ESSAYS IN EMPIRICAL MICROECONOMICS
By
Yu-Wei Luke Chu

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

ABSTRACT
ESSAYS IN EMPIRICAL MICROECONOMICS
By
Yu-Wei Luke Chu
Eighteen states and the District of Columbia have passed laws that allow individuals to use
marijuana for medical purposes. There is an ongoing heated policy debate over whether these laws
have increased marijuana use among non-patients. In Chapter 1, I address that question empirically
by studying marijuana possession arrests in cities from 1988 to 2008. I estimate fixed effects
models with city-specific time trends that can condition on unobserved heterogeneities across
cities in both their levels and trends. I find that these laws increase marijuana arrests among adult
males by about 15–20%. These results are further validated by findings from data on treatment
admissions to rehabilitation facilities: marijuana treatments increased by 10–15% after the passage
of medical marijuana laws.
Medical marijuana laws generate significant policy debates regarding drug policy. In
particular, if marijuana is a complement or a gateway drug to hard drugs, these laws would increase
not only the usage of marijuana but hard drugs such as cocaine and heroin. In Chapter 2, I
empirically study the relationships between marijuana and cocaine or heroin by analyzing data on
drug possession arrests and treatment admissions. I find that medical marijuana laws increase
marijuana usage by 10–20%. However, there is no evidence that cocaine and heroin usage
increases after the passage of medical marijuana laws. In fact, the estimates on cocaine and heroin
arrests or treatments are uniformly negative. From the arrest data, the estimates indicate a 0–20%
decrease in possession arrests for cocaine and heroin combined. From the treatment data, the

estimates show a 20% decrease in heroin treatments but no significant effect on cocaine treatments.
These results suggest that marijuana could be a substitute for heroin.
Are work values a cause (Weber) or consequence (Marx) of the economic environment?
The collapse of the Soviet Union at the end of 1991 provides a unique opportunity to investigate
this link. In Chapter 3, using data collected from an employee survey conducted in over 340
workplaces in Armenia, Azerbaijan and Russia, I and Susan Linz investigate generational
differences in adherence to Protestant work ethic (PWE). The results indicate that Marx was “right”
about the link between work values and economic environment. That is, despite economic and
cultural differences emerging during the transformation process, in all three countries,
participating workers born after 1981 adhere more strongly to PWE than workers born before 1977.
Moreover, the estimate magnitudes are very similar across these economically and culturally
diverse countries.

To my loving wife Hao-Yu: Thank you for your never-ending love and support, without
which this would not have been possible.

iv

ACKNOWLEDGMENTS

I would like to sincerely thank Gary Solon for his invaluable guidance and support, both
in terms of comments and suggestions on this dissertation and in my development as a
researcher. I would also like to thank the other members of my guidance committee: Todd
Elder and Jeff Biddle for their assistance with this and other work. This dissertation also
benefited greatly from support from Susan Linz. Thanks also go to other faculty and
graduate students at Michigan State University including Soren T. Anderson, Michael
Conlin, Stacy Dickert-Conlin, Steven Haider, Sheila Royo Maxwell, Leah Lakdawala,
Quentin Brummet, Stacey Lynn Miller, among many others. Their combined effort has
helped immensely in the completion of this dissertation. All remaining errors are my
responsibility.

v

TABLE OF CONTENTS

LIST OF TABLES……………………………………………………………………...viii
LIST OF FIGURES……………………………………………………………………….x
CHAPTER 1
The Effects of Medical Marijuana Laws on Illegal Marijuana Use..............................…...1
1. Introduction…………………………………………………………………………1
2. Background…………………………………………………………………………7
2.1. Medical Marijuana Laws………………….………………………….7
2.2. Impact on Law Enforcement…….……………………………………9
3. Analysis of the Uniform Crime Reports…………………………………………..11
3.1. UCR Data………………………………………………….………...11
3.2. Empirical Model………………………………………….…………15
3.3. Results………………………………………………….……………18
4. Analysis of the Treatment Episode Data Set……………………………………...34
4.1. TEDS Data………………………………………………….……….34
4.2. Results………………………….……………………………………38
5. Discussion of Results and Conclusion…………………………………………….43
APPENDICES…………………………………………………………………………...46
Appendix A: State Medical Marijuana Laws………………………………………..47
Appendix B: Different Sample Constructions from the Uniform Crime Reports…...49
Appendix C: Effects of Medical Marijuana Laws on Juveniles Aged 15–17………..51
Appendix D: State-Level Data from the National Survey on Drug Use and Health...55
REFERENCES………………………………………………………………………….61
CHAPTER 2
Do Medical Marijuana Laws Increase Hard Drug Use?....................................................66
1. Introduction………………………………………………………………………..66
2. Medical Marijuana Laws………………………………………………………….70
3. Literature Review………………………………………………………………….72
4. Analysis of Arrest Data………………………………………………………...74
4.1. UCR Data………………………………….………………………...74
4.2. Results………………………………………….……………………80
5. Analysis of Treatment Data…………………………………………………….93
5.1. TEDS Data………………………………………….……………….93
5.2. Results……………………………………………….………………96
6. Discussion of Results and Conclusion…………………………………………...106
APPENDIX………………………………………………………………….………….109
REFERENCES.…………………………………………………………………..….…111
CHAPTER 3
Weber, Marx, and Work Values: Evidence from Transition Economies (with Susan Linz)

vi

……..................................................................................................................................117
1. Introduction……………………………………………………………………117
2. Socialist and Post-Socialist Economic and Workplace Environments………..…123
3. Data and Work Ethic Measure….……………………………………………...126
3.1. Sample Characteristics…………………….……………………….128
3.2. Work Ethic Measure…………………………………...………….128
4. Methodology……………………………………………………………………..132
5. Results and Discussion…………………………………………………………..134
5.1. Individual Measures of Work Ethic……………………………...…138
5.2. Generation versus Age……………………………………………..139
6. Conclusion…………………………………………………………………….….141
APPENDIX……………………………………………………………………………..143
REFERENCES……………………………………………...………………………….146

vii

LIST OF TABLES

Table 1.1: UCR Descriptive Statistics (1988–2008)……………………………………..14
Table 1.2: Effects of Medical Marijuana Laws on Marijuana Possession Arrests………19
Table 1.3: Effects of Medical Marijuana Laws (Restricted Sample) ……………………23
Table 1.4: Dynamic Responses of Marijuana Arrest Rates to Legalization……………..27
Table 1.5: UCR Descriptive Statistics in Each Male Age Group (1988–2008)…………28
Table 1.6: Effects of Medical Marijuana Laws in Each Age Group…………………….29
Table 1.7: Smaller Legalization Effects in States with Lower Perceived Risks…………30
Table 1.8: Indirect Tests for Changes in Unobservables………………………………...32
Table 1.9: TEDS Descriptive Statistics (1992–2008)……………………………………37
Table 1.10: Effects of Medical Marijuana Laws on Marijuana Treatments………..……39
Table 1.11: Effects of Medical Marijuana Laws on First-time Marijuana Treatments….40
Table A1: Medical Marijuana Laws……………………………………………………..47
Table B1: Robustness Checks…………………………………………………………....50
Table C1: TEDS Descriptive Statistics for Male Juveniles Aged 15–17 (All Referrals)..52
Table C2: Effects of Medical Marijuana Laws on Male Juveniles Aged 15–17………...53
Table D1: Descriptive Statistics for the NSDUH State-level Data………………………56
Table D2: Effects of Medical Marijuana Laws on Ages 18–25…………………………58
Table D3: Effects of Medical Marijuana Laws on Ages 26+……………………………59
Table 2.1: UCR Descriptive Statistics (1992–2008)……………………………………..78
Table 2.2: Effects of Medical Marijuana Laws on Marijuana Possession Arrests………82
Table 2.3: Effects of Medical Marijuana Laws on Cocaine and Heroin Possession Arrests
………………………………………………………………………………..83

viii

Table 2.4: Dynamic Responses of Cocaine and Heroin Arrests to Legalization………...88
Table 2.5: Robustness Checks…………...………………………………………………89
Table 2.6: UCR Descriptive Statistics for Blacks and Whites (1992–2008)…………….91
Table 2.7: Effects on Cocaine and Heroin Arrests among Blacks and Whites…………..92
Table 2.8: TEDS Descriptive Statistics (1992–2008)………………………………...….96
Table 2.9: Effects of Medical Marijuana Laws on Marijuana Treatments………………..98
Table 2.10: Effects of Medical Marijuana Laws on Cocaine Treatments……………........99
Table 2.11: Effects of Medical Marijuana Laws on Heroin Treatments……………….100
Table 2.12: TEDS Descriptive Statistics for Multidrug Treatments……………………104
Table 2.13: Effects of Medical Marijuana Laws on Multidrug Treatments…………....105
Table E: Descriptive Statistics for Control Variables…………………………………..110
Table 3.1 Sample Characteristics……………………………………………………….129
Table 3.2: Work Ethic Components (by Generation)…………………………………..131
Table 3.3: Determinants of Work Ethic………………………………………………...136
Table 3.4: Generation and Individual Work Ethic Components (by Country)…………139
Table 3.5: Generation versus Age (by Country)………………………………………..140
Table F1: Sample Means and Standard Deviations for Self-Reported Performance…..144
Table F2: Sample Means and Standard Deviations for Personality Trait and Components
……………………………………………………………………………….144
Table F3: Sample Means and Standard Deviations for Expected Rewards……………145

ix

LIST OF FIGURES

Figure 1.1: Marijuana Arrest Rates, Prices and Daily Use Rates 1988–2008………………6
Figure 1.2: Marijuana Arrest Rates and Police Officer Rates………………………….....22
Figure 1.3: Marijuana Arrest Rates Before and After the Passage of Laws.........................25
Figure 1.4: Marijuana All Treatment Ratios (upper) and First-Time Treatment Ratios
(lower) Before and After the Passage of Laws……………………………...42
Figure 2.1: Arrests and Prices of Marijuana and Cocaine 1992–2008……………..….….79
Figure 2.2: Arrest Ratios of Marijuana and Cocaine in California and Colorado……....…85
Figure 2.3: Cocaine Arrest Ratios Before and After the Passage of Laws…………...……87
Figure 2.4: Cocaine Primary Treatment Ratios Before and After the Passage of Laws….101
Figure 2.5: Heroin Primary Treatment Ratios Before and After the Passage of Laws.....102

x

CHAPTER 1
The Effects of Medical Marijuana Laws on Illegal Marijuana Use
“By characterizing the use of illegal drugs as quasi-legal, state-sanctioned, Saturday afternoon
fun, legalizers destabilize the societal norm that drug use is dangerous…Children entering drug
abuse treatment routinely report that they heard that ‘pot is medicine’ and, therefore, believed it
to be good for them.” Andrea Barthwell, M.D., Former Deputy Director of the White House Office
of National Drug Control Policy, in an editorial in The Chicago Tribune, February 17, 2004
1.

Introduction
Medical marijuana legislation represents a major change in U.S. policy towards marijuana

in recent years. As of March 2013, 18 states and the District of Columbia had passed laws that
allow individuals with designated symptoms to use marijuana for medical purposes. There is little
doubt that these laws will increase overall marijuana usage, at least with respect to the intensive
margin, because some existing users will become legal patients and they will be able to increase
their consumption safely and easily. However, as the number of legal patients was relatively small,
the major policy debate is whether these laws also increase illegal use among non-patients.
It is a popular belief among public media that legalization has increased illegal marijuana
use (Leger, 2012; O'Connor, 2011). Some evidence suggests that the leaking of medical marijuana
from legal patients or dispensaries may be common (Salomonsen-Sautel et al., 2012; Thurstone et
al., 2011). Moreover, these laws could send a “wrong message” to the public and increase social
acceptance for marijuana use. Khatapoush and Hallfors (2004) find that people in California
perceived less harm from smoking marijuana after medical marijuana legalization. Based on the
notion that illegal use has increased, federal agencies such as the Drug Enforcement

1

Administration (DEA) remain firmly opposed to these laws and continue to list marijuana as a
Schedule I drug with no accepted medical value (Drug Enforcement Administration, 2011).
Empirically, there is a strong correlation among medical marijuana legislation, the
perceived risk of marijuana, and marijuana use. Drawing on public-use data from the NSDUH for
the years 2002 through 2008, Wall et al. (2011) find that legalization was associated with a higher
prevalence rate and a lower perceived risk of marijuana use among juveniles. (See also Table D1
in Appendix D.) Cerdá et al. (2012) also find a similar correlation among adults from the National
Epidemiologic Survey on Alcohol and Related Conditions (NESARC). Despite the strong
correlation, the causal link appears to be weak after accounting for existing state differences.
Most of the existing studies focus on juveniles. O'Keefe and Earleywine (2011) do not find
any change in juvenile marijuana usage in a before-and-after comparison using data on the Youth
Risk Behavior Surveillance System (YRBSS). Harper et al. (2012) show that the findings from
Wall et al. (2011) are quite sensitive to the inclusion of state fixed effects. Using a number of
datasets that cover a longer period, including the YRBSS, Treatment Episode Data Sets (TEDS),
and National Longitudinal Survey of Youth 1997 (NLSY97), a recent working paper from
Anderson et al. (2012a) still finds no evidence of an increase in marijuana use among teenagers.
Few studies focus on adults, even though the marijuana prevalence rate is actually higher
among young adults than among juveniles (see Table D1 in Appendix D). Gorman and Huber
(2007) use a time series framework and do not find any significant change in marijuana use among
arrestees from the Arrestee Drug Abuse Monitoring data (ADAM) for the years 1995–2002. But
their data were limited to a small portion of arrestees with available urine test samples from only
four cities in a short time span. Based on the public-use NSDUH data, the estimates from Harper
et al. (2012) are positive but insignificant for young adults aged 18–25.

2

Adding to the still-limited literature, this paper focuses on adults and adopts a more robust
difference-in-difference (DD) research design to estimate the effects of medical marijuana laws on
illegal marijuana use among non-patients. Specifically, I use marijuana possession arrests at the
city level from the Uniform Crime Reports (UCR) for the years 1988–2008. As in the standard DD
type approach, I estimate reduced-form models for the effects of medical marijuana laws on male
arrests, controlling for city and year fixed effects. To relax the parallel trends assumption in the
standard DD approach, I also control for city-specific time trends (linear or quadratic) to allow for
different trends of arrests in each city. Assuming that unobservables related to marijuana arrest,
such as law enforcement, do not deviate from a city trend when states enact medical marijuana
laws, this approach will uncover the causal effect of these laws.
In principle, these medical marijuana laws will not affect enforcement towards illegal use
because they only protect legal patients. However, these loosely worded laws have created a grey
legal area that might affect the practice of law enforcement directly or indirectly. In general, it
could bias the estimates in either direction. To address this concern, I supplement the analysis by
using the state-level marijuana treatment admissions from the Treatment Episode Data Sets (TEDS)
for the years 1992–2008. I focus on treatment admissions not referred by the criminal justice
system, so the estimates from the TEDS are not biased by potential changes in law enforcement.
I find strong evidence in both datasets that the main effect of these laws on adult males was
to increase illegal marijuana usage. From the UCR, medical marijuana laws, on average, are
associated with a 15–20% increase in marijuana possession arrests among adult males. The effect
is stronger among younger males aged 18–29, while there is no obvious effect in age groups over
40. The results from the TEDS are consistent with the findings from the arrest data, indicating a
10–15% increase in marijuana treatments among adult males. The estimates from first-time

3

treatments that exclude potential recidivism are even larger and show a 20% increase among adult
males.
This paper advances the literature in several important ways. First, I estimate models that
can condition on empirically important unobserved cross city/state heterogeneity in both the level
and trend of determinants of marijuana-related behaviors. Although these unobserved year and
city/state effects are particularly important in the current context and this framework has become
the standard in the economics and policy evaluation literature, except for Anderson et al. (2012a)
and Harper et al. (2012), none of the existing studies adopt this framework.
The data also offer several advantages. First, these two datasets cover a period during
which 12 states legalized medical marijuana, while most of the previous works covers only a short
time period, leading to imprecise estimates based on a small number of state-level law changes.
Second, by focusing on drug arrestees and treatment patients rather than the general population,
these datasets provide more observations at the city/state levels than many representative datasets
and therefore reduce potential imprecision from small sample sizes. In fact, the UCR arrest data
remain the single most widely available indicator of illicit drug activity within and across
population aggregates in the United States. As the arrest data are available at the city level, I have
a large sample size of over 700 cities that enables me to estimate the effects of medical marijuana
laws more precisely even in the presence of flexible specifications.
In addition, arrest and treatment data represent objective measures, and they do not suffer
from the self-reporting bias that is common in survey data (Golub et al., 2005; Harrison et al.,
2007). Since medical marijuana laws are expected to change social acceptance and perception of
marijuana, changes in reporting behavior are of particular concern in the current context (Miller
and Kuhns, 2011). Finally, as arrest and treatment data represent frequencies rather than

4

individuals, conceptually, they capture changes not only in the extensive margin but also in the
intensive margin. The intensive margin is largely neglected in existing studies, but it is important
especially for adults because their initiation rates are low. Indeed, Anderson et al. (2012b) shows
that the prices of high-quality marijuana are decreasing over time after legalization. As
consumption may respond to price at both margins, the small-to-none estimated effects in some
studies could be a result of ignoring intensive margin. An example of a policy affecting
consumption primarily at the intensive margin is that zero-tolerance laws only decrease heavy
drinking while having no effect on participation in drinking (Carpenter, 2004).
A 10–20% increase is a large effect, but it is plausible for heavy users. Not only treatment
patients but marijuana arrestees are concentrated on heavy users. The marijuana arrests are highly
correlated with marijuana treatments, with correlation coefficients around 0.3–0.5.1 Graphically,
at the national level, Figure 1.1 shows that the marijuana possession arrest rates move closely with
daily use rates but opposite to marijuana prices.2 This is consistent with the finding that daily
marijuana use rates among arrestees (of all offenses) from the ADAM data are twice as high as
they are among general population (Golub and Johnson, 2002). Therefore, the estimates should be
interpreted as a 10–20% increase among heavy users, probably turning from casual users, rather
than a dramatic increase in marijuana users among the general population.

1

To calculate these correlation coefficients, I normalize the state level averages of marijuana arrest
rates or ratios to mean zero and standard deviation one in each state. The treatment ratios are also
normalized in each state.
2
In Figure 1, the marijuana arrests are the yearly averages of arrest rates from my sample, the
daily marijuana use rates are among ages 19–28 from the Monitoring The Future (MTF), and
marijuana prices are from the 2012 National Drug Control Strategy Data Supplement. All series
in Figure 1 are normalized to mean zero and standard deviation one.
5

2
1
0

-1
-2
1988

1992

1996

Year

2000

UCR Arrest
Price

2004

2008

MTF Daily Use

Figure 1.1: Marijuana Arrest Rates, Prices, and Daily Use Rates 1988–2008
(For interpretation of the references to color in this and all other figures, the reader is referred to
the electronic version of this dissertation.)

Currently, there are 11 states with pending legislation to legalize medical marijuana
(ProCon.org, 2013a). As the discussion about medical marijuana laws has become very popular in
the U.S., this paper addresses the heated policy debate on these laws by presenting evidence for an
increase in illegal use among non-patients. By using data reflecting effects on heavy users, this
research is more relevant to the design of policy because heavy usage is associated with negative
health and social outcomes, such as developing dependence and the future use of hard drugs (Chen
et al., 1997; Fergusson et al., 2006; Gruber et al., 2003).
The paper proceeds as follows: Section 2 describes these medical marijuana laws and their
potential impact on law enforcement. I discuss the data and results from the UCR arrests in Section
3 and those from the TEDS treatments in Section 4. Section 5 concludes.

6

2.

Background

2.1.

Medical Marijuana Laws
In the late 1980s and the early 1990s, smokable marijuana was discovered to have a positive

effect on patients suffering from nausea, a common symptom among cancer patients and the
increasing number of AIDS patients (Pacula et al., 2002). With growing evidence of positive
medical effects and lobbying by marijuana legalization advocacy groups such as the National
Organization for the Reform of Marijuana Laws (NORML), many states have joined in passing a
new wave of medical marijuana legislation since 1996. Table A1 in Appendix A provides an
overview of each state’s medical marijuana laws (for legal documents, see ProCon.org, 2013b).
These laws permit patients with legally designated diseases and syndromes to use
marijuana as a means of treatment. The designated symptoms and conditions typically include
AIDS, anorexia, arthritis, cachexia, cancer, chronic pain, glaucoma, migraines, persistent muscle
spasms, severe nausea, seizures, and sclerosis. Some laws, such as the one in California, even
allow use for “any other illness for which marijuana provides relief.” In most states, it is mandatory
to register as a qualified medical marijuana patient or caregiver and to renew this registration every
year.3 Patients can legally possess marijuana up to a fixed amount, but the amount differs by state
(ProCon.org, 2013b).
Not only can patients cultivate marijuana for their own use, these laws also allow
“caregivers” (most of whom are patients as well) to grow and provide marijuana to patients on a

3

California created a registration program in 2004 but registration was voluntary. Colorado allows
patients who do not join the registry to use the “affirmative defense of medical necessity” if they
are arrested on marijuana charges. Maine passed an amendment in November 2009 that created a
registration program and required mandatory registration starting January 1, 2011. Washington
does not have a registration program.
7

not-for-profit basis. Some marijuana dispensaries with grey legal status exist, but how prevalent
they are largely depends on the attitude of the local government (often at the city level) and the
actions of local law enforcement, which could change from time to time. This is because the state
medical marijuana laws do not directly allow marijuana dispensaries in order to conform to federal
regulations in which marijuana remains a Schedule I drug.4
Because even designated syndromes such as chronic pain can be defined subjectively, such
legislation does provide a way for recreational marijuana users to become legal patients. However,
before 2009, the number of legal patients remained relatively small except in California.5 A very
imprecise estimate from ProCon.org (2012) indicates that, as of January 2009, the total number of
legal patients was about 270,000 people, or 0.19% of the population in medical marijuana states.
In 2009, the Obama administration stated that the federal government would no longer seek to
arrest medical marijuana users and suppliers so long as they conformed to state laws. Since then,
the number of registered patients and dispensaries has increased significantly (Caplan, 2012;
Mikos, 2011; Sekhon, 2009). For example, Colorado had only 5,051 registered patients in January
2009, but the number skyrocketed to 99,902 by July 2010, implying that about 2.6% of adults were
legal patients. Although this statement appeared to largely resolve the legal dispute between state
and federal governments, the Obama administration’s medical marijuana policy began to reverse
in 2011, and there have been several cases of DEA raids on medical marijuana dispensaries that
arguably conform to state laws (Dickinson, 2012).

4

Laws and amendments passed after 2009 began to set specific regulations on dispensaries, and a
very small number of state-licensed dispensaries came into being.
5
There is no official number of patients for states without registration. However, based on the
large number of dispensaries, it is believed that California has many more patients than other
medical marijuana states.
8

2.2.

Impact on Law Enforcement
In principle, these medical marijuana laws only provide legal protection for patients and

caregivers, and do not change the legal status of the non-medical use of marijuana. However, there
is a huge grey area and the legal boundary is blurred by these loosely worded laws (Cohen, 2010).
For example, the Colorado attorney general, John W. Suthers, has said, ― “But in Colorado it’s
not clear what state law is.” (Johnson, 2009) The loose wording of these laws is probably done
intentionally by the legalization lobbyists behind these laws such as the NORML, who consider
such legislation the first step towards full legalization. In fact, two medical marijuana states,
Colorado and Washington, successfully passed referenda to legalize marijuana for recreational use
in November 2012.
The impacts on the actions of enforcement towards the supply side largely depend on local
attitudes. Although marijuana can be legally grown and sold under the protection of medical
purposes, it is still a Schedule I drug under federal regulations. In fact, most dispensaries are not
strictly legal even under state laws; for example, they usually operate on a for-profit basis. Many
dispensaries and caregivers are considered to be legal covers for illegal drug dealing and are
constantly being raided by the DEA. On the other hand, as the DEA often needs cooperation from
local law enforcement organizations, its actions are inevitably constrained by local attitudes, which
can differ greatly even within a state. For example, although San Diego County failed to challenge
the state medical marijuana law in court, it is able to set a very restrictive policy towards
dispensaries and its law enforcement organizations actively cooperate with the DEA. Even the
only county-licensed dispensary was forced to close in 2012 (Anderson, 2012). In contrast, there
are more marijuana dispensaries than Starbucks coffee shops or CVS pharmacies in Los Angeles
and San Francisco (Coté et al., 2008). In general, due to the increase in marijuana supply and

9

partial legal protection, it is clear that more legal resources are required to keep the previous level
of enforcement. As legal resources are mostly limited, enforcement towards marijuana supply, on
average, is likely to decrease.
In comparison to the impacts on the supply side, the direct impacts on enforcement towards
low-level possession offenses appear to be small due to a small number of legal patients (at least
prior to 2009). Still, there might be some negative impacts that could lower the level of law
enforcement. Based on a report done by the U.S. General Accounting Office (GAO, 2002), these
laws do create some confusion for enforcement towards possession offenses. For example,
California law only requires patients possessing a "written or oral recommendation" from their
physician, thus not requiring the recommendation to be documented. Even in states with
mandatory registration systems, some law enforcement officials and district attorneys have said
that they were less likely to pursue marijuana cases where someone has an amount of marijuana
within the medical use limit, and would probably be approved for being a legal patient if they did
apply (GAO, 2002). Because of the limited resources, they would rather pursue marijuana cases
that qualify for felony charges or other drugs, like crack cocaine or methamphetamines, that are
often associated with violent crimes. Moreover, law enforcement organizations believe there has
been a general softening in public attitude toward marijuana. For example, Denver passed a
referendum to legalize marijuana possession in 2005; Seattle passed an initiative in 2003 that
requires authorities to make cases involving marijuana offenses the city's lowest law enforcement
priority. Even though these laws are legally ineffective since they directly violate the state laws,
they may still affect the actions of local law enforcers. In fact, in a letter response to the GAO
report, the Department of Justice strongly complained that the GAO report failed to consider the

10

deteriorating relations between federal and local law enforcement (the letter is in Appendix V in
GAO 2002).

3. Analysis of the Uniform Crime Reports
3.1.

UCR data
The Uniform Crime Reports (UCR) arrest data is an administrative series of monthly police

records from state and local police agencies across the U.S compiled by the FBI. It provides
information on marijuana possession arrest counts by age, gender, and race along with agency
populations (estimated from the Census).6 Because a person may be arrested several times, each
arrest count does not necessarily represent a single individual. The UCR arrest data has a hierarchy
rule, which only records arrests according to the most serious offense. As a result, arrestees
classified under marijuana possession do not simultaneously commit other more serious crimes
(such as cocaine possession or other violent or property crimes). Because the FBI reviews and
checks the data using annual arrest totals (Akiyama and Propheter, 2005), I use the yearly
aggregated arrest data provided by the Inter-university Consortium for Political and Social
Research (ICPSR) for the years 1988 through 2008. 7 I use data starting from 1988 to avoid

6

Another marijuana arrest category is marijuana sale/manufacture. To be recorded as a sale arrest,
the amount must exceed some minimum with intention to sell. Because marijuana transactions
often involve small quantities, and sale intention is hard to prove, sale arrest is often due to largescale transactions. In fact, some marijuana possession arrestees are probably low-level sellers
(Jacobson, 2004).
7
2008 was the latest data available when I began this study. Although data through 2010 became
available recently, looking at the period prior to 2009 has an advantage in that the number of legal
patients was relatively small, and the federal policy was fairly uniform prior to the Obama
administration. In addition, severe economic recession may affect drug use, and theoretically the
direction is ambiguous (Bretteville-Jensen, 2011).
11

potential influences from decriminalization. Eleven states decriminalized marijuana in the 1970s,
though there are only minor differences across non-decriminalized and decriminalized states in the
late 1980s (Pacula, Chriqui, and King 2003; Pacula et al. 2010).8
Since participation in the UCR program is generally voluntary, many agencies do not report
every month or every year, and they may not report data in all categories. Although it is not
possible to distinguish a true zero from missing data, the FBI communicates regularly with
agencies of more than 50,000 city residents to ensure data quality (Akiyama and Propheter, 2005),
and most missing data is from agencies with small populations and those that do not report for a
whole year (Lynch and Jarvis, 2008). Therefore, I use police agencies located in cities with
populations greater than 50,000; as population size is generally increasing over time, I include
earlier observations of these cities to make the panel more balanced. (I exclude 233 city-year
observations that have populations less than 25,000). 9 Similar to Carpenter (2007), and as is
common in the criminology literature, I focus on adult male arrests and use observations only if
the agencies report arrests for marijuana possession for at least six months in that year.10 The final

8

Nevada also decriminalized marijuana possession in 2001. “Decriminalization” here is better
termed as depenalization since marijuana possession is still legally a crime and subject to arrest. It
is different from recent decriminalization in California and Massachusetts that removes the
criminal status of marijuana possession. Empirically, depenalization has little or no effect on
marijuana use (MacCoun et al., 2009; MacCoun, 2010; Pacula et al., 2005).
9
For agencies in MSAs with more than 50,000, about 70% of the population lives in cities. Also,
70% of the observations of all MSA agencies are city agencies. I restrict the sample to cities
because marijuana transactions and arrests are concentrated in cities. On average, marijuana arrest
rates in cities are about twice as large as in non-cities.
10
I include 213 city-year observations that report only in December since some agencies report
annually; their means and standard deviations are similar to observations that report for least six
months. I only consider males both to be consistent with the existing literature and because males
are much more likely to be in the criminal justice system than are females. For example, the
average arrest rate for adult males in my sample is seven times that for adult females.
12

panel consists of 751 cities and 12,157 city-year observations in which about half of the cities are
observed in at least 20 years. The sample covers eleven medical marijuana states that passed laws
before July 2008, including Alaska, California, Colorado, Hawaii, Maine, Montana, Nevada, New
Mexico, Oregon, Rhode Island, and Washington. Vermont is not in the sample because no city
from Vermont in the UCR has a population greater than 50,000. (Michigan passed its law in
November 2008 and is coded as a non-medical marijuana state.)
I create three different measures of marijuana arrest for adult males: adult male marijuana
possession arrest rates per 100,000 city residents, the ratios of marijuana possession arrests to all
offense arrests among adult males, and the ratios of marijuana possession arrests to all drug
possession arrests, also among adult males. 11 Although the arrest rate is straightforward and
commonly used, these two measures of arrest ratios can partially account for unobserved changes
in local law enforcement and measurement errors from estimated populations (Carpenter, 2007;
Fryer et al., 2010). In addition, as the resources of law enforcement are mostly limited, these arrest
ratios can capture fluctuations in arrests due to changes in total resources available or the resources
allocated to illicit drug activities. One limitation in these arrest ratios is that missing data in nonmarijuana arrests could introduce substantial measurement errors, especially in the ratios of
marijuana arrests to all drug arrests.
Table 1.1 lists the means and standard deviations of these different arrest measures. The
first row is for all states, the second row is for states without effective medical marijuana laws
before July 2008, and the third row is for states with effective medical marijuana laws before July

11

The UCR provides the category of total drug possession arrests as well as three subcategories
other than marijuana: 1. Opium, cocaine, and their derivatives; 2. Truly addicting synthetic
narcotics; 3. Other dangerous nonnarcotic drugs (methamphetamine is in this category).
13

2008, excluding California and Colorado. These two states are separated in the last two rows.
California has many more observations than any other states, and Colorado has the second largest
number of observations among the medical marijuana states, so I will study these two states
separately to see if there are any heterogeneous effects. In particular, the penalty in these two states
for low-level possession was the lowest in the U.S. with only a $100 maximum fine (Pacula et al.,
2010); as most dispensaries were located in California and Colorado prior to 2009, the legalization
effects and reactions of law enforcement could be different from other medical marijuana states.
Table 1.1 shows that the marijuana possession arrest rates/ratios are significantly lower in medical
marijuana states. Because marijuana use rates are higher in medical marijuana states based on
survey data (see Table D1 in Appendix D), it suggests that the level of law enforcement could be
lower in medical marijuana states. In the next section, I will propose an empirical model that is
able to account for the difference of local law enforcement not only in levels but also in trends.
Table 1.1: UCR Descriptive Statistics (1988-2008)
Marijuana Possession Arrest for Adult Males ( Ages 18+)
MJ Arrest Rate
MJ Arrests to All
MJ Arrests to All
per 100k City
Drug Arrests Ratio
Arrests Ratio (%)
Residents
(%)
Mean
All States
Non-MJ
States
MJ States
w/o CA &
CO

SD.

