Professional Documents
Culture Documents
Daniel B. Firoozi
Undergraduate Student, UC San Diego
Revised: March 2017
ABSTRACT
This paper uses panel data and a difference-in-differences model to analyze local economic
conditions and the racial composition of US counties and states in relation to variation in age-
adjusted drug overdose mortality rates per 100,000 residents. Geographic and year fixed
effects are included, and the model is structured so as to prioritize analysis of variation over
time, rather than variation between panels. Drug overdose mortality rates are demonstrated to
be countercyclical relative to median household income, the unemployment rate, and the
poverty rate at the county-level. A rise in the share of African Americans, Hispanics, and Asian
Americans as a percentage of the population at the county-level corresponds with fewer fatal
overdoses. At the state-level, a 1 percentage point increase in the unemployment rate is
associated with 0.2 more annual overdose deaths per 100,000 residents, and a 1 percentage
point increase in the health insurance coverage rate from the ACA corresponds to a 3.9%
decrease in the per capita rate of overdose deaths. Likewise, a weak instrument legalization of
medical marijuana at the state-level is associated with an increase in annual non-marijuana
drug overdose fatalities, but the relationship is not casual. Evidence on “Good Samaritan” laws
and naloxone access laws are mixed, with the former appearing in most specifications to be
negatively associated with overdoses. The relationship between population growth of a county
and its age-adjusted drug overdose fatality rate evolves over the course of the period of the
Great Recession of 2007-2009 from significantly positive to significantly negative.
Acknowledgements
Special thanks to Professors Kate Antonovics, Julian Betts, and Ross Starr for ideas, advice, and
guidance while working on this paper.
I. Introduction
Since the turn of the 21st Century, the age-adjusted drug mortality rate has more than
doubled in the United States, including a tripling of deaths from opioid-related overdoses (Rudd,
Aleshire, Zibbell, & Gladden, 2016). Nearly half a million Americans died from drug overdoses
between 2000 and 2014, including 47,000 in 2014 alone1. Lethal drug overdoses have now
overtaken vehicle crashes as a leading cause of death and kill four times as many Americans
every year as homicides1 (FBI, 2014). Opioids, a broad classification of pain-relief drugs
including heroin, fentanyl, morphine, hydrocodone, and oxycodone, are responsible for upwards
of 60% of overdose mortality, with cocaine and benzodiazepines responsible for a large share of
the remaining deaths1. A recent paper by Christopher Ruhm (2016) suggests that several factors
certificates. Moreover, the spike in mortality rates from opioids is tied strongly to rates of abuse
and dependence on the same drugs, suggesting that the overdose deaths are acting as a proxy of a
broader growth in drug consumption (Jones, Logan, Gladden, & Bohm, 2015). The balance of
evidence confirms that a disproportionate share of Americans living with drug abuse disorders,
particularly with respect to heroin and cocaine, were previously prescribed opioids, but only a
small fraction of the total population that uses pain relievers goes on to initiate heroin use2
(Cicero, Ellis, Surratt, & Kurtz, 2014; Muhuri, Gfroerer, & Davies, 2013). The demographic
breakdown of the crisis has undergone significant shifts as well; drug use is trending toward rural
areas and 90% of new unprescribed opioid use observed among white Americans 3.
2
Despite the significance of this topic to public health policymakers, the recency of
climbing rates of drug mortality has resulted in a dearth of both data and research, particularly in
the interaction between economic conditions and the use of drugs like opioids, cocaine, and
employment, and poverty on drug consumption is critical to developing policies that will put
downward pressure on drug addiction and mortality in the long run. Because the literature is
understandably thin, it is first helpful to analyze research on the impact of the business cycle on
three related subjects: (1) suicide, (2) alcohol consumption, and (3) tobacco consumption.
overdoses, a substantial share stems from suicide (CDC, 2015). Similarly, the psychological
pathways that posit drug use as a reaction to economic shocks would likely be related to the
mechanisms linking economic conditions with suicide. There is a substantial degree of consensus
between several papers testing suicide rates against unemployment data, which argue that
increases in long-term unemployment and sudden mass-layoffs can increase the rate of suicide
(Ruhm, 2000; Classen & Dunn, 2012; Snipes, Cunha, & Hemley, 2012). When decomposing
these findings by gender and region, the results diverge somewhat by demographic subgroup.
While Yang (1992) demonstrates that the pattern between unemployment and suicide only holds
for white males in the raw data, her controls for sociological conditions yield a result that
parallels the work of Classen and Dunn (2012), suggesting that American males as well as
females experience an increase in the probability of suicide resulting from an increase in state-
level unemployment. Snipes, Cunha, and Hemley (2012) point to variations in the suicide rates’
3
posit intuitive reasoning for the differences and virtually no parallel work to either confirm or
Drug abuse may also share properties with alcohol use, particularly in the extent to which
consumption of both may be subject to swings in consumers’ disposable income and each
product may be used to relieve stress. There is a broadly accepted pattern of pro-cyclical alcohol
consumption in OECD countries reflected in the data for liver disease mortality (Ruhm, 2000;
Ruhm, 2006), the frequency of drinking (Dee, 2001; Ruhm & Black, 2002), self-reported drunk
driving (Ruhm & Black, 2002), and household demand for packaged alcohol (Cotti, Dunn, &
Tefft, 2014). Yet, when the literature narrows in focus to instances of binge drinking, gauged
through survey data, studies demonstrate that the phenomenon is countercyclical with respect to
African American males (Bor, Basu, Coutts, McKee, & Stuckler, 2013; Davalos, Fang, &
French, 2012; Pacula, 2011). There are a few counterexamples, like a paper testing employment
rates against binge drinking by Tekin, McClellan, & Minyard (2013), which find statistically
insignificant results or point estimates too low to be of value, but they tend to represent a
minority view within the field. Furthermore, when Pacula (2011) examines a series of nine
papers on the topic, she notes that most results in developed countries with high disposable
income distinguish between heavy drinkers, who consume significantly less alcohol in
downturns, and light drinkers, who tend to drink slightly more as the economy falters. As she
further notes, a key implication for analyzing the cyclicality of drug abuse is that researchers
must include data on the differential response rates to economic conditions across the
demographic spectrum. While findings on the pro-cyclicality of total alcohol consumption and
the counter-cyclicality of binge drinking may offer some insight into the impact of local
4
economic conditions on drug abuse, alcohol’s status as a legal, taxed, and regulated product
weaken the argument that it is directly applicable to the use of illicit drugs like nonprescription
drug abuse, because of the addictive properties nicotine shares with opioids. Short-run decreases
in work hours and higher local unemployment rates are associated with declines in individuals’
tobacco consumption (Ruhm, 2000; Ruhm, 2005; Colman & Dave, 2014). Tobacco consumption
mirrors that of alcohol and falls most significantly among heavy smokers, dispelling the notion
that increasingly intense addiction to the product might limit or negate the effects of its income
elasticity4. Colman and Dave (2014) notice the disparate effects of unemployment between men
and women in the Great Recession, with men reducing cigarette consumption by a greater degree
in response to regional unemployment, but they caution against drawing a conclusion in absence
of more evidence given that the industries most severely impacted by the recession, durable
goods manufacturing and construction, employed far more males than females. Because the lag
time on tobacco-related deaths may extend significantly longer than that for overdoses from
drugs5 and knowing that little empirical work has been done on tobacco-related mortality, it
should be understood that the comparison between tobacco consumption and drug mortality is an
inexact but, nonetheless, useful tool for understanding the mechanisms influencing drug
overdoses.
