
expectation that the conventional method tends to give more favorable outcome
towards the IMF countries and less favorable outcome for Malaysia, while the time-
shifted approach takes the opposite stand.
97
There is no difference between Malaysia and other countries in term of their foreign
reserves’ accumulation during the short run. However, in the medium and long runs,
Malaysia’s foreign reserves are found to be lower than its three foreign counterparts.
The conventional approach gives us the lower bound of the effect. It shows that on
average Malaysian foreign reserves are 24% lower in the medium run and 39% lower
in the long run. The time-shifted approach gives slightly better outcomes for
Malaysia. It shows Malaysian foreign reserves to be 19% less in the medium run
and 37% less in the long run when comparing to the three IMF countries.
Irrespective of the estimation method, the coefficient γ becomes more negative as
the treatment period increases. What this means is that Malaysia progressively loses
more foreign reserves, while Indonesia, Korea, and Thailand continue to accumulate
theirs as time goes by. These results appear to refute the view that the imposition of
capital controls should free Malaysia from having to defend the Ringgit from further
depreciating. Because under this view, there is no need for Malaysia to sell its
foreign reserves in order to fight against speculative attacks as commonly
experienced in Indonesia, Korea, and Thailand. However, there may be a few
possible explanations for these outcomes. First, Indonesia, Korea, and Thailand
receive additional financial supports from the IMF by following its instructions.
These monetary transfers are part of these three countries’ foreign reserves. On the
other hand, Malaysia receives no external funding from abroad, which may explain
why Malaysian foreign reserves are lower than its three neighboring countries.
Second, the imposition of capital controls may indeed generate more capital flight in
98
Malaysia, leading to a huge withdrawal of foreign currencies once the investors are
free to do so.
The third row of Table 13 and 15 report the results when the dependent variable is
interest rate. Under the time-shifted approach, we find interest rates in Indonesia,
Korea, and Thailand to be greater than the levels prevailed during more normal time.
The largest increase of interest rate occurs during the first year after they seek help
from the IMF (i.e. β = 17.887). For Malaysia, its interest rate is consistently lower
than the three comparator countries. Our estimates reveal that γ is statistically
significant for all periods, and ranges from -21.332 in the short run to -3.981 in the
long run. These point estimates, which are also the upper bound of the effect,
suggest that capital controls are more effective than the average of the IMF programs
at lowering interest rate. We discover somewhat different results when the
conventional approach is used to perform the analysis. From the third row of Table
15, we find the post-September 1998 interest rates in the three IMF countries to be
either the same or less than the levels prevailed before September 1998. Malaysian
interest rate relative to its three foreign counterparts is found to range from being
lower to being higher as the treatment period progresses. These results, which are
also the lower bound of the effect, suggest that there is a mixed performance of
Malaysian capital controls relative to the IMF programs when it comes to stabilizing
the interest rate. In particular, capital controls seem to be more effective in the short
run, equally effective in the medium run, and less effective in the long run.
99
When inflation is entered as the dependent variable, the outcomes are quite
comparable to the interest rate regressions. The coefficient β is positive and
significant for all treatment periods under the time-shifted approach while it is
negative and significant under the conventional approach. These results suggest that
Indonesia, Korea, and Thailand experience higher inflation after undertaking the
IMF prescription while on average their post-September 1998 inflation levels are
lower than their pre-September 1998 levels. As for the coefficient γ , we find it to be
negative and significant during all treatment periods under the time-shifted approach
but when the conventional method is used the coefficient γ is found to be
insignificant across all treatment periods. We can make the following conclusion
about the upper and lower bounds of the effect. On the upper side, Malaysian capital
controls are better at combating inflation than the average of the IMF programs. On
the lower side, we do not find any inflation differential between adopting capital
controls and undertaking the IMF program. The real effect of the controls should fall
between these two bounds.