Mean

SD.

Mean

SD.

137.41 (125.73)

3.33

(2.39)

47.42

(26.09)

162.06 (133.77)

3.85

(2.35)

57.87

(21.53)

118.12 (84.64)

2.86

(1.89)

47.92

(20.58)

# of
Obs. States
12,157
50
(751 cities)
8,007
39
(514 cities)
715
(48 cities)

9

3,203
1
(174 cities)
232
Colorado
127.96 (88.22)
2.78
(1.46)
65.22
(16.90)
1
(15 cities)
Note.― Medical marijuana states include only states that passed laws before July 2008; states that
passed laws afterward are in non-MJ states (including D.C.). Vermont is not in the sample.
California

80.78

(91.07)

2.20

(2.20)

14

19.89

(16.02)

3.2.

Empirical Model
Many studies, such as Conlin et al. (2005) and Fryer et al. (2010), use arrests as a proxy

for drug use. Arrests are constantly reported by government agencies as an indicator for illegal
drug use, but they are often criticized for potential bias from police actions. In contrast, some other
studies use arrests as a measure of law enforcement to estimate its effect on drug use (DeSimone
and Farrelly, 2003; Farrelly et al., 2001; Pacula et al., 2010). I illustrate the relationship between
arrests, law enforcement, and marijuana use by a simple decomposition below, and use this
decomposition as the basis of my empirical model.
Marijuana arrests in a particular city-year can be modeled as follows:
(1)

A = ∑N Fj * P(Xj),
j=1

where N is the number of marijuana users, Fj is individual j’s transaction or use frequencies; P(Xj)
is the probability of being arrested per transaction, a function of Xj, including city specific factors
such as local law enforcement and individuals’ characteristics, such as age and race. As heavy
users have higher use frequencies, and they also probably face a higher arrest probability, arrests
are concentrated on heavy users.12 Since only city-level data are available, for simplicity, I assume
the probability of being arrested is the same for every j and ignore potential heterogeneity. Letting
̅ be the average of Fj and taking logs, then in a particular city-year:
𝐹
(2)

̅
log(A) = log(P) + log(𝐹 ) + log(N).

12

Unlike the market for cocaine or heroin, marijuana transactions are embedded in social
networks and very “safe.” For example, based on the NSDUH, Caulkins and Pacula (2006) find
that most people obtain marijuana indoors (87%), from a friend or relative (89%), and for free
(58%). This suggests that the probability of being arrested may be very low for most casual users
and most marijuana arrestees are heavy users who make regular transactions.
15

Differentiate both sides of (2), and the percentage change in arrests can be decomposed
into the percentage change in arrest probability and the percentage change in marijuana use, either
̅
from the extensive or intensive margins. So, because of the last two terms, log(𝐹 ) and log(N), using
arrests as an explanatory variable to estimate the effect of law enforcement on marijuana use or
price may introduce substantial bias.13 Indeed, as law enforcement is a major reason for shifts in
supply, if marijuana arrests were reflecting the strength of law enforcement, they should move
positively with price rather than in the opposite direction. However, as shown in Figure 1,
marijuana possession arrests move positively with daily use and negatively with price, which looks
exactly like a supply curve moving along a downward sloping demand curve.
When arrest is used as a proxy for marijuana use, arrest probability, log(P), can be treated
as a source of measurement errors in a dependent variable. In general, in addition to law
enforcement, there are other sources of measurement errors, such as the hierarchical recording rule
of the UCR: arrestees who possesses marijuana but also commit other more serious crimes will
not be counted as marijuana possession arrestees. As seen in Figure 1, at least nationally, these
measurement errors appear to be averaged out.
In this paper, I adopt a flexible specification to account for the measurement errors in
marijuana arrests, and estimate city- and year-specific arrests as a function of whether the state has
an effective medical marijuana law in place in that year. Specifically, for city i in state s and year
t, I estimate the following model by OLS:

13

DeSimone and Farrelly (2003) and Farrelly et al. (2001) divide arrests by the number of
marijuana users. Although it may account for bias from the extensive margin (N), it is not able to
account for bias from the intensive margin (F) and introduces additional division bias in their
context.
16

(3)

log(Aist) = β Lawst + City fixed effectsi + Year fixed effectst
+ City linear or quadratic time trendsit + εist,

where Lawst is a dummy variable indicating whether a state s had an effective medical marijuana
law during year t.14 In addition to city and year fixed effects, I include city-specific linear or
quadratic time trends to capture the time-varying unobservables within a city such as law
enforcement. In the main specification, I do not include any control variables because city-specific
time trends and fixed effects have already accounted for any smooth-trending variables.
Throughout this paper, the estimated standard errors are clustered at the state level and therefore
are robust to serial correlation, within-state spatial correlation, and heteroskedasticity.
The OLS estimator of β will be unbiased if the residuals of log(P) (after partialling out
fixed effects and specific trends), on average, are not a function of Lawst. In other words, as
nonsystematic fluctuations in law enforcement or other measurement errors will be averaged out,
this model can account for any existing difference in the levels and trends of enforcement across
cities. On the other hand, the estimates of β would be biased, for instance, if the law enforcement
endogenously responded to medical marijuana laws. As the discussion in Section 2.2 indicates,
although heterogeneous responses of enforcement towards marijuana sale/manufacture arrest may
exist, the direct impact of these laws on possession arrests seems to be moderate.

14

For the first year, Lawst equals 1 if the law is effective before July 1st, and equals 0 otherwise.
I code the law based on the effective date rather than the passing date (it significantly differs only
for Nevada) as there was an instance (Arizona in 1996) that the referendum was vetoed by the state
government. Note that there could be a huge time lag between the law being legally effective and
the marijuana program starting to operate and accept applications.
17

3.3.

Results
Table 1.2 shows the estimates on three different measures of marijuana arrest. The

dependent variables are the logarithm of the arrest rate in the upper panel, the logarithm of the
ratio of marijuana possession arrests to all offense arrests in the middle panel, and the logarithm
of the ratio of marijuana possession arrests to all drug possession arrests in the lower panel. The
first two columns, Columns (1) and (2), show the estimates of β based on Equation (3). The
estimates are small and insignificant for the arrest rate, but positive and highly significant for the
two arrest ratios. If we interpret the log points as a percentage change, medical marijuana laws, on
average, among adult males, result in a 10.0–12.1% increase in the ratio of marijuana to all arrests
and a 14.1–14.8% increase in the ratio of marijuana to all drug arrests.
Note that each observation is a city-year while Lawst only varies at the state level. Therefore,
the estimates of β are essentially weighted least square estimates on state-level averages, where
the weights are given by the numbers of city-years in each state. To see whether the above results
are driven by states with larger observations, I estimate the effects of Lawst on state level average
by OLS, so each state receives the same weight regardless of its number of city-years. Columns
(3) and (4) show these results. Except for all drug arrest ratios, the estimates from the state-level
averages of arrest rates and all arrest ratios are much larger than the estimates in Columns (1) and
(2). This suggests some large states are driving the above results. As California and Colorado
account for more than 80% of observations in medical marijuana states, the smaller estimates in
Columns (1) and (2) are probably driven by heterogeneous effects of medical marijuana laws in
these two states. I separately estimate the effects of California and Colorado laws by including CA
Law and CO Law, two interaction terms of Lawst and dummies for California or Colorado.

18

Table 1.2: Effects of Medical Marijuana Laws on Marijuana Possession Arrests
(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Arrest Rates (per 100k) for Adult Males
Law

0.010
(0.050)

0.056
(0.034)

0.307** 0.140
(0.149) (0.101)

0.261*** 0.201** 0.314*** 0.189
(0.096) (0.079) (0.097) (0.116)

CA Law

-0.289*** -0.164** -0.215** -0.087
(0.095) (0.079) (0.098) (0.112)

CO Law

-0.632*** -0.376*** -0.560*** -0.384***
(0.097) (0.081) (0.099) (0.113)

Arrest Ratio (all arrests) for Adult Males
Law

0.121*** 0.100*** 0.355*** 0.206** 0.314*** 0.224*** 0.449*** 0.252***
(0.045) (0.035) (0.119) (0.083) (0.097) (0.075) (0.121) (0.094)

CA Law

-0.215** -0.137*
(0.098) (0.075)

-0.345*** -0.119
(0.122) (0.090)

CO Law

-0.560*** -0.375*** -0.600*** -0.326***
(0.099) (0.077) (0.122) (0.090)

Arrest Ratio (all drug possession arrests) for Adult Males
Law

0.148*** 0.141*** 0.114*
(0.034) (0.026) (0.068)

0.0889
(0.060)

0.186** 0.145** 0.146** 0.104
(0.082) (0.062) (0.071) (0.064)

CA Law

-0.016
(0.084)

0.016
(0.060)

0.032
(0.071)

0.094
(0.065)

CO Law

-0.376*** -0.257*** -0.354*** -0.259***
(0.084) (0.068) (0.069) (0.050)

Obs.
12,157 City-years
# of States
50

955 State-years
12,157 City-years
955 State-years
50
50
50
Linear Quadratic
Linear Quadratic
Time trends Linear Quadratic
Linear Quadratic
(state)
(state)
(state)
(state)
Note.― All specifications include city (state) and year fixed effects. Robust standard errors are
reported in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

These results are presented in Columns (5) and (6). The estimates for Lawst are
significantly larger than Columns (1) and (2) and more similar to Columns (3) and (4) (except for
the lower panel). Based on the specification with quadratic city time trends, conditional on
California and Colorado, medical marijuana laws, on average, result in a 20.1% increase in the

19

arrest rate, a 22.4% increase in the ratio of marijuana arrests to all arrests, and a 14.5% increase in
the ratio of marijuana arrest to all drug arrests, among adult males.15 These estimates are similar
across three different measures, suggesting that they are not driven by unobserved fluctuations in
overall arrests or overall drug possession arrests. In Appendix B, Table B1, I check the robustness
of the results in Columns (5) and (6) based on different constructions of the sample.
In the last two columns, (7) and (8), I again estimate the specification by separately
controlling for CA Law and CO Law but using the state-level averages. For the specification with
quadratic time trends, the estimates for Lawst from the state-level regression in (7) and (8) are
quantitatively similar to the estimates in (5) and (6), suggesting that these results are not driven by
states with more cities, and the estimated legalization effects are fairly homogenous after
conditioning on California and Colorado.
As implied by the small estimates from Columns (1) and (2), nearly all of the estimates for
CA Law and CO Law are negative and significantly different from the estimates of Lawst; the only
exceptions are the estimates for CA Law from all drug arrest ratios. The estimates indicate that the
legalization effect is positive in California, but the magnitudes vary with different measures. Based
on the specification with quadratic trends, the estimated legalization effect in California is around
a 3.6% (0.201−0.164) increase in the marijuana arrest rate (with t-stat = 1.48), an 8.7% increase in
the all-arrest ratio (with t-stat = 2.98), and a 16.1% increase in the all-drug arrest ratio (with t-stat

15

The slightly smaller estimates from the ratio of marijuana to all drug arrests are expected.
Because marijuana arrests account for about half of drug arrests and appear in both numerator and
denominator, there are fewer variations in this measure. The estimates raise to around 0.2 if I use
the ratio of marijuana arrests to non-marijuana drug arrests as a dependent variable.
20

= 7.69). Since the estimate is largest from the all-drug arrest ratio, it seems that the fluctuations in
overall drug arrests in California can account for the differences in estimates across measures.
For Colorado, the estimates show a decrease in marijuana arrests by 11.2−17.5%,
regardless of the measures. Although I cannot rule out the possibility that marijuana use did decline
in Colorado, the fluctuations in police officer rates could be another explanation. In Figure 1.2, in
the upper graph, I plot the yearly averages of police officer rates per 100,000 city residents from
the UCR Law Enforcement Officers Killed and Assaulted series and marijuana possession arrest
rates in Colorado from my sample (each series is normalized to mean zero and standard deviation
one). It shows that the police officer rate and marijuana arrest rate temporarily dropped around the
year 2001, in which Colorado enacted its medical marijuana law, but both series increased again
after 2002. The graphs constructed using two arrest ratios show a similar pattern. Even though
these fluctuations could come from some unobserved factors unrelated to the medical marijuana
law, the strongly positive correlation between arrests and police rates implies that arrests in
Colorado is probably not a valid measure for marijuana use.16 For comparison, the lower graph in
Figure 1.2 plots the yearly averages of police officer rates and marijuana arrest rates in California
and other medical marijuana states. Both series in California move closely with those series in
other medical marijuana states, and their arrest rates are not positively correlated with police rates.
Because the estimates in Table 1.2 are similar across three different measures without
California and Colorado, I focus on the marijuana possession arrest rates that are most commonly

16

Note that, in Figure 1.1, the national averages of arrest rate are also lower and the prices are
higher around the early 2000s. It is possible that marijuana use in Colorado indeed decreased
around the early 2000s and was unrelated to changes in the police rate; or it could be a combination
of both the decrease in marijuana use and the level of law enforcement. In fact, the marijuana
treatments in Colorado from the TEDS data show a similar pattern.
21

used and exclude these two states from my sample hereafter. The rest of the results in this section
based on the two arrest ratios are qualitatively similar and available upon request.

-2

-1

0

1

2

Colorado

1988

1992

1996

Year

MJ Arrest Rate

2000

2004

2008

Police Rate

-2

-1

0

1

2

3

California and Other Medical Marijuana States (w/o CO.)

1988

1992

1996

Year

MJ Arrest Rate in CA
Police Rate in CA

2000

2004

2008

MJ Arrest Rate in MJ States
Police Rate in MJ States

Figure 1.2: Marijuana Arrest Rates and Police Officer Rates

22

Table 1.3: Effects of Medical Marijuana Laws (Restricted Sample)
(1)
(2)
(3)
(4)
(5)
(6)
Arrest Rates (per 100k) for Adult Males
Law

Model
Controls
Obs.

0.282***
(0.095)

0.225***
(0.076)

Log-Linear
No
No
8,722
8,722

0.301***
(0.071)

0.239***
(0.064)

Log-Linear
Yes
Yes
7,884
7,884

0.168***
(0.060)

0.198***
(0.042)

FE Poisson
No
No
8,722
8,722

# of States

48

48

48

48

48

48

Time
trends

Linear

Quadratic

Linear

Quadratic

Linear

Quadratic

Note.― All specifications include city and year fixed effects. Robust standard errors are reported
in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1. Control
variables include city police officer rates, and state level variables: black male rates,
unemployment rates, state and local government police expenditures, state and local government
health and hospital expenditures, and state 0.08 BAC laws.
In Table 1.3, the first two columns, (1) and (2), show the estimates of β from Equation (3),
with California and Colorado excluded. They are highly significant and indicate a 22.5−28.5%
increase in the marijuana possession arrest rate among adult males. Although not reported here,
the estimates indicate a positive and significant effect of around 17% on adult females. In the next
two columns, (3) and (4), I include city police officer rates per 100,000 city residents and other
state-level controls, including black male population rates, unemployment rates, per capita local
and state expenditures on police protection, per capita local and state expenditures on health and
hospital expenditures, and state 0.08 Blood Alcohol Content laws. The sample size is smaller
because 2001 and 2003 government expenditures were not developed by the Census Bureau due
to sample redesigning. These estimates are nearly identical to results in (1) and (2). Since most
state-level controls are actually poorly estimated, it seems that fixed effects and city-specific time
trends have accounted for most of the variations from these controls. Therefore, I use the
23

specification without any controls hereafter as it includes more years in the sample. To check
whether the results are sensitive to functional forms, in Columns (5) and (6), I estimate a fixedeffect Poisson model in which the dependent variable is the marijuana possession arrest rate. With
quadratic time trends, the point estimate from the fixed-effect Poisson model (the partial effect on
the logarithm of the conditional mean), 0.198, is very close to the estimate from the log-linear
model, 0.225 (the partial effect on the conditional mean of the log arrest rate).
Figure 1.3 provides graphical evidence for the effect of medical marijuana laws on arrests.
The upper graph shows the average adult male marijuana arrest rates (in logarithms) before and
after the passage of medical marijuana laws, where the X-axis measures the year relative to the
state’s law change, with 0 denoting the year of enacting the law, 1 denoting the following year,
and so on. To create a synthetic control group, I compute an average of the log arrest rates in nonmedical marijuana states for each year, and then take a weighted average of these yearly averages,
in which the weights come from the relative composition of years in the treatment group (medical
marijuana states). For instance, in “Year 0,” 58% of observations in the treatment group are from
Oregon and Washington, which passed the laws in 1998 (coded as 1999); 2% of observations are
from Maine, which passed the laws in 1999 (coded as 2000); and so forth. So the weight put on
the year 1999 average arrest rate in the control group is 0.58; the weight put on the year 2000
average arrest rate is 0.02, and so on. In other words, in “Year 0,” 58% of the observations in the
control group are selected from 1999, 2% are from 2000, and etc. The treatment group shows a
significant jump in the arrest rate from “Year -1” to “Year 0.” The arrest rate seems to decline in
“Year 3” and “Year 4”; however, it is a coincidence because most observations in “Year 3” and
“Year 4” are from the early 2000s, when the arrest rates were relatively lower nationally (see
Figure 1.1). To illustrate this directly, in the lower graph, I create a graph similar to the upper one,

24

5
4.9

4.8
4.7
4.6
4.5

-3

-2

-1
0
1
2
3
4
5
Year Relative to Pass Law
Other States
Medical Marijuana States

-4

-3

-2

-.5

-.4

-.3

-.2

-.1

0

-4

-1
0
1
2
3
4
5
Year Relative to Pass Law
Medical Marijuana States
Other States

Figure 1.3: Marijuana Arrest Rates Before and After the Passage of Laws
but I remove the national trend by using the residuals of log arrest rates that partial out year fixed
effects. After removing the national trend, the control group is nearly a horizontal line. In contrast,

25

the treatment group shows a persistent jump after legalization. Note that the magnitude of the jump
is about 0.15, which is close to the regression results above.
In Table 1.4, I investigate the dynamic responses of the adult male arrest rate to the
adoption of medical marijuana laws. In the first two columns, I replace Lawst with a set of dummy
variables, Years 0–1 through Years 8–9 (the maximum lag), which indicate each two-year interval
after the medical marijuana laws were enacted. Note that the estimates for later years are driven
mostly by Oregon and Washington. The estimated standard errors become larger when squared
city time trends are included, but the magnitudes stay similar. (The estimated standard errors from
arrest ratios are around 40–60% smaller and these estimates become significant). Although the
estimated effects on marijuana arrests seem to be increasing over time, a Wald test cannot reject
the null hypothesis that the estimates for Years 0–1 through Years 6–7 are identical. (It is able to
reject the null hypothesis when the estimates for Years 8–9 are included.) Therefore, the restriction
in Lawst of a constant legalization effect should be reasonable, and it is consistent with the
graphical evidence in Figure 3. The latter two columns include an additional dummy, Years
(neg.1–2), which indicates the two-year interval before the passage of the laws. The estimates for
this dummy are small and insignificantly different from zero, and the estimates remain similar for
post-law dummies, which indicates that policy endogeneity is not a serious concern in this context.
The results are quantitatively similar if I include another dummy that indicates years three and four
before the passage of laws (not reported).
Table 1.5 lists the means and standard deviations of marijuana possession arrest rates per
100,000 city residents in each age group for states with and without medical marijuana laws. States
with medical marijuana laws have lower arrest rates in all age groups, but the relative distributions

26

Table 1.4: Dynamic Responses of Marijuana Arrest Rates to Legalization
Arrest Rates (per 100k) for Adult Males
Years
(neg. 1 – 2)

0.036
(0.108)

0.012
(0.095)

Years
0–1

0.322***
(0.074)

0.298***
(0.086)

0.347**
(0.131)

0.310**
(0.149)

Years
2–3

0.373**
(0.150)

0.301
(0.216)

0.403**
(0.192)

0.318
(0.274)

Years
4–5

0.401*
(0.202)

0.298
(0.299)

0.437*
(0.231)

0.319
(0.354)

Years
6–7

0.507**
(0.211)

0.384
(0.368)

0.549**
(0.251)

0.411
(0.442)

Years
8–9

0.723***
(0.228)

0.669
(0.505)

0.770***
(0.278)

0.703
(0.594)

Obs.
8,722
8,722
8,722
8,722
# of States
48
48
48
48
City time trends
Linear
Quadratic
Linear
Quadratic
Note.― All specifications include city and year fixed effects. Robust standard
errors are reported in parentheses, and they are clustered at the state level. ***
p<0.01, ** p<0.05, * p<0.1.

of arrests among age groups are similar. The arrest rates are highest among those aged 18–24 and
declines with age, which is generally consistent with underlying marijuana use from survey data.
Table 6 reports the effects of medical marijuana laws in each age group, in which I control for city
quadratic time trends along with city and year fixed effects.17 Because there are many zero values
in each age group, the results are estimated from a fixed-effect Poisson model. The relative
magnitudes of these estimates are consistent with the age distribution in Table 5. They are larger

17

In the lower panel, age 45+ is estimated only with linear time trends. Because nearly 20% of
observations are zeros, there is not enough variation after controlling for linear trends and fixed
effects.
27

Table 1.5: UCR Descriptive Statistics in Each Male Age Group (1988–2008)
MeanMedical marijuana
All states
Other states
difference
states
Tests
Mean
SD.
Mean
SD.
Mean
SD.
t-stat
Ages
18 – 20
21 – 24
25 – 29
30 – 34
35 – 39

47.54
39.41
28.29
17.51
11.62

(38.50)
(34.66)
(25.90)
(16.39)
(11.73)

34.70
27.64
20.15
13.26
9.53

(29.71)
(21.13)
(15.41)
(10.37)
(7.69)

48.69
40.47
29.01
17.89
11.81

(38.99)
(35.43)
(26.52)
(16.77)
(12.01)

-9.36
-9.53
-8.80
-7.26
-5.00

40 – 44
45+

7.29
6.79

(8.00)
(8.61)

6.31
6.54

(6.09)
(7.56)

7.38
6.81

(8.15)
(8.70)

-3.41
-0.82

Obs.

8,722 (562 cities)

715 (48 cities)

8,007 (514 cities)

Note.― Medical marijuana states include only states that passed laws before July 2008; states
that passed laws afterward are in "other states." California, Colorado, and Vermont are not in
the sample. The t-statistics are from mean-difference tests for medical marijuana states and other
states.

among people under age 30 and decrease with age, and as expected, the estimates for the oldest
age groups (ages 40–44 and age 45+) are small and insignificant. Although not reported here, I
also find a positive effect on juveniles aged 15–17.18
Medical marijuana laws could increase marijuana use through an increase in availability or
a decrease in perceived risk. In principle, these two factors can affect each other and they are
probably codetermined by unobservables such as local public attitudes. For example, people may
perceive a lower risk and have greater availability if there are a lot of dispensaries around their

18

The estimates on juveniles aged 15–17 are around 10–18%. However, because the data on
juvenile crime and custody rates are much less complete than the associated data for adults,
interpretations based on these estimates should be cautious. Also, the juvenile justice system is
very different from the adult system in areas such as its procedures, incentives, and sanctions
(Carpenter, 2007; Levitt, 1998; Terry-McElrath et al., 2009). In fact, although the adult arrest rates
in medical marijuana states are much lower than in non-medical marijuana states, the juvenile
(ages 12–17) arrest rates are similar.
28

Table 1.6: Effects of Medical Marijuana Laws in Each Age Group
Marijuana Possession Arrest Rate per 100k
Age
18 – 20
21 – 24
25 – 29
Law

0.215***
(0.034)

0.226***
(0.049)

0.271***
(0.068)

30 – 34

35 – 39

40 – 44

45+

0.180**
(0.085)

0.177***
(0.058)

0.052
(0.058)

-0.015
(0.074)

Obs.
8,722
8,722
8,722
8,722
8,722
8,722
8,722
# of States
48
48
48
48
48
48
48
Note.― All specifications include city and year fixed effects and city quadratic time trends. (Only
linear time trends for ages 45+.) Robust standard errors are reported in parentheses, and they are
clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

neighborhoods, while the number of dispensaries depends on the actions of local police. On the
other hand, as noted in the previous sections, the number of legal patients and dispensaries was
relatively small in the sample period, especially without California and Colorado in the sample.19
Therefore, if marijuana arrests indeed increase due to these laws, changes in perceived legal and
health risks should be an important factor.
In Table 1.7, I indirectly test whether the increase in marijuana arrests is consistent with a
decline in perceived risks. If the perceived risks of marijuana are already low in a state, then the
implied effect on marijuana use in that state will be smaller. A higher referendum passage rate (see
Table A1) can be a proxy for lower perceived risk and a more open attitude towards marijuana.
Because of the different legal processes, it is also plausible that referendum states are more liberal
towards marijuana than lawmaker states. In the left panel of Table 7, Law× Referendum is the
interaction term of Lawst and a dummy denoting referendum states, and Law× Pro Rate is the

19

The ideal measure for availability would be the number of dispensaries in each city, but these
data are generally unavailable. On the other hand, the number of marijuana dispensaries was
actually small before 2005 even in California.
29

Table 1.7: Smaller Legalization Effects in States with Lower Perceived Risks
Referendum

Possession Limits

Law

0.537***
(0.175)

0.446***
(0.086)

Law

0.583***
(0.180)

0.363**
(0.174)

Law* Pro Rate

-0.006*
(0.003)

-0.005**
(0.002)

Law* Plant
Limit

-0.020**
(0.009)

-0.009
(0.009)

Law

0.567***
(0.174)

0.473***
(0.078)

Law

0.474***
(0.141)

0.302**
(0.131)

Law*
Referendum

-0.376*
(0.189)

-0.301***
(0.109)

Law* Ounce
Limit

-0.014*
(0.007)

-0.005
(0.007)

Obs.
# of States

8,722
48

8,722
48

Obs.
8,722
8,722
# of States
48
48
City time
City time trends
Linear
Quadratic
Linear
Quadratic
trends
Note.― All specifications include city and year fixed effects. Robust standard errors are
reported in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05,
* p<0.1.

interaction term of Lawst and passage rate (%) of referenda (0 for lawmaker states). As shown in
the table, the estimates for Law× Pro Rate and Law× Referendum are negative and significant.
Therefore, they support the above prediction that the arrest rate in referendum states with a high
passage rate will increase less than states with a low passage rate or lawmaker states.
Higher possession limits on the amount of marijuana could be another proxy for greater
availability or lower perceived risk. Except for California, which was without a limit before 2004,
all medical marijuana states specified the amount allowance in ounces and numbers of plants when

30

the referenda or senate bills were signed into effective laws.20 In the right panel of Table 7, Law×
Plant Limit is the interaction term of Lawst and the state limits on marijuana plant possessed, and
Law× Ounce Limit is the interaction term of Lawst and the state limits on ounces possessed. The
estimates for Law× Plant Limit and Law× Ounce Limit are negative (significant under the linear
trend specification), suggesting that states allowing larger amounts of marijuana possessed have
smaller legalization effects on illegal marijuana use. This implies that the legalization effects are
smaller in states with potentially lower perceived risks. As these possession limits may also reflect
availability, these estimates are consistent with the idea that an increase in availability is not the
major mechanism for the positive effects of these medical marijuana laws.
In Table 1.8, I further investigate whether the increase in marijuana arrests could be due to
other unobserved factors. Before 2009, all states that initiated the legalization process successfully
passed these laws except for South Dakota, which failed twice in 2006 and 2010. Many states
passed laws through referenda, suggesting that medical marijuana states could be different from
other states in many aspects. In fact, medical marijuana states are geographically concentrated in
the West Coast and the Northeast, and they tend to be more liberal (Hall and Schiefelbein, 2011).
It is a concern that the increase in marijuana arrests is driven by common unobservables in states
initiating the legalization processes. Interestingly, there is one state, Arizona, which did pass a
referendum to legalize medical marijuana in 1996 that did not lead to an effective law.21 Arizona

20

Before 2008, Washington only had a vague “60-day supply” limit. The limit was fixed to 24
ounces and 15 plants on November 2008; I use this limit for Washington. Some states distinguish
between mature and immature plants, while some other states do not. I use the possession limit on
the total number of plants regardless of plant maturity.
21
Arizona passed a referendum in 1996 (Proposition 200) that legalized medical marijuana under
doctors’ prescription. However, the state legislature dismantled it through the terminology —
31

Table 1.8: Indirect Tests for Changes in Unobservables

Effect of Arizona 1996 Law
-0.198***
(0.027)

-0.127***
(0.025)

Effect of Laws on the
Proportion of Blacks
0.002
(0.006)

-0.001
(0.044)

8,722
8,722
8,711
8,711
Obs.
# of States
48
48
48
48
City Time
Linear
Quadratic
Linear
Quadratic
Trends
Note.― All specifications include city and year fixed effects. Robust standard errors
are reported in parentheses, and they are clustered at the state level (48 states). ***
p<0.01, ** p<0.05, * p<0.1.

could serve as a counterfactual for other medical marijuana states as if they did not legalize and
set up medical marijuana programs. If the increases in marijuana arrests were not due to
legalization but some common unobservables, we would also observe a similar increase in Arizona.
However, the estimates in the first two columns in Table 1.8 show that the marijuana arrest rate in
Arizona actually decreased after the passage of its 1996 referendum.22
I also test whether the racial composition changes after legalization, which would be a
“smoking gun” that the increase in marijuana arrests is a result of police actions. In the next two
columns of Table 8, the dependent variables are the ratios of marijuana arrests of adult African
Americans to marijuana arrests of all adults.23 Because there are some zero values, the results are

“prescribe.” Because marijuana is a Schedule I drug, federal law prohibits physicians from
“prescribing" it. This made Proposition 200 ineffective.
22
The negative estimates for Arizona also suggest that the positive legalization effects are not
from the publicity surrounding the legal processes. This is consistent with the results from Table
1.4 and Table 1.7; if the publicity is the major cause, the estimated effects should be decreasing
over time and larger in referendum states.
23
The ratio of black arrestees includes females since the UCR does not separate gender within
races.
32

estimated by a fixed-effect Poisson model in these two columns. It is well documented that
minorities, especially African Americans, are much more likely to be arrested for marijuana
possession (Ramchand et al., 2006). It could be that African Americans tend to live in
disadvantaged neighborhoods that attract more police attention, or it could be due to potential
racial profiling. As a result, African Americans are often disproportionally affected by police
actions. A controversial instance that attracts much attention is New York City’s “stop and frisk”
practice, which exhibits significant racial disparities in low-level marijuana possession arrests
(Fellner, 2009; Golub et al., 2007). So, if the increase in marijuana possession arrests was a result
of police targeting low-level marijuana offenses, then police would patrol disadvantaged
neighborhoods or make stops on suspicious individuals more frequently, which should have a
greater impact on blacks than whites. However, in Table 1.8, I do not find that the proportion of
black arrestees increased after states enacted their medical marijuana laws.
As discussed in the model section, the estimates above from Tables 1.2–1.8 are robust to
any systematic changes in law enforcement or other measurement errors that are smoothly trending.
The estimates from Table 1.2 suggest that the positive estimated effects are not a result of changes
in total resources available to law enforcement. And the estimates from Table 1.3 are robust to
controlling for the number of police officers. In addition, the test based on racial composition from
Table 1.8 does not show any direct evidence of changes in police behaviors. Nonetheless, it is still
a concern that other aspects of law enforcement might be affected by legalization. For example,
since the legal boundary is blurred due to legalization, law enforcement needs more resources to
fight marijuana dealers because some of them are now under legal cover. Limited by the overall
resources available, some legal resources may be reallocated from other offenses to marijuana
sale/manufacture offenses. If such reallocation occurs, my estimates based on marijuana

33

possession arrests could be biased upward due to a spillover effect on low-level possession
offenses. To address this concern, I supplement the analysis by studying substance treatment
admissions that are not directly influenced by law enforcement. As Figure 1 indicates, marijuana
arrests move closely with daily marijuana use. Because heavy users are likely to be associated with
dependence and need for treatment, I use marijuana treatment patients to provide direct evidence
that medical marijuana laws increase heavy use in the next section.