In a larger context, the evidence from suicide, alcohol, and tobacco would appear to point
the results for drug overdose mortality in two different directions, with smoking and total alcohol
4 Ruhm, 2005
5 Note that the term “drugs” in this paper, unless otherwise specified, refers to addictive substances such
as opioids, cocaine, benzodiazepines, hallucinogens, etc. that produce euphoria and may result in death
from overdose. Critically, marijuana and other substances that cannot result in overdose are not include.
5
consumption rising and falling in sync with the business cycle as binge drinking and suicide rates
move in the opposite direction. However, it is possible to reconcile the consensus on these topics
by considering the distinction between the two groups of activities. Extreme health-adverse
activities like suicide and binge drinking represent cases where the psychological mechanisms of
a weak economy dominate, whereas the regular use of legal substances such as alcohol and
cigarettes represent an example where the income elasticity of normal goods proves more
influential. Since drug overdose deaths can be either intentional or unintentional from legally
prescribed or black market sources, a coherent story does not arise from the related literature as
to whether the data should yield results more similar to the extreme health-adverse activities or
II. Theory
The conceptual framework most applicable to drug overdose deaths is the theory of
rational addiction developed by Becker and Murphy (1988). Maintaining that individuals are
equipped with perfect information about a good’s costs as well as utility and that consumers
understand that they will become dependent on an addictive substance, they argue that the choice
to purchase an addictive good is a rational decision to maximize the present discounted value of
utility over the long run. Notwithstanding the controversial assumptions, rational addiction
theory remains the most prominent theory of addiction and provides a variety of useful
implications for studying addictive behavior. The theoretical model’s most important conclusions
maintain “that ‘cold turkey’ is used to end strong addictions, that addicts often go on binges, that
addicts respond more to permanent than to temporary changes in prices of addictive goods, and
that anxiety and tensions can precipitate an addiction” (Becker & Murphy, 1988), which have
some key consequences in the context of addictive drug use and the business cycle. Because it is
6
generally assumed that the price of a given good relative to a consumer’s income can influence
the quantity of the good purchased and that consumer income fluctuates in economic cycles, a
permutation of rational addiction theory might contradict the short-run notion of an income effect
in the case of addictive drugs, contingent on how recessions influence individual’s anxieties and
tensions. For example, assuming that a prospective drug consumer’s mental health deteriorates as
a consequence of a recession, the individual might initiate drug use despite a fall in disposable
income, because the utility from drug use is magnified by economic distress and exceeds the
associated long-run costs of addiction. Thus, the derived results from rational addiction theory
for the question of how economic conditions impact drug consumption hinge on whether or not
potential consumers are psychologically impacted by a downturn such that the utility of
addiction overwhelms the short-run loss of disposable income available for drug purchases6.
There are three core weaknesses stemming from Becker and Murphy’s model that are
worth briefly mentioning. First, the model assumes that because people experiencing substance
abuse disorders are long-run welfare maximizers, they are only likely to quit or reduce
consumption of an addictive good in response to long-term shifts in the prices of the good to
which they are addicted. This may not necessarily be the case if addiction tends to weaken a
consumer’s ability to effectively maximize their long-run utility by weakening their mental
faculties, or if addiction leads the consumer of an addictive substance to increase the discount
rate they apply to long-run outcomes. Second, it assumes that people with substance abuse
disorders are rational actors, or at least as rational as the average consumer in the market. A
growing body of evidence out of medical research, however, suggests that a disproportionately
high share of people with addictions may be suffering from a mental illness (Williams &
7
Ziedonis 2004) and, thus, are not necessarily capable of approaching potential long-run utility
bundles from a purely rational standpoint. Finally, the notion of the “self control problem”, an
inability to carry out optimal strategies for consumption due to short term symptoms like
withdrawal pain, presents a problematic challenge to rational addiction theory’s assumption that
quitting “cold turkey” is a viable option for addicts (Gruber & Koszegi, 2002; Gruber &
Mullainathan, 2002). Although these flaws remain unresolved, Becker and Murphy’s model
remains the most compelling model of addiction, and thus its implications are taken at face value
in this paper.
In a larger sense, researchers have melded the theory of rational addiction and existing
economic literature to illustrate two overarching pathways for local employment, income, and
poverty to affect an individual’s consumption of drugs: psychological and economic. The former,
relying on the assumption that economic shocks can alter psychological perceptions, holds that a
weaker economy necessarily puts upward pressure on the rate of drug use and mortality, whereas
the latter, acting through several dimensions of the business cycle, has either an upward or
Underneath the psychological pathway exist two distinct concepts: proximity and self-
medication. Arkes (2011) contends that a weakening economy incentivizes a greater number of
people to sell drugs as formerly employed individuals seek to replace lost income, which would
likely expand the accessibility of drugs and increase the fraction of the local population that
know an individual that sells drugs on the black market. Assuming consumers have imperfect
information about the location and quality of the supply of drugs, a rise in the accessibility and
familiarity of black market sellers could assuage concerns about the quality of the product and
make it easier to contact and transact in the local black market without getting caught by law
8
enforcement. This in turn, would mean a greater quantity of drug purchases as both the supply
and demand curve shift outward. By contrast the most prominent rationalization of the link
between economic conditions and drug use, self-medication theory, rests solely on the demand
side of the equation. Building off rational addiction, it logically follows that seeing or
experiencing job loss, the threat of job loss, and the rising demands of paid work during a
recession could spur stress, anxiety, and mental health problems, inducing a greater fraction of
the population to “self-medicate” with drugs that relieve pain or produce euphoria7 (Chalmers &
Ritter, 2011).
pattern of drug abuse and overdoses, depending on the assumption specifications an individual
researcher is willing to make: net utility, health insurance access, and income effects. The net
utility of activities other than drug use, such as job searches, may decline as the economy
deteriorates, pushing the number of overdoses higher7. Conversely, it is also possible that the
opportunity cost of drug use rises during a downturn, as it becomes more difficult to find
alternative employment and as the lower productivity that may result from drug use makes a
worker more likely to be laid off7. Comparably, the rise in health insurance rates during
recoveries may either lead to a fall in drug overdose mortality as a growing share of the
population can afford rehabilitation programs (Maclean, Cantor, and Pacula, 2015) or an
upswing in mortality as a growing number of people can afford to buy prescription opioids and
subsequently transition to the black market. Finally, there could be either a positive or negative
relationship between real median household income and drug abuse contingent on whether
7 Arkes (2011)
9
III. Applied Research
What little applied work has been done on the link between economic conditions and
drug use can be classified into three methodological groups with unique insights and problems:
natural experiments involving youth, questionnaire based research, and natural experiments
involving adults. The methods, findings, and weaknesses of each will be elaborated.
Three papers analyze the impact of state-level unemployment rates on self-reported drug
use in health surveys among American youth. Arkes (2007; 2011) employs data from the
National Longitudinal Survey of Youth in a panel data model and discovers a countercyclical
trend in the previous year cannabis use rate among young people ages 14 to 24 on the order of
roughly 1.4% for each 1 percentage point increase in the state-level unemployment rate.