The last two rows of Table 13 and 15 report the estimation results when the
dependent variables are merchandise export and merchandise import
19
. For these
two dependent variables, the point estimates from the time-shifted approach closely
resemble those of the conventional approach. The coefficient γ is positive and
19
We include merchandise export and import instead of total export and import because only
quarterly data are available for the latter set of variables.
100
significant in both export and import regressions. These point estimates demonstrate
that Malaysia experiences higher level of export and import than its three
neighboring countries during all treatment periods. Thus, it is fair to conclude that
imposing capital controls leads to more trade between Malaysia and the rest of the
world than the counterfactual of implementing the IMF programs. Because the
estimates from the time-shifted and conventional approaches are similar, the upper
and lower bounds of the effect tend to be very close in this case.
When Indonesia, Korea, and Thailand are used as comparator countries, our
estimates from the time-shifted approach always favor the controls policy as a better
crisis’s solution than the IMF programs during the short and medium runs. When the
conventional method is used, we find Malaysian capital controls to be either more
effective than the IMF programs or as effective as the IMF programs. The result
varies depending on the economic measure of interest. This remark is true for most
aspects of the economy except foreign reserves.
Now, we turn our attention to Table 14 and 16 where Korea is the only comparator
country. In general, the estimation results are quite similar to those reported on
Table 13 and 15. In the short and medium runs, Malaysia is either more effective
than Korea or as effective as Korea at stabilizing its exchange rate, controlling
inflation, and promoting export and import. The “more effective” outcomes usually
come from the time-shifted approach while the “equally effective” outcomes often
101
derive from the conventional method. Regardless of the estimation method, foreign
reserves are found to be lower in Malaysia than Korea. This result matches our
finding on Table 13 and 15.
When interest rate is employed as the dependent variable, using only one comparator
country can lead to very different results from using all the three comparators. With
three comparator countries, Malaysian interest rate is found to be less than or equal
to the rates prevailed in its three foreign counterparts, implying that Malaysian
capital controls are at least as effective as the IMF programs if not more effective
when it comes to lowering interest rate. Quite the opposite, when the comparator
country only consists of Korea, the conventional difference-in-difference estimation
shows Malaysian interest rate to be higher than the rate experienced in Korea for all
treatment periods. This result tells us that Malaysian capital controls are inferior to
the IMF program in Korea at stabilizing the interest rate. Thus, with respect to
interest rate, it takes the most successful case of the IMF, Korea, to beat the
performance of Malaysia.
Table 14 and 16 also report regression results for two additional dependent variables.
These dependent variables are industrial production index and share price index.
The estimated coefficients from the time-shifted approach suggest that, after
following the IMF program, Korea yields significantly lower industrial production in
the short and medium runs (i.e. β = -0.048 to -0.158) compared to its pre-crisis
102
level. For Malaysia, we discover that its industrial production in the short and
medium runs are persistently higher than Korea (i.e. γ = 0.058 to 0.092).
Nevertheless, in the long run, the coefficient γ turns out to be both negative and
significant. These outcomes suggest that capital controls are more effective at
preventing the reduction in industrial production during the short and medium runs
but become less effective in the long run compared to the IMF program in Korea.
When the conventional difference-in-difference estimation is used to perform the
analysis, we derive the opposite outcomes. The post-September 1998 industrial
production in Korea is found to be slightly lower in the short run and significantly
higher in the medium and long runs when comparing to its pre-September 1998
level. For Malaysia, our estimate of γ is insignificant during one and two years of
treatment period then turns out to be negative and significant thereafter. Thus, in
term of generating industrial production, Korea appears to be a better performer than
Malaysia. These results clearly support our conjecture that the time-shifted approach
tends to give more favorable outcome towards Malaysian capital controls and less
favorable outcome for the IMF program and vice versa when the conventional
method is used.