4. Analysis of the Treatment Episode Data Set
4.1.

Data
The treatment data is from the Substance Abuse and Mental Health Services

Administration’s (SAMHSA) Treatment Episode Data Set (TEDS) for the years 1992 through
2008. The TEDS collects admission data from all substance-abuse treatment facilities that receive
public funding in each state. Some states collect data on all patients in these publicly funded
facilities, but some other states only collect data on publicly funded patients. For each admission,
the data identifies the primary, secondary, and tertiary substance abuse problem of the patient,
his/her demographics such as gender and age, referral sources, and the number of prior treatments
the patient had received. Similar to the UCR, each admission does not represent an individual, but
it is possible to create a measure representing individuals by using admissions without any prior
treatment. To be consistent with the previous analysis of the UCR arrests, I also focus on adult
(above age 18) male admissions and exclude California and Colorado.24 About 40% of treatment

24

The estimates for the California law are negative but not always significant; the estimated
effects of medical marijuana laws are quantitatively similar if California is included. The estimates
for Colorado show an implausible decrease in marijuana treatment ratios of 30–50%, which is
partly due to a dramatic increase (around 80%) in alcohol primary treatments in 2002. However,
34

admissions are referred by the criminal justice system, 30% are referred by patients themselves or
other individuals, and around 20% are referred by health care providers and alcohol or drug abuse
care providers.25 I exclude admissions referred by the criminal justice system and therefore the
data are not directly affected by changes in law enforcement. I also separately consider admissions
referred only by health or substance abuse care providers as these medical professionals reflect the
use of professional criteria of marijuana abuse that are not biased by the general public’s perception
of risk.
Since some states only collect data on publicly funded patients, probably due to changes in
available funding, the total number of admissions greatly fluctuates in some state-years. For
instance, the total number of admissions in Washington after 1998 was only about half of the
previous level. Therefore, as commonly used by the SAMHSA, I create ratios of marijuana
treatments to all substance treatments within non-criminal justice referrals or professional referrals
for each state. I define marijuana-related treatment admissions as such if marijuana is identified as
the primary, secondary, or tertiary abuse problem, and marijuana-primary treatment admissions as
such if marijuana is recorded only as the primary abuse substance. The sample includes all medical
marijuana states that passed laws before July 2008; except for Alaska, data from which is missing
for most years, they have data in every year.26 The upper two panels in Table 1.9 present means

the estimates still show around a 10% decrease in Colorado using treatment ratios excluding
alcohol-primary treatments. In fact, the yearly changes in marijuana treatments in Colorado are
quite similar to those in marijuana arrests. Both treatment and arrests drop in the early 2000s and
increase afterward.
25
The remaining 10% of admissions are referred by community or religious organizations, and
self-help groups such as Alcoholics Anonymous (AA). School referrals are very small as I focus
on adults.
26
Alaska does not report referral sources for the years 1998–2003, and it does not report any data
for the years 2004–2007.
35

and standard deviations for these treatment ratios (in percentage points) among non-criminal
justice referrals and professional referrals. Nearly one-third of patients have marijuana abuse
problems, but fewer than 8% of patients have marijuana as their primary problem. 27 This is
consistent with the notion that, while marijuana is the most commonly abused illegal substance,
marijuana itself is not strongly addictive.
To obtain a measure representing individuals, I also construct ratios using only first-time
marijuana treatments.28 This measure can avoid potential bias from recidivism that is a problem
particularly for using treatment data (Anderson, 2010). A drawback of this measure is that
information on the number of previous treatments is largely missing in some state-years.
Fortunately, except for Washington, which does not report this information for the years 1992–
1999, it is very complete in medical marijuana states, with an average missing rate of 1.7%. I
restrict the sample to state-years that are missing less than 50% of the information on the number
of previous treatments, and scale the treatment ratios by the proportion of reporting data in each
state-year.29 In Table 1.9, the lower two panels show the descriptive statistics for these first-time
treatment (scaled) ratios among non-criminal justice referrals and professional referrals (in
percentage points). Note that the denominators are the same as the upper panels, so it suggests that
roughly half of marijuana treatment patients are first-time patients.

27

Among non-criminal justice referrals, 50% of their primary problems are alcohol, and around
30% are cocaine and heroin.
28
It is not possible to observe whether a patient has had prior treatment episodes for a particular
substance; only the number of previous treatment episodes a patient has had for any drug or alcohol
problem is available.
29
I also exclude 20 state-years (including Rhode Island in 2003 and 2004) that report zero firsttime treatments for any substances as they are probably missing data. The regression results in
Table 11 are slightly greater without scaling (but the estimated standard errors are 10–20% larger).
36

Table 1.9: TEDS Descriptive Statistics (1992–2008)
All states

Medical
marijuana states

Other states

Marijuana Treatment Ratios among Adult Males
Mean
SD.
Mean
SD.
Mean
SD.
Marijuana31.19 (9.23)
32.39 (9.25)
30.88 (9.21)
related
Marijuana8.13
(3.57)
7.36
(2.86)
8.33
(3.71)
primary
Marijuana Treatment Ratios among Adult Males (Professional Referrals)

Meandifference
Tests
t-stat
1.85
-3.09

Mean
Marijuanarelated
Marijuanaprimary
Obs.

SD.

Mean

SD.

Mean

SD.

t-stat

32.34

(9.03)

34.40

(9.01)

31.82

(8.97)

3.24

7.69

(3.75)

7.46

(3.53)

7.75

(3.81)

0.86

787 (49 States)

160 (10 States)

627 (39 States)

First-time Marijuana Treatment Ratios among Adult Males
Mean
SD.
Mean
SD.
Mean
SD.
t-stat
Marijuana14.46 (6.64)
14.40 (6.33)
14.48 (6.73)
-0.13
related
Marijuana4.75
(2.39)
4.17
(2.02)
4.91
(2.46)
-3.36
primary
First-time Marijuana Treatment Ratios among Adult Males (Professional Referrals)
Mean
SD.
Mean
SD.
Mean
SD.
t-stat
Marijuana13.11 (7.21)
14.34 (7.25)
12.77 (7.16)
2.36
related
Marijuana3.87
(2.33)
3.85
(2.35)
3.88
(2.33)
0.15
primary
Obs.
690 (47 States)
150 (10 States)
540 (37 States)
Note.― The means and standard deviations are in percentage points. Medical marijuana states
include only states that passed laws before July 2008; states that passed laws afterward are in
"other states." California and Colorado are not in the sample. The sample for first-time treatments
only includes state-years that the information on prior treatment is missing less than 50%. The tstatistics are from mean-difference tests for medical marijuana states and other states.

In contrast with the arrest data, the cross-sectional difference in treatments between
medical marijuana states and other states seems to be consistent with survey data. For both all
treatments and first-time treatments, except for one instance, the marijuana-related treatment ratios

37

are significantly higher in medical marijuana states than in other states. This is consistent with the
higher prevalence rates in medical marijuana states from survey data. Interestingly, marijuanaprimary treatments are lower in medical marijuana states, but the differences are only significant
when individual referrals are included. The occurrence that marijuana is less likely to be the
primary abuse problem for only individual referrals could reflect the lower perceived risks in
medical marijuana states.

4.2.

Results
To examine the effects of medical marijuana laws on marijuana treatments, I estimate the

following model by OLS:

(4)

Yst = β Lawst + State fixed effectss + Year fixed effectst + State time trendsst + εst ,

where Yst is the logarithm of marijuana treatment ratio in state s and year t. As in the UCR analysis,
I do not include any controls to keep a larger sample size. The results in this section are nearly
identical when the same set of state-level controls is included or estimated by a fixed-effect Poisson
model (not reported).
Table 1.10 shows the estimates of β from Equation (4) for all treatment ratios (with any
number of previous treatment episodes). The upper panel is for non-criminal justice referrals, and
the lower panel is for professional referrals. The first two columns show the estimates on
marijuana-related treatments. The estimates are very similar in non-criminal justice referrals and
professional referrals. In terms of percentage change, on average, medical marijuana laws
increased the marijuana-related treatment ratio by 9.3–11.0% among non-criminal justice referrals

38

Table 1.10: Effects of Medical Marijuana Laws on Marijuana Treatments
Marijuana-related

Marijuana-primary

Marijuana Treatment Ratios
Law

0.110***
(0.030)

0.093**
(0.038)

0.145**
(0.059)

0.139**
(0.068)

0.095
(0.057)

0.173***
(0.065)

Marijuana Treatment Ratios (Professional Referrals)
Law

0.090**
(0.034)

0.121***
(0.027)

Obs.
787 ( 49 States)
787 ( 49 States)
Time Trends
Linear
Quadratic
Linear
Quadratic
Note.― All specifications include state and year fixed effects. Robust standard
errors are reported in parentheses, and they are clustered at the state level. ***
p<0.01, ** p<0.05, * p<0.1.

and by 9.0–12.1% among professional referrals. In the next two columns, the estimated effects are
larger (but noisier) for marijuana-primary treatments. For non-criminal justice referrals, these laws
are associated with a 13.9–14.5% increase in the marijuana-primary treatment ratio. For
professional referrals, perhaps due to smaller observations, the estimate magnitudes are a little
sensitive to time trend specification. They indicate a 9.5–17.3% increase in the marijuana-primary
treatment ratio. Note that the marijuana-primary treatment ratio and the ratio of marijuana arrests
to all drug arrests are roughly comparable measures, and the estimated effects of 14% in marijuanaprimary treatment ratios are very close to the estimates from the all-drug arrest ratio in Table 2.30
Although not reported here, I also find similar legalization effects on secondary and tertiary

30

To make the treatment ratios more comparable to all drug arrest ratios, I also create marijuanaprimary treatment ratios that exclude alcohol-primary treatment admissions, and the estimates are
quantitatively similar (not reported).
39

marijuana treatments. In contrast to previous studies that use measures for general use rates among
juveniles and do not find any effects from these laws, the estimates in Appendix C indicate a 7.1–
14.5% increase in juvenile treatments.
Table 1.11 shows the estimates based only on first-time marijuana treatments. Since a
proportion of addictive patients will repeatedly enter treatment, we would expect the estimates
based on first-time treatments to be smaller than estimates from all treatments. The popular notion
that marijuana is a gateway drug also suggests a smaller estimate from first-time patients as the
proportion of patients reporting cocaine and heroin abuse is actually monotonically increasing with
the number of previous treatments.31 Somewhat surprisingly, all of the estimates from Table 1.11

Table 1.11: Effects of Medical Marijuana Laws on First-Time Marijuana
Treatments
Marijuana-related

Marijuana-primary

First-Time Marijuana Treatment Ratios
Law

0.257**
(0.111)

0.182**
(0.070)

0.208
(0.145)

0.157**
(0.075)

First-Time Marijuana Treatment Ratios (Professional Referrals)
Law

0.215**
(0.086)

0.218***
(0.061)

0.172
(0.122)

0.227*
(0.114)

Obs.
690 (47 states)
690 (47 states)
Time Trends
Linear
Quadratic
Linear
Quadratic
Note.― All specifications include state and year fixed effects. Robust standard
errors are reported in parentheses, and they are clustered at the state level. ***
p<0.01, ** p<0.05, * p<0.1.

31

For first-time marijuana-related treatment admissions, 37% of patients also report cocaine abuse
and 6% report heroin abuse. On the other hand, among patients with at least one previous treatment,
the proportion that reports cocaine and heroin abuse increases to 49% and 11%, respectively.
40

are nearly twice as great, regardless of the referral sources. Specifically, on average, medical
marijuana laws are associated with a 15.7–25.7% increase in first-time treatments. Note that the
estimates from first-time treatments are around twice as great and first-time treatments also
account for half of all marijuana treatments, which implies that the estimates in Table 1.10 are
entirely driven by first-time treatments. In fact, I estimate the effects of laws on patients with at
least one previous treatment episode, and the results are nearly zero (not reported). Because firsttime treatments represent individuals, it suggests that medical marijuana laws increase the number
of new heavy users by around 20%.
It is straightforward to see graphically that the estimates in Table 1.10 and 1.11 are driven
by first-time treatments. Figure 1.4, constructed in the same way as Figure 1.3, shows the effects
of laws on marijuana-related treatment ratios (in logarithm) among non-criminal justice referrals.
The upper graph is from all treatments, and the lower graph is from first-time treatments. The scale
in the lower graph is twice as great as the scale in upper the graph. (The scale in the upper graph
is the same as Figure 1.3). Both graphs show similar patterns of increase in marijuana-related
treatments after the passage of laws in medical marijuana states, but the magnitude of increase is
much greater in first-time treatments.
In summary, consistent with the results from the UCR marijuana arress, the estimates from the
TEDS indicate a 10–20% increase in marijuana treatments after medical marijuana legalization.
This positive effect largely comes from the increase in first-time marijuana treatments, which has
some interesting implications. First, it suggests that medical marijuana laws do not have a
significant effect on strongly addictive patients who repeatedly enter treatment. These patients
could be “always-takers” who would be heavy marijuana users regardless of marijuana’s legal

41

status. Second, consistent with existing medical evidence, it implies that marijuana is not strongly

-1.5

-1.4

-1.3

-1.2

-1.1

-1.0

addictive. Finally, because repeated patients consist of a greater proportion of cocaine and heroin

-4

-3

-2

-1
0
1
2
3
Year Relative to Pass Law

5

Other States

-2.5

-2.3

-2.1

-1.9

-1.7

-1.5

Medical Marijuana States

4

-4

-3

-2

-1
0
1
2
3
Year Relative to Pass Law

Medical Marijuana States

4

5

Other States

Figure 1.4: Marijuana All Treatment Ratios (upper) and First-Time Treatment Ratios
(lower) Before and After the Passage of Laws

42

users, it does not support the popular belief that the use of marijuana increases usage of hard drugs.
In fact, there could even be a substitution between hard drugs and marijuana.

5.

Discussion of Results and Conclusion
In this paper, I estimate the effects of medical marijuana laws on illegal marijuana use

based on marijuana possession arrests. To address potential bias from changes in law enforcement,
as those who are arrested for marijuana possession are likely to be heavy users, I also use marijuana
treatments that are not referred by the criminal justice system as another proxy for heavy usage.
Based on an empirical model that is able to account for state unobservables both at their levels and
trends, I find that medical marijuana laws are associated with a 10–20% increase in marijuana
arrests and treatments, suggesting a positive legalization effect on illegal marijuana use.
Based on existing studies, MacCoun (2010) suggests that the non-price effect of marijuana
decriminalization is around a 35% increase in general use rate (use in past month). Although
medical marijuana laws represent a less dramatic change than decriminalization, a 10–20%
increase is not particularly large for heavy users, as previous research suggests that heavy users
could be disproportionately responsive to legal changes (Becker and Murphy, 1988; Carpenter,
2004). This magnitude is also comparable to policy changes regarding alcohol and their associated
substitution effects with marijuana (Conlin et al., 2005; DiNardo and Lemieux, 2001). Because
“heavy use” is a mix of measures for both the intensive and extensive margins, conceptually, a
significant part of the effect can be viewed as an increase in the intensive margin.
One limitation of this study is that the arrest and treatment data are not able to separately
identify the extensive and intensive margins. Although the still very limited literature suggests a
small or nonexistent effect of these laws on the extensive margin, this conclusion should be treated

43

with great caution, since the estimates in existing studies often come with large estimated standard
errors. In fact, the noisy estimates are probably a result of the fact that some of the datasets or
measures are not handled appropriately. For instance, in Appendix C, I illustrate that the noisy
estimates based on the TEDS data for juveniles from Anderson, Hansen, and Rees (2012a) are
possibly a result of using population as a denominator for their marijuana treatment measures. In
addition, in Appendix D, I show that the estimates for adults from Harper et al. (2012) are
somewhat sensitive to the inclusion of state time trends, and they are also sensitive to whether
2009 data are included and how these medical marijuana laws are coded. Actually, based on the
model specification in this paper, I find some evidence of an increase of roughly 6% in marijuana
use rate among people aged 18–25 from the public-use NSDUH. At minimum, those estimates
suggest that the fixed-effect estimates from the public-use NSDUH are not very robust, which is
likely because the public-use NSDUH only provides two-year moving averages for the state-level
marijuana use rates. Moreover, as Anderson et al. (2012a) point out in their analysis of the National
Longitudinal Survey Youth 97, the sample sizes in smaller states are often quite small in many
representative datasets. In fact, to obtain a larger sample size and therefore to increase precision is
one of the main reasons that the NSDUH only provides the state-level estimates as two-year
moving averages (Wright, 2004).
In summary, due to the preliminary stage of the literature on medical marijuana laws, this
paper alone cannot be taken as definitive evidence; rather, it provides evidence that some indicators
of heavy marijuana use do respond to these medical marijuana laws. Since this paper cannot
provide direct information on whether these medical marijuana laws increase initiation rates
among general populations, future research will contribute to this ongoing policy debate,
particularly by separately identifying changes in either the extensive or intensive margins. Put

44

differently, as there is a large heterogeneity among drug users, the policy effects could be different
on different types of drug users. Future studies should be more careful when framing research
questions and correspondingly choose appropriate data and measures.
Although the estimates in this paper may only be appropriate for inference on heavy users,
they are still relevant to policy because heavy marijuana users are often associated with negative
health and social outcomes, such as developing dependence and the need for treatment. A 20%
increase in heavy users, as indicated by both arrests and first-time treatments, represents a
nontrivial cost to society. On the other hand, based on the estimates from all treatments, the net
effect on treatment is somewhat smaller, and therefore there could be substitution between
marijuana and other substances. This substitution can be viewed as a benefit of medical marijuana
laws. Other additional benefits may exist. For example, Anderson et al. (2012b) and Anderson et
al. (2012c) provide some evidence for a decrease in drunk driving and suicide. Therefore,
evaluating the effects of medical marijuana laws requires a more complete cost/benefit analysis
that is beyond the scope of this study.

45

APPENDICES

46

Appendix A: State Medical Marijuana Laws1
Table A1: Medical Marijuana Laws
State

Pass/Effective date

Pass Rate

Registration

Possession Limit

Alaska

Nov. 3, 1998
/Mar. 4, 1999

58%
(Ballot Measure 8)

Yes

1 oz/ 6 plants
(3 mature,
3 immature)

Arizona

Nov. 2, 2010

50.13%
(Proposition 203)

Yes

2.5 oz/ 12 plants

California

Nov. 5, 1996
/Nov. 6, 1996

56%
(Proposition 215)

Yes
(Voluntary since
Jan. 1, 2004)

Colorado

Nov. 7, 2000
/Jun. 1, 2001

54%
(Ballot Amendment 20)

Yes

8 oz/
6 mature or
12 immature
2 oz/ 6 plants
(3 mature,
3 immature)

96-51 House; 21-13
Senate
(HB 5389)
13-0 vote
(Amendment Act B18622)
27-14 House; 17-4
Senate
(Senate Bill 17)
32-18 House; 13-12
Senate
(Senate Bill 862)

Connecticut

May 31, 2012

D.C

May 21, 2010
/Jul. 27, 2010

Delaware

May 13, 2011
/Jul. 1, 2011

Hawaii

Jun. 14, 2000
/Dec. 28, 2000

Maine

Nov. 2, 1999
/Dec. 22, 1999

61%
(Ballot Question 2)

Massachusetts

Nov. 6,2012
/Jan. 1, 2013

63%
(Ballot Question 3)

Yes

Not specified yet

Michigan

Nov. 4, 2008
/Dec. 4, 2008

63%
(Proposal 1)

Yes

2.5 oz/ 12 plants

Montana

Nov. 2, 2004

62%
(Initiative 148)

Yes

1 oz/4 plants
(mature)

1

Yes

Not specified yet

Yes

2 oz/ Not
specified yet

Yes

6 oz

Yes

3 oz/ 7 plants
(3 mature,
4 immature)

Yes
(Mandatory after 2.5 oz/ 6 plants
Dec. 31, 2010)

For legal documents and detail, please see “18 Legal Medical Marijuana States and D.C.” from
ProCon.org website: http://medicalmarijuana.procon.org/view.resource.php?resourceID=000881.
47

Table A1 (cont’d)
Nevada

Nov. 7, 2000
/Oct. 1, 2001

New Jersey

Jan. 18, 2010

New Mexico

Mar. 13, 2007
/Jul. 1, 2007

Oregon

Nov. 3, 1998
/Dec. 3, 1998

Rhode Island

Jan. 3, 2006

Vermont

May 26, 2004
/Jul. 1, 2004

Washington

Nov. 3, 1998

65%
(Ballot Question 9)
48-14 House; 25-13
Senate
(Senate Bill 119)
36-31 House; 32-3
Senate
(Senate Bill 523)
55%
(Ballot Measure 67)
52-10 House; 33-1
Senate
(Senate Bill 0710)
82-59 House; 22-7
Senate
(Senate Bill 76)
59%
(Initiative 692)

48

Yes

1 oz/ 7 plants
(3 mature,
4 immature)

Yes

2 oz/ Not
specified yet

Yes

Yes

6 oz / 16 plants
(4 mature,
12 immature)
24 oz/ 24 plants
(6 mature,
18 immature)

Yes

2.5 oz/ 12 plants

Yes

2 oz/ 9 plants
(2 mature,
7 immature)

No

24 oz/ 15 plants

Appendix B: Different Sample Constructions from the Uniform Crime Report
In Appendix B, I check the robustness of the main results based on different constructions
of the sample. I estimate the effects of medical marijuana laws from a log-linear model or a FixedEffect Poisson model, and I separately estimate the legalization effects of California and Colorado
as Columns (5) and (6) in Table 1.2. The dependent variables are the arrest rate in the upper panel,
the ratio of marijuana possession arrests to all arrests in the middle panel, and the ratio of marijuana
possession arrests to all drug possession arrests in the lower panel [all dependent variables are in
logarithm except for Columns (5) and (6)].
In Columns (1) and (2), following Carpenter (2007), I scale arrest counts by a factor that
equals the fraction reported of a year (12 divided by the number of months reported) using agencies
that report at least six months (agencies that only report in December are excluded). In Columns
(3) and (4), I include city agencies that report any number of months without scaling. Since a
particular problem for the UCR data is that it is not able to distinguish a true zero from missing
data, in Columns (5) and (6), I create a sample based on the same criteria of city populations that
are greater than 25,000 in any year and 50,000 for at least one year. However, I treat marijuana
possession arrests from city agencies that report any positive adult male arrests for any drug
possession but marijuana as true zeros.2 Because of these zeros, I estimate a fixed-effect Poisson
model in Columns (5) and (6). In the last two columns, Columns (7) and (8), the dependent
variables are the state level arrest rates/ratios for adult males. I sum up marijuana possession arrests
to the state level from all available agencies, including any cities and non-cities that report to the
UCR, and create state-level arrest rates and arrest ratios. All of these estimates show a similar

2

I also use agencies report any positive arrests or any positive adult male arrests to estimate the
legalization effects on arrest rates and all arrest ratios. The results are quantitatively similar.
49

pattern to the estimates in Table 1.2. The estimates for Lawst are positive and indicate roughly a
20% increase in marijuana arrests. The legalization effect in California is generally positive, but
the magnitudes are sensitive to which measures were used. Colorado always shows a negative
legalization effect.
Table B1: Robustness Checks
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Cities reporting at Cities reporting any Cities reporting any All city and county
least 6 months with number of months arrests to the UCR agencies aggregated
scaling
without scaling
to state level
Arrest Rates (per 100k) for Adult Males
Law
0.229*** 0.178*** 0.333** 0.179*
0.142* 0.173** 0.298**
0.160
(0.082) (0.066) (0.142) (0.106) (0.080) (0.071) (0.126) (0.151)
CA Law -0.720*** -0.370*** -0.628*** -0.351*** -0.251*** -0.216*** -0.359** -0.147
(0.081) (0.068) (0.145) (0.109) (0.087) (0.077) (0.137) (0.160)
CO Law -0.233*** -0.0880 -0.387*** -0.192* -0.313*** -0.175** -0.518*** -0.259*
(0.079) (0.066) (0.144) (0.112) (0.079) (0.070) (0.113) (0.131)
Arrest Ratios (all arrests) for Adult Males
Law
0.282*** 0.200*** 0.370*** 0.193**
(0.079) (0.063) (0.131) (0.085)
CA Law -0.156* -0.061 -0.304** -0.151
(0.079) (0.062) (0.136) (0.091)
CO. Law -0.615*** -0.360*** -0.605*** -0.363***
(0.080) (0.065) (0.137) (0.091)

0.172** 0.170** 0.255*** 0.146*
(0.086) (0.076) (0.075) (0.082)
-0.124
-0.119 -0.160* -0.068
(0.094) (0.081) (0.082) (0.093)
-0.226** -0.174** -0.448*** -0.268***
(0.088) (0.078) (0.075) (0.073)

Arrest Ratios (all drug possession arrests) for Adult Males
Law
0.156** 0.123** 0.243*
0.171
0.124**
(0.075) (0.046) (0.139) (0.106) (0.062)

0.118**
(0.056)

CA Law

0.033
0.090** -0.111
-0.055
0.094
0.092
(0.073) (0.043) (0.142) (0.108) (0.069) (0.060)
CO Law -0.413*** -0.257*** -0.422*** -0.298** -0.291*** -0.256***
(0.075) (0.054) (0.145) (0.114) (0.062) (0.058)
Time
Linear Quadratic Linear Quadratic Linear Quadratic
Trends
(City)
(city)
(City)
(city)
(City)
(city)
Obs.

11,944

12,676

Note.― *** p<0.01, ** p<0.05, * p<0.1.
50

13,498

0.109*
(0.057)

0.087*
(0.044)

0.009
0.004
(0.059) (0.042)
-0.162** -0.140***
(0.061) (0.043)
Linear Quadratic
(state )
(state)
1,006

Appendix C: Effects of Medical Marijuana Laws on Juveniles Aged 15–17
In Appendix C, I estimate the effects of laws on male juveniles aged 15–17 based on the
TEDS data. Table C1 shows the descriptive statistics for marijuana-related and marijuana-primary
treatment ratios among male juveniles aged 15–17. I construct the sample based on all referrals,
including criminal justice referrals. The descriptive statistics from non-criminal justice referrals
are nearly identical, and the estimates are also quantitatively similar but slightly noisier. The main
reason for including criminal justice referrals is to compare with the results in the TEDS from
Anderson, Hansen, and Rees (2012a). Also, for comparison, California and Colorado are included.
Because some state-years have very small numbers of juvenile admissions, I exclude 17 stateyears with total admissions of all substances less than 20. As a result, Delaware is not in the
sample.3 As we would expect, marijuana is the most common abuse problem for juvenile treatment
patients; nearly 80% of juvenile patients report marijuana abuse, and nearly 60% of them have
marijuana as the primary problem. Similar to adults, marijuana-related treatment ratios among
juveniles are significantly higher in medical marijuana states. However, for juveniles, marijuanaprimary treatment ratios are not different from each other (even including California and Colorado).
Table C2 shows the estimated effects of medical marijuana laws on male juveniles aged
15–17 based on Equation (4). The marijuana-related treatments are in the upper panel, and the
marijuana-primary treatments are in the lower panel. In Columns (1) and (2), I use marijuana
treatment rates per state residents as the dependent variable in order to compare with the results
from Anderson, Hansen, and Rees (2012a), in which they also conduct an analysis for teenagers

3

Data in Alaska for 1998–2003 are in the sample because I can include state-years that have
missing data in referral sources. The estimates below are not sensitive to using a larger threshold
of excluding small numbers of treatment admissions for all substances.
51

Table C1: TEDS Descriptive Statistics for Male Juveniles Aged 15–17 (All Referrals)
MJ-related Treatment
MJ-primary
StateRatio (%)
Treatment Ratio (%)
year
All States
Non-MJ States
MJ States w/o CA &
CO
California

Mean
79.25

Std. Dev.
(14.72)

Mean
58.42

Std. Dev.
(17.16)

Obs.
803

77.96

(15.39)

58.47

(18.02)

606

83.36

(12.35)

57.23

(14.77)

163

85.18

( 3.99)

63.67

(7.97)

17

Colorado
79.87
(8.37)
62.74
(12.13)
17
Note.― MJ states include only states that passed laws before July 2008; states that passed laws
afterward are in non-MJ states. The sample includes both non-criminal justice referrals and
criminal justice referrals. I exclude 17 state-years that have less than 20 treatment admissions
for any substances. (Delaware is not in the sample.)

using the TEDS data. The estimates based on treatment rates are qualitatively similar to the results
from Anderson, Hansen, and Rees (2012a); they are small or even negative with very large
estimated standard errors.4
However, population is not an appropriate denominator for substance treatments from the
TEDS data. Because some states only collect data on publicly funded patients, the number of
admissions fluctuates greatly in some state-years, probably due to changes in available funding. A
large proportion of the variation in treatment rates will come from the changes in total treatment
admissions rather than changes in marijuana treatment admissions. This explains the large
estimated standard errors in Columns (1) and (2). In the next two columns, (3) and (4), I estimate
the effects of laws on treatment ratios of marijuana treatments to all substance treatments. The

4

The model and specification here are slightly different from those in Anderson, Hansen, and
Rees (2012a). Specifically, their dependent variables are the logarithm of marijuana-related
treatment gender specific rates per 100,000 of the population aged 15–17. Their specification
includes only state linear time trends and some state-level control variables. They also find similar
results for patients aged 18–20.
52

Table C2: Effects of Medical Marijuana Laws on Male Juveniles Aged 15–17
(1)
(2)
(3)
(4)
(5)
(6)
MJ Treatment Rates
(per state residents)

MJ Treatment Ratios

MJ Treatment Ratios

Marijuana-related Treatment for All Referrals Aged 15–17
0.087**
(0.035)

0.095
(0.057)

CA Law

-0.133***
(0.045)

-0.046
(0.060)

CO Law

-0.029
(0.030)

-0.049
(0.053)

0.145**
(0.058)

0.119**
(0.057)

CA Law

-0.251***
(0.064)

0.057
(0.059)

CO Law

-0.054
(0.061)

-0.044
(0.058)

Law

-0.018
(0.108)

-0.015
(0.123)

0.071**
(0.031)

0.087*
(0.048)

Marijuana-primary Treatment for All Referrals Aged 15–17
Law

0.026
(0.119)

0.017
(0.120)

0.116**
(0.054)

0.119**
(0.048)

Obs.
803 ( 50 States)
Time Trends
Linear
Quadratic
Linear
Quadratic
Linear
Quadratic
Note.― All specifications include state and year fixed effects. Robust standard errors are reported
in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

estimates show a 7.1–8.7% increase in the marijuana-related treatment ratio and an 11.6–11.9%
increase in the marijuana-primary treatment ratio among male juveniles aged 15–17. The last two
columns, (5) and (6), show the estimates separately for California and Colorado. Although they
are still negative, most of them are not significantly different from the estimates of Lawst. Unlike
the results for adults, the estimated effects of medical marijuana laws on male juveniles aged 15–
17 are quite similar with or without California and Colorado.

53

Although not reported, I also find a positive effect of 10–18% on the marijuana arrest rate
(per 100,000 city residents) and the ratio of marijuana arrests to all arrests among male juveniles
aged 15–17 (excluding California and Colorado). But the estimates are small and insignificant for
the ratio of marijuana arrests to all drug arrests among male juveniles aged 15–17. However, as
discussed in the paper (see Note 18), juvenile arrests could be largely determined by unobserved
heterogeneities across administrative areas and they should be treated with caution.