Interestingly, his results for drugs other than marijuana claim a 1% marginal increase in past year
use for 14 to 19 year olds versus no statistically significant increase for 20 to 24 year olds. Arkes’
data also show no statistically significant racial gaps, conflicting with Pabiliona (2014), whose
work on the Youth Time Risk Survey claims a 1.9% higher likelihood of past month marijuana
use solely for young African American males per percentage point rise in state-level
unemployment. Still, the countercyclical nature of cannabis consumption among youth seems to
be somewhat consistent. The caveat with testing economic conditions among the youth is that
young people, particularly teenagers, are more exposed to recessions and unemployment than
older cohorts, because they are more likely to work in sectors prone to larger cyclical swings,
making the probability of employment more exogenous to their actions relative to adults.
Moreover, they may underestimate or discount the long-run value of their actions relative to the
group of adults who make up the vast majority of people with substance abuse disorders.
10
The second group of papers relies heavily on hypothetical questions posed to a group of
people who self-identify as having histories of drug abuse and suffers from a number of potential
pitfalls in both internal and external validity. After identifying a sample of approximately 100
people with a history of abusing cocaine, heroin, or alcohol, two studies pose a series of
interview questions regarding how potential fluctuations in income would affect a participant’s
consumption of illicit drugs and find that respondents plan rising consumption of illicit drugs as a
share of their income as disposable income rises in several steps (Petry, 2000; Roddy &
Steinmiller, 2011). The information collected from the survey is then used to map respondents’
demand for illicit drugs against the hypothetical income they are assigned in each question to
find the income elasticity of drug demand. Notably, Roddy & Steinmiller (2011) find some
evidence that heroin consumers’ “purchasing repertoire is very cost-effective”, that they live very
close to their black market suppliers, and that only a strong external shock to income or the
probability of arrest may shift consumption preferences. Taken together, the information that
emerges from each paper implies that recreational drugs like cocaine and heroin are subject to a
positive income effect and, consequently, are normal goods. However, the problems of self-
selection bias, incredibly small sample sizes, and the absence of hard data constitute serious
flaws in this methodology and render the data as an extension of anecdotal evidence, rather than
Research papers employing cross-sectional, time-series data on adult drug abuse are the
most effective at evaluating the relationship between the economic cycle and drug abuse, but the
existing literature is limited. Ruhm (2013) mentions that the trend for unintentional opioid
overdoses has shifted to become robustly countercyclical in recent years, but the data available
did not allow for demographic controls or decomposition. Chalmers and Ritter (2011), by
11
distinction, provide a detailed explanation of their regression of cannabis use on both Australian
state-level income and unemployment, revealing that past year participation in cannabis use
varies counter-cyclically with state-level per capita income. However, the dual drawbacks of
focusing wholly on cannabis, which cannot result in overdose deaths, and the Australian dataset,
which the authors admit spans a time period without a single recession, undermine the external
validity when applying the findings to American drug abuse and mortality. Carpenter, McClellan,
and Rees (2016), acting on information provided by the US National Survey on Drug Use and
Health, are the most successful in employing survey data and describe a pattern of pro-cyclical
LSD use and countercyclical consumption of analgesics and ecstasy. The authors further
establish that the highest rates of drug abuse are among white, working-age males without a
college education, confirming recent reports that highlight this group as the most at risk for drug
overdose mortality. Likewise with previous work, the paper’s dependence on survey data, which
may be prone to asymmetric response bias and social desirability bias, undercuts the reliability
and precision of the point value estimates of the coefficient of unemployment’s impact on drug
abuse. To bypass the weaknesses of survey data, Maclean, Cantor, and Pacula (2015) analyze the
relationship between three measures of unemployment and admissions rates for drug treatment
programs and observe that entrance to substance abuse programs fall as unemployment climbs.
Yet, the intuition remains unclear on whether treatment program membership varies inversely
with drug consumption or is roughly a fixed proportion of the total population with drug abuse
The current paper will seek to build on the existing literature about the cyclicality of drug
abuse and distinguish itself by (1) incorporating county-level data as well as state-level data for a
more accurate reflection of the economic conditions individuals face on a daily basis, (2) testing
12
health insurance coverage rates as a right hand side variable to partially evaluate the affects of
the implementation of the Affordable Care Act and to disentangle the connection between
economic conditions and access to prescription opioid coverage, (3) employing drug overdose
mortality rates as a proxy of drug abuse to overcome the issues innate to survey data, (4)
experimenting with racial composition controls and racial-economic interaction terms, (5) testing
the enactment of state-based drug laws including medical marijuana laws, “Good Samaritan”
laws8, and naloxone access laws9 to see if implementation of such laws relates to drug overdose
mortality, (6) measuring the relationship between county population growth and drug overdose
deaths, (7) utilizing lagged unemployment to test for the statistical significance of the duration of
economic malaise on overdose mortality, and (8) using alternate measures of economic
wellbeing to supplement unemployment data including the poverty rates and real median
household income.
IV. Data
The data used for this paper is primarily publicly available from four divisions of the US
federal government: the Census Bureau, the Centers for Disease Control, the Bureau of Labor
Statistics, and the St. Louis Federal Reserve. Each public institution provides some of the most
reliable data in their respective field, along with a bevy of statistics available for public use.
Observations from each dataset were matched to one another by Federal Information Processing
Service (FIPS) code and data year, creating a strongly balanced panel for state-level data and a
strongly balanced panel for county-level data, with a few exceptions. For the purposes of this
8 “Good Samaritan” laws, in this case, are laws that provide legal protections for people who intervene to
rescue someone at imminent risk of death from a drug overdose.
9 Naloxone is an emergency treatment medication for opioid drug overdoses that can prevent an overdose
from being fatal. Naloxone access laws are laws that make the medication more readily available or
accessible to either civilians or emergency response personnel.
13
analysis, parishes and independent cities were treated as effectively the same as counties.
Alaska’s geographic subdivisions were excluded due to frequent border changes and
reorganizations, meaning that the county-level data represent the results from 49 states and the
District of Columbia. The state-level data, by contrast, include all 50 states and exclude the
District of Columbia. There is no intuitive reason to believe that this difference will
fundamentally alter the results of the regression. Only two counties outside of Alaska were
excluded from the dataset for all years, both due to abnormal circumstances. Kalawao County,
Hawaii (FIPS code: 15005) is omitted due to lack of data and its unusual demographics as a
sparsely populated former quarantine zone for people with leprosy. Likewise, the independent
city of Bedford City, Virginia (FIPS code: 51515) is omitted, because of its loss of independent
status during the timeframe and its incorporation within Bedford County. In total, 3,112 counties
are included in the finalized dataset with observations from the 12 year-long timespan of 2003 to
2014. In total, less than 1% of county observations are excluded from the sample.
The Centers for Disease Control’s data on age-adjusted drug poisoning mortality is
estimated from the cause of death listed on US death certificates and is available for every county
and state annually throughout the aforementioned time period. At both aggregation levels, the
values are roughly normally distributed. The age adjusted overdose death rate is used rather than
the crude overdose death rate, because the age distribution of a locale may jointly determine
economic wellbeing and crude overdose deaths, resulting in omitted variable bias. Each
observation came paired with population by year in the county-level dataset, specifically.
Critically, the county-level and state-level data differ in the method of expressing drug overdose
mortality rates. For states, the CDC lists the precise value of the drug overdose mortality rate,
whereas only an estimated range is provided for each county. Moreover, at the county-level, drug
14
overdose mortality rates are sorted into 11 bins from 0 to 2 deaths per 100,000 residents per year
to 20+ deaths per 100,000 residents per year, making the county-level value a discrete dependent
variable.