When share price index is entered as the dependent variable, we find the coefficient
β to be negative and significant for almost all treatment periods under the time-
shifted approach while it is mostly positive and significant under the conventional
approach. These results suggest that Korea experiences lower share price after the
103
adoption of the IMF policy while its post-September 1998 share price tends to be
higher than its pre-September 1998 level. As for the coefficient γ , we find it to be
positive and significant in the short run and insignificant for the rest of the treatment
periods under the time-shifted regression. However, when the conventional method
is used, the coefficient γ is found to be mostly negative and significant. We can
make the following conclusion about the upper and lower bounds of the effect. On
the upper side, Malaysian capital controls seem to be more effective than the IMF
program in Korea at inducing higher share price in the short run. On the lower side,
the controls policy is inferior to the Korean IMF program during all treatment
periods when it comes to stimulating higher share price.
Table 17 summarizes our findings on the effectiveness of Malaysian capital controls
relative to the IMF programs for all aspects of the economy, all treatment periods,
both estimation methods, and both sets of comparator countries. Finally, we perform
the F-test of joint significance for all regressions on Table 13 to 16. All of them
reject the null hypothesis of zeros’ coefficients, implying that the explanatory
variables in model (2.4.1)-(2.4.4) are jointly significant for all regressions.
104
Table 17 Effectiveness of Malaysian Capital Controls Relative to the IMF
Program(s)
Comparator Countries:
Indonesia, Korea, Thailand
Comparator Country: Korea
Variable
Treatment
Period Time-Shifted
Approach
(Upper
Bound)
Conventional
Approach
(Lower
Bound)
Time-Shifted
Approach
(Upper
Bound)
Conventional
Approach
(Lower
Bound)
Short Run
More Equal More Equal
Medium Run
More More More Equal
Exchange
Rate
Long Run
More More More Equal
Short Run
Equal Equal More Equal
Medium Run
Less Less Less Less
Foreign
Reserves
Long Run
Less Less Less Less
Short Run
More More More Less
Medium Run
More Equal More
Less
Interest Rate
Long Run
More Less Less Less
Short Run
More Equal Equal Equal
Medium Run
More Equal Equal Equal
Inflation
Rate
Long Run
More Equal Equal Equal
Short Run
Equal More Equal Equal
Medium Run
More More Equal More
Export
Long Run
More More Equal Equal
Short Run
More More More More
Medium Run
More More More Equal
Import
Long Run
More More More Less
105
Table 17 Effectiveness of Malaysian Capital Controls Relative to the IMF
Program(s), Continued
Comparator Countries:
Indonesia, Korea, Thailand
Comparator Country: Korea Variable
Treatment
Period
Time-Shifted
Approach
(Upper
Bound)
Conventional
Approach
(Lower
Bound)
Time-Shifted
Approach
(Upper
Bound)
Conventional
Approach
(Lower
Bound)
Short Run
- - More Equal
Medium Run
- - More Less
Industrial
Production
Index
Long Run
- - Less Less
Short Run
- - More Equal
Medium Run
- - Equal Less
Share Price
Index
Long Run
- - Equal Less
Note: (1) “More” implies Malaysian capital controls are “more effective” than the IMF program(s)
(2) “Equal” implies Malaysian capital controls are “equally effective” as the IMF program(s)
(3) “Less” implies Malaysian capital controls are “less effective” as the IMF program(s)
2.6 Conclusion
This paper assesses the effectiveness of capital controls in restoring Malaysian
economy as compared to the IMF programs used in Indonesia, Korea, and Thailand.
We analyze how capital controls help in stabilizing various aspects of the economy
under the short, medium, and long runs. We employ monthly data from January
1993 to June 2005, and use both the time-shifted and conventional difference-in-
difference estimations in this study. These two estimation methods allow us to
obtain the upper and lower bounds of the impact of Malaysian capital controls,
106
which in turn enables us to capture the real effect of the controls policy that falls
between these two bounds.
When using Indonesia, Korea, and Thailand as comparator countries, the upper
bound from the time-shifted approach usually specifies capital controls policy to be
more effective than the IMF programs while the lower bound from the conventional
method often shows it to be equally effective as the IMF programs during the short
and medium runs. With the exception of foreign reserves, our point estimates
consistently show Malaysia to be similar to or better than its three foreign
counterparts when it comes to stabilizing the exchange rate and interest rate,
controlling inflation, and promoting export and import.