54

Appendix D: State-Level Data from the National Survey on Drug Use and Health
In Appendix D, I estimate the effects of medical marijuana laws using the state-level
estimates of past-month marijuana use and perceived risks of marijuana from the NSDUH data
provided by the Substance Abuse and Mental Health Services Administration (SAMHSA). I
estimate the effects based on Equation (4) in the paper and compare with the results from Harper,
Strumpf, and Kaufman (2012).
The state-level estimates of marijuana use and perceived risks of marijuana are only
available from 2002 and they are reported as two-year moving averages. 5 Specifically, the
SAMHSA estimates a logistic model using two years of data together with a list of predictors such
as racial composition, arrests for drugs and other crimes, treatment rates, and local economic
indicators. The predicted values from the model are reported as the state-level measures of drug
usage (Wright 2004). For example, the measure in the 2008 report is a predicted probability using
both the 2008 and 2007 data. The NSDUH is a national representative sample, but it oversamples
younger populations in order to obtain more precise estimates on drug use behaviors for youths. 6
The state-level estimates are separately available for three age groups that are equally sampled:
ages 12–17, 18–25, and 26 and above. Table D1 shows the descriptive statistics of past-month
marijuana use rates (in percentage points) and the percentage of people who perceive a great risk

5

For the first year available, 2002, the state-level estimates are based on only that year rather than
two years. The state level estimates are also available for 1999–2001; however, the SAMHSA
changed the survey procedure in 2002 and the response rates and substance prevalence rates were
significantly higher than previous years. Therefore, these data from 1999–2001 are not comparable
with later years.
6
Except for eight large states, the number of observations in each year is 900 in the other 42 states
and D.C.
55

Table D1: Descriptive Statistics for the NSDUH State-Level Data
All states

Medical marijuana
states

2002 to 2007–2008, Use in Past Month (%)
Mean Std. Dev.
Mean
Age12–17
7.64
(1.63)
9.23
Age 18–25
17.42
(4.35)
21.21
Age 26+
4.20
(1.17)
5.58

Std. Dev.
(1.62)
(4.40)
(0.99)

Other states

Mean
7.16
16.26
3.77

Meandifference
Tests

Std. Dev.
(1.29)
(3.62)
(0.86)

2002 to 200–2008, Perceptions of Great Risk of Smoking Marijuana Once a Month (%)
Mean Std. Dev.
Mean Std. Dev.
Mean Std. Dev.
Age 12–17
33.73
(4.49)
29.42
(2.76)
35.05
(4.07)
Age 18-25
23.19
(4.96)
19.53
(4.67)
24.31
(4.49)
Age 26+
41.00
(6.15)
35.12
(4.75)
42.81
(5.35)
Obs.
357
84
273
2002–2003 to 2008–2009, Use in Past Month (%)
Mean Std. Dev.
Mean Std. Dev.
Age 12–17
7.58
(1.62)
9.10
(1.53)
Age 18–25
17.47
(4.39)
21.14
(4.31)
Age 26+
4.25
(1.23)
5.65
(1.07)

Mean
7.06
16.21
3.78

Std. Dev.
(1.28)
(3.66)
(0.86)

t-stat
12.13
10.39
16.28

t-stat
-11.88
-8.45
-11.82

t-stat
13.28
11.31
17.88

2002–2003 to 2007–2008, Perceptions of Great Risk of Smoking Marijuana Once a Month (%)
Mean Std. Dev.
Mean Std. Dev.
Mean Std. Dev.
t-stat
Age 12–17
33.58
(4.53)
29.54
(2.78)
34.96
(4.18)
-12.32
Age 18–25
22.86
(4.94)
19.36
(4.42)
24.06
(4.51)
-9.21
Age 26+
40.59
(6.19)
35.06
(4.74)
42.48
(5.46)
-12.34
Obs.
357
104
304

in smoking marijuana once a month for these three age groups.7 I created two samples. In the
upper panel of Table D1, to be consistent with the UCR and TEDS analysis in the paper, I use the
2002 state estimates through the 2007–2008 state estimates. To compare with the results from
Harper, Strumpf, and Kaufman (2012), I also use the 2002–2003 estimates through the 2008–2009

7

The survey question is: How much do people risk harming themselves physically and in other
ways when they smoke marijuana once a month?
56

estimates (they do not use the 2002 data). In the lower panel, because Michigan passed a law in
November 2008, it is counted as a medical marijuana state. Note that only four other states,
Montana, New Mexico, Rhode Island, and Vermont changed their laws during the sample period.
It is clear from the table that medical marijuana states have higher usage rates and lower perceived
risks in all age groups.
As in the paper, I estimate a log-linear model, with or without controlling for state-specific
time trends, and the standard errors are clustered at the state level. I do not report the estimates for
juveniles aged 12–17 for brevity, as they are generally very noisy and similar to Harper, Strumpf,
and Kaufman (2012). In Table D2, I estimate the effect of medical marijuana laws on the marijuana
use rate and perceived risks among young adults aged 18–25. In the left panel, the coding of
medical marijuana laws is the same as the Lawst in the paper; the first year of legalization is coded
based on the effective date. In the right panel, the first year of legalization is coded based on the
passing date as in Harper, Strumpf, and Kaufman (2012). However, only 2004 Montana and 2008
Michigan are changed. The results from past-month use rates are in the upper two panels. For pastmonth use rates among ages 18–25, in both samples, the estimates are not sensitive to alternative
coding of the laws, but they are somewhat sensitive to time trends specifications. In the higher
upper panel, the 2002 through 2007–2008 sample, the estimates show around a 5.9–9.6% increase
in use rates without time trends or with quadratic time trends, but they are smaller and insignificant
with linear time trends. In the lower upper panel, for the 2002–2003 through 2008–2009 sample,
the estimates without time trends are qualitatively similar to the results in Harper, Strumpf, and
Kaufman (2012); they are positive but small and insignificant.8 On the other hand, the estimates

8

In Appendix D, all estimates from a level specification are qualitatively similar, and I can
successfully replicate the results from Harper, Strumpf, and Kaufman (2012).
57

Table D2: Effects of Medical Marijuana Laws on Ages 18–25
Law

Law (based on Harper et al. coding)
2002 to 2007–2008, Use in Past Month
0.096*
0.059***
0.023
(0.053)
(0.021)
(0.028)

0.064**
(0.025)

0.028
(0.036)

0.029
(0.038)

2002–2003 to 2008–2009, Use in Past Month
0.055*
0.065
0.027
0.062***
(0.031)
(0.052)
(0.036)
(0.022)

0.088*
(0.048)

0.076
(0.050)

2002 to 2007–2008, Perceptions of Great Risk of Smoking Marijuana Once a Month
-0.008
-0.015
-0.095
0.004
0.014
0.0032
(0.055)
(0.097)
(0.067)
(0.043)
(0.064)
(0.078)
2002–2003 to 2008–2009, Perceptions of Great Risk of Smoking Marijuana Once a Month
-0.026
-0.043
-0.038
-0.012
-0.006
0.024
(0.038)
(0.066)
(0.073)
(0.033)
(0.050)
(0.086)
Time Trends
No
Linear
Quadratic
No
Linear
Quadratic
Note.― All specifications include state and year fixed effects. Robust standard errors are reported
in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

show around a 5.5–7.6% increase with time trends (significant under linear trends). In the lower
two panels, the estimates are uniformly negative based on my coding of medical marijuana laws,
but none of the estimates are significant and the estimated standard errors are very large. However,
the estimates based on the alternate coding from Harper, Strumpf, and Kaufman (2012) tend to
show positive signs.
Table D3 shows the estimates for adults aged 26 and above. For the past-month use rates,
based on the specification with time trends and my coding of laws, the estimates are generally
positive and show roughly a 5% increase, but they are never significant with large estimated
standard errors. Also, these estimates are really sensitive to alternative coding and they become
negative in the two upper right panels. There is some evidence showing a decrease in perceived

58

Table D3: Effects of Medical Marijuana Laws on Ages 26+
Law

Law (based on Harper et al. coding)

0.006
(0.044)

-0.045
(0.051)

2002 to 2007–2008, Use in Past Month
0.055
-0.009
-0.066
(0.115)
(0.046)
(0.046)

-0.027
(0.121)

-0.000
(0.042)

2002–2003 to 2008–2009, Use in Past Month
0.050
0.046
-0.012
0.000
(0.061)
(0.060)
(0.046)
(0.067)

-0.038
(0.097)

2002 to 2007–2008, Perceptions of Great Risk of Smoking Marijuana Once a Month
-0.011
-0.049
-0.120***
0.010
0.023
-0.034
(0.014)
(0.043)
(0.036)
(0.027)
(0.061)
(0.059)
2002–2003 to 2008–2009, Perceptions of Great Risk of Smoking Marijuana Once a Month
-0.020
-0.040**
-0.084**
-0.001
0.003
-0.013
(0.025)
(0.020)
(0.034)
(0.023)
(0.039)
(0.075)
Time Trends
No
Linear
Quadratic
No
Linear
Quadratic
Note.― All specifications include state and year fixed effects. Robust standard errors are reported
in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

risks for ages 26 and above, at least based on my coding. Under the specifications with time trends,
there is a 4.0–12.0% decrease in people aged 26 and above who perceive a great risk in smoking
marijuana once a month. However, this effect disappears when using the alternative coding of laws
from Harper, Strumpf, and Kaufman (2012).
In summary, the estimates suggest an increase of roughly 6% in past-month marijuana use
for young adults aged 18–25, and there is some weak evidence showing that perceived risks
decrease among adults. Generally speaking, these estimates are very noisy, and they are sensitive
to different coding of laws, model specifications, and what years are covered. There are at least a
few reasons. First, these state-level measures are two-year moving averages, so they are designed
to reduce variations across years. Nevertheless, these estimates with state fixed effects are

59

identified through these yearly variations within a state. Second, these two-year moving averages
make the first-year coding of medical marijuana laws arbitrary. Furthermore, these above problems
are amplified due to the fact that the sample only covers a few years in which only a handful of
states changed their laws. Therefore, the fixed-effect estimates based on these two-year moving
averages are unreliable and should be treated with great caution.

60

REFERENCES

61

REFERENCES

Akiyama, Y., Propheter, S. K., 2005. Methods of data quality control: for Uniform Crime
Reporting programs. Federal Bureau of Investigation, Criminal Justice Information Services
Division.
Anderson, D. M., 2010. Does information matter? the effect of the meth project on meth use
among youths. Journal of Health Economics 29 (5), 732-742.
Anderson, D. M., Hansen, B., Rees, D. I., 2012a. Medical marijuana laws and teen marijuana
use. IZA Discussion Paper No. 6592.
Anderson, D. M., Hansen, B., Rees, D. I., 2012b. Medical marijuana laws, traffic fatalities, and
alcohol consumption. Journal of Law and Economics (forthcoming).
Anderson, D. M., Rees, D. I., Sabia, J. J., 2012c. High on life? medical marijuana laws and
suicide. IZA Discussion Paper No. 6280.
Anderson, E., 2012. "Mother earth" medical marijuana clinic prepares to close. KPBS Public
Broadcasting. San Diego, CA.
Becker, G. S., Murphy, K. M., 1988. A theory of rational addiction. The Journal of Political
Economy 96 (4), 675-700.
Bretteville-Jensen, A. L., 2011. Illegal drug use and the economic recession—what can we learn
from the existing research? The International Journal on Drug Policy 22 (5), 353-359.
Caplan, G., 2012. Medical marijuana: a study of unintended consequences. McGeorge Law
Review 43, 127-146.
Carpenter, C., 2004. How do zero tolerance drunk driving laws work? Journal of Health
Economics 23 (1), 61-83.
Carpenter, C., 2007. Heavy alcohol use and crime: evidence from underage drunk‐driving laws.
Journal of Law and Economics 50 (3), 539-557.
Caulkins, J. P., Pacula, R. L., 2006. Marijuana markets: inferences from reports by the household
population. Journal of Drug Issues 36 (1), 173-200.
Cerdá, M., Wall, M., Keyes, K. M., Galea, S., Hasin, D., 2012. Medical marijuana laws in 50
states: investigating the relationship between state legalization of medical marijuana and
marijuana use, abuse and dependence. Drug and Alcohol Dependence 120 (1-3), 22-27.
Chen, K., Kandel, D. B., Davies, M., 1997. Relationships between frequency and quantity of
marijuana use and last year proxy dependence among adolescents and adults in the United
States. Drug and Alcohol Dependence 46 (1-2), 53-67.

62

Cohen, P. J., 2010. Medical marijuana 2010: it's time to fix the regulatory vacuum. The Journal
of Law, Medicine & Ethics 38 (3), 654-666.
Conlin, M., Dickert‐Conlin, S., Pepper, J., 2005. The effect of alcohol prohibition on illicit‐drug‐
related crimes. Journal of Law and Economics 48 (1), 215-234.
Coté, J., Allday, E., Buchanan, W., Lagos, M., 2008. Feds say S.F. has more pot clubs than
Starbucks, but it might not add up. San Francisco Chronicle.
DeSimone, J., Farrelly, M. C., 2003. Price and enforcement effects on cocaine and marijuana
demand. Economic Inquiry 41 (1), 98-115.
Dickinson, T., 2012. Obama's war on pot. Rolling Stone.
DiNardo, J., Lemieux, T., 2001. Alcohol, marijuana, and american youth: the unintended
consequences of government regulation. Journal of Health Economics 20 (6), 991-1010.
Drug Enforcement Administration (DEA). 2011. The DEA position on marijuana 2011
http://www.justice.gov/dea/marijuana_position.pdf
Farrelly, M. C., Bray, J. W., Zarkin, G. A., Wendling, B. W., 2001. The joint demand for
cigarettes and marijuana: evidence from the National Household Surveys on Drug Abuse.
Journal of Health Economics 20 (1), 51-68.
Fellner, J., 2009. Race, drugs, and law enforcement in the United States. Stanford Law & Policy
Review 20 (2), 257-292.
Fergusson, D. M., Boden, J. M., Horwood, L. J., 2006. Cannabis use and other illicit drug use:
testing the cannabis gateway hypothesis. Addiction 101 (4), 556-569.
Fryer, R., Heaton, P., Levitt, S., Murphy, K., 2010. Measuring crack cocaine and its impact.
Economic Inquiry. 51 (3), 1651-1681.
Government Accountability Offfice (GAO), 2002. Marijuana: early experiences with four states'
laws that allow use for medical purposes. Washington, DC.
Golub, A., Johnson, B. D., 2002. The misuse of the ‘gateway theory’ in US policy on drug abuse
control: a secondary analysis of the muddled deduction. International Journal of Drug Policy
13 (1), 5-19.
Golub, A., Johnson, B. D., Dunlap, E., 2007. The race/ethnicity disparity in misdemeanor
marijuana arrests in New York City. Criminology & Public Policy 6 (1), 131-164.
Golub, A., Liberty, H. J., Johnson, B. D., 2005. Inaccuracies in self-reports and urinalysis tests:
impacts on monitoring marijuana use trends among arrestees. Journal of Drug Issues 35 (4),
941-966.
Gorman, D. M., Huber, C. J., 2007. Do medical cannabis laws encourage cannabis use?
International Journal of Drug Policy 18 (3), 160-167.

63

Gruber, A. J., Pope, H. G., Hudson, J. I., Yurgelun-Todd, D., 2003. Attributes of long-term
heavy cannabis users: a case-control study. Psychological Medicine 33 (8), 1415-1422.
Hall, J. C., Schiefelbein, J., 2011. The political economy of medical marijuana laws. Atlantic
Economic Journal 39 (2), 197-198.
Harper, S., Strumpf, E. C., Kaufman, J. S., 2012. Do medical marijuana laws increase marijuana
use? replication study and extension. Annals of Epidemiology 22 (3), 207-212.
Harrison, L. D., Martin, S. S., Enev, T., Harrington, D., 2007. Comparing drug testing and selfreport of drug use among youths and young adults in the general population. Substance
Abuse and Mental Health Services Administration, Office of Applied Studies: Rockville,
MD
Jacobson, M., 2004. Baby booms and drug busts: trends in youth drug use in the United States,
1975-2000. The Quarterly Journal of Economics 119 (4), 1481-1512.
Johnson, K., 2009. States pressed into new role on medical marijuana. The New York Times.
Khatapoush, S., Hallfors, D., 2004. "Sending the wrong message": did medical marijuana
legalization in California change attitudes about and use of marijuana? Journal of Drug
Issues 34 (4), 751-770.
Leger, D. L., 2012. Survey: 1 in 15 high school seniors smoking pot. USA TODAY.
Levitt, S., 1998. Juvenile crime and punishment. The Journal of Political Economy 106 (6),
1156-1185.
Lynch, J. P., Jarvis, J. P., 2008. Missing data and imputation in the Uniform Crime Reports and
the effects on national estimates. Journal of Contemporary Criminal Justice 24 (1), 69-85.
MacCoun, R., Pacula, R. L., Chriqui, J., Harris, K., Reuter, P., 2009. Do citizens know whether
their state has decriminalized marijuana? assessing the perceptual component of deterrence
theory. Review of Law & Economics 5 (1), 347-371.
MacCoun, R. J., 2010. Estimating the non-pricee effect of legalization on cannabis consumption.
Rand Corporation Working paper WR-767-RC.
Mikos, R. A., 2011. A critical appraisal of the Department of Justice's new approach to medical
marijuana. Stanford Law & Policy Review 22 (2), 633-670.
Miller, R. N., Kuhns, J. B., 2011. Exploring the impact of medical marijuana laws on the validity
of self-reported marijuana use among juvenile arrestees over time. Criminal Justice Policy
Review 23 (1), 40-66.
O'Connor, A., 2011. Marijuana use growing among teenagers. The New York Times.
O'Keefe, K., Earleywine, M., 2011. Marijuana use by young people: the impact of state medical
marijuana laws. Marijuana Policy Project: Washington, DC.

64

Pacula, R. L., Chriqui, J. F., Reichmann, D. A., Terry-McElrath, Y. M., 2002. State medical
marijuana laws: understanding the laws and their limitations. Journal of Public Health Policy
23 (4), 413-439.
Pacula, R. L., Kilmery, B., Grossmanz, M., Chaloupka, F. J., 2010. Risks and prices: the role of
user sanctions in marijuana markets. The B.E. Journal of Economic Analysis & Policy 10 (1),
1-36.
Pacula, R. L., MacCoun, R., Reuter, P., Chriqui, J., Beau Kilmer, K. H., Paoli, L., Schäfer, C.,
2005. What does it mean to decriminalize marijuana? a cross-national empirical examination.
In: Grossman, M., Lindgren, B. (eds), Substance abuse: Individual behaviour, social
interactions, markets and politics. Elsevier: Amsterdam.
ProCon.org. 2012. How many people in the US use medical marijuana?
http://medicalmarijuana.procon.org/view.answers.php?questionID=001199 (04.12.2012.)
ProCon.org. 2013a. 11 states with pending legislation to legalize medical marijuana
http://medicalmarijuana.procon.org/view.resource.php?resourceID=002481 (04.12.2012.)
ProCon.org. 2013b. 18 legal medical marijuana states and D.C.
http://medicalmarijuana.procon.org/view.resource.php?resourceID=000881 (04.12.2012.)
Ramchand, R., Pacula, R. L., Iguchi, M. Y., 2006. Racial differences in marijuana-users’ risk of
arrest in the United States. Drug and Alcohol Dependence 84 (3), 264-272.
Salomonsen-Sautel, S., Sakai, J. T., Thurstone, C., Corley, R., Hopfer, C., 2012. Medical
marijuana use among adolescents in substance abuse treatment. Journal of the American
Academy of Child and Adolescent Psychiatry 51 (7), 694-702.
Sekhon, V., 2009. Highly uncertain times: An analysis of the executive branch's decision to not
investigate or prosecute individuals in compliance with state medical marijuana laws.
Hastings Constitutional Law Quarterly 37 (3), 553-564.
Terry-McElrath, Y. M., Chriqui, J. F., Bates, H., McBride, D. C., 2009. Do state policies matter
in prosecutor-reported juvenile marijuana case disposition? Crime & Delinquency
(Forthcoming). http://cad.sagepub.com/content/early/2009/11/30/0011128709348457
Thurstone, C., Lieberman, S. A., Schmiege, S. J., 2011. Medical marijuana diversion and
associated problems in adolescent substance treatment. Drug and Alcohol Dependence 118
(2-3), 489-492.
Wall, M. M., Poh, E., Cerdá, M., Keyes, K. M., Galea, S., Hasin, D. S., 2011. Adolescent
marijuana use from 2002 to 2008: higher in states with medical marijuana laws, cause still
unclear. Annals of epidemiology 21 (9), 714-716.
Wright, D., 2004. State estimates of substance use from the 2002 national survey on drug use and
health. Substance Abuse and Mental Health Services Administration, Office of Applied
Studies: Rockville, MD.

65

CHAPTER 2
Do Medical Marijuana Laws Increase Hard Drug Use?
“I believe that marijuana is a gateway drug.” John McCain, U.S. Senator, meeting at Milton, New
Hampshire, August 11, 2007
“I believe marijuana should be illegal in our country. It is the pathway to drug usage by our
society, which is a great scourge—which is one of the great causes of crime in our cities.” Mitt
Romney, JD, former Governor of Massachusetts, speaking to students at St. Anselm Institute of
Politics in Manchester, New Hampshire, October 4, 2007
1.

Introduction
The notion that marijuana is a gateway drug, or more generally a complement to hard drugs,

is an important but controversial justification for marijuana prohibition and has had a strong
influence on U.S. drug policy. The literature on the gateway hypothesis in past decades is huge,
but the causal link is still not well established. A key difficulty in identifying the causal effect of
marijuana on hard drug use is to find a mechanism that generates arguably exogenous variation in
marijuana consumption. Ethical and legal constraints prevent running controlled experiments on
illegal drugs using human subjects, but even the evidence from animal experiments is not
conclusive (Ellgren et al., 2006; Solinas et al., 2004). Therefore, the original proposer of the
gateway hypothesis, Denise B. Kandel, concludes that, without a clear neurological mechanism,
the evidence is at best mixed (Kandel, 2003).
Medical marijuana legislation represents a major change in U.S. policy towards marijuana
in recent years. As of March 2013, 18 states and the District of Columbia had passed laws that
allow individuals with designated symptoms to use marijuana for medical purposes. Although the
direct effects of these medical laws are limited to legal patients, it is a popular belief that

66

legalization has increased illegal marijuana use among non-patients as well (Leger, 2012;
O'Connor, 2011). Not only might the associated stigma diminish, but people may perceive lower
health and legal risks of smoking marijuana. Some evidence also suggests that the leaking of
medical marijuana from legal patients or dispensaries may be common (Salomonsen-Sautel et al.,
2012; Thurstone et al., 2011). In fact, lobby groups behind these laws such as the National
Organization for the Reform of Marijuana Laws (NORML) consider such legislation to be the first
step towards full legalization, and two medical marijuana states, Colorado and Washington,
successfully passed referenda to legalize marijuana for recreational use in November 2012.
The potential effects of legalization on marijuana and hard drug use are not only policy
relevant, but they can provide evidence on the relationship between marijuana and other substances.
Some empirical evidence suggests that marijuana consumption has increased after medical
marijuana legalization. For example, Anderson et al. (2012b) find that the prices of high-quality
marijuana are decreasing over time. Chu (2012) shows that these medical marijuana laws are
associated with an increase in marijuana possession arrests and treatment referrals. The notion that
marijuana is a gateway drug leads many people to be concerned that the use of hard drugs, such as
cocaine and heroin, will consequently increase. In fact, it is one of the major reasons for federal
agencies like the Drug Enforcement Administration (DEA) and the Office of National Drug
Control Policy (ONDCP) firmly opposing such medical marijuana laws and to continue listing
marijuana as a Schedule I drug (Drug Enforcement Administration, 2011).
To examine whether medical marijuana laws––and the associated increase in marijuana
use––affect cocaine and heroin usage, I study drug possession arrests from the Uniform Crime
Reports (UCR). As the arrest data do not distinguish cocaine and heroin, and they could be
potentially biased by changes in law enforcement, I also study treatment admissions that are not

67

referred by the criminal justice system from the Treatment Episode Data Set (TEDS). Although
these datasets do not follow individuals over time, these data have many advantages. First, by
focusing on drug arrestees and treatment patients rather than the general population, these data
provide much more observation of hard drug users. Based on the National Survey on Drug Use
and Health (NSDUH), the past month’s prevalence rates in the U.S. are only around 1–2% for
cocaine and 0.2% for heroin. The low prevalence rates suggest that the sample sizes at the state
level in most representative datasets are probably not large enough to have statistical power. For
instance, even for the largest survey, the NSDUH, the sample sizes in most states are only 900
people (600 for adults). In contrast, the UCR arrest data are available even at the city level, and
the TEDS data contain 1.5 to 2 million substance admissions each year in which cocaine and heroin
account for 40% of the admissions. In addition, these data are objective measures and they do not
suffer from the self-reporting bias that is common in survey data (Golub et al., 2005; Harrison and
Hughes, 1997). This is a particular concern in the current context. These medical marijuana laws
are expected to change the perception of marijuana. Indeed, Miller and Kuhns (2011) find that
arrestees report marijuana usage more honestly after the passage of medical marijuana laws. If
these laws also reduced the stigma on other illicit drugs, there could be a spurious relationship
between marijuana and cocaine or heroin due to people changing their reporting behaviors.
I estimate reduced-form models for the effects of these laws, controlling for city/state and
year fixed effects as well as city/state-specific time trends. Assuming that unobservables related
to arrests or treatments do not deviate from a city or a state’s trend when states enact medical
marijuana laws, this approach will uncover the causal effects of these laws. To preview the results,
I find strong evidence supporting the popular belief that marijuana use has increased after the
passage of medical marijuana laws. However, in contrast to what the gateway theory predicts, I do

68

not find any evidence that cocaine and heroin arrests or treatments have increased. In fact, all of
the estimates show negative signs, suggesting that medical marijuana laws could have a negative
effect on hard drug use. Specifically, the results from the UCR arrest data indicate that, on average,
medical marijuana laws are associated with about a 15–25% increase in marijuana possession
arrests among adult males. The estimates on possession arrests for cocaine and heroin combined
are uniformly negative, while the magnitudes fluctuate from close to zero to a 10–25% decrease,
depending on model specifications. From the TEDS treatment admission data, I also find a similar
increase in marijuana treatment admissions of roughly 10%. On the other hand, I find a 15–20%
decrease in heroin treatment admissions but no significant effect on cocaine treatment admissions.
This research is important for several reasons. First, this paper employs a new policy tool—
medical marijuana laws—for detecting the effects of marijuana on hard drug use. Most of the
previous studies either use instrumental variables, such as marijuana penalty and state excise taxes
on beer, or try to model individual heterogeneity econometrically. All of these approaches have
clear limitations in the context of drug consumption. Second, the causal effects of medical
marijuana laws on marijuana and hard drug usage are at the core of the current policy debate. In
addition, as treatment patients are heavy users who are associated with negative health and social
outcomes, understanding the causal effects among this population is particularly relevant to the
design of policy. Finally, the results indicate some direct costs incurred by these medical marijuana
laws, such as the increase in marijuana treatments, while they also suggest that some unintended
positive externalities may exist. Future cost and benefit analysis may utilize these findings to
obtain more precise estimates for the impacts of these medical marijuana laws.

69

The paper proceeds as follows: Section 2 briefly describes the medical marijuana laws and
Section 3 reviews the relevant literature. I discuss the data and results from the UCR arrests in
Section 4, and those from the TEDS treatment admissions in Section 5. Section 6 concludes.

2.

Medical Marijuana Laws
Medical marijuana laws permit patients with legally designated diseases and syndromes to

use marijuana as a treatment. The designated symptoms are often as follows: AIDS, anorexia,
arthritis, cachexia, cancer, chronic pain, glaucoma, migraines, persistent muscle spasms, severe
nausea, seizures, and sclerosis. Some laws, such as the one in California, however, allow use for
“any other illness for which marijuana provides relief” (Cohen, 2010). Patients can legally possess
marijuana up to a fixed amount. In many states, they can cultivate marijuana on their own. These
laws also allow “caregivers” (most of whom are patients as well) to grow and provide marijuana
to patients on a not-for-profit basis. In most states, it is mandatory to register and renew the
registration every year to be a qualified medical marijuana patient or caregiver.1
In principle, these medical marijuana laws only provide legal protection for patients and
caregivers, and do not change the legal status of non-medical use of marijuana. However, there is
a huge grey area and the legal boundary is blurred by these loosely worded laws (Cohen, 2010). It
is probably done intentionally by the legalization lobbyists behind these laws, such as the NORML,
who consider such legislation the first step towards full legalization. A significant example of the

1

California created a registration program in 2004 but registration is voluntary. Colorado allows
patients who do not join the registry to use the "affirmative defense of medical necessity" if they
are arrested on marijuana charges. Maine passed an amendment in November 2009 that created a
registration program and required mandatory registration starting January 1, 2011. Washington
does not have a registration program.
70

inherent grey legal area in these laws is marijuana dispensary. As most state medical marijuana
laws do not directly authorize marijuana dispensaries, they only exist under the name of caregiver
or patient cooperative, but how prevalent they are largely depends on the attitude of the local
government. For example, San Diego County has very a restrictive policy towards dispensaries
and its law enforcement organizations actively cooperate with the DEA; even the only countylicensed dispensary was forced to close in 2012 (Anderson, 2012). In contrast, there are more
marijuana dispensaries than Starbucks coffee shops or CVS pharmacies in Los Angeles and San
Francisco (Coté et al., 2008). Yet local attitude and law enforcement could change from time to
time. For instance, Los Angeles ordered the closure of over 70% of the 638 dispensaries then
operating in the city in June 2010.
Some ambiguities for low-level marijuana possession offense also exist. For example,
California only requires patients to possess a "written or oral recommendation" from their
physician, thus not requiring the recommendation to be documented. In general, there has been a
softening in public attitude toward marijuana in medical marijuana states, and the federal agencies
complain that the cooperative relations between federal and local law enforcement are
deteriorating (GAO, 2002). For instance, cities like Denver or Seattle passed referenda to either
legalize marijuana or make marijuana possession offenses the lowest enforcement priority. On the
other hand, direct impacts on enforcement towards low-level possession offenses appear to be
small due to a small number of legal patients. Except for California, the number of legal patients
and marijuana dispensaries remained relatively small prior to 2009 (ProCon.org, 2012). 2 To

2

Although there is no official number of patients for states without registration, based on the large
number of dispensaries, it is believed that California has many more patients than other medical
marijuana states.
71

alleviate the tension between the federal and state governments, in 2009, the Obama administration
stated that the federal agencies will no longer seek to arrest medical marijuana users and suppliers
so long as they conform to state laws. Since then, the number of registered patients and dispensaries
increased significantly (Caplan, 2012; Mikos, 2011; Sekhon, 2009). Although this statement
appeared to largely resolve the legal dispute between state and federal governments, the Obama
administration’s medical marijuana policy began to reverse in 2011, and there have been several
cases of DEA raids on medical marijuana dispensaries that conform to state laws (Dickinson, 2012).

3.

Literature Review
An extensive literature on the relationship between marijuana and hard drugs has yielded

many hypotheses but little consensus. I do not attempt to provide a comprehensive survey of the
literature but rather to focus on some more recent studies that adopt different methodologies.
A straightforward question for economists is whether marijuana is a substitute or a
complement to other hard drugs. For example, Saffer and Chaloupka (1999) and Grossman and
Chaloupka (1998) estimate demand functions for marijuana and cocaine, and they find that cocaine
price is negatively correlated with marijuana use and the status of marijuana decriminalization is
positively associated with cocaine use, suggesting that they are complements. On the other hand,
recent studies based on laboratory control experiments show a more complex pattern, even though
external validity could be a concern due to the small sample sizes. The relationships between drugs
seem to depend on different types of drug users. Jofre-Bonet and Petry (2008) find that marijuana
is a complement to heroin for heroin addicts but a substitute for heroin for cocaine addicts. Petry
(2001) finds that marijuana consumption is independent from cocaine for alcoholics, while Petry
and Bickel (1998) find marijuana is a substitute for heroin for opioid-dependent patients.