The US Census Bureau’s Small Area Income and Poverty Estimates (SAIPE) Report
offers poverty rates and nominal median household income values for every state and county
running for the full timespan. The nominal median household income values were then adjusted
for the Consumer Price Index level, retrieved from the St. Louis Federal Reserve website, and
baselined to 2014 US dollars to generate real median household income. In addition to poverty
and income data, the US Census Bureau’s Current Population Survey (CPS) takes an annual
sample of the population in each state and county to create intercensal estimates of racial
composition, which were pulled and included for all years at the state-level and the period
The Bureau of Labor Statistics’ offers several statistics describing the vital indicators of
the local labor force. The Local Area Unemployment Statistics (LAUS) Program provides
information on total employment, labor force participants, and the unemployment rate for every
county with the exception of 7 Louisiana parishes in the immediate aftermath of Hurricane
Katrina, 2005-2006. At the state-level, the LAUS program offers, in addition to the overall
unemployment rate, the unemployment rate for all sufficiently large racial and gender groups.
Health insurance coverage rates for both states and counties were pulled from the Census
Bureau’s Small Area Health Insurance Estimate (SAIHE) Report for the years 2006 to 2014, the
last three quarters of the timespan of this paper. Critically, this data overlaps with the beginning
of the implementation of the Affordable Care Act (ACA), which spanned 2011 to 2014 and
extended health insurance coverage and guaranteed benefits for drug rehabilitation programs.
15
Information on health insurance coverage rates were available for all states and counties for the
Data on state-level drug laws were acquired through the National Organization to Reform
Marijuana Laws (NORML), an organization centered around the advocacy of liberalizing drug
laws. Although the organization takes a clear stance on the issue, all data on years of enactment
and implementation of key drugs laws were corroborated via external links to state government
websites and relevant legislation. The data, however, act as a weak proxy for legalization,
because they have little relation to the opening of dispensaries or the accessibility of cannabis.
Appendix Table A lists all relevant information on medical marijuana laws by state and district.
“Good Samaritan” law and naloxone access law enactment and implementation
information were acquired directly from a recent working paper (Rees, Sabia, Argys, Latshaw, &
Dave, 2017). Appendix Tables B and C are pulled directly from corresponding tables in the
working paper so as to leverage the data in this paper for a replication study.
Descriptive statistics for each dataset are available in Appendix Tables D and E.
Based on the data available, the econometric model best equipped to gauge the
relationship between local economic conditions, race, and drug overdose mortality is an OLS
regression with locality and year fixed effects as well as standard errors clustered by locale.
Standard errors are clustered at the locale level (state or county for the state-level and county-
level regressions, respectively), because measurement errors in death certificates are likely to be
related by county over time, meaning that standard errors are likely to be correlated by locale
over time. Because this is a natural experiment using panel data, controlling for time trends and
the unobserved qualities that differ from state to state or county to county is essential to isolating
16
the difference-in-difference estimators for the association between the right hand side variables
and drug overdose mortality rates. Each of the regressions on state and county data will have the
The overdose death rate is presented as a function of a constant (α), a vector of dummy
variables for each of 12 years (yy), a vector of dummy variables for each county FIPS code (ci), a
vector of continuous local economic wellbeing variables (ei,y), a vector of continuous local racial
composition variables (ri,y), a vector socio-economic interaction term(s) (xi,y), a vector of social
variables (si,y), a vector of dummy variables for enactment of relevant state-wide drug laws (li,y),
In this model, the deviation in a right hand side variable for a state or county relative to
the state or county’s mean value of that variable over the given timespan is taken relative to the
deviation of the average state or county from its mean value over the given timespan. Thus, the
regression measures the impact of a marginal change in a right hand side variable relative to the
national deviation from the national long-term mean. Because of the structure of the regression,
the model works to measure the impact the right hand side variables have in a fixed geographic
area as they vary over time, rather than attempting to explain the differences between geographic
areas. This approach is justifiable because baselining the geographies to their respective means
and regressing the variation relative to the mean accounts for the many unobservable
characteristics that vary between the geographies and may affect overdose mortality, but are
likely to remain fixed or nearly constant over time. While this approach soaks up a significant
degree of the variation in the data, it avoids the possibility of Simpson’s Paradox and helps
17
effectively identify the coefficient on a change in the right hand side variables being tested,
rather than capturing the impact of unobserved differences between geographies, exogenous time
VI. Results
Prior to testing various regressions based on the general form model, the correlations
between right hand side variables must be considered. Variables highly correlated with one
another should not be included in the same regression, so as to avoid the problem of
multicollinearity. Thus, two correlation tables are appended below for each dataset, with dashed
lines denoting the separate groups of economic variables, demographic variables, interaction
pov 1
rinc -0.7549 1
ln_pop -0.1683 0.4698 0.4576 0.2939 0.0896 -0.2000 0.1389 0.0902 0.4718 1
marij -0.0513 0.0993 0.1109 -0.0910 0.1050 -0.0452 -0.1708 0.1713 0.2149 0.0508 1
NAL -0.0371 0.1284 0.1377 0.0200 0.1414 -0.1051 -0.0419 0.1506 0.2467 0.1524 0.2566 1
GSL 0.0562 -0.0392 -0.0332 -0.0910 0.0048 -0.1431 -0.0423 0.2304 -0.0081 -0.0092 0.2205 -0.0095 1
Note: Correlations with an absolute value greater than 0.500 are bolded.
18
Table 4: Correlations between State-Level Right Hand Side Variables
pov rinc inc ln_inc insure u_rate wu wmu white black hisp marij NAL GSL
pov 1
rinc -0.7532 1
black 0.4329 -0.0761 -0.0698 -0.0938 -0.0808 0.2426 -0.0828 -0.1197 -0.4725 1
hisp 0.1593 0.1413 0.1662 0.1789 -0.4389 0.1536 0.2236 0.1665 -0.5939 -0.1187 1
marij -0.0930 0.2432 0.2929 0.2976 -0.0082 0.1623 0.2537 0.2461 -0.1974 -0.2994 0.2620 1
NAL -0.0740 0.2563 0.2563 0.3204 0.1737 0.1190 0.1463 0.1191 -0.1292 -0.0326 0.1972 0.2207 1
GSL 0.0924 0.0800 0.1666 0.1622 0.0648 0.0640 0.0775 0.0626 -0.1969 -0.0539 0.3262 0.3085 0.5188 1
Note: Correlations with an absolute value greater than 0.500 are bolded.
Because the racial and ethnic minority group variables are not particularly strongly
correlated with one another but could result in near perfect multicollinearity if all are included in
the same regression, separate regressions are run for whites and major racial/ethnic minority
groups. That the overall unemployment rate is closely tied to the unemployment rate for white
Americans and white males at the state-level forecloses the possibility of including these terms in
model with variables that both satisfy an intuitive understanding of the issue of drug overdose
19
The two datasets for counties and states will be tested separately, with three different left
hand side variables regressed over a series of several models. For county-level data the left hand
side variable to be regressed will be the discrete bin value of drug overdose mortality. Each
county for each year is assigned a value ranging from 1 to 11, with one representing a range of 0
to 2 overdose deaths per 100,000 residents and 11 representing more than 20 overdose deaths for
the same population. The two state-level left hand side variables for overdose mortality will be
the age adjusted drug overdose rate per 100,000 residents and the natural logarithm of the same
value. For each left hand side variable there will be a regression on each of the economic
variables and interaction terms alone as well as a multiple regression for each of the
demographic groups and the unemployment rate. Additional models will include variables such
as the natural logarithm of population, drug laws, and health insurance coverage.