When comparing Malaysia to only Korea, the overall results are not as clear. For
exchange rate, inflation, export and import, the upper and lower bounds of the effect
show Malaysian capital controls to be at least as effective as the IMF program in
Korea. For other aspects of the economy such as interest rate, industrial production,
and share price index, the effect of the capital controls policy is found to vary
depending on the estimation method and treatment period considered. In particular,
in the short and medium runs, the time-shifted approach usually reveals Malaysia to
be more effective while the conventional method often shows it to be less effective.
Thus, there is no obvious conclusion whether Malaysian capital controls perform
better or worse than the Korean IMF program. However, one should bear in mind
107
that Korea is regarded as the most successful recipient of the IMF in this region,
making this comparison a very demanding test for Malaysia.
In the long run, our results show no clear indication whether Malaysian capital
controls are more or less effective than the IMF programs. Furthermore, we believe
that in the long run another important variable that needs to be examined is foreign
direct investment
20
. Some economists argue that, due to capital controls, foreign
investors may lose their confidence on the country and ultimately withdraw their
long-term investments from Malaysia. This can be worrisome since Malaysia is one
of the world largest FDI’s recipients during the past few decades. Appendix 4 gives
FDI inflows and outflows for all four countries. The decrease in FDI inflows and
outflows are commonly experienced by all countries at some point after the financial
crisis. However, the largest drop occurs in Indonesia. Therefore, there is no clear
evidence that capital controls have negative impact on Malaysian FDI in the long run
when comparing to the countries that adopt the IMF programs.
We believe that this paper’s findings are crucial for policy implications. In
particular, the majority of our point estimates show that in the short and medium runs
capital controls are at least as effective as the IMF programs at stabilizing the
economy. As a result, if the controls are implemented primarily to prevent short-
20
No regression is performed on FDI because only annual data are available for FDI.
108
term speculative attacks as well as allow foreign direct investment to be unaffected,
then this policy may be beneficial to a country both in the short and medium runs.
109
Bibliography
ABCI (2002), “Australia Illicit Drug Report,” Discussion paper, Australia Bureau of
Criminal Intelligence.
ABS (2003a), “Consumer Price Index 14
th
Series: by Region, All Groups,” Australia
Bureau of Statistics, Cat. No. 640101b.
ABS (2003b), “Australian National Accounts: State Accounts,” Australia Bureau of
Statistics, Cat. No. 5220.0.
ACC (2003), “Australia Illicit Drug Report 2001-02,” Australia Crime Commission.
Amemiya, T. (1985), Advanced Econometrics, Harvard University Press.
Australia Institute of Health and Welfare (2002), “2001 National Drug Strategy
Household Survey First Results,” Drug Statistics Series No 9.
Australia Institute of Health and Welfare (2002), “National Drug Strategy Household
Survey 2001 Technical Report,” http://assda.anu.edu.au/codebooks/d1050/
d1050techrep.pdf.
Bartel, A. (1989), “Where do the New US Immigrants Live?,” Journal of Labor
Economics, 7, 371-391.
Becker, S. and A. Ichino (2002), “Estimation of Average Treatment Effects based on
Propensity Scores,” Stata Journal, 2, 358-377.
Becker, G. and K. Murphy (1988), “A Theory of Rational Addiction,” Journal of
Political Economy, 96, 675-700.
110
Cameron, C. and P. Trivedi (2005), Microeconometrics: Methods and Applications,
Cambridge University Press.
Cameron, L. and J. Williams (2001), “Cannabis, Alcohol and Cigarettes: Substitutes
or Complements?,” The Economic Record, 77, 19-34.
Clements, K. (2004), “Three Facts about Marijuana Prices,” The Australian Journal
of Agricultural and Resource Economics, 48, 271-300.