72

Based on the classical framework from Becker and Murphy (1988), the gateway hypothesis
is a special case of intertemporal complementarity between marijuana and hard drugs. Popularized
by Kandel’s (1975) influential paper published in Science, the gateway hypothesis is based on one
of the most robust empirical observations: most hard drug users have started with less dangerous
drugs and there seems to be a “staircase” stepping from marijuana (or legal substances like alcohol)
to cocaine and heroin. A gateway effect might be indeed causal through physiological or
psychological demand for stronger drug-induced pleasures and experiences, which Becker and
Murphy (1988) call consumption capital (of addictive drugs). In addition, a gateway effect could
come from social interactions like gaining access to hard drugs due to participation in the illegal
drug market (MacCoun, 1998).3
The infeasibility of running controlled experiments makes it extremely difficult to
empirically establish the causality due to unobserved heterogeneity. DeSimone (1998) uses
marijuana penalties, beer taxes, and the presence of alcoholic parents as instrumental variables,
and finds strong evidence for marijuana being a gateway drug for cocaine. Fergusson et al. (2006a)
also find strong evidence using longitudinal data and controlling for individual fixed effects (see
also comments from Kandel et al. (2006), Maccoun (2006) and Fergusson et al. (2006b). Due to
the difficulty of finding a valid instrument, some studies try to econometrically model unobserved
heterogeneity. These studies generally find that unobserved heterogeneity is an important factor,
but whether marijuana is a gateway drug remains unclear. For example, Pudney (2003) does not
find a gateway effect after accounting for unobserved heterogeneity, while some other studies find

3

Note that if a gateway effect is working through social interactions, then legalization of soft drugs
and separating their markets from hard drugs would be a better policy. This is actually the rationale
behind the policy in the Netherlands that allows legal sale of marijuana in “coffee-shops.”
73

marijuana is a gateway drug for cocaine (Bretteville-Jensen et al., 2008; Melberg et al., 2010; van
Ours, 2003). Another strand of studies from epidemiology utilizes data on twins and finds a
positive relationship between early marijuana use and the use of other illicit drugs (Agrawal et al.,
2004; Lynskey et al., 2006; Lynskey et al., 2003). However, as in the critiques made by Bound
and Solon (1999), one potential problem in these twin studies is that the causes for observably
identical twins to make different choices are unlikely to be exogenous. Since even evidence from
animal experiments is not conclusive (Ellgren et al., 2006; Solinas et al., 2004), Kandel (2003)
concludes that the existing evidence for the gateway effect is at best mixed due to the lack of a
clear neurological mechanism.

4. Analysis of Arrest Data
4.1. UCR Data
The data on drug possession arrests are from the FBI's Uniform Crime Reports (UCR) for the
years 1992 through 2008. Although variation in drug arrests is subject to law enforcement, arrest
data remain the single most widely available indicator of illegal drug use within and across
jurisdictions in the United States. Studies from criminology show that drug arrests are generally
valid measures for illicit drug activities, especially for cocaine and heroin (Moffatt et al., 2012;
Rosenfeld and Decker, 1999; Warner and Coomer, 2003). The UCR arrest data provide monthly
information on arrest counts by age, gender, and race in each crime category along with agency
populations (estimated from the Census) for state and local police agencies. Note that each arrest
count does not necessarily represent a single individual since a person may be arrested multiple

74

times. So, conceptually, the measure reflects changes in both the intensive and extensive margins.4
There are four categories in drug possession arrests, including one category for marijuana and one
for powder cocaine, crack cocaine, heroin, and other opium derivatives together.5 In this section,
I denote the latter category by "cocaine" arrest for brevity as cocaine is the most popular drug
among these drugs. As the crack epidemic ended around the early-to-mid 1990s (Drug
Enforcement Administration, 1991; Fryer et al., 2010), and to be consistent with the starting point
in the TEDS data, I use data on possession arrests from the years 1992 through 2008.6
I use yearly aggregated arrest data provided by the Inter-university Consortium for Political
and Social Research (ICPSR), as the FBI also reviews and checks the data using annual arrest
totals (Akiyama and Propheter, 2005). Since participation in the UCR program is generally
voluntary, many agencies do not report in every month or every year; even when an agency reports,
it may not report data in all categories. One problem is that it is not possible to distinguish a true
zero from missing data. Empirically, however, most missing data is from agencies with small
populations and those that do not report for a whole year (Lynch and Jarvis, 2008). In this paper,

4

We can model arrests as follows: A = ∑N Fj * Pj, where Fj is individual j’s transaction or use
j=1

frequency, N is the number of users, and Pj is the probability of being arrested per transaction or
̅
per use for an individual j. Suppose Pj = P and take logarithm, then log(A) = log(P) + log(𝐹 ) +
log(N), where ̅ is the average of Fj. Therefore, arrests reflect effects on both the extensive and
𝐹
intensive margins.
5
The other two subcategories are "truly addicting synthetic narcotics" and "other dangerous nonnarcotic drugs" (most of this category are methamphetamines).
6
Although data through 2010 became available recently, looking at the period prior to 2009 has
an advantage in that the number of legal patients was relatively small, and the federal policy was
fairly uniform prior to the Obama administration. In addition, severe economic recession may
affect drug use, and theoretically the direction is ambiguous (Bretteville-Jensen, 2011). As of
December 2012, most states that passed laws after 2008 have not yet accepted patient applications.
(Only Arizona began to accept patient applications since April 2011).
75

I focus on police agencies located in cities of more than 50,000 residents because the FBI regularly
checks and communicates with these agencies to ensure data quality (Akiyama and Propheter,
2005). Since population is generally increasing over time, I include earlier observations of the
above cities to make the panel more balanced if their populations are no less than 25,000. Similar
to Carpenter (2007), and as is common in the criminology literature, I focus on adult male arrests,
and use city-years only if a city reports arrests for marijuana or cocaine possession for at least six
months in that year.7 (I include city-year observations that report only in December since some
agencies may report annually.) The sample covers eleven medical marijuana states that passed
laws before July 2008, including Alaska, California, Colorado, Hawaii, Maine, Montana, Nevada,
New Mexico, Oregon, Rhode Island, and Washington; Vermont is not in the sample because no
city from Vermont in the UCR has a population greater than 50,000. (Michigan passed its law in
November 2008 and is coded as a non-medical marijuana state).
Similar to Carpenter (2007) and Fryer et al. (2010), I create arrest ratios of marijuana or
cocaine possession arrests to all offense arrests among adult males. Although the arrest rate is
straightforward and commonly used, the measure of arrest ratios can partially account for
unobserved changes in local law enforcement and measurement errors in arrest rates from
estimated populations. In addition, as the resources of law enforcement are mostly limited, the
arrest ratios can capture fluctuations in arrests due to changes in total resources available.

7

I only consider males both to be consistent with the existing literature and because males are
much more likely to be in the criminal justice system than are females. For example, the possession
arrest rates for adult males in my sample are four to seven times those for adult females. Cocaine
and heroin use among juveniles is relatively low. In addition, the juvenile justice system is very
different from the adult system in areas such as its procedures, incentives, and sanctions (Carpenter,
2007; Levitt, 1998; Terry-McElrath et al., 2009).
76

Table 2.1 lists the means and standard deviations of the arrest ratios of marijuana and
cocaine possession among adult males (age 18 and above). The upper panel is for marijuana and
the lower panel is for cocaine. In both panels, the first rows are for all states, the second rows are
for states without effective medical marijuana laws before July 2008, and the third rows are for
states with effective medical marijuana laws before July 2008, excluding California and Colorado.
In the upper panel, the marijuana arrest ratio is lower in medical marijuana states than in nonmedical marijuana states. Because marijuana prevalence rates are higher in medical marijuana
states from survey data such as the NSDUH, the lower arrest ratio suggests that the level of law
enforcement towards marijuana is probably lower in these states. Medical marijuana states
excluding California and Colorado also show lower cocaine arrests. However, even though cocaine
usage rates in medical marijuana states tend to be higher from the NSDUH, it is not clear to what
extent the lower arrest ratios are attributed to law enforcement. As the past-month cocaine use rate
is only around 2%, the state-level estimates from the NSDUH may not be precise enough for
comparisons. Note that marijuana arrest ratios are only roughly 1.5 times higher than cocaine
arrests, while the prevalence rates of marijuana are five to seven times higher than those of cocaine.
It suggests those marijuana arrestees may be quite different from general marijuana users in the
survey data. As the marijuana arrests are highly correlated with marijuana treatments, with
correlation coefficients around 0.3–0.5, many of the marijuana arrestees are probably also heavy
users who make regular transactions and therefore are subjected to higher arrest risks.
The descriptive statistics of California and Colorado are separated in the last two rows. As
most dispensaries were located in California and Colorado prior to 2009, and the penalty in these
two states for low-level possession was the lowest in the U.S. with only $100 maximum fine
(Pacula et al., 2010), the legalization effects and potential reactions of law enforcement could be

77

Table 2.1: UCR Descriptive Statistics (1992−2008)
Marijuana Possession Arrest Ratios (%) for Adult Males
Mean
SD.
City-year Obs.
All States
3.60
2.44
10,032 (746 cities)
Non-MJ States
4.17
2.37
6,559 (510 cities)
MJ States w/o CA&CO
3.07
1.95
580 (47 cities)
California
2.38
2.29
2,710 (174 cities)
Colorado
3.15
1.40
183 (15 cities)
Cocaine and Heroin Possession Arrest Ratios (%) for Adult Males
All States

Mean
3.12

SD.
2.47

City-year Obs.
8,811 (721 cities)

Non-MJ States

2.73

2.32

5,386 (489 cities)

MJ States w/o CA&CO

1.80

1.78

488 (43 cities)

California

4.22

2.50

2,791 (174 cities)

Colorado
1.26
1.64
146 (15 cities)
Note.― MJ states include only states that passed laws before July 2008;
states that passed laws afterward are in non-MJ states.

different from other medical marijuana states. In fact, California has the highest drug possession
arrest ratios among all states, but its marijuana arrest ratios are some of the lowest, suggesting that
enforcement in California is probably very lenient towards marijuana but focuses on other hard
drugs. The situation might be similar in Colorado. In fact, Colorado attorney general, John W.
Suthers, even said about the medical marijuana law in Colorado, “In Colorado it’s not clear what
state law is” (Johnson, 2009). Since California has many more observations than any other states,
and Colorado has the second largest observations among the medical marijuana states, I will study
these two states separately to see if there are any heterogeneous effects.
Do marijuana and cocaine possession arrests represent underlying drug use? Graphical
evidence at the national level suggests that they do. Figure 2.1 plots the yearly averages of the ratio
of marijuana and cocaine possession arrests to all offense arrests along with marijuana or cocaine

78

-2

-1

0

1

2

Marijuana Arrests and Prices

1992

1996

2000
Year

Marijuana Arrest Ratio

2004

2008

Marijuana Price

-2

-1

0

1

2

Cocaine Arrests and Prices

1992

1996

2000
Year

Cocaine Arrest Ratio

2004

2008

Cocaine Prices

Figure 2.1: Arrests and Prices of Marijuana and Cocaine 1992–2008

prices per gram. The marijuana prices are from the 2012 National Drug Control Strategy Data
Supplement, and the cocaine prices are the median prices in each year from the DEA's System to

79

Retrieve Information from Drug Evidence (STRIDE).8 (All of these series are normalized to mean
zero and standard deviation one.) In both figures, the prices move in the opposite direction to the
arrests, which is consistent with a supply curve moving along a downward sloping demand curve.

4.2. Results
My primary empirical strategy involves estimating city- and year-specific drug possession
arrests as a function of whether the state has an effective medical marijuana law in place in that
year. I begin by estimating the following model by OLS:
(5)

Yist = β Lawst + City fixed effectsi + Year fixed effectst + City time trendsit + εist ,

where Yist is the logarithm of marijuana or cocaine arrest ratios among adult males for city i in
state s and year t. Lawst is a dummy variable indicating whether a state s had a medical marijuana
law during year t.9 In addition to city and year fixed effects, I include city-specific time trends to
capture the time-varying unobservables like law enforcement within a city. As there are many
small values, especially for cocaine arrests, I also estimate the same specification by a fixed effect
Poisson model to check the robustness of functional form. Because city-specific time trends and
fixed effects already account for any smooth-trending variables, and there are missing data in some

8

The cocaine prices are calculated by the author. The average cocaine prices that exclude some
extreme values are similar to median prices. See Horowitz (2001) and Arkes et al. (2008) for
discussions on the STRIDE data.
9
There is normally a time lag between passing a referendum and it becoming an effective law.
There were examples that the referendum was delayed (ex: Nevada) or even vetoed (ex: Arizona
in 1996) by the state government. Throughout this paper, the coding of Lawst is based on the
effective date. For the first year, Lawst equals 1 if the law is effective before July 1st, and equals
0 otherwise.
80

control variables, I do not include control variables in the main specification. Throughout this
paper, the estimated standard errors are clustered at the state level and therefore are robust to serial
correlation, within-state spatial correlation, and heteroskedasticity.
Table 2.2 shows the estimates for medical marijuana laws on marijuana arrests among adult
males. The first two columns, Columns (1) and (2), show the estimates of β based on Equation (1).
The estimates are positive and highly significant. If we interpret the log points as a percentage
change, medical marijuana laws, on average, result in a 14.2–14.5% increase in the ratio of
marijuana to all arrests among adult males. As California and Colorado account for more than 80%
of observations in medical marijuana states, it is a concern that the positive estimates in Columns
(1) and (2) are entirely driven by the effects of medical marijuana laws in these two states. So I
separately estimate the effects of California and Colorado laws by including CA Law and CO Law,
two interaction terms of Lawst and dummies for California and Colorado. These results are
presented in Columns (3) and (4). The estimates for CA Law and CO Law are actually negative,
and the estimates of Lawst are significantly larger than estimates in Columns (1) and (2).
Specifically, based on the specification with quadratic city time trends, conditional on California
and Colorado, medical marijuana laws, on average, result in a 26.3% increase in the ratio of
marijuana arrests to all arrests, among adult males. To check robustness, Columns (5) and (6) show
the estimates of marijuana and cocaine arrests when a set of city- and state-level control variables
are included. 10 Since these results are nearly identical, it suggests that fixed effects and cityspecific time trends have accounted for most of the variations from these controls. In the last

10

The city-level control is the logarithm of city police officer rates per city residents (from the
UCR). The state-level controls include black male rates, unemployment rates, per capita local and
state expenditures on police protection (in logarithm), per capita local and state expenditures on
81

Table 2.2: Effects of Medical Marijuana Laws on Marijuana Possession Arrests
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Marijuana Arrest Ratios among Adult Males
Law

0.145*** 0.142*** 0.273** 0.263*** 0.167*** 0.271*** 0.121*** 0.203***
(0.037) (0.037) (0.106) (0.068) (0.035) (0.064) (0.031) (0.056)

CA Law

-0.138
(0.106)

-0.132*
(0.076)

CO Law

-0.421*** -0.374***
(0.109) (0.069)

-0.102
(0.074)

-0.0825
(0.061)

-0.519***
(0.095)

-0.293***
(0.060)

Model
Log Linear
Log Linear
Log Linear
FE Poisson
Obs.
10,032
10,032
8,420
10,032
# of States
50
50
50
50
Controls
No
No
No
No
Yes
Yes
No
No
Time
Linear Quadratic Linear Quadratic Quadratic Quadratic Quadratic Quadratic
trends
Note.― All specifications include city and year fixed effects. Robust standard errors are reported
in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

two columns, (7) and (8), to check whether the results are sensitive to functional forms, I estimate
the effects of medical marijuana laws by a fixed-effect Poisson model. The results are similar and
they are not sensitive to different functional forms.
In Table 2.3, I estimate the legalization effects on cocaine arrests; all columns are based on
the same specifications as those in Table 2.2. Somewhat surprisingly, all of the estimates of β are
negative for cocaine arrests in Table 2.3. It does not support the gateway hypothesis or the notion
that marijuana and cocaine or heroin are complements. In Columns (1) and (2), based on the full
sample that includes California and Colorado, the estimated effects of laws on cocaine arrests are
negative, small, and insignificant, suggesting no effect of these medical marijuana laws on cocaine

health and hospital expenditures (in logarithm), and state 0.08 blood alcohol content laws. The
sample sizes are smaller because 2001 and 2003 data on government expenditures were not
developed by the Census Bureau due to sample redesign. (Appendix Table E shows the descriptive
statistics.)
82

Table 2.3: Effects of Medical Marijuana Laws on Cocaine and Heroin Possession Arrests
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Cocaine and Heroin Arrest Ratio
-0.135*
(0.068)

-0.123
(0.093)

CA Law

0.088
(0.087)

0.063
(0.122)

CO Law

0.463*** 0.462***
(0.071) (0.090)

Law

-0.043
(0.044)

-0.045
(0.054)

-0.030
(0.051)

-0.111
(0.095)

-0.088** -0.222***
(0.035) (0.080)

0.092
(0.126)

0.149*
(0.090)

0.281**
(0.130)

0.301***
(0.080)

Model
Obs.

Log Linear
8,811

Log Linear
8,811

Log Linear
7,316

FE Poisson
8,811

# of States
Controls

50

50

50

50

Time
trends

No

No

No

Linear

Quadratic

Linear

No

Yes

Yes

No

No

Quadratic Quadratic Quadratic Quadratic Quadratic

Note.― All specifications include city and year fixed effects. Robust standard errors are reported
in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

arrests. In Columns (3) and (4), the estimated legalization effects are more negative when
California and Colorado are excluded, but the estimated standard errors are also larger. If we take
the face value of these estimates, conditional on these two states, medical marijuana laws are
associated with a 12.3−13.5% decrease in the ratios of cocaine arrests to all arrests. The next two
columns, (5) and (6), shows quantitatively similar estimates when the city and state level control
variables are included. In the last two columns, (7) and (8), the estimates from fixed-effect Poisson
models are still negative while the magnitudes are even larger.
From Table 2.2, the estimates indicate that the legalization effects on marijuana arrests in
California are positive but smaller than in other medical marijuana states. For example, based on
Column (4), the estimates indicate a 13.1% (0.263–0.132) increase in the marijuana arrest ratio.
However, as the estimates for CA Law are generally not significant, the legalization effects in

83

California on marijuana are not statistically different from other medical marijuana states. The
legalization effects on cocaine arrests show a similar pattern. The estimates for California are
smaller (in absolute terms) but insignificantly different from other medical marijuana states. From
Table 2.3, the estimates indicate a 6.0–4.7% decrease in cocaine arrests after California passed its
medical marijuana law. In Figure 2.2, the upper graph shows the yearly averages of marijuana and
cocaine ratios in California. Marijuana arrests were steadily increasing over time until 2000 and
then slightly decreased until 2004. Although there was no significant change in marijuana arrests
around the time (1996) in which California passed its medical marijuana law, marijuana arrests
did show a significant increase beginning in 2005, which is consistent with the time that many
dispensaries began to exist in California. On the other hand, consistent with the small estimated
effects, cocaine arrests are roughly stable overtime, although they also show a slightly decrease
after 2005.
In contrast, in Colorado, the legalization effects are negative on marijuana arrests and
positive on cocaine arrests, and they are significantly different from other medical marijuana states.
For example, based on Column (4) in Table 2.2 and Table 2.3, the estimates indicate an 11.1%
decrease in marijuana arrest ratios and a 33.9% increase in cocaine arrest ratios. Although I cannot
rule out that marijuana use decreased and cocaine use increased in Colorado, the negative (positive)
estimates in marijuana (cocaine) arrests could be driven by some unobservables such as changes
in law enforcement. To see the change in marijuana and cocaine arrests overtime, in Figure 2.2, I
plot the yearly averages of marijuana and cocaine arrest ratios in Colorado. We can see that the
ratios of marijuana arrests to all arrests temporarily dropped around 2001, when Colorado enacted
its medical marijuana laws. Nevertheless, marijuana arrests indeed increased again after 2002.
Interestingly, cocaine arrest ratios in Colorado do move opposite to marijuana arrest ratios.

84

1

2

3

4

Colorado

1992

1996

2000
Year
Marijuana

2004

2008

Cocaine

1

2

3

4

5

California

1992

1996

2000
Year
Marijuana

2004

2008

Cocaine

Figure 2.2: Arrest Ratios of Marijuana and Cocaine in California and Colorado

Figure 2.3 presents graphical evidence on the effects of medical marijuana laws on cocaine
arrests. The upper graph is based on the restricted sample without California and Colorado. The

85

graph shows the averages of cocaine arrest ratios before and after the medical marijuana laws
became effective, where the X-axis measures the year relative to the state’s law change, with 0
denoting the first year of passing the law, 1 denoting the following year, and so on. To create a
synthetic control group, I compute the average arrest ratios in non-medical marijuana states for
each year, and then take a weighted average of these yearly averages, in which the weights come
from the relative composition of each year in the treatment group (medical marijuana states). For
example, for “Year 0” in the upper figure, 55% of observations in the treatment group are from
Alaska, Oregon, and Washington, which passed the laws in 1998 (coded as 1999), so the weight
put on the average of year 1999 in the control group is 0.55. In other words, in “Year 0,” 55% of
the observations in the control group are selected from year 1999. In the upper graph, the arrest
ratios in both the control and treatment groups are fairly flat before legalization; after the passage
of medical marijuana laws, the arrest ratios seem to be decreasing over time in medical marijuana
states (without California and Colorado), while they are roughly stable in other states. For
completeness, in the lower graph of Table 2.2, I also create the same graph for the full sample
including California and Colorado. However, as the treatment group is mostly driven by California,
the graph looks almost identical to Figure 2.2 and it does not show any significant effect on cocaine
arrests.
To further investigate the dynamic responses of cocaine arrests to the adoption of medical
marijuana laws, in Table 2.4, I replace Lawst by a set of dummy variables, Years 0–1 through
Years 6–7, which indicate each two-year interval after the medical marijuana laws were enacted,
and a dummy, Years 8+, for the eighth year and above. To capture potential endogeneity, I also
include a dummy, Years (neg. 1–2), which indicates a two-year interval before passing laws. If
drug users need some time to progress from the gateway drug, marijuana, to cocaine or heroin,

86

1.0

1.5

2.0

2.5

3.0

Cocaine Arrest Ratios (w/o CA. & CO.)

-5

-4

-3

-2

-1
0
1
2
Year Relative to Pass Law

MJ States (w/o CA. & CO.)

3

4

5

Non-MJ States

2.5

3.0

3.5

4.0

4.5

Cocaine Arrest Ratios

-5

-4

-3

-2

-1
0
1
2
Year Relative to Pass Law

MJ States

3

4

5

Non-MJ States

Figure 2.3: Cocaine Arrest Ratios Before and After the Passage of Laws

estimating the dynamics could allow me to detect such lagged positive effects. In the first two
columns, the estimates are from the full sample, and they indicate that these laws have negative

87

Table 2.4: Dynamic Responses of Cocaine and Heroin Arrests to
Legalization
Years
(neg. 1–2)

-0.016
(0.074)

-0.099
(0.068)

Years
0–1

-0.076
(0.049)

-0.096
(0.123)

-0.086
(0.073)

-0.192
(0.129)

Years
2–3

-0.067
(0.053)

-0.092
(0.145)

-0.098
(0.100)

-0.233
(0.146)

Years
4–5

-0.146**
(0.071)

-0.176
(0.172)

-0.256
(0.170)

-0.421**
(0.160)

Years
6–7

-0.171**
(0.079)

-0.203
(0.185)

-0.120
(0.283)

-0.314
(0.232)

Years
8+

-0.222**
(0.096)

-0.255
(0.197)

0.174
(0.408)

-0.052
(0.333)

Obs.
8,811
5,874
Note.― All specifications include city and year fixed effects and city quadratic
time trends. Robust standard errors are reported in parentheses, and they are
clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

effects on cocaine arrests that are generally decreasing over time; the estimate for Years (neg. 1–
2) is small and insignificant, suggesting policy endogeneity is not a particular concern. In the latter
two columns, the estimates become quite noisy when California and Colorado are excluded, but
they are qualitatively similar to the estimates in the first two columns. Overall, the estimates in
Table 2.4 are consistent with Figure 2.3, and there is no evidence supporting a gateway hypothesis
that marijuana use will increase future hard drug use.
In Table 2.5, I check the robustness of the main results from Tables 2.2 and 2.3 based on
different constructions of samples. The upper panel is for marijuana arrests and the lower panel is
for cocaine arrests. For comparison across different samples, in Table 2.5, I estimate a fixed-effect
Poisson model with quadratic time trends for both the samples with or without California and
Colorado. In the first two columns, (1) and (2), I use all city-year observations regardless of the
88

(1)

Table 2.5: Robustness Checks
(2)
(3)
(4)

Cities reporting any
number of months

(5)

(6)

Cities reporting any
drug arrests to the UCR

State-level averages
of arrests

0.125*** 0.217***
(0.030)
(0.053)

0.140***
(0.036)

0.241***
(0.075)

0.164**
(0.066)

0.210***
(0.073)

10,366
Yes

10,503
Yes

7,485
No

764
Yes

730
No

-0.063*
(0.036)

-0.103
(0.072)

-0.107
(0.088)

-0.153
(0.124)

Marijuana All Arrest Ratios
Law
Obs.
CA. & CO.

7,363
No

Cocaine and Heroin All Arrest Ratios
-0.041
(0.063)

-0.229***
(0.075)

Obs.
9,968
6,964
10,503
7,485
722
688
CA. & CO. Yes
No
Yes
No
Yes
No
Note.― Columns (1) and (2) include cities that report marijuana or cocaine possession
arrests in any number of months. Columns (3) and (4) include cities that report any drug
possession arrests to the UCR. Columns (5) and (6) use state level averages of arrest
ratios as dependent variables. All columns are estimated by a fixed-effect Poisson model.
All specifications include city (state) and year fixed effects and city (state) quadratic time
trends. Robust standard errors are reported in parentheses, and they are clustered at the
state level. *** p<0.01, ** p<0.05, * p<0.1.

number of reporting months. The results are similar to the estimates from the last two columns in
Table 2.2 and Table 2.3 but slightly noisier. One major drawback of the UCR arrest is the inability
to distinguish true zeros from missing data. In all previous tables, I treat any city-year observations
with zero adult male arrests as missing data. To test the robustness, in Columns (3) and (4), I treat
marijuana or cocaine possession arrests from cities that report any drug possession arrests but zero
marijuana or cocaine possession arrests as true zeros. There are many zeros, especially for cocaine
arrests. (I use all city-years regardless of the number of reporting months.) The estimates are not
sensitive to how to treat missing data and they are quantitatively similar to those in Columns (1)
and (2) or those in Tables 2.2 and 2.3. As previously mentioned, since Lawst varies only at the

89

state level while each observation is a city-year, states with small populations receive little weight.
As shown in Tables 2.2 and 2.3, the estimates are generally different when California and Colorado
are excluded. To ensure that the results are not driven solely by larger states, in the last two
columns, (5) and (6), I average marijuana or cocaine arrest ratios to the state level, so each state
receives equal weight regardless of the number of city-years (based on the samples from Tables
2.2 and 2.3). For the full sample, as expected, the estimates based on the state averages are larger
(in absolute terms). However, for the restricted sample, the estimates from state-level averages are
quantitatively similar to those in Tables 2.2 and 2.3, suggesting that the estimated effects are
relatively homogenous when California and Colorado are excluded. In summary, I find a quite
robust estimated effect of 10–20% increase in marijuana arrests. Although the estimate magnitudes
for cocaine arrests cover a wide range, I do not find any evidence supporting the notion that
marijuana is a gateway drug or a complement to cocaine/heroin.
One major disadvantage of the arrest data is that the estimates could be driven by
unobserved changes in law enforcement. To address this concern, although indirectly, I examine
the effects of medical marijuana laws separately for blacks and whites. If there is a considerable
racial difference in the estimated effects, this would be a “smoking gun” that the negative estimates
for cocaine arrests are caused by actions of law enforcement. It is well documented that African
Americans are much more likely to be arrested for drug possession. Even though hard drug usage
rates, especially for crack cocaine, tend to be higher among African Americans, a nontrivial
proportion of the racial difference in arrest risk can be attributed to law enforcement (Beckett et
al., 2006; Dannerbeck et al., 2006; Donohue III and Steven D. Levitt, 2001; Gross and Barnes,
2002; Hernández-Murillo and Knowles, 2004; Parker and Maggard, 2005). In addition to potential
racial profiling, for example, African Americans often engage in risky purchasing behaviors such

90

as making transactions in open places, or they tend to live in disadvantaged neighborhoods that
attract more police attention and therefore increase their likelihood of arrest (Beckett et al., 2005;
Ramchand et al., 2006). Therefore, cocaine arrests among African Americans are expected to be
more sensitive to changes in police behaviors.
Table 2.6: UCR Descriptive Statistics for Blacks and Whites (1992−2008)
Cocaine and Heroin Possession Arrest Ratios (%)
Adult Blacks

Adult Whites

City-year
Mean
SD.
Mean
SD.
Obs.
All States
3.37
2.82
2.98
2.68
8,790
Non-MJ States
3.28
2.72
2.41
2.48
5,368
MJ States w/o CA&CO
2.53
2.69
1.72
1.66
488
California
3.79
2.98
4.42
2.68
2,788
Colorado
1.74
2.19
1.10
1.46
146
Note.― MJ states include only states that passed laws before July 2008; states that
passed laws afterward are in non-MJ states.

To account for the fact that non-drug offense rates and arrest risks are also higher among
African Americans, I create all arrest ratios for cocaine separately for adult blacks and whites.
(Other racial categories in the UCR are Asians and Native Alaskans or American Indians.) Table
2.6 shows the summary statistics of these ratios. Similar to Table 2.1, medical marijuana states
that exclude California and Colorado show lower cocaine arrest ratios than other states for both
blacks and whites. In line with findings from other studies, Table 2.6 shows that the proportion of
cocaine arrestees is higher among blacks even conditional on race (except for California). Note
that, however, the difference in racial differences between medical marijuana states and nonmedical marijuana states is very small [(3.28 − 2.41) − (2.53 − 1.72) = -0.06%]. So, at least for
cocaine possession offenses, the law enforcement in medical marijuana states does not
discriminate more or less against African Americans than other states.

91

Table 2.7: Effects on Cocaine and Heroin Arrests among Blacks and Whites
Cocaine and Heroin All Arrest Ratios among Blacks
Law

-0.081**
(0.033)

-0.094**
(0.041)

-0.148
(0.100)

-0.188***
(0.072)

-0.158
(0.098)

-0.182*
(0.105)

Cocaine and Heroin All Arrest Ratios among Whites
Law

-0.130***
(0.039)

-0.083***
(0.032)

CA. & CO.
Yes
Yes
No
No
Time Trends
Linear
Quadratic
Linear
Quadratic
Note.― All specifications include city and year fixed effects. Robust standard errors are
reported in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05,
* p<0.1.

Table 7 presents the estimated effects of medical marijuana laws on the cocaine arrest ratio
for blacks and whites. Because around 10% of black ratios are zeros, I estimate a fixed-effect
Poisson model with quadratic time trends based on both the samples with or without California
and Colorado. Because African Americans are more subjected to the strength of law enforcement,
we would expect a racial difference in the response of cocaine arrests if police behaviors were the
major driving force. However, in both the restricted sample and the full sample, the estimates do
not exhibit significant racial differences and they are similar to the results in Table 3. The estimates
from the restricted sample are larger but much noisier (insignificant) than the full sample.
Specifically, these estimates indicate a 6.5−14.1% decrease in white arrest ratios, and a 7.5−13.4%
decrease in black arrest ratios. The similar magnitudes of estimates among blacks and whites
suggest that changes in enforcement are unlikely to be the major reason for the decline in cocaine
arrests.
Based on the UCR data, although marijuana arrests have increased after the passage of
medical marijuana laws, there is no evidence that cocaine and heroin arrests have increased.
Nevertheless, in addition to potential bias from law enforcement, another disadvantage of the UCR

92

arrest data is that it does not separate cocaine and heroin. Cocaine is a stimulant and its neurological
effects are fundamentally different from depressants like heroin. Although marijuana is hard to
classify, many of its neurological effects are more similar to depressants (Abood and Martin, 1992;
Domino, 1971). For instance, both marijuana and heroin can relieve pain. Stimulants and
depressants are often complements; for example, heroin can reduce the depression that ensues after
the "high" from cocaine and help with sleeping. In fact, anecdotal evidence suggests that cocaine
dealers are often also heroin dealers. If there is really a substitution with marijuana, as those
negative estimates indicate, it is more likely between marijuana and heroin rather than between
marijuana and cocaine. In the next section, to further evaluate the impacts of medical marijuana
laws separately on cocaine and heroin, I employ data on substance treatment referrals from
rehabilitation facilities that are not referred by the criminal justice system.