Beginning with the county-level dataset, the drug overdose mortality bin is regressed on
the poverty rate, real median household income in 2014 dollars, nominal median household
income, and the unemployment rate. Then the unemployment rate is tested jointly in series of
combinations with various drug laws, population size, and the share of the county population that
is non-hispanic white, non-hispanic black, hispanic, and non-hispanic asian. Results are listed
below in Tables 5a and 5b. Table 5c then presents various models incorporating health insurance
coverage over a more limited timespan of 2006 to 2014. Table 5d concludes with models
20
Table 5a: County Fixed Effects Model of Drug Overdose Mortality (2003-2014)
Variable 1 2 3 4 5 6
0.016***
pov
(3.95)
-0.006*
rinc
(-1.83)
-0.023***
inc
(-6.55)
-0.661***
ln_pop
(-2.76)
0.113**
marijuana
(2.29)
0.057
NAL
(1.48)
-0.103**
GSL
(-2.18)
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
21
Table 5b: County Fixed Effects Model of Drug Overdose Mortality (2003-2010)
Variable 7 8 9 10 11 12
0.042*** 0.049***
white
(6.33) (7.15)
0.147*** 0.161***
marijuana
(2.89) (3.16)
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
22
Table 5c: County Fixed Effects Model of Drug Overdose Mortality (2006-2014)
Variable 13 14 15 16 17 18
0.001
pov
(0.43)
-0.001
rinc
(-0.66)
-0.010***
inc
(-4.32)
0.016*** 0.016***
u_rate
(4.91) (3.18)
0.021
marijuana
(0.49)
0.045
NAL
(1.40)
-0.109***
GSL
(-2.89)
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
23
Table 5d: County Fixed Effects Model of Drug Overdose Mortality with Lags
Variable 19 20 21 22 23
0.006* 0.011***
lag2u
(1.19) (3.01)
-0.003
lag3u
(-0.64)
0.028***
rollavg3u
(3.55)
0.023**
rollavg4u
(2.56)
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
Because the county-level data set has a dramatically larger sample size than the state-
level data and much greater variation within each variable, it is better able to identify statistically
significant results for the coefficients on each of the righthand side variables. Moreover, if one of
the operating assumptions of the paper, that individuals are more likely to experience economic
conditions reflected by the statistics of their county than the statistics of their state, the results
should be stronger for the county-level data set. The findings from the first four models
interval it can be stated that fatal drug overdoses vary directly with the poverty rate, nominal
24
household income, and the unemployment rate. Congruously, rising real median household
income is associated with a falling drug overdose mortality rate at a 90% confidence interval.
Table 5b offers intriguing new findings on drug addiction, but this evidence is tempered
against inconsistencies over the time period. More specifically, Table 5b includes racial
coefficients from racial data that was only available at the county-level from 2003 to 2010 and,
therefore, tests a different set of observations than Table 5a. While the relationship between
unemployment and drug overdoses remains virtually unchanged between the 2003 to 2010 and
2010 to 2014 period, the relationship between population growth and overdoses does not. The
results suggest that the counties facing the brunt of overdose mortality experienced above
average population growth during the growth period of 2003 to 2007 and experienced subpar
population growth during the 2010 to 2014 post-Great Recession period. This conclusion can be
drawn from the observation that when a regression identical to that of model 5 in Table 5a is run
over the restricted time span of 2003 to 2010 used in Table 5b, the coefficient on the natural
population in models 9 and 11. These findings on population are accompanied by fascinating
revelations on the relationship between drugs laws and overdose mortality in both Table 5a and
Table 5b. Curiously, the weak proxy of marijuana legalization is positively associated with
overdoses from drugs other than marijuana in models in both tables, whereas naloxone access
laws (NALs) had no statistically significant impact, and “Good Samaritan” laws (GSLs)
corresponded to a significant decrease in per capita fatal overdoses. The results for the latter two
laws were the reverse of those found by Rees, Sabia, Argys, Latshaw, and Dave (2017), who
discovered a statistically significant reductions in overdose mortality from NALs and statistically
25
Models 7-12 further confirm what existing literature has found about the racial fault lines
of the drug addiction crisis. The more rapid the growth in the share of the county population that
is non-hispanic white and the lower the share of the population that is either non-hispanic black,
non-hispanic Asian American, or hispanic, the higher the rate of drug poisoning deaths. In each
of the joint racial-economic models the economic variable and demographic variable are
Table 5c begins to explore a series of models which account for the impact of health
insurance coverage on overdose rates. During the time period 2006 to 2014 the most substantial
shifts in coverage rates stemmed from the implementation of the ACA and the accompanying
Medicaid expansion. The newfound insurance coverage for prescription opioids and drug
rehabilitation represent conflicting pressures upward and downward, respectively, on the rate of
drug poisoning deaths. However, at the county-level none of the regression coefficients on health
insurance coverage are statistically significant. Controlling for health insurance coverage did
seem to render the effect of medical marijuana laws statistically insignificant, suggesting that the
results for medical marijuana legalization were spurious. This is potentially the case if state
governments with an ideological predisposition to enact medical marijuana laws also have a
Finally, Table 5d illustrates a set of regressions testing various lagged economic variables
and rolling averages to identify whether or not a longer duration of elevated unemployment
places upward pressure on overdose mortality. Statistically significant results for longer lag times
would be consistent with the media portrayal of such fatalities as “despair deaths”, because local
economic conditions presumably would have to remain noticeably dismal for an extended period
of time for conditions like depression to set in. As Models 19, 20, 22, and 23 demonstrate, there
26
is some merit to the idea, given that one year lagged unemployment has a positive significant
impact on per capita mortality, and given that three and four year rolling averages have roughly
Taken together, the observations from county-level data provide a solid reflection of the
economic conditions an individual may face and a clear indication of the sign of the coefficient
of the relationship between right hand side variables and overdose deaths. However, the benefits
of a greater ability to identify statistically significant results are counterbalanced by the key
disadvantage that the point estimate for each beta coefficient does not have a direct, meaningful
interpretation. The knowledge that each beta represents the change in expected drug mortality
bin contingent on a 1 unit increase of the right hand side variable must, therefore, be
supplemented with data that can convey whether or not the results are significant in a policy
The state-level drug overdose mortality dataset affords just such an opportunity for
analysis. In the initial seven models in Table 6a we regress a state’s precise drug overdose
mortality rate on five economic variables and two interaction terms: the poverty rate, real
median household income, nominal household income, the natural logarithm of household
income, the unemployment rate, the unemployment rate for white residents, and the
unemployment rate for white male residents. Next, the unemployment rate and nominal median
household income are placed in joint regressions with demographic variables and dummy
variables for state-level drug laws in Table 6b. Table 6c then presents five additional models
27
Table 6a: State Fixed Effects Model of Drug Overdose Mortality Rate (2003-2014)
Variable 1 2 3 4 5 6 7
0.162
pov
(0.61)
-0.151
rinc
(-1.61)
0.226**
u_rate
(2.06)
0.212*
wu
(1.82)
0.245**
wmu
(2.39)
-0.183**
inc
(-2.22)
-8.323
ln_inc
(-1.55)
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
28
Table 6b: State Fixed Effects Model of Drug Overdose Mortality Rate (2003-2014)
Variable 8 9 10 11 12
-0.182**
inc
(-2.20)
0.133
white
(1.29)
0.018
black
(0.09)
-0.170
hisp
(-1.32)
-0.533
NAL
(-0.83)
-0.504
GSL
(-0.64)
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
29
Table 6c: State Fixed Effects Model of Drug Overdose Mortality Rate (2006-2014)
Variable 13 14 15 16 17
0.028
inc
(0.22)
2.690*** 2.936***
marijuana
(3.45) (3.19)
0.285
NAL
(0.43)
-0.735
GSL
(-0.99)
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
Unsurprisingly, the more limited sample size and variation within the state-level data
result in fewer instances of statistically significant coefficients. Among the economic variables
and interaction terms, only the unemployment rate, nominal household income, the white male
unemployment rate, and the white unemployment rate are statistically significant at the 95%,
95%, 95%, and 90% confidence intervals respectively. For each measure of unemployment, an
increase of 1 percentage point equates to between roughly a 0.15 and 0.25 person increase in
number of fatal drug overdoses per 100,000 residents. Although the coefficients on the poverty
rate and real median household income are not statistically significant at a 90% confidence
30
interval it is notable that the sign of each coefficient is consistent with the findings from the
county-level data. This information coupled with the statistically significant results from two
new racial-economic interaction terms add further evidence to the notion that drug overdose
mortality is countercyclical, and that economic conditions are experienced more strongly at the
Table 6c offers some interesting insights which may explain patterns in the existing
literature that are inconsistent with medically observed patient demographics. At the state-level,
the coefficients on the non-hispanic white, non-hispanic black, and hispanic share of the
population do not pass the threshold of statistical significance at a 90% confidence interval.
