Data on suicide and homicide deaths were obtained from the Multiple-Cause of Death files for 1990–2004, collected by the National Center for Health Statistics. Files containing individual-level data were obtained through the National Bureau of Economic Research (http://www.nber.org/data/multicause.html
). Data on deaths between 1990 and 2004 were compiled; 1990 was chosen as the start date because it was a census year, thus enabling a precise estimate of the population at-risk for the beginning of the observation period. The year 2004 was the last year for which publicly accessible, geographically identified mortality data was available.
In order to estimate risk ratios, it was necessary to merge death record data with corresponding records on the living population. Population data were culled from 1% samples of individual-level data from the 1990 U.S. Census, the 2000 U.S. Census, and the 2004 American Community Survey (ACS). The ACS is an ongoing population survey conducted by the U.S. Census Bureau, phased in as a replacement to the long form of the U.S. census. The 2004 sample contained data on 353,220 individuals in our targeted age-range, or 0.44% of the U.S. population. All population data were obtained from the Integrated Public Use Microdata Series maintained by the Minnesota Population Center (Ruggles et al., 2010
Our analytical approach relies on differences in policy within birth cohorts, and requires that there are changes in policy over time. Therefore, we limited analyses to the population who approached age of majority during the period in which minimum legal drinking ages were in flux. This population consists of individuals who turned 18 during the years 1967 to 1990, corresponding to birth years 1949 to 1972. We also limited our analyses to records for individuals who could be classified as non-Hispanic White, non-Hispanic Black, or Hispanic of any known or unknown origin. We limited analyses to these large race/ethnicity categories to maximize comparability across death certificate codings from different jurisdictions and different eras. Inclusion of an “other” category would have resulted in a group that was heterogeneous with respect to race and, more problematically, would have changed significantly over time, both as a result of coding changes in the vital statistics system and demographic changes in the United States. We included only records for individuals born in the 50 U.S. states and the District of Columbia in order to facilitate estimation of MLDA exposure. Likewise, we tabulated only deaths occurring in those areas.
We used two approaches for estimating the “control” population, that is, the living population, and those who died from causes other than suicide and homicide. For our core analyses, we used the 1990 U.S. Census 1% micro-extract as the control population. Although this approach includes decedents from suicide and homicide as part of the control group, this represents only 0.25% of the population, so the error introduced by this approximation is negligible. The advantage of this approach is that the census extract is a precise 1% sample of the population born between 1949 and 1972 and alive at the beginning of our 1990–2004 observation window. The disadvantage of using single-year population data is that secular trends with respect to year-of-death (period effects) cannot be modeled because the controls yield no information about the population in the years beyond 1990. Hence, for regression analyses that included period as a covariate, we used an alternative approach to estimating the control population. In this second approach, we used estimates of the population for every year between 1990 and 2004 to produce a “time-averaged” population sample. The sample was assembled by combining the 1990 and 2000 U.S. census samples with the 2004 American Community Survey sample. Linear interpolation was used to estimate population data for intercensal years. Further details are provided in the Supplemental Material
Outcomes, Covariates, and Population Estimates
From the complete set of death records occurring between 1990 and 2004, we selected records for which either suicide or homicide was listed as the underlying cause, or among the contributing causes of death using ICD-9 and ICD-10 codes (World Health Organization, 1980
; World Health Organization, 1992
). For 1999–2004 records, ICD-10 codes for suicide include X60-X84 and Y87, corresponding to intentional self-harm and sequelae of intentional self-harm; ICD-10 codes for homicide included X85-Y09 and Y89 (assault and sequelae of assault). For 1990–1998, suicide records corresponded to ICD-9 codes of E950–E959, and homicide codes included E960–E969. ICD-9 to ICD-10 comparability ratios for both of these outcomes are very close to 1.0 (Anderson et al., 2001
Covariates included variables that were available in both the mortality and population data sources: year of birth, state of birth, whether the subject resided in their birth-state at time of death or census, year of death or year of census, sex, educational attainment, and race/ethnicity. Educational attainment was collapsed into a binary variable indicating more than 12 years of education (equivalent to a high school degree) or 12 or fewer years of education. Race/ethnicity was categorized as non-Hispanic White, non-Hispanic Black, and Hispanic categories. A dichotomous variable to indicate whether an individual resided in their birth state at time of death or census was created by comparing state of birth and state of residence variables. This variable was used as a measure of the likelihood that a person resided in their birth state when they were between 18- and 21-years-old. Those who resided in their state of birth as adults were assumed to be more likely to also have dwelled in their home state during early adulthood, although this does not preclude the possibility that such individuals left their state temporarily. We refer to these subjects as “likely non-movers”, and some analyses were limited to this subset of the data in order to test the robustness of our results to approximations described below.