5. Analysis of Treatment Data
5.1. TEDS Data
The treatment data is from the Substance Abuse and Mental Health Services Administration’s
(SAMHSA) Treatment Episode Data Set (TEDS) for the years 1992 through 2008. The TEDS
collects admission data from all substance-abuse treatment facilities that receive public funding in
each state. Some states collect data on all patients in facilities, while other states only collect data
on publicly funded patients. Similar to the UCR, each admission does not uniquely identify an
individual. For each admission, the data report at most three substance abuse problems of the
patient, his/her demographics, such as gender and age, and the sources of referral. About 40% of
treatment admissions are referred by the criminal justice system, 30% are referred by patients
themselves or other individuals, and around 20% are referred by health care providers and alcohol

93

or drug abuse care providers.11 I exclude criminal justice referrals in order to obtain estimates that
are not directly affected by potential changes in law enforcement.
In the TEDS, the total number of admissions greatly fluctuates in some state-years, which
might be due to changes in available funding or reporting practice. For example, the total number
of treatments reported was only about half of its previous levels in Alaska and Washington since
1999. To account for the fluctuations in total admissions and capacity constraints of rehabilitation
facilities, as commonly used by the SAMHSA, I create ratios of cocaine or heroin treatments to
all substance treatments for each state as measures. Because each admission lists at most three
drugs, I define marijuana-/cocaine-/heroin-related treatment admissions as such if they are
identified as the primary, secondary, or tertiary abuse problem; and marijuana-/cocaine-/heroinprimary treatment admissions as such if they are recorded only as the primary abuse substance. As
juvenile hard drug treatments are rare, and to be consistent with the UCR arrests, I only use adult
(above age 18) treatment admissions. On the other hand, since criminal justice referrals are not
included and therefore the potential gender differences in arrest risks are not a particular concern.
I use both male and female admissions to keep more observations; the results from only male
admissions in this section are nearly identical. The sample includes all medical marijuana states
that passed laws before July 2008; except for Alaska, for which data are missing for many years,
they have data in every year.12

11

The remaining 10% are referred by community or religious organizations, and self-help groups
such as Alcoholics Anonymous.
12
Alaska does not report referral sources for the years 1998–2003, and it does not report any data
for the years 2004–2007.
94

The summary statistics of marijuana, cocaine, and heroin treatment ratios are in Table
2.8.13 The related treatment ratios are in the upper panel, and the primary treatment ratios are in
the lower panel. Consistent with its popularity, 29.9% of treatment patients report a marijuana
abuse problem; but its low addictiveness makes marijuana account for only 7.8% of all primary
abuse problems. On the other hand, consistent with its strong addictiveness, although only 14.1%
of treatments are heroin-related, it accounts for 11.8% of primary abuse problems. Cocaine is kind
of in between; it accounts for 16.3% of primary problems, and 32.7% of treatments are cocainerelated. As treatments for hard drugs are disproportionally present, clearly, treatment patients in
the TEDS data are heavy users and they are different from general drug users. Although not shown
in the table, the alcohol-related (-primary) treatment ratio is about 70% (50%). The marijuana
treatment ratios in states with and without medical marijuana laws are similar. For cocaine
treatments, although these treatments do not include criminal justice referrals and therefore they
are not directly affected by law enforcement, the cocaine treatment ratios in medical marijuana
states are still much lower than non-medical marijuana states as in the UCR data. It suggests that,
at least for the subpopulation at risk, the cocaine use rates may actually be lower in medical
marijuana states. Heroin treatment ratios are similar in medical marijuana states and other states.
California has an extremely high proportion of heroin treatments. Colorado has a very low
percentage of drug treatments; in fact, 80% of Colorado primary treatments are alcohol treatments.

13

The sample size is only 810 for heroin treatments because Tennessee does not report any heroin
treatment for years after 1997.
95

Table 2.8: TEDS Descriptive Statistics (1992−2008)
Related Treatment Ratios (%)
Marijuana
Cocaine
Heroin
Obs.
29.90
32.71
14.06
All States
821
(8.98)
(14.62)
(15.30)
30.11
35.42
13.1
Non-MJ States
627
(8.85)
(14.87)
(15.56)
MJ States w/o
31.81
24.48
15.84
160
CA. & CO.
(8.08)
(9.54)
(12.32)
19.81
28.69
41.08
California
17
(2.61)
(6.35)
(9.88)
14.34
14.31
5.56
Colorado
17
(2.65)
(3.69)
(1.41)
Primary Treatment Ratios (%)
Marijuana
Cocaine
Heroin
Obs.
7.78
16.26
11.84
All States
821
(3.46)
(10.43)
(13.94)
8.15
18.69
11.12
Non-MJ States
627
(3.60)
(10.61)
(14.12)
MJ States w/o
7.18
8.35
12.60
160
CA. & CO.
(2.43)
(4.17)
(11.23)
4.39
10.31
38.22
California
17
(1.97)
(0.91)
(9.18)
3.42
6.97
4.28
Colorado
17
(0.67)
(2.06)
(1.04)
Note.― Standard deviations are in the parenthesis. Medical marijuana states
include only states that passed laws before July 2008; states that passed laws
afterward are in Non-MJ states. Heroin treatment is missing in Tennessee for
years after 1997, and the number of observations for Heroin is only 616 in NonMJ states.

5.2. Results
To evaluate the effects of medical marijuana laws on drug treatments, I estimate a model
that is similar to the one in the UCR analysis:
(6)

Yst = β Lawst + CA Law + CO Law + State fixed effectss + Year fixed effectst
+ State time trendsst + ɛst ,

96

where Yst is the marijuana, cocaine, or heroin treatment ratios (or their logarithms) in state s and
year t. I estimate Equation (2) by OLS or a fixed-effect Poisson model. As in the previous analysis,
I focus on the specifications without controls to keep a larger sample size, and the estimated
standard errors are clustered at the state level.
Table 2.9 shows the estimated effects on marijuana treatment ratios. The estimates from
log-linear models are in the left panel, and the estimates from a fixed-effect Poisson model are in
the right panel. In the first columns in each panel, I exclude the interaction terms, CA Law and CO
Law, and estimate the legalization effects based on all states. Consistent with marijuana arrests in
the UCR, the estimated effects of medical marijuana laws are positive. The estimates suggest a
6.9–9.1% increase in marijuana, but they are quite noisy in marijuana-primary treatments. In the
latter two columns in each panel, conditional on California and Colorado, the estimates for Lawst
are positive and highly significant for marijuana-related treatments. Although the estimates are
still noisier in primary treatments, in terms of percentage change, the magnitudes are very similar
to those in marijuana-related treatments. Specifically, conditional on California and Colorado,
medical marijuana laws are associated with a 9.1–11.9% increase in marijuana-related treatment
ratios or an 11.2–13.0% increase in the marijuana-primary treatment ratio.
Even though the treatments do not include criminal justice referrals, the estimates of CA
Law and CO Law are qualitatively similar to those from the UCR marijuana arrest. The legalization
effect in California seems to be smaller than in other medical marijuana states, but it is a little
sensitive to time trend specification and not always significantly different. However, estimates for
Colorado are always negative and significantly different from the estimates for Lawst. Although I
cannot rule out that marijuana use has actually decreased in Colorado, it could be that people are

97

Table 2.9: Effects of Medical Marijuana Laws on Marijuana Treatments
Marijuana Related Treatment
Log Linear Model
FE Poisson
Law

0.080*
(0.041)

0.119***
(0.033)

0.093***
(0.034)

CA Law

-0.087**
(0.043)

CO Law

-0.390***
(0.041)

0.091***
(0.029)

0.111***
(0.028)

0.092***
(0.032)

-0.064
(0.044)

-0.084**
(0.036)

-0.083**
(0.038)

-0.412***
(0.042)

-0.360***
(0.032)

-0.417***
(0.036)

Marijuana Primary Treatment
Log Linear Model
FE Poisson
Law

0.069
(0.068)

0.130**
(0.065)

0.112*
(0.059)

CA Law

-0.210***
(0.068)

CO Law

-0.496***
(0.075)

0.084
(0.053)

0.116**
(0.054)

0.118**
(0.052)

0.013
(0.068)

-0.225***
(0.062)

-0.060
(0.060)

-0.496***
(0.065)

-0.495***
(0.062)

-0.514***
(0.058)

Obs.

821

821

821

821

821

821

Time
Trends

Linear

Linear

Quadratic

Linear

Linear

Quadratic

Note.― All specifications include state and year fixed effects. Robust standard errors are
reported in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

less likely to seek treatments due to a lower perceived risk or treatment facilities giving the priority
to other substance patients.14
Table 2.10 and Table 2.11 present the effects of medical marijuana laws on cocaine and
heroin treatment ratios, respectively. In both Table 2.10 and Table 2.11, the estimates are
quantitatively similar with or without California and Colorado. In Table 2.10, the estimates on

14

The extremely large negative estimates for Colorado are partially due to a large increase in
alcohol treatments after 2001. The number of alcohol treatments in Colorado was almost doubled
from 2001 to 2002. However, the estimated legalization effects on marijuana treatments in
Colorado still show a 10% decrease even when alcohol treatments are excluded.
98

Table 2.10: Effects of Medical Marijuana Laws on Cocaine Treatments
Cocaine Related Treatment
Log Linear Model
FE Poisson
Law

-0.010
(0.058)

-0.010
(0.066)

-0.066
(0.069)

CA Law

0.140**
(0.068)

CO Law

-0.194**
(0.075)

-0.004
(0.058)

-0.012
(0.069)

-0.080
(0.066)

0.110
(0.070)

0.108
(0.068)

0.121*
(0.068)

-0.190**
(0.075)

-0.148**
(0.075)

-0.174**
(0.069)

Cocaine Primary Treatment
Log Linear Model
FE Poisson
Law

-0.005
(0.056)

-0.017
(0.059)

-0.004
(0.073)

CA Law

0.250***
(0.070)

CO Law

-0.188***
(0.069)

-0.013
(0.070)

-0.031
(0.081)

-0.057
(0.079)

0.285***
(0.076)

0.211**
(0.085)

0.289***
(0.081)

-0.248***
(0.077)

-0.137
(0.084)

-0.207***
(0.080)

Obs.
821
821
821
821
821
821
Time
Linear
Linear
Quadratic
Linear
Linear
Quadratic
Trends
Note.― All specifications include state and year fixed effects. Robust standard errors are
reported in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1.

cocaine treatments are negative but very small and never significant. In contrast, in Table 2.11, the
estimates on heroin treatments are negative and quite large in absolute terms. After the passage of
medical marijuana laws, on average, the heroin-related treatments decreased by 15.2–18.4%, and
the heroin-primary treatments decreased by 19–25.9%. These estimates are quite noisy, especially
in log-linear models. As indicated by the large standard deviations in Table 2.8, the heroin
treatment ratios are somewhat dispersed with many tiny values, and therefore a linear model may
fit the data poorly. Unlike in the UCR arrests, Colorado often shows negative effects on both
cocaine and heroin treatments, while California sometimes shows positive effects on these
treatments. However, the interpretation of the estimates for CA Law and CO Law should be treated
99

Table 2.11: Effects of Medical Marijuana Laws on Heroin Treatments
Heroin Related Treatment
Log Linear Model
Law

-0.153
(0.108)

-0.152
(0.128)

-0.183*
(0.097)

CA Law

0.146
(0.139)

CO Law

-0.205
(0.127)

-0.173***
(0.055)

-0.184**
(0.072)

-0.156**
(0.072)

0.229**
(0.108)

0.065
(0.078)

0.149**
(0.076)

-0.161
(0.110)

-0.177**
(0.078)

-0.192**
(0.082)

Heroin Primary Treatment
Log Linear Model
Law

-0.216*
(0.116)

FE Poisson

-0.245*
(0.138)

-0.259**
(0.102)

CA Law

0.215
(0.160)

CO Law

-0.209***
(0.053)

FE Poisson
-0.223***
(0.074)

-0.190**
(0.083)

0.324**
(0.123)

0.051
(0.084)

0.180*
(0.094)

0.080
(0.134)

0.094
(0.112)

0.001
(0.079)

-0.042
(0.089)

Obs.

810

810

810

810

810

810

Time
Trends

Linear

Linear

Quadratic

Linear

Linear

Quadratic

Note.― All specifications include state and year fixed effects. Robust standard errors are
reported in parentheses, and they are clustered at the state level. *** p<0.01, ** p<0.05, * p<0.1

with caution as the estimated standard errors are probably biased in these estimates. It is because
the estimates for CA Law and CO Law are identified based on only one state while the estimated
standard errors are clustered at the state level.
As in the previous analysis in the UCR arrests, the results in Table 2.10 and 2.11 support
neither the gateway hypothesis nor the complementarity between marijuana and cocaine or heroin.
In fact, there could be a decline in heroin usage while there is no change in cocaine usage. To see
it graphically, Figures 2.4 and 2.5, constructed in the same way as Figure 2.3, illustrate the changes
in cocaine- (Figure 2.4) and heroin- (Figure 2.5) primary treatment ratios before and after the

100

5

10

15

20

Cocaine Primary Treatment Ratios (w/o CA. & CO.)

-5

-4

-3

-2

-1
0
1
2
Year Relative to Pass Law

MJ States (w/o CA. & CO.)

3

4

5

Non-MJ States

5

10

15

20

Cocaine Primary Treatment Ratios

-5

-4

-3

-2

-1
0
1
2
Year Relative to Pass Law

MJ States

3

4

5

Non-MJ States

Figure 2.4: Cocaine Primary Treatment Ratios Before and After the Passage of Laws

passage of laws for both the samples with or without California and Colorado. (The graphs based
on related treatments are nearly identical.) Note that the scales are different between Figure 2.4

101

10

12

14

16

18

Heroin Primary Treatment Ratios (w/o CA. & CO.)

-5

-4

-3

-2

-1
0
1
2
Year Relative to Pass Law

3

4

5

Non-MJ States

MJ States (w/o CA. & CO.)

10

12

14

16

18

Heroin Primary Treatment Ratios

-5

-4

-3

-2

-1
0
1
2
Year Relative to Pass Law

MJ States

3

4

5

Non-MJ States

Figure 2.5: Heroin Primary Treatment Ratios Before and After the Passage of Laws

and Figure 2.5. For cocaine, in both the states with and without medical marijuana laws, the
treatment ratios indicate the same pattern of smoothly decreasing over time, suggesting no effect

102

of legalization on cocaine treatments. For heroin, the treatment ratios were gradually increasing
before the passage of laws in both medical marijuana states and other states. In contrast, after the
passage of laws, the treatment ratios have shown a decrease in medical marijuana states while they
keep increasing slightly in other states.
Multidrug abuse is common among cocaine and heroin treatment patients. In my sample,
30% of cocaine-primary treatment patients report marijuana abuse, and 40% of heroin-primary
treatment patients report cocaine abuse. Moreover, the relationship is not symmetric: only 17% of
marijuana-primary treatment patients report cocaine abuse, and 5% of cocaine-primary treatment
patients report heroin abuse. This observational fact that patients who use harder drugs are more
likely to use softer drugs (but not vice versa) is a major basis for the gateway hypothesis. Although
the estimates in previous tables indicate that, on average, marijuana could be a substitute to heroin
but has no direct relationship with cocaine, there could be heterogeneous effects of these laws. As
some experimental studies suggest, the substitution and complementarity between drugs may vary
by different types of drug users (Chalmers et al., 2010; Jofre-Bonet and Petry, 2008; Petry, 2001;
Petry and Bickel, 1998). To investigate potential heterogeneous legalization effects, I focus on the
two most common combinations in the TEDS: cocaine-primary with marijuana and heroin-primary
with cocaine. Cocaine-primary with marijuana treatments are the subset of cocaine-primary
treatments in which marijuana are either secondary or tertiary abuse problems. Heroin-primary
with cocaine treatments are defined in the same way. In addition, I also utilize the information on
the routes of drug use to create "speedball" treatment ratios, a subset of heroin-primary with
cocaine treatments, in which heroin injection is the primary problem and cocaine injection is the
secondary or tertiary problem. (The number of treatments for which cocaine injection is the
primary problem and heroin injection is secondary/tertiary problems is very small). The direct

103

injection of cocaine and heroin together is the strongest and most dangerous way to combine them,
called "speedball" or alternatively known as "powerballing," and it has caused many celebrity
deaths, including the deaths of John Belushi and River Phoenix along with many others. One
potential concern about previous results from Table 2.11 is that they might be driven by treatment
facilities' unobserved reactions to medical marijuana laws rather than real changes in underlying
drug use. For instance, rehabilitation facilities might give priority to marijuana addicts, and
therefore indirectly reduce the enrollment of heroin patients due to capacity constraints. The results
on "speedball" treatments can partially address this concern as "speedball" is one of the hardest
drugs and its treatment admissions are less likely to be affected by these unobservable factors.
Table 2.12: TEDS Descriptive Statistics for Multidrug Treatments

Cocaine Primary w/
Marijuana

Treatment Ratios (%)
All states
MJ states
5.29
2.79
(3.61)
(1.42)

Non-MJ states
5.93
(3.72)

Obs.

786

160

626

Heroin Primary w/
Cocaine

4.53
(5.73)
760

4.06
(3.38)
160

4.65
(6.20)
600

1.48
(1.99)
721

1.77
(1.69)
160

1.40
(2.06)
561

Obs.
Speedball
Obs.

Note.― MJ states include only states that passed laws before July 2008; states
that passed laws afterward are in non-MJ states.

The summary statistics of these treatment ratios are presented in Table 2.12. For brevity, I
focus on the restricted sample without California and Colorado; the observations are smaller in
non-medical marijuana states due to missing data in non-primary drugs or routes of use. Table 2.13

104

shows the estimated effects on treatment ratios of these drug combinations. The estimates on
cocaine-primary with marijuana treatments are actually positive. Although the estimated standard
errors are large, and only the estimate from the fixed-effect Poisson model with linear time trends
is significant, in terms of percentage changes, the estimates from the log specification are similar
to those in marijuana treatments in Table 2.9. They indicate roughly a 10% increase in treatments
in which cocaine is the primary abuse problem and marijuana is the secondary abuse problem. The
positive effect of laws on this set of cocaine users implies that the gateway hypothesis may indeed
exist for some particular types of drug users. One might be worried that the estimates on marijuana
treatments in Table 2.9 are driven by these cocaine-primary users, so the increase in marijuana is
Table 2.13: Effects of Medical Marijuana Laws on Multidrug Treatments
Cocaine Primary w/ Marijuana
Log Linear Model
Law

0.106
(0.074)

FE Poisson

0.096
(0.113)

0.155**
(0.075)

0.094
(0.105)

Heroin Primary w/ Cocaine
Log Linear Model
Law

-0.228*
(0.114)

FE Poisson

-0.240**
(0.107)

-0.154
(0.108)

-0.163*
(0.090)

Speedball
Log Linear Model
Law

-0.333*
(0.169)

FE Poisson

-0.307
(0.183)

-0.185
(0.218)

-0.152
(0.157)

Time
Linear
Quadratic
Linear
Quadratic
Trends
Note.― All specifications include state and year fixed effects. Robust standard errors
are reported in parentheses, and they are clustered at the state level. *** p<0.01, **
p<0.05, * p<0.1

actually a byproduct of the increase in cocaine-primary with marijuana treatments. However, the
results in marijuana treatments are nearly the same even if I exclude admissions that report any

105

cocaine use (not reported). For heroin-primary treatment patients who also report cocaine abuse,
the estimated effects of the laws are negative and often significant, with a similar magnitude of
decrease by around 20% as in Table 2.10. Therefore, the decline in heroin treatment patients is
similar regardless whether they use cocaine together or not. For "speedball" treatments, although
the estimates are somewhat sensitive to model specifications and quite noisy due to the small
number of "speedball" treatments, in terms of percentage changes, they indicate a decrease of about
15 –30%. As the estimates for “speedball” treatments are similar to those for heroin treatments,
they provide further evidence that the decline in heroin treatments is indeed caused by medical
marijuana laws.

6. Discussion of Results and Conclusion
In this paper, I use data on drug possession arrests and treatment admissions, and estimate
reduced-form models for the effects of medical marijuana laws on marijuana, cocaine, and heroin
usage. My results indicate a 10−20% increase in marijuana arrests and treatments after the passage
of medical marijuana laws. Although it is a widely accepted belief that marijuana is a complement
or even a gateway drug to hard drugs, I do not find strong evidence supporting such relationships
between marijuana and cocaine or heroin. In contrast, the possession arrests of cocaine and heroin
do not appear to significantly change or even decrease after legalization. In addition, there is no
racial difference in the estimated effects, suggesting that the estimates are not driven by changes
in police behaviors. To address the potential bias from law enforcement and to examine the effects
separately for cocaine and heroin, I use treatment admissions that are unrelated to criminal justice
from the TEDS data. I also find a similar magnitude of about 20% decrease in heroin treatment
admissions, but no significant change in cocaine treatment admissions. Although these findings

106

are fairly unexpected, they are consistent with some qualitative studies that report medical
marijuana patients substituting alcohol and other illegal drugs for marijuana (Harris et al., 2000;
Reiman, 2007, 2009). Interestingly, Anderson et al. (2012a) actually find a large reduction in
cocaine use among teenagers, even though they do not find any increase in marijuana use.
A 10–20% change in marijuana and heroin use represents a large legalization effect. As
this study is one of the first to assess the effects of medical marijuana laws on hard drug usage, it
is difficult to compare the specific outcomes with the results from the existing literature. However,
the magnitude is not implausibly large in comparison to other policy effects on substance use found
in existing studies. For example, Conlin et al. (2005) find that legal changes in the availability of
alcohol have a substitution effect on arrests of marijuana and hard drugs with an order of magnitude
of 10–30%. Conceptually, since arrests and treatment admissions do not represent individuals but
frequencies, they capture changes in both the intensive and extensive margins. Because hard drugs
like heroin are highly addictive, and the substitution effect seems to exist even among people using
one of the hardest drugs, "speedball," it is unlikely for these drug users to just stop using these
drugs. Therefore, a significant part of the substitute effects is probably on the intensive margin.
For instance, some anecdotal evidence suggests that marijuana can help to ease the craving for
heroin.
Since arrests and treatments tend to reflect heavy drug use, these results might not be
simply applied to the general drug users; there is no doubt that treatment patients are different from
casual marijuana users. As Chu (2012) points out, the differences underlying drug users could be
one reason that some studies based on national representative samples tend to find little or no
effects of these medical marijuana laws on marijuana use. On the other hand, for hard drugs,
especially for heroin, such distinctions may be less important as heroin users are generally heavy

107

users. For policy evaluations, however, these estimates are still relevant to the design of policy
because the major concern of these laws is whether heavy marijuana use will cause hard drug use.
Results in this study suggest that, on average, marijuana could be a substitute for heroin,
and it does not have a strong relationship with cocaine. However, as this paper focuses on drug
usage at the aggregated level, it may ignore potential heterogeneous effects on different types of
individuals, especially multidrug users. In fact, I intentionally avoid using terms like "elasticity of
substitution" to interpret my reduced-form results. For example, the estimated effects are positive
on patients who use cocaine with marijuana but negative on patients who use cocaine with heroin.
As suggested by some experimental studies, the relationships between substances may depend on
different types of users, and it is possible that the gateway effect is indeed causal for some
particular types of drug users. To further evaluate the impacts of medical marijuana laws and to
conduct a more complete cost-benefit analysis, future research has to pay more attention to these
potential heterogeneous effects. Due to constraints such as sample sizes from currently available
datasets, qualitative studies with detailed and extensive descriptions of drug use behaviors may be
equally important as quantitative studies.

108

APPENDIX

109

Table E: Descriptive Statistics for Control Variables
Mean

SD.

Obs.

Police Officer Rates per 100k City
Residents

179.61

70.11

9,539

State Percentage of Blacks among
Males (%)

4.84

3.02

10,032

State Unemployment Rates (%)

5.58

1.46

10,032

State 0.02 Blood Alcohol Content
Law

0.63

0.48

10,032

State Health and Hospital
Expenditures per capita (in
thousands)

497.38

178.02

8,852

State Police Expenditures per capita
(in thousands)

215.22

74.79

8,852

110

REFERENCES

111

REFERENCES

Abood, M. E., Martin, B. R., 1992. Neurobiology of marijuana abuse. Trends in Pharmacological
Sciences 13 (0), 201-206.
Agrawal, A., Neale, M. C., Prescott, C. A., Kendler, K. S., 2004. Cannabis and other illicit drugs:
comorbid use and abuse/dependence in males and females. Behavior Genetics 34 (3), 217-228.
Akiyama, Y., Propheter, S. K., 2005. Methods of data quality control: for Uniform Crime
Reporting programs. Federal Bureau of Investigation, Criminal Justice Information Services
Division.
Anderson, D. M., Hansen, B., Rees, D. I., 2012a. Medical marijuana laws and teen marijuana use.
IZA Discussion Paper No. 6592.
Anderson, D. M., Hansen, B., Rees, D. I., 2012b. Medical marijuana laws, traffic fatalities, and
alcohol consumption. Journal of Law and Economics (Forthcoming).
Anderson, E., 2012. "Mother earth" medical marijuana clinic prepares to close. KPBS Public
Broadcasting. San Diego, CA.
Arkes, J., Pacula, R. L., Paddock, S. M., Caulkins, J. P., Reuter, P., 2008. Why the DEA STRIDE
data are still useful for understanding drug markets. NBER Working Paper 14224.
Becker, G. S., Murphy, K. M., 1988. A theory of rational addiction. The Journal of Political
Economy 96 (4), 675-700.
Beckett, K., Nyrop, K., Pfingst, L., 2006. Race, drugs, and policing: Understanding disparities in
drug delivery arrests. Criminology 44 (1), 105-137.
Beckett, K., Nyrop, K., Pfingst, L., Bowen, M., 2005. Drug use, drug possession arrests, and the
question of race: lessons from Seattle. Social Problems 52 (3), 419-441.
Bound, J., Solon, G., 1999. Double trouble: on the value of twins-based estimation of the return to
schooling. Economics of Education Review 18 (2), 169-182.
Bretteville-Jensen, A. L., 2011. Illegal drug use and the economic recession—what can we learn
from the existing research? The International Journal on Drug Policy 22 (5), 353-359.
Bretteville-Jensen, A. L., Melberg, H. O., Jones, A. M., 2008. Sequential patterns of drug use
initiation – can we believe in the gateway theory? The B.E. Journal of Economic Analysis &
Policy 8 (2), (Contributions), Article 1.
Caplan, G., 2012. Medical marijuana: a study of unintended consequences. McGeorge Law
Review 43, 127-146.
Carpenter, C., 2007. Heavy alcohol use and crime: evidence from underage drunk‐driving laws.
Journal of Law and Economics 50 (3), 539-557.

112

Chalmers, J., Bradford, D., Jones, C., 2010. The effect of methamphetamine and heroin price on
polydrug use: a behavioural economics analysis in Sydney, Australia. International Journal of
Drug Policy 21 (5), 381-389.
Chu, L. Y.-W., 2012. Medical
http://ssrn.com/abstract=2164778.

marijuana

laws

and

illegal

marijuana

use.

Cohen, P. J., 2010. Medical marijuana 2010: It's time to fix the regulatory vacuum. The Journal of
Law, Medicine & Ethics 38 (3), 654-666.
Conlin, M., Dickert‐Conlin, S., Pepper, J., 2005. The effect of alcohol prohibition on illicit‐drug‐
related crimes. Journal of Law and Economics 48 (1), 215-234.
Coté, J., Allday, E., Buchanan, W., Lagos, M., 2008. Feds say S.F. has more pot clubs than
Starbucks, but it might not add up. San Francisco Chronicle.
Dannerbeck, A., Harris, G., Sundet, P., Lloyd, K., 2006. Understanding and responding to racial
differences in drug court outcomes. Journal of Ethnicity in Substance Abuse 5 (2), 1-22.
DeSimone, J., 1998. Is marijuana a gateway drug? Eastern Economic Journal 24 (2), 149-164.
Dickinson, T., 2012. Obama's war on pot. Rolling Stone.
Domino, E. F., 1971. Neuropsychopharmacologic studies of marijuana: some synthetic and natural
THC derivatives in animals and man. Annals of the New York Academy of Sciences 191 (1),
166-191.
Donohue III, John J., Steven D. Levitt. 2001. The impact of race on policing and arrests. Journal
of Law and Economics 44 (2), 367-394.
Drug Enforcement Administration (DEA). 1991. DEA
http://www.usdoj.gov/dea/pubs/history/1985-1990.html

history

book

1875–1990

Drug Enforcement Administration (DEA), 2011. The DEA Position on marijuana.
Ellgren, M., Spano, S. M., Hurd, Y. L., 2006. Adolescent cannabis exposure alters opiate intake
and opioid limbic neuronal populations in adult rats. Neuropsychopharmacology 32 (3), 607615.
Fergusson, D. M., Boden, J. M., Horwood, L. J., 2006a. Cannabis use and other illicit drug use:
testing the cannabis gateway hypothesis. Addiction 101 (4), 556-569.
Fergusson, D. M., Boden, J. M., Horwood, L. J., 2006b. Testing the cannabis gateway hypothesis:
replies to Hall, Kandel et al. (2006) and Maccoun (2006). Addiction 101 (4), 474-476.
Fryer, R., Heaton, P., Levitt, S., Murphy, K., 2010. Measuring crack cocaine and its impact.
Economic Inquiry 51 (3), 1651-1681.
Government Accountability Office (GAO), 2002. Marijuana: early experiences with four states'
laws that allow use for medical purposes. Washington, DC.

113

Golub, A., Liberty, H. J., Johnson, B. D., 2005. Inaccuracies in self-reports and urinalysis tests:
impacts on monitoring marijuana use trends among arrestees. Journal of Drug Issues 35 (4),
941-966.
Gross, S. R., Barnes, K. Y., 2002. Road work: racial profiling and drug interdiction on the highway.
Michigan Law Review 101 (3), 651-754.
Grossman, M., Chaloupka, F. J., 1998. The demand for cocaine by young adults: A rational
addiction approach. Journal of Health Economics 17 (4), 427-474.
Harris, D., Jones, R. T., Shank, R., Nath, R., Fernandez, E., Goldstein, K., Mendelson, J., 2000.
Self-reported marijuana effects and characteristics of 100 San Francisco medical marijuana
club members. Journal of Addictive Diseases 19 (3), 89-103.
Harrison, L., Hughes, A., 1997. The validity of self-reported drug use: improving the accuracy of
survey estimates. National Institute on Drug Abuse.
Hernández-Murillo, R., Knowles, J., 2004. Racial profiling or racist policing? bounds tests in
aggregate data. International Economic Review 45 (3), 959-989.
Horowitz, J. L., 2001. Should the DEA’s STRIDE data be used for economic analyses of markets
for illegal drugs? Journal of the American Statistical Association 96 (456), 1254-1262.
Jofre-Bonet, M., Petry, N. M., 2008. Trading apples for oranges? results of an experiment on the
effects of heroin and cocaine price changes on addicts’ polydrug use. Journal of Economic
Behavior & Organization 66 (2), 281-311.
Johnson, K., 2009. States pressed into new role on medical marijuana. The New York Times.
Kandel, D., 1975. Stages in adolescent involvement in drug use. Science 190 (4217), 912-914.
Kandel, D. B., 2003. Does marijuana use cause the use of other drugs? JAMA: The Journal of the
American Medical Association 289 (4), 482-483.
Kandel, D. B., Yamaguchi, K., Klein, L. C., 2006. Testing the gateway hypothesis. Addiction 101
(4), 470-472.
Leger, D. L., 2012. Survey: 1 in 15 high school seniors smoking pot. USA TODAY.
Levitt, Steven D., 1998. Juvenile crime and punishment. The Journal of Political Economy 106
(6), 1156-1185.
Lynch, J. P., Jarvis, J. P., 2008. Missing data and imputation in the Uniform Crime Reports and
the effects on national estimates. Journal of Contemporary Criminal Justice 24 (1), 69-85.
Lynskey, M., Vink, J., Boomsma, D., 2006. Early onset cannabis use and progression to other drug
use in a sample of Dutch twins. Behavior Genetics 36 (2), 195-200.
Lynskey, M. T., Heath, A. C., Bucholz, K. K., Slutske, W. S., Madden, P. A. F., Nelson, E. C.,
Statham, D. J., Martin, N. G., 2003. Escalation of drug use in early-onset cannabis users vs.
co-twin controls. JAMA: The Journal of the American Medical Association 289 (4), 427-433.