Dehejia, R. and S. Wahba (1999), “Causal Effects in Nonexperimental Studies:
Reevaluating the Evaluation of Training Programs,” Journal of the American
Statistical Association, 94, 1053-1062.
DiNardo, J. and T. Lemieux (2001), “Alcohol, Marijuana, and American Youth: the
Unintended Consequences of Government Regulation,” Journal of Health
Economics, 20, 991-1010.
Duan, N., W. Manning Jr., C. Morris, and J. Newhouse (1983), “A Comparison of
Alternative Models for the Demand for Medical Care,” Journal of Business
and Economic Statistics, 1, 115-126.
Duan, N., W. Manning Jr., C. Morris, and J. Newhouse (1984), “Choosing between
the Sample-Selection Model and the Multi-Part Model,” Journal of Business
and Economic Statistics, 2, 283-289.
Duncombe, W., M. Robbins, and D. Wolf (2001), “Retire to where? A Discrete
choice model of residential location,” International Journal of Population
Geography, 7, 281-293.
Feridhanusetyawan, T. and M. Kilkenny (1996), “Rural/Urban Residence Location
Choice,” Center for Agricultural and Rural Development, Iowa State
University, Working Paper.
111
Frenkel, A. (2001), “Why High-Technology Firms Choose to Locate in or near
Metropolitan Areas,” Urban Studies, 38, 1083-1101.
Gabriel, S. and S. Rosenthal (1989), “Household Location and Race: Estimates of a
Multinomial Logit Model,” The Review of Economics and Statistics, 71, 240-
249.
Hartwell, C. (2001), “The Case against Capital Controls, Financial Flows, Crises,
and the Flip Side of the Free-Trade Argument,” Cato Institute Policy Analysis
No. 403.
Heckman, J. and R. Robb (1985), “Alternative Methods for Evaluating the Impact of
Interventions,” in Longitudinal Analysis of Labor Market Data, edited by J.
Heckman and B. Singer. New York: Cambridge University Press.
Hsiao, C., Li, Q., and J. Racine (2006), “A Consistent Model Specification Test with
Mixed Discrete and Continuous Data,” IEPR Working Paper 06.47.
Kaplan, E. and D. Rodrik (2000) “Did the Malaysia Capital Controls Work?,”
Harvard University working paper.
Kittiprapas, S. and P. McCann (1999), “Industrial Location Behaviour and Regional
Restructuring within the Fifth Tiger Economy: Evidence from the Thai
Electronics Industry,” Applied Economics, 31, 37-51.
Krugman, P. (1998) “Capital Control Freaks Synopsis: How Malaysia Got Away
with Economic Heresy,” Slate, September 27, 1999.
Li, Q. and J. Racine (2007), Nonparametric Econometrics Theory and Practice,
Princeton University Press.
NDSHS (2001), Computer Data Files, National Drug Strategy Household Surveys
2001, Social Science Data Archives, Australian National University,
Canberra.
112
Neeley, C. (1999) “An Introduction to Capital Controls,” Federal Reserve Bank of
St. Louis Review.
Newey, W., J. Powell, and J. Walker (1990), “Semiparametric Estimation of
Selection Models: Some Empirical Results,” American Economic Review, 80,
324-328.
Pacula, R. (1998), “Does Increasing the Beer Tax Reduce Marijuana
Consumption?,” Journal of Health Economics, 17, 557-585.
Pacula, R., J. Chriqui, and J. King (2003), “Marijuana Decriminalization: What does
it mean in the United States?,” NBER Working Paper No 9690.
Rosenbaum, P. and D. Rubin (1983), “The Central Role of the Propensity Score in
Observational Studies for Causal Effects,” Biometrika, 70, 41-55.
Saffer, H. and F. Chaloupka (1995), “The Demand for Illicit Drugs,” NBER
Working Paper No 5238.
Saffer, H. and F. Chaloupka (1998), “Demographic Differential in the Demand for
Alcohol and Illicit Drugs,” NBER Working Paper No 6432.