Paired with the fact that the coefficients on the white male, white, and total unemployment rates
do not statistically differ from one another, the evidence suggests that aggregation at a higher
level than county data does not capture enough variation to provide a reliable estimation of the
relationship between race and the drug overdose crisis. This is notable, because virtually all
papers utilizing geography-based panel data to date have exclusively employed state-level data.
However, the state-level results for medical marijuana legalization do at least provide strong
evidence to confirm county-level patterns observed in Table 5a and Table 5b. Specifically, the
enactment of medical marijuana laws at the state-level, during the period 2003-2014, was
associated with an increase in overdose deaths from drugs other than marijuana on the order of
roughly 1.6 deaths per 100,000 residents. Both naloxone access laws and “Good Samaritan” laws
seem to have a negative relationship of approximately 0.5 fewer deaths per 100,000 residents,
The final group of regressions in Table 6c reflect, perhaps, the most significant findings
for public health policymakers. At the state-level, increases in health insurance coverage were
31
linked strongly, at the 99% confidence interval, with dramatic reductions in drug overdose
mortality at a rate of 0.5 deaths per 100,000 residents for each 1 percentage point increase in
health insurance coverage. Given that most of the variation in health insurance came during the
ACA’s dramatic expansion of coverage, which occurred during the latter end of the timespan of
the paper, this would seem to suggest that the early phase of the ACA put downward pressure on
the overdose mortality crisis. Whether this trend continued into the later phases of the ACA’s
implementation is unclear, but will be a critical point of future investigations when more recent
For the final series of models, the natural logarithm of the drug overdose mortality rate is
leveraged as the left-hand side variable to yield beta coefficients that reflect the percentage
change in the overdose mortality rate resulting from a one unit increase in each right hand side
variable. The same approach to generating regression models was taken as in the previous state-
level data models with the addition of the natural logarithm of real median household income
and health insurance coverage as righthand side variables. Results are listed below in Table 7.
32
Table 7: State Fixed Effects Model of the Natural Log of Drug Overdose Mortality Rate
Variable 1 2 3 4 5 6 7 8 9
0.017
pov
(0.95)
-0.589
ln(rinc)
(-1.40)
0.010
u_rate
(0.94)
-0.039***
insure (-5.45)
0.010
wmu
(1.10)
0.004
white
(0.72)
0.000
black
(0.01)
-0.000
hisp
(-0.11)
Sample
600 600 600 600 450 600 600 600 600
Size
R2
0.542 0.545 0.545 0.541 0.355 0.541 0.546 0.545 0.546
Within
R2
0.080 0.020 0.025 0.221 0.094 0.291 0.004 0.020 0.015
Between
R2
0.204 0.165 0.173 0.181 0.137 0.187 0.120 0.164 0.158
Overall
Note: The t-statistic for each point estimate is included in parentheses below the numeric value of the coefficient.
The symbols of *, **, and *** signal statistical significance at the 90%, 95%, and 99% confidence intervals,
respectively. Dummy variables for year and county are excluded from the table.
Model 5 confirms evidence in Table 6c, and suggests that every 1 percentage point
increase in health insurance coverage at the state-level during the period 2006 to 2014
33
corresponded to approximately a 3.9% reduction in the rate of fatal drug overdoses per capita. If
this relationship persisted with the same point-estimate at the national-level through 2017 and,
assuming it was a causal relationship10, this would suggest that the ACA’s expansion of health
insurance access by about 6 percentage points (Cohen & Martinez, 2016) would be holding
today’s drug overdose mortality rate roughly 23.4% lower than where it would without
Despite the absence of statistically significant findings for the remaining eight of the nine
regressions, the estimated coefficients do offer promise for further research. The point estimates
are significant for public policy, but appear statistically insignificant due to a dearth of
observations. Should the roughly 1 to 2% decrease in the drug poisoning mortality rate for a one
percentage point increase in the poverty rate, the unemployment rate, the white unemployment
rate, the white male unemployment rate, or a $1,000 increase in real median household income
be found in papers leveraging different data sources, they would prove extraordinarily significant
for public health policymakers. Moreover, if the results of Table 7, Model 3 prove consistent
across future work a -0.589 income elasticity of fatal drug overdoses could begin to establish a
body of academic literature suggesting drugs like opioids are effectively inferior goods.
VII. Conclusion
Admittedly, there are a number of ways in which this paper’s methodology is limited.
Both datasets only span the timeframe of 2003 to 2014, which includes just a single economic
downturn, the Great Recession of 2007 to 2009. This somewhat tempers the external validity of
applying the results to future shocks, which are likely to be significantly less severe and may hit
the housing and financial services sectors less intensely. With an absence of obvious and readily
34
available data on instrumental variables there was no apparent solution to issues of simultaneous
causality between drug use and local economic prospects or omitted variable bias, particularly
with respect to education. The innate trouble with reliably estimating black market supply,
prices, and quality of illicit drugs likely make isolating shifts in the demand for illicit drugs
impractical. Likewise, the relocation of workers in search of better job prospects from county to
county or state to state may introduce additional biases for which it is difficult to correct.
Evidence from county-level regressions suggest differences in internal population growth and
migration patterns that correlate with drug overdoses over the period, which are extraordinarily
difficult to disentangle.