114

MacCoun, R., 1998. In what sense (if any) is marijuana a gateway drug. The FAS Drug Policy
Analysis Bulletin (4).
Maccoun, R. J., 2006. Competing accounts of the gateway effect: the field thins, but still no clear
winner. Addiction 101 (4), 473-474.
Melberg, H., Jones, A., Bretteville-Jensen, A., 2010. Is cannabis a gateway to hard drugs?
Empirical Economics 38 (3), 583-603.
Mikos, R. A., 2011. A critical appraisal of the department of justice's new approach to medical
marijuana. Stanford Law & Policy Review 22 (2), 633-670.
Moffatt, S., Wan, W.-Y., Weatherburn, D., 2012. Are drug arrests a valid measure of drug use? a
time series analysis. Policing: an International Journal of Police Strategies & Management 35
(3), 458 - 467.
O'Connor, A., 2011. Marijuana use growing among teenagers. The New York Times.
Pacula, R. L., Kilmery, B., Grossmanz, M., Chaloupka, F. J., 2010. Risks and prices: the role of
user sanctions in marijuana markets. The B.E. Journal of Economic Analysis & Policy 10 (1),
1-36.
Parker, K. F., Maggard, S. R., 2005. Structural theories and race-specific drug arrests: what
structural factors account for the rise in race-specific drug arrests over time? Crime &
Delinquency 51 (4), 521-547.
Petry, N. M., 2001. A behavioral economic analysis of polydrug abuse in alcoholics: asymmetrical
substitution of alcohol and cocaine. Drug and Alcohol Dependence 62 (1), 31-39.
Petry, N. M., Bickel, W. K., 1998. Polydrug abuse in heroin addicts: a behavioral economic
analysis. Addiction 93 (3), 321-335.
ProCon.org. 2012. How many people in the US use medical marijuana?
http://medicalmarijuana.procon.org/view.answers.php?questionID=001199 (Accessed
4.12.2012.)
Pudney, S., 2003. The road to ruin? sequences of initiation to drugs and crime in Britain. The
Economic Journal 113 (486), 182-198.
Ramchand, R., Pacula, R. L., Iguchi, M. Y., 2006. Racial differences in marijuana-users’ risk of
arrest in the United States. Drug and Alcohol Dependence 84 (3), 264-272.
Reiman, A., 2007. Medical cannabis patients: Patient profiles and health care utilization patterns.
Complementary Health Practice Review 12 (1), 31-50.
Reiman, A., 2009. Cannabis as a substitute for alcohol and other drugs. Harm Reduction Journal
6 (1), 35.
Rosenfeld, R., Decker, S. H., 1999. Are arrest statistics a valid measure of illicit drug use? the
relationship between criminal justice and public health indicators of cocaine, heroin, and
marijuana use. Justice Quarterly 16 (3), 685-699.
115

Saffer, H., Chaloupka, F., 1999. The demand for illicit drugs. Economic Inquiry 37 (3), 401-411.
Salomonsen-Sautel, S., Sakai, J. T., Thurstone, C., Corley, R., Hopfer, C., 2012. Medical
marijuana use among adolescents in substance abuse treatment. Journal of the American
Academy of Child and Adolescent Psychiatry 51 (7), 694-702.
Sekhon, V., 2009. Highly uncertain times: an analysis of the executive branch's decision to not
investigate or prosecute individuals in compliance with state medical marijuana laws. Hastings
Constitutional Law Quarterly 37 (3), 553-564.
Solinas, M., Panlilio, L. V., Goldberg, S. R., 2004. Exposure to [delta]-9-tetrahydrocannabinol
(THC) increases subsequent heroin taking but not heroin's reinforcing efficacy: a selfadministration study in rats. Neuropsychopharmacology 29 (7), 1301-1311.
Terry-McElrath, Y. M., Chriqui, J. F., Bates, H., McBride, D. C., 2009. Do state policies matter in
prosecutor-reported juvenile marijuana case disposition? Crime & Delinquency (Forthcoming).
http://cad.sagepub.com/content/early/2009/11/30/0011128709348457
Thurstone, C., Lieberman, S. A., Schmiege, S. J., 2011. Medical marijuana diversion and
associated problems in adolescent substance treatment. Drug and Alcohol Dependence 118 (2–
3), 489-492.
van Ours, J. C., 2003. Is cannabis a stepping-stone for cocaine? Journal of Health Economics 22,
539-554.
Warner, B. D., Coomer, B. W., 2003. Neighborhood drug arrest rates: are they a meaningful
indicator of drug activity? a research note. Journal of Research in Crime and Delinquency 40
(2), 123-138.

116

CHAPTER 3
Weber, Marx, and Work Values: Evidence from Transition Economies (with Susan Linz)
1.

Introduction
Prior to the publication of Weber’s Protestant Ethic and the Rise of Capitalism (1904-

1905/1958), one of the more detailed descriptions of the emergence of capitalism was found in
Marx’s Capital (1867/1976). Interestingly, although both Marx and Weber were influenced by
German social thought, generally, and by Hegel, in particular (Birnbaum, 1953; Mommsen, 1977;
Mueller, 1986), they differed in their assessment of the role of values in a society. According to
Marx, societies are in a continual, inevitable process of change. At any given time, a society is
characterized by a set of economic institutions defined by that society’s production technology,
and one’s position in the economic structure accounts for one’s values. Thus, Marx implies that
values depend upon and are defined by the existing economic environment (Anthony, 1977).
In contrast, Weber focused on how values contribute to the emergence of a particular
economic environment (Hill, 1996). More specifically, how did the adoption of and adherence to
“Protestant work ethic” (PWE) – where PWE encompasses general orientation toward hard work,
industriousness, individual gain, need for achievement, as well as negative attitude towards time
waste (Furnham and Rajamanickham, 1992; Johassen, 1947; Jones, 1997) – contribute to the rise
of the particular economic structure of capitalism.1 While recent studies provide mixed results

1

Numerous studies indicate that PWE is not unique to Protestants (Abdalla, 1997; Arslan, 2000,
2001; Furnham and Muhiudeen, 1984), but that it does appear to be stronger among individuals
who adhere strongly to religious beliefs (Aygun et al., 2008; Giorgi and Marsh, 1990). Although
not focused explicitly on PWE, Torgler (2011), using 1999/2000 European Value Survey data for
16 Western European countries and 15 Eastern European countries, finds a positive link between
Protestantism and “work should always come first,” and documents a positive link between
religiosity, more generally, and work centrality.
117

regarding the causality suggested by Weber (Arrunada, 2010; Becker and Woessman, 2009; 2010;
Cavalcanti et al., 2007; Delacroix and Nielson, 2001; Doepke and Zilibotti, 2008), no study
examines whether Weber or Marx was “right” in terms of the causal link between work values and
economic environment. Somewhat ironically, the collapse of the Soviet Union at the end of 1991
provides a unique natural experiment to investigate this link.
Ideally, longitudinal panel data collected from a representative sample of workers in the
pre-transition, transition, and post-transition economies which included information on a wide
variety of worker characteristics and work-related values would be utilized to analyze the link.
Since such data are not available, testing explicitly for a causal link between work values and
economic environment is not possible. However, we are able to investigate in some detail whether
differences in work ethic among those with and without work experience in the socialist economic
system are consistent with results that would be predicted regarding the link between work values
and economic environment if Marx were “right” or Weber were “right.” Moreover, we document
the extent to which our results have internal validity. Thus, while rather exploratory in nature due
to the quality of the data actually available, this study takes a preliminary step in investigating the
link between work values and economic environment by examining whether the results are
consistent with the position held by Marx or the position put forth by Weber.
The transformation of socialist economies to capitalist economies at the end of the 20th
century was nothing like the rise of capitalism described by Marx and Weber. Their analysis relates
to the gradual emergence of capitalism from a feudal society. Even though both socialist and feudal
societies are considered collectivist in nature (Hofstede, 1980), the socialist economies that
initiated the transformation to market-oriented economies in the early 1990s were not feudal
societies when the transition began, nor was their transformation as gradual as the emergence of

118

capitalism in the 17th century. However, given that transition economies currently are populated
by individuals both with and without work experience in the former socialist economy, it is
possible to empirically investigate the link between values and economic environment.
If Marx is “right” in transition economies, older generation workers (individuals who were
trained and worked in the former socialist economy) would have a different work ethic than young
generation workers (individuals whose training and work experience is limited to the marketoriented economy). In particular, older generation workers would likely adhere less strongly to a
“capitalist” work ethic, such as PWE, characterized by individual gain associated with hard work.
Older workers’ experiences and values were formed in the socialist economy, where pay
differentials were modest and opportunities to generate wealth severely restricted. In socialist
economies, the economic and workplace environment often is characterized by “we pretend to
work, and they pretend to pay us” (Granick, 1987; Gregory, 1987; Lane, 1986). Adherence to PWE
among workers in socialist economies, and thus older generation workers in transition economies,
would likely be low.2
Continuing this logic and taking into account the wealth of studies that link PWE to
capitalism, young generation workers, because they were trained and have work experience
exclusively in the market-oriented environment, would likely adhere more strongly to PWE

2

Partial support for the proposition that older workers will adhere to PWE less strongly than young
workers is found, for example, in van den Broek and de Moor (1994), who report that work values
in East European countries coincide with less appreciation for work initiative, achievement, and
responsibility than in West European countries. We note that East European countries were
developed capitalist economies before socialism was imposed, thus difference in work values
between East and West Europe would likely be less than what one might find when focusing on
countries that had no prior experience with capitalism. In a more general study, Schwartz and Bardi
(1997) document differences in values between countries with no socialist experience and
countries with socialist experience (communist rule), as well as between countries with different
durations of socialist experience.
119

(Halman, 1996; Torgler, 2011). Indeed, the proposition that older and young generation workers
in transition economies will have different work values is indirectly supported by survey results
that indicate significant generational differences in attitudes about market institutions, transition
outcomes, and nostalgia for the socialist era (Denisova et al., 2007; EBRD, 2007; Pew Global
Attitudes Project, 2011).
If we find that older generation workers do exhibit significantly weaker PWE adherence
than younger generation workers, we believe this sends a relatively strong signal in support of our
proposition. First, the generational differences in work values that we hypothesize among workers
in former socialist economies are opposite the results reported for developed market economies.
That is, studies suggest that in developed market economies, work ethic is weaker among the
younger generation (frequently referred to as Generation Y or Millennial) than among older
workers (Cennamo and Gardner, 2008; Gursoy et al., 2008; Keepnews et al., 2010; Macky et al.,
2008; Smola and Sutton, 2002). Second, if Marx is “right,” older generation workers who remain
employed in the emerging market-oriented economy are likely to adopt work ethics that tend to
coincide with capitalist economies. 3 Since our employee survey was conducted more than a
decade after the initial shock of the transformation process occurred, it may be that value adaptation
among older generation workers had already taken place. Moreover, older generation workers with
weaker PWE adherence may have dropped out of the workforce by this time.

3

Adaptation of values is documented among workers in Germany after the reunification (Torgler,
2003). However, other studies show that values, like personalities, are formed relatively early in
life (Caspi, 2000; Costa and McCrae, 1997; Hattrup et al., 2007; Nave et al., 2010), and are slow
to change (Schwartz and Bardi, 1997). More generally, in the empirical literature on moral values,
Torgler (2007) finds a linear relationship (positive association) between values and age.
120

As mentioned previously, since we do not have comparable work ethic data collected from
different age cohorts in the pre-transition economies, we cannot directly partial out potential age
effects in our analysis.4 We note, however, that panel data, even if available, would not be very
useful for our purpose, because cohort (generation) status does not change over time: older
generation workers had experience with the socialist economy, younger generation workers did
not. While longitudinal panel data would allow one to trace the change in an individual’s work
values or work ethic over time, for example, workers adjusting their work ethic in response to
changes in the economic environment; it is not possible to separate the age effect from any time
effect because age, like time, changes in annual increments.
Do older workers in formerly socialist economies adhere less strongly to PWE than young
workers? We analyze the link between work values and economic environment using data
collected from employees in three transition economies: Armenia, Azerbaijan, and Russia. All
three countries were part of the former Soviet Union, and thus shared a common socio-economic
environment. Since the early 1990s, however, the three countries have exhibited both economic and
cultural differences. In 2008, Armenia and Azerbaijan had per capita income levels just over onethird of Russia’s; unemployment totaled some 6% in Russia and Azerbaijan, but reached over 28%
in Armenia (World Bank). While Armenia and Azerbaijan each have ethnically homogenous
populations, with more than 98% of Armenians adhering to Christian faith and more than 93%
Azeris adhering to Islamic practices, Russia is populated by a greater diversity of ethnic groups,
with about half of the population participating in either Christian or Islamic practices, and half who

4

Like the majority of studies, we are not able to determine, for example, whether work values
depend upon age per se or career stage.

121

are either non-practicing or non-believers (NationMaster, 2012). Thus including employees from
these three countries in our analysis allows us to assess whether the predicted differences between
young and older generation workers in PWE adherence is sustained across different cultural and
economic environments.
Because financial constraints precluded selecting nationally representative samples in each
country, we view this analysis as an exploratory effort to document generational differences in
work values. Given our convenience rather than random sample, we are careful to restrict our
discussion and interpretation of results only to those workers participating in the survey. Knowing
who adheres to PWE helps to shape managerial strategies to motivate workers. This is particularly
true when the information is drawn from employees across a wide variety of workplaces in
different sectors of the economy, rather than from students, or from a single workplace, as is the
case in many work ethic studies. Indeed, one of the strengths of our study is the extensive set of
worker and workplace controls we are able to include because of the detailed nature of our data.
Knowing who adheres to PWE among employees in formerly socialist economies is important
because capturing work ethic, even if only at one point in time, provides a foundation for better
understanding the link between values and behavior. Similarities that emerge in these diverse
environments likely signal results that will contribute to developing a more global perspective of
factors influencing worker performance. Moreover, documenting adherence to PWE among
employees in former socialist economies extends the work ethic literature to include countries that
previously have been ignored, despite their growing importance in the global community.
Our analysis of PWE adherence by young and older generation workers in three formerly
socialist economies proceeds as follows: Section 2 summarizes characteristics of socialist and
post-socialist economic and workplace environments. Section 3 describes the data and work ethic

122

measure used in this analysis. The methodology used is presented in Section 4. Results and
discussion are provided in Section 5. In terms of generational differences in work ethic, we find
that in all three countries young generation workers participating in our study adhere more strongly
to PWE than older generation workers, and the magnitudes are similar across these three countries.
More generally, we find PWE adherence to be stronger among participating workers with an
internal locus of control, as well as among workers who hold supervisory positions. Gender
differences in PWE adherence are evident, but not consistent across countries. Section 6 offers
concluding remarks.

2.

Socialist and Post-Socialist Economic and Workplace Environments
A number of characteristics of the former Soviet economy contributed to the likelihood of

weak PWE adherence among the majority of workers. First, the combination of “parasite laws”
(labor codes that required all able-bodied adults work full time) and “job rights” (the institution
that guaranteed each individual a job regardless of their performance or the performance of their
workplace) meant that the labor force was comprised of individuals who effectively had no choice
about whether or not to work, and who were not fired for poor performance (Granick, 1987;
Gregory, 1987). 5 This situation, sustained by planners’ policies of “soft budget constraints”
(Kornai, 1980) and using firms as the loci for providing such social services as housing, medical
clinics, child care, sports facilities, and subsidized meals (Linz, 1995), not only contributed to low

5

Because unemployment was essentially illegal in the Soviet economy, no institutions were in
place to provide unemployment compensation; nor were there job retraining or job placement
centers because firms did not release workers made redundant by technological advance (Gregory
and Collier, 1988).

123

labor productivity, but also to the situation of workers and managers viewing employment and the
corresponding benefits as an entitlement (Lane, 1986; Standing, 1996).
Second, the collectivist nature of the socialist society contributed to a reward structure not
focused on individual performance. Centrally-determined wages, differentiated by occupation,
industry, region and workplace conditions, served to allocate labor in accordance with planners’
preferences rather than to reward performance (Chapman, 1963; Moskoff, 1984). The relative
equality of reward likely weakened worker motivation to expend effort or strive to achieve, a
proposition supported in general terms by Schwartz and Bardi (1997). In addition, given the
planner-determined environment where fulfilling quantity targets was the primary objective,
managers of state-owned firms were unconcerned with either economic or financial success. Thus,
unlike market economies where work values are important contributing factors to a firm’s success
(Judge et al., 2001), promoting work values consistent with PWE in the former Soviet economy
was largely ignored.
Third, legal restrictions in the Soviet economy not only precluded individuals from owning
property, but also undermined any activity that would generate non-labor income. “Speculative”
behavior continued to be illegal even as market-oriented conditions were introduced under
perestroika (Jones and Moskoff, 1991; Linz, 1995). Moreover, state control of media and religious
practices greatly influenced the cultural environment (Millar, 1987). Paternalism dominated both
social and economic realms (Feher, 1982), fostering passivity and undermining ambition (Lane,
1986; Millar, 1987). Consequently, the links between values, effort and reward described by
Weber were not evident in the official Soviet economy. Instead, research articles and popular press
document behaviors consistent with the adage “we pretend to work, they pretend to pay us”
(Granick, 1987; Gregory 1987).

124

At the end of 1991, the Soviet Union ceased to exist; each republic became a separate
country. Each initiated a transformation from socialism to capitalism. What is common across
Armenia, Azerbaijan and Russia is the severity of the economic impact of the transition process;
a process well-documented by data and research provided by the World Bank, the European Bank
for Reconstruction and Development, and a multitude of academic scholars, among others. Also
common are the market-oriented institutions and behaviors adopted during the transition. In
particular, privatization and enterprise restructuring, combined with new labor codes that expanded
individual and firm rights in the employment realm, created conditions where employment fell (as
a consequence of lower labor force participation rates and growing unemployment), performancebased pay prevailed (especially in privately-owned organizations), and opportunities expanded to
develop non-labor incomes. Thus, following Hofstede (1980), it is likely that values consistent
with capitalist market economies, such as the Protestant work ethic, became more prevalent among
individuals trained and working in the new economic environment. Indeed, PWE adherence is
stronger in economic environments where individual freedom, personal autonomy, and
performance-based pay dominates, and where individual ambition and personal gain is promoted
(Halman, 1996; Mann, 2010; Torgler, 2011).
These three countries differ in terms of the economic environments that emerged as their
market economies developed. One need only glance briefly at any number of data sources to
underscore this observation (see, for example, World Bank, EBRD, United Nations, OECD, or
World Values Survey). For example, macroeconomic conditions vary rather substantially across
country. Per capita income is much lower in Armenia, and reliance upon agriculture much higher,
than in Azerbaijan and Russia. While labor force participation rates are largely similar across
countries, employment in industry and the percent of self-employed differs significantly. If, as

125

some speculate, PWE adherence would likely be stronger among self-employed, then PWE
adherence should be higher in Azerbaijan and Armenia than in Russia.
Cultural differences in the post-Soviet period, particularly between Armenia and
Azerbaijan, also are well-documented (CIA World Factbook, 2010). Ethnic homogeneity
characterize Armenia and Azerbaijan, but different religious beliefs dominate: Christianity in
Armenia, Islam in Azerbaijan. In contrast, ethnic diversity characterizes Russia. According to the
CIA World Factbook, ethnic Russians comprise about 80% of population; 15-20% of population
are Russian Orthodox, 10-15% are Muslim, and 2% Christian. Documenting commonalities in
PWE adherence across these three countries, particularly among young and older generation
workers, begins an exploratory effort to expand the development of a global perspective of factors
influencing worker performance.

3.

Data and Work Ethic Measure
As part of a larger study investigating worker performance in formerly socialist economies,

an employee survey was conducted at more than 340 workplaces in Armenia, Azerbaijan and
Russia. Financial constraints precluded the selection of a nationally representative sample of
workers in each country, or a representative sample of workplaces in a particular locale. Instead,
local project coordinators used personal and professional connections and snowballing technique
to contact organizations about participating in the study.6 In organizations granting permission to

6

Because local project coordinators were instructed to include a diversity of workplaces (where
diversity is measured by ownership structure, sector, size) and a diversity of workers (young, older;
men, women; managers, workers), issues related to selection bias associated with under-coverage
should be minimal. However, because our country convenience samples were voluntary, we may
face selection bias associated with non-response (employees with weak work ethic do not
126

conduct the survey, the questionnaire was distributed to employees in common areas or at specific
job sites in the workplace because financial constraints also precluded selecting a representative
sample of workers in each organization. Employees were informed about policies to ensure
confidentiality and anonymity, and those who agreed to participate had the option of turning in a
complete or incomplete questionnaire.
Although our convenience sample includes 5,954 employees, we restrict it to include only
those young and older generation participants who answered all relevant questions for this
analysis.7 We define young generation workers as those born after 1981; individuals who were ten
years old at the time the transition began, and thus who had no real work experience in the former
socialist economy. We define older generation workers as those born before 1977. Older
generation workers were trained in the former socialist economy, were old enough to work prior
to the transition, and confronted “parasite laws” that obliged all able-bodied adults to work. Thus,
while we did not ask specifically about work experience in socialist times, given labor force
participation rates in the Soviet Union in excess of 96% (Gregory, 1987), we presume they worked.
To facilitate comparison between workers in the older and young generations, individuals
born between 1977 and 1981 (n = 1,340) are not included in this analysis. Our resulting
convenience sample includes a total of 3,340 observations. For simplicity, we utilize country name
to refer to participating workers for that country.

participate), or with voluntary response (those who agree to participate may be those with strong
work ethic).
7
The most frequent missing variable was monthly income. However, we found no significant
difference between those who elected to answer this question and those who did not in terms of
age, gender, education or job tenure.
127

3.1

Sample characteristics
Table 3.1 provides summary sample statistics by country and by generation. As seen in

Table 3.1, while there is little generational difference in years of schooling, generational
differences are evident, as expected, for marital status, job tenure, holding a supervisory position,
and earnings.8 Young and older participants were equally likely to be working multiple jobs at the
time the survey was conducted; young generation workers were at least twice as likely to have
reported experience unemployment. At least a third of the participants were employed in stateowned organizations (except for young Azeri participants). In all three countries, manufacturing,
health/education, and retail and other services tend to account for the majority of participating
workplaces; relatively few were employed in finance organizations (except for young Azeri
participants) or in construction / transportation (except for young Russian participants).

3.2

Work Ethic Measure
Empirical studies of work values tend to focus on Protestant work ethic (PWE) – a

commitment to the values of hard work, achievement, thrift, discipline, and self-reliance (Mann,
2010; Meriac et al., 2009). Our work ethic measure, based on Blood (1969), includes eight
components. 9 As seen in Table 3.2, four components are positively worded, with participants
asked to respond using a scale ranging from 1 (strongly disagree) to 5 (strongly agree). Four

8

The relatively high level of education among participants is likely caused by the nature of the
survey project – the level of reading required to complete the questionnaire, and the individual’s
willingness to participate in a “research project.”
9
We were granted permission to use part of a questionnaire originally used in a survey conducted
in 1995 of Russian and Polish retail workers (Huddleston and Good 1999), which was developed
by conducting a pilot study and focus groups. The work ethic measure used here comes from that
questionnaire.
128

Table 3.1: Sample Characteristics
Armenia
Azerbaijan
Young Older
All
Young Older

Russia
All
All Young
Worker Characteristics
Age (at time
40.4
22.4
42.7
33.2
25.1
43.5 40.42 20.9
of interview)
(11.68) (1.73) (10.33) (10.71) (2.86) (7.84) (11.06) (1.5)
15.1
14.6
15.2
13.6
13.5
13.7 14.31 13.45
Years of schooling
(2.09) (1.40) (2.15) (2.78) (2.30) (3.30) (2.84) (2.17)
Job tenure
7.6
2.0
8.3
5.4
2.1
9.5
10.5
1.5
(years at current (7.87) (1.27) (8.06) (7.40) (1.90) (9.46) (9.1) (1.1)
workplace)
Women (%)
52
55
52
39
41
36
71
69
Married (%)
52
15
57
50
27
81
62
14
Supervisor (%)
40
12
43
29
16
47
38
14
Holds multiple
15
12
15
14
14
15
14
10
jobs (%)
Unemployment
27
51
24
45
58
28
17
54
experience (%)
88871 78220 90228
364
342
392
5238 4114
Own earnings
(local currency) (75753) (67394) (76674) (267)
(265)
(267) (4848) (2611)
1.17
1.08
1.18
1.11
0.96
1.30
1.25 0.91
Own to peers'
earnings ratio
(0.96) (0.84) (0.97) (0.70) (0.48) (0.87) (0.94) (0.60)
Workplace Characteristics
State-owned (%)
35
32
35
25
17
36
45
42
Manufacturing (%) 24
13
25
45
50
40
41
33
Education/Health
18
25
17
20
14
27
29
33
Care (%)
Retail and other
31
41
30
16
16
15
19
11
services (%)
Finance (%)
3
3
3
9
12
5
0
0
Public sector (%)
20
13
21
6
5
6
5
9
Construction/
3
3
3
5
3
8
6
15
Transportation (%)
Average Firm
11.1
11.1
11.1
5.7
5.8
5.7
8.5
8.5
Earnings (log)
(0.57) (0.46) (0.57) (0.41) (0.43) (0.36) (0.40) (0.39)
1,222
138
1,084
713
400
313
1405 123
Obs.
Armenia: employee survey conducted in Yerevan (2005) and Shirak region (2008).
Azerbaijan: employee survey conducted in Baku, Sumgait, Shabran, Sabirabad (2011).
Russia: employee survey conducted in Rostov region (2002) Sverdlovsk region
Bashkortostan autonomous republic (2005).

Older
42.3
(9.69)
14.40
(2.88)
11.37
(9.06)
71
67
40
15
14
5346
(4998)
1.28
(0.96)
46
42
29
19
0
5
5
8.4
(0.42)
1,282

(2003),

components are negatively worded, using the same 5-point scale, which we reverse-coded. We
sum the eight components into a single work ethic measure, with a minimum value of eight and

129

maximum value of forty, where higher scores indicate stronger adherence. For informational
purposes, we include the reliability statistic (Cronbach alpha), noting that the relatively low scores
match those found in other PWE studies (Abdalla, 1997; Furnham, 1990).10
We acknowledge that our data, like most data used in other PWE studies, are not derived
from nationally representative samples, so it is difficult to assess whether the low alpha is due to
the internal validity of the PWE measure or due to our sampling. For example, in our survey, only
around 5% of the participating Russian workers strongly agree with the statement that “hard
workers makes one a better person,” but among Armenian and Azeri participants, the number is
around 30%. In contrast, more than 50% of the participating workers in all three countries strongly
agree with the statement that “wasting time is as bad as wasting money.” Although we do not have
nationally representative samples, these differences could reflect cultural differences.11 We focus
on commonalities among participating workers in these culturally and economically diverse
countries, not to gloss over the issue, but rather to take a step toward better understanding the link
between work values and economic environment.
As seen in Table 3.2, there is little difference in the mean PWE score by country, and only
among Armenian participants is there a statistically significant difference in PWE adherence
between the young and older generation workers. Differences are evident in the component parts,
however. For example, young generation workers are more likely than older generation workers
to agree that “hard work makes one a better person” (in Armenia and Russia, although the opposite

10

The Cronbach alpha scores are not any better for either the group of positively-worded
statements or the group of negatively-worded statements.
11
Studies suggest that values associated with Western culture are not always appropriate measures
for work ethic in other cultures (Abdallah, 1997; Ali, 1988, 1992; Niles, 1999)

130

Table 3.2: Work Ethic Components (by Generation)
(1= strongly disagree, 5 = strongly agree)
Armenia

Azerbaijan

Russia

Positively-worded
statements

All

Young

Old

All

Hard work makes
one a better person

3.34
(1.44)

3.54**
(1.39)

3.31
(1.45)

3.64
(1.29)

3.38 3.97*** 1.96 2.51*** 1.90
(1.31) (1.19) (1.24) (1.31) (1.22)

Wasting time is as
bad as wasting
money

4.15
(1.03)

4.20
(1.03)

4.15
(1.03)

4.41
(1.03)

4.49
(0.96)

A good indication of
a person's worth is 3.72
how well his/her job (1.05)
is done

3.72
(1.11)

3.72
(1.05)

3.93
(0.99)

3.74 4.18*** 3.82 3.55 3.85***
(1.00) (0.92) (1.11) (1.10) (1.11)

3.46
(1.15)

3.51
(1.15)

3.46
(1.15)

2.99
(1.47)

3.09**
(1.48)

2.86 3.15 3.20
3.15
(1.44) (1.27) (1.18) (1.28)

4.22
(1.06)

4.17
(1.07)

4.22
(1.06)

3.73
(1.35)

3.83**
(1.34)

3.61 4.18 4.36** 4.16
(1.35) (1.09) (0.92) (1.10)

The principal
purpose of a
3.87
person's job is to
provide a means for (1.10)
enjoying free time

3.62 3.90*** 3.99
(1.07) (1.10) (1.20)

3.92
(1.22)

4.07* 3.89 3.72 3.91**
(1.17) (1.18) (1.20) (1.17)

Whenever possible,
3.62
a person should
relax and accept life (1.22)
as it is

3.68
(1.13)

If all other things are
equal, it is better to
have a job with a lot
of responsibility
Negatively-worded
statements
When the work day
is finished, a person
should forget his/her
job

Young

Old

All Young

Old

4.31 4.23 4.05 4.24**
(1.11) (1.04) (1.19) (1.02)

3.61
(1.23)

2.59
(1.52)

2.46 2.76*** 3.67 3.30 3.71***
(1.45) (1.60) (1.25) (1.33) (1.23)

People who do
3.09
3.09
3.09
things the easy way
are the smart ones (1.29) (1.31) (1.29)
PWE composite
23.87 24.41** 23.80
measure
(3.23) (3.54) (3.18)

3.80
(1.43)

3.70
(1.46)

3.91** 2.74 2.75
2.74
(1.37) (1.36) (1.35) (1.36)

24.87

24.79

24.97 22.67 23.19* 22.62

(4.25)

(4.51)

(3.91) (4.09) (4.18) (4.08)

0.57

0.65

Cronbach alpha

0.56

0.45

0.56

131

0.44

0.45

0.49

0.45

is true among Azeri participants) and “when workday is finished, a person should forget his/her
job and enjoy” (Azerbaijan and Russia). Older generation workers tend to agree that “a good
indication of a person’s worth is how well his/her job is done” (Azerbaijan, Russia) and disagree
that “whenever possible a person should relax and accept life as it is, rather than striving for
unreachable goals” (Azerbaijan, Russia), and “principal purpose of a person’s job is to provide a
means for enjoying free time” (Armenia, Azerbaijan).
While data collected from nationally representative samples would be preferred, these data
indicate that in former socialist economies generational differences in PWE adherence are likely.
We turn to regression analysis to more explicitly examine this proposition.

4.

Methodology
The individual components of our PWE measure are categorical variables (with values

from 1 to 5), and the composite measure takes on values from eight to forty.12 Consequently, we
employ fractional logit regression analysis that accounts for the bounded nature of our dependent
variable (Papke and Wooldridge, 1996). Because the fractional logit model requires that the
dependent variable, PWE, be in the unit interval, we subtract eight and divide by thirty-two. We
report the average marginal (partial) effects, that is, the averages of the predicted effects on
conditional means.
To evaluate whether PWE adherence differs between young and older generation workers
in these three formerly socialist economies, we consider both a basic and an extended specification.
In both the basic and extended specifications, the main variable of interest, young, is a dummy

12

As is common with any composite measure, we treat responses to each component as a cardinal
measure, despite their categorical (ordinal) nature.
132

variable equal to one if the participant was born after 1981. Because middle-aged workers, those
participants born between 1977 and 1981, are excluded in this analysis, we compare PWE
adherence between young and older generations directly. In both specifications we cluster by firm
to take into account the likelihood that a firm’s policies or workplace environment might influence
employee work ethic; that is, there might be within firm correlation.
In the basic specification, we control for a variety of worker characteristics: age, years of
schooling, and workplace tenure; with dummy variables equal to one if married, female, supervisor,
holds multiple jobs, experienced unemployment in previous 5 years. To account for the possibility
that work ethic may be correlated with own performance, we include a measure of self-reported
performance (see Appendix Table F1). To account for the possibility that work ethic may be
correlated with performance of others at the workplace, we include a measure of relative earnings
(log value of the ratio of own earnings to workplace average earnings), which is possible to
construct because we have employer-employee linked data.13
We also control for a variety of workplace characteristics. That is, we include dummy
variables equal to one if the organization is state-owned, and for different sectors: education/health
care, retail and other services, finance, public, construction/transportation (manufacturing is the
omitted sector). As a final workplace control, we include average firm earnings (logarithm),
calculated by using the earnings of all others at the participant’s workplace.

