Data
TEDS is a national admission-based data system administrated by the Center for Behavioral Health Statistics and Quality of the Substance Abuse and Mental Health Services Administration (SAMHSA). Since 1992, the TEDS system has compiled data from each state to track annual discharges and admissions to public and private substance abuse facilities that receive public funding. Treatment facilities that receive public funds or are licensed or certified by state substance abuse agencies are required to report data on all clients, regardless of health insurance status. The TEDS system comprises two major components: the Admissions Data Set and the Linked Discharge Data Set. Both data sets provide demographic, clinical, substance service characteristics and settings, employment status, presence of psychiatric problems, prior history, route of administration, insurance status, and source of payment for all patients 12 years of age and older. In addition, three (primary, secondary, and tertiary) substances of abuse, their route of administration, frequency of use, age at first use, and a source of referral to treatment are recorded for each admission. TEDS classifies substance use into seven categories: alcohol, marijuana/hashish, opiates (heroin, non-prescription methadone, and other opiates and synthetics), cocaine, stimulants (methamphetamine, other amphetamines, and other stimulants), other drugs, and none reported. We used the TEDS admissions data set from 1993 to 2016, excluding the “none reported” category.
Measurement
Dependent variables
State identifiers in TEDS were used to derive the state-level aggregate number of treatment admissions for six categories of primary substance use (alcohol, marjuana/hashish, opiates, cocaine, stimulants and other drugs) for each year individually. Additionally, we used a broader outcome for the state-level aggregated number of treatment admissions, including primary, secondary, and tertiary diagnosis for any each category of substance use.
Independent variables
Our main dependent variable was economic condition. We used state unemployment rates to represent the economic condition. State unemployment rates and median household income (in thousand dollars) for each state were obtained from the Bureau of Labor Statistics [17,18,19]. We also obtained state laws on medical marijuana laws fromProCon.org [20], state alcohol taxes from the tax policy center [21], and the insurance coverage rate for each state from the U.S. Census Bureau [22]. We followed the Census Bureau’s recommendations for obtaining insurance coverage rates and used the Health Insurance Historical Tables - Original Series for 1993 through 1998, the Current Population Survey Annual Social and Economic Supplement (CPS ASEC) data to estimate 1999 through 2007, and the American Community Survey (ACS) for rates after 2007. These two estimates differ slightly but parallel in change between 2009 and 2012. Additional state-level characteristics including the log of population, mean age, percentage of the state population that is male, and percentage of the population that is white were calculated using U.S. Population Data through the National Cancer Institute (NCI) [23]. The inflation-adjusted beer excise tax was measured in each state at the 2018 price level.
We created a dichotomous indicator, economic trend, to test whether economic conditions had an asymmetric effect on substance use treatment admissions in economic upturns and downturns. When unemployment was higher than in the previous period, the economy experienced a downward trend. Otherwise, the economy was under expansion. We additionally created a recession indicator. The two periods, 2001 and 2008–2009, were considered recessions with negative economic growth, in accordance with the National Bureau of Economic Research, Inc assessment [24].
Statistical analysis
We used difference-in-difference (DID) models to estimate whether changes in economic conditions were associated with changes in substance use treatment admission. Generally, DID is a quasi-experimental design used to estimate the effect of a specific policy or intervention (state unemployment rate in our study) by comparing the changes in outcomes over time between the treatment group and control group [25]. The outcome variable, the state-level aggregate number of treatment admissions, was log-transformed in order to address potential skewness. Multivariable linear regressions models were used to assess the association between the economic condition (state unemployment rate) and substance use treatment admissions. Model 1 adjusted all listed state-level characteristics, including log of population, mean age, percentage of state population that is male, percentage of state population that is white, state insurance coverage rate, state median household income (in thousands), medical marijuana laws, and survey year. Model 1 additionally adjusted for census division fixed effects to capture unobserved confounders that cluster among neighboring states at the division level. State beer taxes were adjusted for alcohol substance use treatment only (see eq. (1)).
$$ {Y}_{st}\sim {\beta}_1\ast Eco{n}_s+{\beta}_2\ast {Year}_t+{\beta}_3\ast {Division}_s+\sum {\beta}_j{X}_{st}+{\epsilon}_{st} $$
(1)
Model 2 adjusted for state fixed effects to control unobserved confounding influences that are time-invariant and state-specific. We used variance inflation factors (VIF) to check for multicollinearity in covariance between the included state-level characteristics and state fixed effects. We elected not to include the log of population and percentage of the state population that is white in Model 2 (VIF > 10) (see eq. (2)).
$$ {Y}_{st}\sim {\beta}_1\ast Eco{n}_s+{\beta}_2\ast {Year}_t+{\beta}_3\ast {State}_s+\sum {\beta}_j{X}_{st}+{\epsilon}_{st} $$
(2)
Model 3 added interaction between state and year to Model 2, in order to allow for a state-unique time trend and control for unobserved state-level factors that evolve at a constant smooth function (see eq. (3)). Models 1–3 were repeated for the broader substance use treatment variable.
$$ {Y}_{st}\sim {\beta}_1\ast Eco{n}_s+{\beta}_2\ast {Year}_t+{\beta}_3\ast {State}_s+{\beta}_4\ast {State}_s\ast {Year}_t+\sum {\beta}_j{X}_{st}+{\epsilon}_{st} $$
(3)
To investigate whether economic conditions have an asymmetric effect on the number of substance use treatment admissions during economic upturns and downturns, following the work by Mocan and Bali [26], Model 4 modified Model 3 by defining the number of substance use admissions as an asymmetric function of two decomposed unemployment rates during economic downturns (state unemployment rate in periods when it is higher than the prior period) and economic upturns (state unemployment rate in periods when it is lower than the prior period) (see eq. (4)).
$$ {Y}_{st}\sim {\beta}_1\ast Eco{n}_{down}+{\beta}_2\ast Eco{n}_{up}+{\beta}_3\ast {Year}_t+{\beta}_4\ast {State}_s+{\beta}_5\ast {State}_s\ast {Year}_t+\sum {\beta}_j{X}_{st}+{\epsilon}_{st} $$
(4)
Econdown and Econup were constructed using state unemployment rates. Econdown equals unemployment rate in periods when it is higher than the prior period, and Econup equals 0 during economic downturns. Econup equals unemployment rate when it is lower than the prior period, and Econdown equals 0 during economic upturns. The Xst is a vector of other covariates included in the model.
$$ {Y}_{st}\sim {\beta}_1\ast {R}_{st}+{\beta}_2\ast {UR}_{st}+{\beta}_3\ast {R}_{st}\ast {UR}_{st}+{\beta}_3\ast {Year}_t+{\beta}_4\ast {State}_s+{\beta}_5\ast {State}_s\ast {Year}_t+\sum {\beta}_j{X}_{st}+{\epsilon}_{st} $$
(5)
To test the moderation effects of economic recessions (2001, 2008–09) on the relationship between economic conditions and substance use treatment, Model 5 modified Model 3 by adding the economic recession indicator (R) and an interaction term between the state unemployment rate (UR) and economic recession indicator (R) (see eq. (5)). All tests were two-sided and used a 5% significance level. All of the analyses were performed using SAS 9.4 (SAS Institute, Inc., Cary, NC) and R 3.5 (R Foundation for Statistical Computing, Vienna, Austria).