Thies, C. and C. Register (1993), “Decriminalization of Marijuana and the Demand
for Alcohol, Marijuana, and Cocaine,” The Social Science Journal, 30, 385-
399.
Toussaint-Comeau, M. and S. Rhine (2004), “The Relationship between Hispanic
Residential Location and Homeownership,” Federal Reserve Bank of
Chicago, Economic Perspectives Third Quarter.
Williams, J. (2004), “The Effects of Price and Policy on Marijuana Use: What can be
learned from the Australian Experience?,” Health Economics, 13, 123-137.
113
Zhao, X. and M. Harris (2004), “Demand for Marijuana, Alcohol and Tobacco:
Participation, Levels of Consumption, and Cross-Equation Correlations,”
Economic Record 80(251): 394-410.
114
Appendices
Appendix 1 Map of Australia
This appendix gives the map of Australia with the names of eight states and
territories in our study. Decriminalized states consist of South Australia, Australia
Capital Territory, and Northern Territory while non-decriminalized states include
New South Wales, Victoria, Queensland, Western Australia, and Tasmania.
Figure 4 Map of Australia
Source: Google Image
115
Appendix 2 Definitions of Variables
This appendix provides the definitions of all variables used in the first chapter of this
dissertation.
Table 18 Definitions of Variables for the First Chapter
Variable Description
y
Decrim
P
MAR
Income
Age1419
Age2024
Age2529
Age3034
Age3539
Age4069
Age70
Male
Married
Divorce
Widow
1 if using marijuana in the last 12 months, otherwise 0
1 if respondent resides in decriminalized states (i.e. SA, ACT,
and NT)
Natural logarithm of real price of marijuana in Australian
Dollars per ounce
Natural logarithm of real household annual income before tax in
Australian Dollars
1 if age is 14 to 19 years old
1 if age is 20 to 24 years old
1 if age is 25 to 29 years old
1 if age is 30 to 34 years old
1 if age is 35 to 39 years old
1 if age is 40 to 69 years old
1 if age is 70 years old and above, it is a reference category and
is omitted from the estimation
1 for male and 0 for female
1 if married
1 if divorced
1 if widowed
116
Appendix 2 Definitions of Variables, Continued
Table 18 Definitions of Variables for the First Chapter, Continued
Variable Description
Never Married
# Depchild
Degree
Working Status
Aboriginal
Unemployment rate
1 if single, it is a reference category and is omitted from the
estimation
Number of dependent children aged 14 or below in the
household
1 if the highest qualification is university degree, and 0
otherwise
1 if respondent is unemployed, and 0 otherwise
1 if respondent has Aboriginal or Torres Strait Islander ethnic
origin
State unemployment rate (%)
117
Appendix 3 Definitions of Variables
This appendix provides the definitions of all variables used in the second chapter of
this dissertation.