Although the methods applied in this paper provide valuable reduced form estimates,
they do not afford reliable structural form estimates to decompose the individual effects of the
competing pressures of less disposable income and greater psychological pressure. This problem
also makes it difficult to assess the mechanisms through which health insurance may reduce drug
overdose mortality. More fundamentally, it is still not clear what is driving the secular trend of
rapid rises in overdose mortality, and the popularization of the “despair death” narrative, which
asserts that these deaths are primarily a result of economic malaise, does not fully explain why
overdoses continue to climb both nationally and across a broad swath of counties well into an
economic upswing. Until the key question of why economic conditions and drug addiction are
associated with one another is addressed in future research, it will be difficult for policymakers to
Yet, several findings uncovered from the recently released CDC data used in this study do
offer some preliminary conclusions on the relationship between economic conditions and drug
use. Namely, that measures of improving economic wellbeing are associated with declining rates
35
of fatal drug overdose, and that drug overdose mortality in the era of the Great Recession varies
counter-cyclically. In relation to previous research, the evidence suggests that fatal drug
incidence, despite being driven by unintentional overdoses, may have more in common with
binge drinking and suicide than the regular consumption of either alcohol or tobacco. Presuming
that drug overdoses correlate strongly with overall drug use, such a finding would suggest that
the psychological influences of economic anxiety may dominate the income effect innate to
normal goods. At the same time, the possibility that heavy drug users’ binges are driving spikes
in overdose mortality should not be discounted in the absence of quality evidence that separates
Racial controls, on balance, confirm public health data, which suggest that white
Americans are experiencing the bulk of recent increases in the drug overdose rate. The results for
interaction terms such as white unemployment and white male unemployment tend not to differ
significantly from corresponding aggregate measures of economic wellbeing like the overall
unemployment rate. Differences in significance contingent on aggregation level likely reflect the
intuition that people are most likely to have their economic wellbeing and outlook reflected in
their county domicile, rather than their state of residence. Thus the findings imply that there are
not detectible gaps between the responsiveness of non-hispanic white Americans, non-hispanic
African Americans, non-Hispanic Asian Americans, and Hispanic Americans to changes in local
economic conditions.
It should be noted, that this paper offers ample evidence that may be of use to
between health insurance coverage, a state’s adoption of key drugs laws, and overdose deaths are
valuable. With the potential repeal of the ACA’s Medicaid expansion and 24 million Americans
36
at risk of losing health insurance access (CBO, 2017), this paper’s findings on the possibility that
such a policy change may result in increasing rates of overdose from nonprescription drugs as
consumers switch to cheaper black market alternatives for pain relief may be worth
consideration. Moreover, the somewhat consistent evidence on “Good Samaritan” laws putting
downward pressure on the overdose fatality rate may illustrate an example of how harm
Finally, policymakers may take care to note the implications of the countercyclical trend
of drug mortality for fiscal federalism. Such a relationship between the business cycle and drug
mortality may make it wise for the federal government to offer states and municipalities
categorical grants for rehabilitation funding, because as of early 2017 forty-nine states are
constitutionally obligated to balance their budgets, making long-run budgets for state drug
rehabilitation programs susceptible to budget cuts at precisely the time that drug overdoses are
spiking. The distribution of naloxone to state and local authorities would likewise benefit from
being targeted at areas particularly hard hit by economic downturns, and whose demographic
patterns fit those of counties facing the brunt of the drug overdose crisis.
37
Bibliography
Arkes, Jeremy (2007). “Does the Economy Affect Teenage Substance Abuse?”
Health Economics, 16: 19-36.
Arkes, Jeremy (2011). “Recessions and the Participation of Youth in the Selling
and Use of Illicit Drugs,” International Journal of Drug Policy, 22(5): 335-340.
Becker, Gary S. & Kevin M. Murphy (1988). “A Theory of Rational Addiction,” The Journal of
Political Economy, 96(4): 675-700.
Bor, Jacob, Sanjay Basu, Adam Coutts, Martin McKee, & David Stuckler (2013). “Alcohol Use
During the Great Recession of 2008-2009,” Alcohol and Alcoholism, 48(3): 343-348.
Carpenter, Christoper S., Chandler B. McClellan, and Daniel I. Rees (2016). “Economic
Conditions, Illicit Drug Use, and Substance Abuse Disorders in the United States,”
NBER Working Paper No. 22051.
Chalmers, Jenny and Alison Ritter (2011). “The Business Cycle and Drug Use in Australia:
Evidence from Repeated Cross-sections of Individual Level Data,” International Journal
of Drug Policy, 22(5): 341-352.
Cicero, Theodore J., Matthew S. Ellis, Hilary L. Surratt, & Steven P. Kurtz (2014). “The
Changing Face of Heroin Use in the United States,” JAMA Psychiatry, 71(7): 821-826.
Center for Disease Control (2015). Division for Violence Prevention. “Suicide Facts at a
Glance”. Online.
Classen, Timothy J., & Richard A. Dunn (2012). “The Effect of Job Loss and Unemployment
Duration on Suicide Risk in the United States: A New Look Using Mass-Layoffs and
Unemployment Duration,” Health Economics, 21: 338-350.
Cohen, Robin A., & Michael E. Martinez (2016). “Health Insurance Coverage: Early Release of
Estimates from the National Health Interview Survey, 2008.” National Health Survey
Early Release Program. National Center for Health Statistics. Centers for Disease Control
and Prevention.
Colman, Gregory, & Dhaval Dave (2014). “Unemployment and Health Behaviors Over the
Business Cycle: A Longitudinal View,” NBER Working Paper No.20748.
Congressional Budget Office (2017). “The American Healthcare Act,” Congressional Budget
Office Cost Estimate. March 13, 2017.
Cotti, Chad, Richard A. Dunn, & Nathan Tefft (2014). “The Great Recession and Consumer
Demand for Alcohol: A Dynamic Panel-Data Analysis of U.S. Households,” American
Journal of Health Economics, forthcoming.
38
Davalos, Maria E., Hai Fang, & Michael T. French (2012). “Easing the Pain of an Economic
Downturn: Macroeconomic Conditions and Excessive Alcohol Consumption,” Health
Economics, 21: 1318-1335.
Dee, Thomas S. (2001). “Alcohol Abuse and Economic Conditions: Evidence form Repeated
Cross-Sections of Individual-Level Data,” Health Economics, 10: 257-270.
Federal Bureau of Investigation (2014). Unified Crime Statistics. “Crime in the United States”.
Expanded Homicide Data Table 8. Online.
Gruber, Jonathan, & Botond Koszegi (2002). “A Theory of Government Regulation of Addictive
Bads: Optimal Tax Levels and Tax Incidence for Cigarette Excise Taxation,” NBER
Working Paper No. 8777.
Gruber, Jonathan, & Sendhil Mullainathan (2002). “Do Cigarette Taxes Make Smokers
Happier?” NBER Working Paper No. 8872.
Jones, Christopher M., Joseph Logan, R. Matthew Gladden, & Michele K. Bohm (2015). “Vital
Signs: Demographic and Substance Use Trends Among Heroin Users — United States,
2002-2013,” Morbidity and Mortality Weekly Report, 64(26): 719-725.
Maclean, Johanna Catherine, Jonathan H. Cantor, and Rosalie Riccardo Pacula (2015).
“Economic Downturns and Substance Abuse Treatment: Evidence from Admissions
Data,” NBER Working Paper No.19115.
Muhuri, Pradip K., Joseph C. Gfroerer, & M. Christine Davies (2013). “Associations of
Nonmedical Pain Reliever Use and Initiation of Heroin Use in the United States,” Center
for Behavioral Health Statistics and Quality Data Review, August 2013.
Pabiliona, Sabrina Wulff (2014). “The Effects of the Great Recession on Teenagers’ Risky Health
Behaviors and Time Use,” NBER Working Paper No. 474.
Pacula, Rosalie Liccardo (2011). “Substance Use and Recessions: What Can be Learned from
Economic Analyses of Alcohol?” International Journal of Drug Policy, 22: 326-334.
Rees, Daniel I., Joseph J. Sabia, Laura M. Argys, Joshua Latshaw, & Dhaval Dave (2017). “With
a Little Help from My Friends: The Effect of Naloxone Access and Good Samaritan
Laws on Opioid Related Deaths,” NBER Working Paper No. 23171.
Roddy, Juliette, & Caren L. Steinmiller (2011). “Heroin Purchasing is Income and Price
Sensitive,” Psychology of Addictive Behaviors, 25(2): 358-364.