Estimation of MLDA Exposure
We considered those who were permitted to drink legally prior to age 21 as having been “exposed” to permissive MLDA. Those who were not permitted to drink legally until age 21 were considered “unexposed”. To precisely assess exposure, we would need to know where individuals resided between the ages of 18 and 21. However, death and census records contain only state of birth and state or residence at time of observation. Therefore, we used state of birth as a proxy for state of residence during the ages of interest. Several tests were conducted to examine the validity of the proxy approach and the robustness of results to alternative approaches. First, we used state-of-residence at time of death or census as the proxy, rather than state-of-birth. Second, we conducted analyses that were limited to “likely non-movers” (individuals who resided in their state of birth at the time of their death or census). Finally, to quantitate the degree of error introduced into the exposure variable by the proxy approach, we culled 18–20 year old respondents from the 1980 and 1990 censuses to examine the percentage of those who lived in their birth states, and the percentages of those whose exposure would be correctly assigned under the assumption that they lived in their birth states, regardless of whether they had moved.
Precise dates of birth were not available from public-use death records, so date-of-birth was estimated as the mid-point in the reported year of birth. MLDA exposure was then estimated from state and year of birth. Month and year of changes in MLDA laws were coded based on data from published research articles and from data provided by the Statewide Availability Data System, a data source for alcohol policy and other alcohol-related epidemiological information (DuMouchel et al., 1987
; O’Malley and Wagenaar, 1991
; Ponicki, 2004
; Wagenaar, 1981–82
Graphical Analysis of Age-adjusted Mortality Risk
In order to present a visual representation of the data, we prepared plots of the relative risks associated with MLDA in each of the 39 states that enacted MLDA changes that affected the birth cohorts analyzed here. These calculations, which were done only for the graphical analysis, were conducted separately for men and women. Specifically, we computed relative risks for suicide and homicide by state and sex. This was done using logistic regression to estimate the association between MLDA and each cause of death with year-of-birth, coded as a continuous variable and included as the only covariate. This covariate was included to account for the slightly lower risk for suicide and slightly higher risk for homicide associated with more recent birth years. The age-adjusted, sex-specific relative risk ratios were then plotted for each state.
Analytical Design and Statistical Methods
The analytical approach exploits changes in policy within states over time, and differences in policy between states at any given time, to estimate the policy effect independently of state and time confounders. (Theory behind this approach is provided in the Supplemental Material
). The general model is outlined in Equation 1, below. It expresses the outcome variable, Y, as a function of (left-to-right): a vector of birth-year fixed effects (λ), a vector of state fixed effects (α), the policy effect (β) where policy (MLDAXst
) is a function of state and birth year, a vector of all other individual-level covariates such as demographic variables (X) and an error or disturbance term (ε).
In this case, Y represents probability of death; MLDAXst is a binary indicator coded as one for subjects who were legally permitted to drink prior to age 21, and zero for those who were not. MLDAX is determined by state (s) and birth year (t), with β being the parameter of interest; namely, the log of the odds ratio for the association between exposure to MLDA<21 and death by suicide or homicide. For some analyses, we also included interactions between the state indicators (α) and birth-year coded as a continuous variable to account for unobserved state-level variables that may change over time.
Parameter estimates and standard errors were calculated using the generalized estimating equation method incorporated into SAS (Version 9, SAS Institute, Cary, NC) using state as the clustering unit (Arellano, 1987
; Liang and Zeger, 1986
). This approach accounts for intra-correlation of observations within clusters (states) in estimating standard errors, which is the promary consideration in standard error estimation when applying this quasi-experimental approach (Angrist and Pischke, 2008
; Bertrand et al., 2004