13

The importance of relative (peers’ or comparison) earnings in worker’s assessment of own wellbeing or job satisfaction is well-documented (Clark et al., 2009; Gao and Smyth, 2010; Linz and
Semykina, 2012; Senik, 2004), suggesting that relative earnings may also play a role in own
performance or work ethic.

133

In the extended specification, we include additional worker and workplace controls. It may
be, for example, that work ethic is correlated with such individual characteristics as personality,
so we control for personality using locus of control (see Appendix Table F2). 14 Moreover,
individual work ethic may be influenced by workplace reward practices. To control for this, we
include variables which capture worker expectations about receiving particular rewards if they do
their job well (see Appendix Table F3).
To offset concerns about unobservable workplace characteristics that might affect work
ethic, and given that we have employer-employee matched data, we also considered firm fixed
effects. For example, if firms with higher PWE are more attractive to young generation workers,
including firm fixed effects can reduce bias from such (unobservable) sorting process. Employing
fixed effects in the extended specification allows us to check the robustness of our results.

5.

Results and Discussion
Our fractional logit regression results are reported in Table 3.3. For each country, the

average partial effects associated with the basic model are reported in Columns (1), the extended
model in Columns (2), and the fixed effects model in Columns (3). Because PWE is restricted to
the unit interval, we scale all estimates by 100 and the estimates can be interpreted as differences
in percentage points.
As seen in the Table 3.3, in all three countries, regardless of specification, young generation
workers adhere more strongly to PWE than older generation workers. That is, young generation

14

Locus of control (Rotter, 1966) is perhaps the most frequently used single personality trait,
especially among economists. Studies indicate a positive correlation between internal LOC and
worker performance (Andrisani, 1977; Coleman and DeLeire, 2003; Semykina and Linz, 2007).
134

workers score 2–4 percentage points higher on the PWE measure than older generation workers, a
result that is statistically significant in all cases but one. Although the absolute magnitude is
moderate, the relative magnitude is quite large in comparison to other control variables. For
example, the estimates on young are five to ten times larger than the estimates on education or
personality. Moreover, because our estimates are based on cross-sectional data, they are likely
lower bound estimates of the true effect of the change in economic environment on PWE adherence.
If, as these results suggest, PWE is shaped by economic environment, PWE adherence among
older generation workers would also be influenced by the economic transformation. Thus, PWE
adherence among older generation workers at the time the survey was conducted is likely stronger
than what would have occurred in the socialist period; older generation workers participating in
this study had been living and working in the new socio-economic environment for nearly two
decades. While longitudinal data would certainly be required to assess changes in PWE adherence
among particular workers as the economic transition proceeds, the richness of these cross-sectional
data allows us to take a preliminary step toward documenting the link between work ethic and
economic environment.
We note that the estimate magnitudes for young are quite similar across countries,
especially when we control for personality and workplace reward practices. In particular, in the
baseline specification, Columns (1), the estimated coefficient for young in Azerbaijan is greater
than in Armenia and Russia. However, in Columns (2), where we control for personality and
workplace reward practices, the estimates become more similar across these three countries. This
suggests that young generation workers in Armenia and Russia are similar to older generation
workers in terms of personality, but that among Azeri participants, personality differences across
generations can explain part of the difference in PWE adherence.

135

Table 3.3: Determinants of Work Ethic
(1)
(2)
(3)
(1)
(2)
(3)
(1)
(2)
(3)
Armenia
Azerbaijan
Russia
2.88** 2.30** 2.29* 4.06** 2.66
2.00 4.39*** 3.70*** 2.98**
Young
(1.14)
(1.10) (1.21) (1.80)
(1.75)
(1.74) (1.46) (1.38) (1.44)
0.05
0.06*
0.05
0.08
0.09
0.11
0.06
0.09** 0.06
Age
(0.03)
(0.03) (0.04) (0.09)
(0.09)
(0.08) (0.04) (0.04) (0.04)
Years of
0.43*** 0.35** 0.34* 0.33
0.14
0.19
0.08
-0.06 -0.00
schooling (0.16)
(0.16) (0.19) (0.24)
(0.15)
(0.15) (0.12) (0.13) (0.13)
-2.03*** -2.00*** -0.67
0.80
2.71*** 3.45*** -0.31 0.03
-0.40
Female
(0.53)
(0.56) (0.59) (1.14)
(1.01)
(1.07) (0.68) (0.73) (0.74)
1.10*
0.89
0.40
0.43
-0.31
-0.41 0.91
0.28
0.75
Married
(0.59)
(0.59) (0.64) (1.00)
(0.97)
(1.10) (0.73) (0.74) (0.75)
-0.78
-0.88
-0.62
1.51
1.85
1.49
2.37*** 1.78** 1.94**
Supervisor
(0.65)
(0.64) (0.73) (1.30)
(1.16)
(1.04) (0.86) (0.90) (0.86)
Workplace -0.09
-0.06
-0.04 -0.09
-0.13
-0.09 0.11** 0.11** 0.10**
tenure
(0.06)
(0.06) (0.05) (0.10)
(0.11)
(0.11) (0.05) (0.05) (0.05)
Self-reported 0.17
0.06
-0.03 0.98*** 0.44
0.28
0.63** 0.48
0.51*
performance (0.16)
(0.16) (0.16) (0.36)
(0.29)
(0.31) (0.26) (0.29) (0.28)
Unemploy. 0.78
0.91
1.11
-4.40*** -2.84*** -2.00** 1.82
1.80
1.29
exp.
(0.77)
(0.72) (0.76) (1.21)
(1.00)
(0.98) (1.19) (1.27) (1.30)
-0.61
-0.36
0.57
-1.56
-0.57
-0.94 1.73
1.33
1.16
Multiple jobs
(0.92)
(0.92) (0.95) (1.07)
(1.00)
(0.96) (1.07) (1.10) (1.07)
Relative
1.48*** 1.18** 1.52** 0.85
-0.21
-0.05 1.19* 0.83
0.86
earnings
(0.57)
(0.56) (0.61) (1.33)
(1.24)
(1.16) (0.61) (0.65) (0.72)
Locus of
0.35*** 0.42***
0.38*** 0.38**
0.44*** 0.38***
control
(0.08) (0.08)
(0.15)
(0.16)
(0.08) (0.07)
Firm
Yes
Yes
No
Yes
Yes
No
Yes
Yes
No
characteristics
Expected
No
Yes
Yes
No
Yes
Yes
No
Yes
Yes
rewards
Firm
No
No
Yes
No
No
Yes
No
No
Yes
fixed effects
Obs.
1,222
1,222
1,222 713
713
713
1,405 1,405 1,405
Standard errors adjusted for clustering at firm level are in parentheses under coefficient estimates.
*** p<0.01, ** p<0.05, * p<0.1
In terms of our initial question about the link between work values and economic
environment, these results suggest that Marx was “right.” First, estimates for young are positive
and significant in all three countries. Second, the estimate magnitudes are nearly identical in all

136

three countries; countries that exhibit substantial economic and cultural diversity. What is common
among these three countries is the transformation of the economic environment from socialist to
capitalist. Thus we view both the sign and similar magnitude of the estimated coefficient to lend
support to Marx’s argument that values are shaped by economic environment.
Our second objective was to document who among the participating workers in these three
transition economies adheres to PWE. Generally, as seen in Table 3.3, the estimates on age are
small and there is no significant effect on PWE after we control for generation. Education seems
to have a positive effect on PWE adherence, although it is only significant among Armenian
participants. Significant country differences are evident in terms of gender: among Armenian
participants, women adhere to PWE less strongly than men; among Azeri participants, the opposite
is true. The estimates on personality are positive and highly significant, and the magnitudes are
similar across all countries. That is, Armenian, Azeri, and Russian participants with an internal
locus of control (those who believe they can influence their outcomes) adhere more strongly to
PWE. Self-reported performance and holding a supervisory position tend to be positively
correlated with PWE. Similarly, PWE adherence tends to be stronger among participants with high
relative earnings. Finally, PWE adherence tends to be stronger among participants employed in
organizations that reward hard work with the chance to develop new skills / learn new things.
When we include firm fixed effects, in Columns (3), the estimates of young in Armenia
and Russia remain roughly the same, while the estimate in Azerbaijan becomes smaller and
insignificant. Although unobservable firm characteristics, such as job sorting or workplace
matching, may be an important contributing factor, it could also be that sampling restrictions are

137

driving the result.15 Access to a representative sample of workers for the participating firms would
facilitate resolving this issue.

5.1

Individual measures of work ethic
We noted earlier that the reliability statistic (Cronbach alpha) for the composite PWE

measure was rather low, although comparable to other studies. However, because of this, as well
as to more systematically explore whether generational differences in work ethic are significant,
we reproduce the regression analysis using the eight individual PWE components as dependent
variables. Because they are categorical variables, ranging from 1 (strongly disagree) to 5 (strongly
agree), we use ordered probit regression analysis for each PWE component, controlling for the
same worker and workplace characteristics as in the extended specification (Table 3.3, Columns
(2)). We repeat the exercise using firm fixed effects.
In Table 3.4 we report the estimates of the generation indicator variable (young) by country.
In each country, the upper rows are based on specification with firm characteristics (as Columns
(2) in Table 3.3), and the lower rows are based on specification with firm fixed effects (as Columns
(3) in Table 3.3). These estimates reflect the marginal effect on the probability that a worker
strongly agrees with the particular statement. The results in Table 3.4 are similar to the simple
mean differences presented in Table 3.2. Young generation workers prefer jobs with more
responsibilities, and tend not to see their job simply as a way to make money so that they can better
enjoy their leisure time. Older generation workers who were trained and had work experience in

15

We note that the average number of workers in a firm is 21 in Armenia (182 firms), 30 in
Azerbaijan (62 firms) and 39 in Russia (87 firms).
138

Table 3.4: Generation and Individual Work Ethic Components (by Country)
Better
Person

Not
Waste
Time

Worth

Responsibility

Forget
Job

Enjoy
Life

Relax

Easy
Way

Armenia
2.2
2.3
-0.7
-2.9
-8.8*
-13.9***
1.1
-4.8
(4.1)
(4.9)
(3.9)
(3.4)
(5.1)
(4.2)
(4.1)
(3.0)
3.6
-2.8
1.1
-1.8
-8.1
-8.6*
1.0
-5.5*
(4.5)
(5.0)
(4.5)
(4.0)
(5.0)
(4.4)
(4.6)
(3.3)
Azerbaijan
-3.1
3.7
-9.9
8.8**
4.2
-19.7*** -0.6
-6.2
(5.4)
(5.9)
(6.1)
(4.3)
(6.2)
(6.1)
(4.0)
(6.8)
-1.1
4.9
-11.3*
6.9
9.3*
-12.3**
-5.7
-3.4
(5.2)
(5.5)
(6.8)
(5.3)
(5.5)
(5.1)
(4.2)
(6.7)
Russia
3.9**
8.7
0.3
9.3***
1.3
-0.5
-3.3
1.5
(1.6)
(5.6)
(4.6)
(2.9)
(5.7)
(5.3)
(4.4)
(2.7)
4.0**
6.5
-0.9
6.5**
1.5
-1.1
-2.9
2.1
(1.8)
(5.6)
(4.8)
(2.9)
(5.7)
(5.3)
(4.8)
(2.9)
For complete description of each work ethic variable, see Table 3.2. In each panel, the first rows
include workplace controls, and the second rows include firm fixed effects. All estimates are based
on the extended specification that include locus of control and expected rewards. *** p<0.01, **
p<0.05, * p<0.1

Soviet economy appear more likely to think smart people are those who take the easy way at work,
and who forget the job when they leave the workplace. Given the limited worker responsibility
and pervasive shirking in Soviet times (Gregory, 1987), these results seem to be consistent with
Marx’s proposition that regime change affects work values.

5.2

Generation versus age
To clarify the link between work ethic, generation, and age among participating workers

in these three transition economies, we repeat our regression analysis including additional age
categories. That is, we include two dummy variables: one for individuals born in the 1960s and
one for individuals born in 1970s. The omitted group is the cohort born before 1960 (this includes

139

only 51 participants who were born before 1940). We control for the same worker characteristics
as reported in the extended model in Table 3.3; conducting the analysis, first, using firm
characteristics and, second, using firm fixed effects. This strategy provides additional insight
considering the fact that our generation cutoff is relatively conservative: participants born between
1970 and 1976 (15–21 years old in 1991) had little time to acquire work experience in the Soviet
regime.
Table 3.5: Generation versus Age (by Country)
Armenia
Azerbaijan
Firm
Controls

Firm
FE

Firm
Controls

Firm
FE

Russia

Firm
Controls

Firm
FE

Young

5.64**
(2.58)

6.95**
(2.70)

1.73
(4.68)

2.89
(4.05)

8.59***
(2.72)

8.19***
(2.77)

Old70

2.73
(1.80)

4.00**
(1.87)

-0.718
(3.36)

0.608
(2.76)

3.98**
(1.80)

4.03**
(1.78)

Old60

1.34
(1.06)

1.44
(1.14)

-0.08
(2.68)

1.16
(2.33)

2.10*
(1.22)

2.72**
(1.15)

Age

0.15**
(0.08)

0.20**
(0.08)

0.06
(0.15)

0.12
(0.12)

Obs.

1,222

713

0.24***
0.22**
(0.08)
(0.08)
1,405

All worker and workplace controls from extended specification included. *** p<0.01, ** p<0.05,
* p<0.1

As seen in Table 3.5, on average, young generation workers in all three countries exhibit
the highest PWE.16 Among Russian participants, the estimates show strongly that PWE adherence
is decreasing with older age groups. This pattern is similar in Armenia, but not in Azerbaijan.
Therefore, at least among Russian participants, PWE adherence among relatively older workers

16

While the table reports the results in comparison to omitted group, we also tested to see if the
difference between the estimated coefficients on Young and Old70 was statistically significant.
Wald test results indicate the affirmative within the 5% confidence interval.
140

appears to respond to the process of economic transition. The stronger effect in Russia is consistent
with the fact that Russia pursued a more rapid transformation program than Armenia or Azerbaijan.
While we acknowledge that the pattern of decreasing PWE with age may be driven by unobserved
cohort characteristics rather than economic regime change, we note that the pattern emerging
among the participating employees in these three transition economies is the opposite of findings
associated with PWE adherence among workers in developed market economies (Twenge, 2010).
We also note that the results presented in Table 3.5 suggest that our estimates in Table 3.3 may
under-estimate the effect of regime change on PWE adherence. Indeed, the estimate for age is
actually positive and often significant in these regressions which implies that, within each age
group, older people have higher PWE.

6.

Conclusion
We use data collected from employees in over 340 workplaces in Armenia, Azerbaijan and

Russia to investigate whether individuals who were trained and worked in the former Soviet
economy (older generation workers) adhere more or less strongly to PWE than individuals with
no work experience in the former socialist economy (young generation workers). We find PWE
adherence most strong among young generation workers, with the estimate magnitudes very
similar across these three culturally and economically diverse countries.
Our results provide robust evidence for generational differences in PWE adherence among
workers in formerly socialist economies, differences which cannot be explained either by worker
personality and or by various firm characteristics. Moreover, the result that young generation
workers adhere more strongly to PWE than older workers stands in contrast to generational
differences reported in studies conducted in developed market economies where older workers

141

adhere more strongly (Smola and Sutton 2002, Twenge 2010). Thus, we view these results as
evidence in support of the link between work values and economic environment described by Marx,
noting that future research, using repeated cross sectional data or panel data to directly estimate
the potential dynamic response of PWE to economic transition, is certainly warranted.

142

APPENDIX

143

Table F1: Sample Means and Standard Deviations for Self-Reported Performance
Armenia
Azerbaijan
Russia
Older
Young
Older
Young
Older
Compared to others doing the Young
same work
3.46
3.60
3.18
3.71
3.33
3.43
The quality and quantity of
(0.69)
(0.73)
(1.02)
(0.77)
(0.63)
(0.60)
my work is
3.50
3.62
3.33
3.77
3.40
3.46
My productivity is
(0.68)
(0.72)
(0.96)
(0.79)
(0.66)
(0.61)
3.49
3.64
3.80
3.87
3.32
3.44
My ability to anticipate /
(0.78)
(0.74)
(0.60)
(0.78)
(0.67)
(0.61)
prevent problems is
Performance
10.45
10.86
10.30
11.35
10.05
10.29
(1.84)
(1.91)
(1.97)
(1.88)
(1.61)
(1.55)
138
1084
400
313
123
1282
Obs.
Mean derived using 5 point scale where: 1 = much worse than others; 5 = much better than
others.

Table F2: Sample Means and Standard Deviations for Personality Trait and Components
Armenia
Azerbaijan
Russia
Personality trait
components
young
older
young
older
young
older
Internal LOC
3.37
3.55
2.70
3.44
3.11
3.32
Success comes from
(1.11)
(1.18)
(1.41)
(1.39)
(1.12)
(1.23)
hard work, not luck
3.98
3.80
3.08
3.36
3.77
3.63
People get respect
(1.11)
(1.10)
(1.33)
(1.33)
(1.22)
(1.26)
deserved
I can make my plans
4.09
3.75
3.26
3.51
3.60
3.47
work
(0.84)
(1.00)
(0.98)
(0.99)
(1.03)
(1.09)
3.74
3.55
3.69
3.79
3.72
3.49
I control what happens
(1.06)
(1.12)
(1.27)
(1.28)
(1.13)
(1.18)
to me
3.33
3.43
2.64
3.31
3.31
3.45
Getting what I want has
(1.02)
(1.07)
(1.25)
(1.39)
(1.11)
(1.17)
little to do with luck
External LOC
3.69
3.73
3.82
3.77
3.34
3.70
Without right breaks,
(1.10)
(1.08)
(1.25)
(1.31)
(1.12)
(1.20)
cannot be good leader
3.26
3.34
3.55
3.42
2.82
3.02
Unhappy outcomes
(1.06)
(1.10)
(1.30)
(1.32)
(1.07)
(1.21)
caused by bad luck
3.65
3.53
3.53
3.44
3.20
3.61
Promotions depend on
(1.23)
(1.13)
(1.26)
(1.30)
(1.25)
(1.24)
luck
3.47
3.48
3.29
3.39
3.15
3.45
Life is controlled by
(1.08)
(1.02)
(1.13)
(1.15)
(1.06)
(1.17)
accidents
2.98
3.26
3.18
3.31
2.64
3.05
I have no influence over
(1.08)
(1.10)
(1.35)
(1.14)
(1.22)
things that happen to me (1.02)

144

Table F2 (cont’d)
1.46
0.75
-2.00
0.08
2.37
0.52
Locus of Control(LOC)
(4.48)
(6.43)
(6.06)
(5.37)
(4.37)
(5.65)
138
1084
400
313
123
1282
Obs.
Scale: 1 = strongly disagree; 5 = strongly agree
LOC is constructed by summing the first 5 components (internal LOC), and the second five
components (external LOC), and then using the formula: internal - external.

Table F3: Sample Means and Standard Deviations for Expected Rewards
How likely is it that each
Armenia
Azerbaijan
Russia
of these things listed
Young Older
Young Older
Young Older
below would happen if
you perform your job
especially well?
3.36
2.98
3.02
3.21
Receive bonus/pay
3.15 2.84
increase
(1.31) (1.37)
(1.20) (1.24)
(1.41) (1.46)
3.75
3.51
3.06
3.50
3.72 3.64
Have better job security
(1.22) (1.18)
(1.31) (1.28)
(1.25) (1.26)
3.45
3.14
3.31
3.19
Be promoted/get better
3.02 2.54
job
(1.32) (1.26)
(1.17) (1.19)
(1.49) (1.34)
3.89
3.30
3.69
3.81
3.37 3.22
Supervisor will praise
(1.06) (1.26)
(1.21) (1.33)
(1.28) (1.30)
4.06
3.56
3.92
4.03
Respected by co3.82 3.83
workers
(0.98) (1.19)
(1.01) (0.97)
(1.15) (1.09)
3.91
3.44
3.22
3.32
Chance to learn new
3.59 3.38
things
(1.06) (1.21)
(1.37) (1.29)
(1.28) (1.32)
3.85
3.45
3.00
3.13
Accomplish something
3.53 3.55
worthwhile
(1.16) (1.15)
(1.47) (1.35)
(1.26) (1.30)
3.75
3.32
3.44
3.29
Freedom at work, on the
2.89 2.89
job
(1.00) (1.19)
(1.38) (1.39)
(1.35) (1.33)
Friendly co-workers
Job makes me feel good
about myself
Obs.

3.77
(1.26)
4.14
(0.95)
138

3.54
(1.20)
3.59
(1.15)
1084

145

3.79
(1.15)
4.04
(0.95)
400

4.02
(1.03)
4.17
(0.97)
313

3.87
(1.12)
3.95
(1.02)
123

3.87
(1.08)
3.91
(1.12)
1282

REFERENCES

146

REFERENCES

Andrisani, P., 1977. Internal-external attitudes, personal initiative, and the labor market experience
of black and white men. Journal of Human Resources 12 (3), 308-328.
Abdalla, I., 1997. Construct and concurrent validity of three Protestant work ethic measures in an
Arabian Gulf society. Journal of Managerial Psychology 12 (4), 251-260.
Ali, A., 1988. Scaling an Islamic work ethic. Journal of Social Psychology 128 (5), 575-583.
Ali, A., 1992. The Islamic work ethic in Arabia. Journal of Psychology 132 (5), 507-519.
Anthony, P., 1977. The ideology of work. Tavistock: London.
Arrunada, B., 2010. Protestants and Catholics: similar work ethic, different social ethic. Economics
Journal 120 (547), 890-918.
Arslan, M., 2000. A cross cultural comparison of British and Turkish managers in terms of
protestant work ethic characteristics. Business Ethics: A European Review 9 (1), 13-19.
Arslan, M., 2001. The work ethic values of protestant British, Catholic Irish, and Muslim Turkish
managers. Journal of Business Ethics 31, 321-339.
Aygun, Z., Arslan, M. and Guney, S., 2008. Work values of Turkish and American university
students. Journal of Business Ethics 80, 205-223.
Becker, S., Woessman, L., 2009. Was weber wrong? a human capital theory of protestant economic
history. Quarterly Journal of Economics 124 (2), 531-596.
Becker, S. Woessman, L., 2010. The effect of Protestantism on education before the
industrialization: evidence from 1816 Prussia. Economics Letters 107, 224-228.
Birnbaum, N., 1953. Conflicting interpretations of the rise of capitalism: Marx and Weber. British
Journal of Sociology 4 (2), 125-141.
Blood, M.R., 1969. Work values and job satisfaction. Journal of Applied Psychology 53, 456-459.
Cavalcanti, R., Parente, S., Zhao, R., 2007. Religion in macroeconomics: a quantitative analysis of
Weber’s thesis. Economic Theory 32, 105-123.
Caspi, A., 2000. The child is father of the man: personality continuities from childhood to
adulthood. Journal of Personality and Social Psychology 78, 158-172.
Cennamo, L., Gardner, D., 2008. Generational differences in work values, outcomes, and personorganization values fit. Journal of Management Psychology 23 (3), 891-906.
Chapman, J., 1963. Real wages in the Soviet Union since 1928. Harvard University Press:
Cambridge.

147

CIA World Factbook, 2010. https://www.cia.gov/library/publications/the-world-factbook/
(Accessed 5.12.2012)
Clark, A., Kristensen, N., Westergard-Nielsen, N., 2009. Job satisfaction and co-worker wages:
status or signal? Economic Journal 119 (536), 430-447.
Coleman, M., DeLeire, T., 2003. An economic model of locus of control and the human capital
investment decision. Journal of Human Resources 38 (3), 701-721.
Costa, P., McCrae, R., 1997. Longitudinal stability of adult personality. In: Hogan, R. et. al.,
Handbook of personality psychology. Academic Press: San Diego.
Delacroix, J., Nielson, F., 2001. The beloved myth: Protestantism and the rise of industrial
capitalism in nineteenth century Europe. Social Forces 80 (2), 509-553.
Denisova, I., Eller, M., Zhuravskaya, E., 2007. What Russians think about transition: evidence
from RLMS survey. CEFIR Working Paper No. 114.
Doepke, M., Zilibotti, F., 2008. Occupational choice and the spirit of capitalism. Quarterly
Journal of Economics 123 (2), 747-793.
EBRD, 2007. Russian attitudes and aspirations: results of focus groups in nine Russian cities.
http://www.ebrd.com/downloads/research/surveys/asp.pdf.
Feher, F., 1982. Paternalism as a mode of legitimation in Soviet-type societies. In: T.H. Righby,
T.H., Feher, F. (eds), Political legitimation in communist states. Macmillan: New York, 6481.
Furnham, A., 1990. A content, correlational, and factor analytic study of seven questionnaire
measures of the Protestant work ethic. Human Relations 43, 383-399.
Furnham, A., Muhiudeen, C., 1984. The Protestant work ethic in Britain and Malaysia. Journal
of Social Psychology 122, 157-161.
Furnham, A., Rajamanickham, R., 1992. The Protestant work ethic and just world beliefs in
Great Britain and India. International Journal of Psychology 27, 401-416.
Giorgi, L., Marsh, C., 1990. The Protestant work ethic as a cultural phenomenon. European
Journal of Social Psychology 20 (6), 499-517.
Gao, W., Smyth, R., 2010. Job satisfaction and relative income in economic transition: status or
signal? the case of urban China. China Economic Review 21 (3), 442-485.
Granick, D., 1987. Job rights in the Soviet Union: their consequences. Cambridge University
Press: New York.
Gregory, P., 1987. Productivity, slack and time theft in the Soviet economy: evidence from the
soviet interview project. In: Millar (eds), Politics, work and daily life: a survey of former
Soviet citizens. Cambridge University Press: New York.

148

Gregory, P., Collier, I., 1988. Unemployment in the Soviet Union: evidence from the Soviet
interview project. American Economic Review 78, (4), 613-632.
Gursoy, D, Maier, T., Chi, C., 2008. Generational differences: an examination of work values and
generational gaps in the hospitality workforce. International Journal of Hospitality
Management 27, 448-458.
Halman, L., 1996. Individualism and the fragmentation of work values: evidence from the
European values survey. Work and Organization Research Center paper 96.07.013, Tilburg
University.
Hattrup, K., Mueller, K., Joens, I., 2007. The effects of nations and organizations on work value
importance: a cross-cultural investigation. Applied Psychology 56 (3), 479-499.
Hill, R., 1996. Historical context of the work ethic, history of work ethic. University of Georgia
http://www.coe.uga.edu/~rhill/workethic/hist/htm.
Hofstede, G., 1980. Culture’s consequences: international differences in work-related values.
Sage: Beverly Hill.
Huddleston, P., Good, L., 1999. Job motivators in Russian and Polish retail firms. International
Journal of Retail & Distribution Management 27 (9), 383-392.
Johassen, C., 1947. The Protestant ethic and the spirit of capitalism in Norway. American
Sociological Review 12, 676-686.
Jones, H. B., 1997. The Protestant work ethic: Weber’s model and the empirical literature.
Human Relations 50, 757-778.
Jones, A., Moskoff, W., 1991. KO-OPS: the rebirth of entrepreneurship in the Soviet Union.
Indiana University Press: Bloomington.
Judge, T., Thoresen, C., Bono, J., Patton, G., 2001. The job satisfaction-job performance
relationship: a qualitative and quantitative review. Psychology Bulletin 127 (3), 376-407.
Keepnews, D., Brewer, C., Kovner, C., Shin, J.H., 2010. Generational differences among newly
licensed registered nurses. Nursing Outlook 58 (3), 155-164.
Kornai, J., 1980. The economics of shortage. North-Holland Publishers: New York.
Lane, D., 1986. Labour and employment in the USSR. New York University Press: New York.
Linz, S.J., 1995. Russian labor market in transition. Economic Development and Cultural
Change 43 (4), 693-716.
Linz, S. J., Semykina, A., 2012. What makes workers happy? anticipated rewards and job
satisfaction. Industrial Relations 54 (4), 811-844.
Macky, K., Gardner, D., Forsyth, S., 2008. Generational differences at work: introduction and
overview. Journal of Managerial Psychology 23 (8), 857-861.

149

Mann, M., 2010. A quantitative and qualitative analysis identifying antecedents of work ethic
beliefs and the relationship between work ethic beliefs and in-role and extra-role work
behavior: new work ethic dimensions and scale introduced. PhD Dissertation, School of
Business, University of Albany, State University of New York.
Marx, K., 1867(2010). Capital: a critique of political economy. Progress Publishers: Moscow.
http://www.marxists.org/archive/marx/works/download/pdf/Capital-Volume-I.pdf.
Meriac, J., Poling, T., Woehr, D., 2009. Are there gender differences in work ethic? an
examination of the measurement equivalence of the multidimensional work ethic profile.
Personality and Individual Differences 47, 209-213.
Millar, J. R., 1987. Politics, work and daily life: a survey of former soviet citizens. Cambridge
University Press: New York.
Mommsen, W., 1977. Max Weber as a critic of Marxism. Canadian Journal of Sociology 2 (4),
373-398.
Moskoff, W., 1984. Labor and leisure in the Soviet economy. Macmillan Press: London.
Mueller, G. H., 1986. Weber and Mommsen: non-Marxist materialism. British Journal of
Sociology 37 (1), 1-20.
NationMaster, 2012. http://www.nationmaster.com/index.php (Accessed 6. 11. 2012.)
Nave, C., Sherman, R., Funder, D., Hampson, S., Goldberg, L., 2010. On the contextual
independence of personality: teachers’ assessments predict directly observed behavior after
four decades. Social Psychology and Personality Science 1 (4), 327-334.
Niles, F. S., 1999. Toward a cross-cultural understand of work-related beliefs. Human Relations
52 (7), 855-867.
OECD Country Statistical Profiles. http://www.oecd-ilibrary.org/economics/country-statisticalprofiles-key-tables-fromoecd_20752288 (Accessed 6. 11. 2012.)
Papke, L., Wooldridge, J., 1996. Econometric methods for fractional response variables with an
application to 401(k) plan participation rates. Journal of Applied Econometrics 11 (6), 619632.
Pew Global Attitudes Project, 2011. Confidence in democracy and capitalism wanes in former
Soviet Union. http://www.pewglobal.org/files/2011/12/Pew-Global-AttitudesFormer-SovietUnion-Report-FINAL-December-5-2011.pdf
Rotter, J., 1966. Generalized expectancies for internal versus external control of reinforcement.
Psychological Monographs 80 (1), 1-28.
Schwartz, S., Bardi, A., 1997. Influences of adaptation to communist rule on value priorities in
Eastern Europe. Political Psychology 18 (2), 385-410.
Semykina, A., Linz, S. J., 2007. Gender differences in personality and earnings: evidence from
Russia. Journal of Economic Psychology 28, 387-410.
150

Senik, C., 2004. When information dominates comparison: a panel data analysis using Russian
subjective data. Journal of Public Economics 88, 2099-2133.
Smola, K., Sutton, C., 2002. Generational differences: revisiting generational work values for the
new millennium. Journal of Organizational Behavior 23, 363-382.
Standing, G., 1996. Russian unemployment and enterprise restructuring: reviving dead souls.
Macmillan Press: London.
Torgler, B., 2003. Does culture matter? tax morale in an East-West-German comparison.
FinanzArchiv/Public Finance Analysis 59 (4), 504-528.
Torgler, B., 2007. Tax compliance and tax morale: a theoretical and empirical analysis. Edward
Elgar: Cheltenham UK.
Torgler, B., 2011. Work values in Western and Eastern Europe. FEEM Working Paper No.
94.2011
Twenge, J., 2010. A review of the empirical evidence on generational differences in work
attitudes. Journal of Business and Psychology 25 (2), 201-210.
van den Broek, A., de Moor, R., 1994. Eastern Europe after 1989. In: Ester, P., Halman, L., de
Moor, R. (eds), The individualizing society: value change in Europe and North America.
Tilburg University Press: Tilburg, 197-228.
Weber, M., 1958. The Protestant ethic and the spirit of capitalism (translated by Talcott Parsons).
Charles Scribner’s Sons: New York. (Originally published as 2 separate essays, 1904-1905).
World Bank Data (by country). http://data.worldbank.org/country (Accessed 6. 11. 2012.)
World Values Survey – Findings. http://www.worldvaluessurvey.org/index_findings. (Accessed
6. 11. 2012.)

151