39
Rudd, Rose A., Noah Aleshire, Jon E. Zibbell, & R. Matthew Gladden (2016). “Increases in
Drug and Opioid Overdose Deaths — United States, 2000-2014,” Morbidity and
Mortality Weekly Report, 64(51-2): 1378-1382.
Ruhm, Christopher J. (2000). “Are Recessions Good for Your Health?*”, The Quarterly Journal
of Economics, May 2000: 617-650.
Ruhm, Christoper J. (2005). “Healthy Living in Hard Times,” Journal of Health Economics, 24:
341-363.
Ruhm, Christoper J. (2006). “Deaths Rise in Good Economic Times: Evidence from the OECD,”
Economics and Human Biology, 4(3): 298-316.
Ruhm, Christoper J. (2013). “Recessions, Healthy No More?” NBER Working Paper No. 19287.
Ruhm, Christoper J. (2016). “Drug Poisoning Deaths in the United States, 1999-2012: A
Statistical Adjustment Analysis,” Population Health Metrics, 14: 2.
Ruhm, Christopher J., & W. E. Black (2002). “Does Drinking Really Decrease in Bad Times?”
Journal of Health Economics, 21(4): 659-678.
Snipes, Michael, Timothy M. Cunha, & David D. Hemley (2012). “Unemployment Fluctuations
and Regional Suicide 1980-2006”, Journal of Applied Economics and Business Research,
2(2): 103-122.
Stevens, Ann Huff, Douglas L. Miller, Marianne E. Page, & Mateusz Filipski (2011). “The Best
of Times, The Worst of Times: Understanding Pro-Cyclical Mortality,” NBER Working
Paper No.17657.
Tekin, Erdal, Chandler McClellan, & Karen Jean Minyard (2013). “Health and Health Behaviors
During the Worst of Times: Evidence from the Great Recession,” NBER Working Paper
No. 19234.
Williams, Jill M., & Douglas Ziedonis (2004). “Addressing Tobacco among Individuals with a
Mental Illness or an Addiction,” Addictive Behaviors, 29: 1067-1083.
Yang, Bijou (1992). “The Economy and Suicide: A Time-Series Study of the U.S.A.,” The
American Journal of Economics and Sociology, 51(1): 87-99.
40
Appendix
Bolded years indicate medical marijuana policy changes taking effect within the timespan of this paper.
41
Table B: Timeline of State and D.C. Naloxone Access Law Implementation 1999-2014
State Naloxone Decriminalization Standing Order Effective Date 1st Full Legal Year
CA Y Y 1/1/2008 2008
CO N Y 5/10/2013 2014
CT N N 10/1/2003 2004
DC Y N 3/19/2013 2014
DE N Y 8/4/2014 2015
GA N Y 4/24/2014 2015
IL Y Y 1/1/2010 2011
KY Y Y 6/25/2013 2014
ME N Y 4/29/2014 2015
MD N Y 10/1/2013 2014
MA Y Y 8/2/2012 2013
MI N N 10/14/2014 2015
MN Y Y 5/10/2014 2015
NJ Y Y 7/1/2013 2014
NM Y Y 4/3/2001 2002
NY Y Y 6/24/2014 2015
NC N Y 4/9/2013 2014
OH N Y 3/11/2014 2015
OK N N 11/1/2013 2014
OR N Y 6/6/2013 2014
PA N Y 11/29/2014 2015
RI Y Y 6/18/2012 2013
TN N Y 7/1/2014 2015
UT N Y 5/13/2014 2015
VT Y Y 7/1/2013 2014
VA N Y 7/1/2013 2014
WA Y Y 6/10/2010 2011
WI Y Y 4/9/2014 2015
Bolded years indicate medical marijuana policy changes taking effect within the timespan of this paper.
Note: This Table is excerpted directly from Rees, Sabia, Argys, Latshaw, & Dave (2017).
42
Table C: Timeline of State and D.C. Good Samaritan Law Implementation 1999-2014
State Extends to Drug Paraphernalia Extends to Alcohol Effective Date 1st Full Legal Year
AK N N 10/8/2014 2015
CA Y N 1/1/2013 2013
CO Y N 5/29/2012 2013
CT Y N 10/1/2011 2012
DC Y Y 3/19/2013 2014
DE Y N 8/31/2013 2014
FL N N 10/1/2012 2013
GA Y N 4/24/2014 2015
IL N N 6/1/2012 2013
LA N N 8/1/2014 2015
MD Y Y 10/1/2014 2015
MA N N 8/2/2012 2013
MN Y N 7/1/2014 2015
NJ Y N 5/2/2013 2014
NM N N 6/15/2007 2008
NY Y Y 9/18/2011 2012
NC Y N 4/9/2013 2014
PA Y N 12/1/2014 2015
RI Y N 6/18/2012 2013
UT Y N 3/20/2014 2015
VT N Y 6/5/2013 2014
WA N Y 6/10/2010 2011
WI Y N 4/9/2014 2015
Bolded years indicate medical marijuana policy changes taking effect within the timespan of this paper.
Note: This Table is excerpted directly from Rees, Sabia, Argys, Latshaw, & Dave (2017).
43
Table D: County-Level Data Summary
Unique Numeric
fips 37,344 1001 56045 Centers for Disease Control
Identifier for Counties
year Observation Year 37,344 2003 2014 Centers for Disease Control
Median Household
med_hh_inc 37,344 $16,868 $125,635 US Census Bureau
Income
1 Year Lagged
lag1u 34,218 1.1 28.9 Bureau of Labor Statistics
Unemployment Rate
2 Year Lagged
lag2u 31,106 1.1 28.9 Bureau of Labor Statistics
Unemployment Rate
3 Year Lagged
lag3u 27,994 1.3 28.9 Bureau of Labor Statistics
Unemployment Rate
44
Variable Description Observations Minimum Maximum Source
Health Insurance
insure 28,008 50.5 97.3 US Census Bureau
Coverage Rate
Hispanic Share of
hisp 24,896 0.0 96.9 US Census Bureau
Population
Non-hispanic White
white 24,896 2.5 99.4 US Census Bureau
Share of Population
Non-hispanic Black
black 24,896 0.0 85.9 US Census Bureau
Share of Population
Non-hispanic Asian
American or Pacific
asian 24,896 0.0 44.8 US Census Bureau
Islander Share of
Population
Non-hispanic Native
American or Alaskan
native 24,896 0.0 94.2 US Census Bureau
Native Share of
Population
45
Table E: State-Level Data Summary
year Observation Year 600 2003 2014 Centers for Disease Control
Age-Adjusted Overdose
death Death Rate per 100,000 600 1.8 36.3 Centers for Disease Control
Residents
Median Household
med_hh_inc 600 $32,397 $73,851 US Census Bureau
Income
Health Insurance
insure 450 72.4 96.2 Census Bureau
Coverage Rate
Non-Hispanic White
wu 600 2.2 13.9 Bureau of Labor Statistics
Unemployment Rate
Hispanic Share of
hisp 600 0.4 46.5 US Census Bureau
Population
Non-hispanic Asian
asian American Share of 600 0.1 47.2 US Census Bureau
Population
46
Variable Description Observations Minimum Maximum Source
Non-hispanic Pacific
pi Islander Share of 600 0.0 14.4 US Census Bureau
Population
Non-hispanic Native
American or Alaskan
native 600 0.0 15.1 US Census Bureau
Native Share of
Population
47