Table 19 Definitions of Variables for the Second Chapter
Variable Description
Exchange Rate
Foreign Reserves
Interest Rate
Inflation
Merchandise Export
Merchandise Import
Industrial Production
Index
Share Price Index
Monthly average of market rates or official rates of the
reporting country in units of national currency per US$
Monetary authorities' claims on nonresidents in the form of
foreign banknotes, bank deposits, treasury bills, short- and
long-term government securities, and other claims usable
in the event of balance of payments need
Money market rate (i.e. the rate on short-term lending
between financial institutions)
Monthly inflation is percentage change of CPI from the
previous month (i.e. the formula is 100 *
1
1
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛ −
−
−
CPI
CPI CPI
t
t t
)
Measured on the "free-on-board" (f.o.b.) basis, that is, by
the value of the goods at the border of the exporting
country
Measured on the “cost-insurance-freight” (c.i.f.) basis, that
is, including the cost of freight and insurance incurred
beyond the border of the exporting country
The coverage of industrial production index comprises
mining and quarrying, manufacturing and electricity, and
gas and water according to the UN International Standard
Industrial Classification (ISIC), and the index is compiled
using the Laspeyres formula; for many developing
countries the index refers to the production of a major
primary commodity such as crude petroleum
Common shares of companies traded on national or foreign
stock exchanges
Source: Introduction to International Financial Statistics
118
Appendix 4 FDI Inflows and Outflows (Millions of US$)
Table 20 FDI Inflows and Outflows
Indonesia
Korea Thailand Malaysia Year
FDI
inflow
FDI
outflow
FDI
inflow
FDI
outflow
FDI
inflow
FDI
outflow
FDI
inflow
FDI
outflow
1993
2003
481
539
1340
1807
234
5741
1063
1994
2108 3283 788 2461 1369 494 4581 2329
1995
4346 1319 1250 3552 2070 887 5815 2488
1996
6194 600 2012 4670 2338 932 7297 3768
1997
4678 178 2640 4449 3882 584 6323 2675
1998
-241 44 5040 4740 7492 132 2714 863
1999
-1865 72 9448 4198 6091 349 3895 1422
2000
-4550 150 8591 4999 3350 -22 3788 2026
2001
-2978 125 3692 2420 3886 346 554 267
2002
145 182 2975 2617 947 106 3203 1905
2003
-597 15 3785 3426 1952 486 2473 1369
2004
1023 107 7687 4792 1064 362 4624 2061
Source: United Nations Conference on Trade and Development (UNCTAD)
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Essays on the econometrics of program evaluation
PDF
Estimation of heterogeneous average treatment effect-panel data correlated random coefficients model with polychotomous endogenous treatments
PDF
Essays on the estimation and inference of heterogeneous treatment effects
PDF
Essays on causal inference
PDF
Maternal full-time employment and childhood obesity
PDF
Essays on treatment effect and policy learning
PDF
Essays on high-dimensional econometric models
PDF
Three essays on linear and non-linear econometric dependencies
PDF
Estimation of treatment effects in randomized clinical trials which involve non-trial departures
PDF
Two essays in econometrics: large N T properties of IV, GMM, MLE and least square model selection/averaging
PDF
Three essays on econometrics
PDF
The development of the prognostic propensity score: an introduction to a method to identify optimal treatment according to individual tailoring variables when heterogeneous treatment effects are ...
PDF
Essays on human capital accumulation -- health and education
PDF
The role of local actors in international development: how incentive types affect foreign aid effectiveness
PDF
The impact of agglomeration policy on CO₂ emissions: an empirical study using China’s manufacturing data
PDF
Essays in panel data analysis
PDF
Essays on econometrics analysis of panel data models
PDF
Three essays on cooperation, social interactions, and religion
PDF
Essays on the luxury fashion market
PDF
Essays on economics of education
Asset Metadata
Creator
Damrongplasit, Kannika
(author)
Core Title
Essays on the econometrics of program evaluation
School
College of Letters, Arts and Sciences
Degree
Doctor of Philosophy
Degree Program
Economics
Publication Date
08/30/2007
Defense Date
06/18/2007
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
average treatment effect,capital controls,decriminalization,difference-in-difference estimation,endogenous probit switching,nonparametric kernel-based test,OAI-PMH Harvest,propensity score matching
Language
English
Advisor
Hsiao, Cheng (
committee chair
), Nichol, Michael B. (
committee member
), Nugent, Jeffrey (
committee member
)
Creator Email
kannikad@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-m802
Unique identifier
UC1190627
Identifier
etd-Damrongplasit-20070830 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-540469 (legacy record id),usctheses-m802 (legacy record id)
Legacy Identifier
etd-Damrongplasit-20070830.pdf
Dmrecord
540469
Document Type
Dissertation
Rights
Damrongplasit, Kannika
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Repository Name
Libraries, University of Southern California
Repository Location
Los Angeles, California
Repository Email
cisadmin@lib.usc.edu
Tags
average treatment effect
capital controls
decriminalization
difference-in-difference estimation
endogenous probit switching
nonparametric kernel-based test
propensity score matching