Association of an Alcohol Abstinence Program With Mortality in Individuals Arrested for Driving While Alcohol Impaired

Abstract

This study examines the associations of 24/7 sobriety programs with reduced alcohol-involved offenses.

Alcohol is the third-leading cause of preventable death in the US, with alcohol-impaired driving alone claiming 11 654 lives in 2020—a 14% increase over the previous year.¹ To address this problem, some jurisdictions are using 24/7 sobriety programs, which combine alcohol abstinence orders with frequent alcohol testing (eg, twice-daily breathalyzers, monitoring bracelets) and swift, certain, and moderate sanctions for noncompliance (typically 1 to 2 nights in jail) to reduce alcohol consumption among those arrested or convicted for alcohol-involved offenses. Individual-level analyses of South Dakota’s 24/7 sobriety program demonstrate lower risk of rearrest among participants relative to nonparticipants.² While county-level analyses also document a negative association with mortality,³ this study, with instrumental variables analysis, was conducted to assess the association between 24/7 sobriety and time-to-mortality using individual-level data.

Methods

We use linked criminal history and mortality data for individuals arrested for driving under the influence (DUI) of alcohol in South Dakota during 2004 to 2011. Our analysis sample includes 11 827 24/7 participants and 48 834 nonparticipants (eMethods in Supplement 1). Mortality data through 2016 provide at least 5 years of follow-up for each individual. We used 3 approaches to examine how an individual’s 24/7 sobriety participation was associated with time to mortality. First, we estimate a Cox model to predict the hazard ratio (HR) associated with individual 24/7 sobriety participation in time to mortality controlling for demographics, criminal history, arrest year, and county fixed effects. This model may overestimate or underestimate the association depending on the nature of selection. Thus, our preferred model is a 2-stage residual inclusion Cox model of time to mortality with individual frailty (2SRI-frailty) assuming a γ distribution.⁴ To address selection, the 2SRI-frailty’s first stage exploits variation in county-level 24/7 sobriety availability to predict individual-level participation, following Kilmer and Midgette²; first-stage residuals are then included in the second-stage estimation of time to mortality.⁴ Sensitivity analyses explore the robustness of 2SRI-frailty results to alternate selection of index arrests among those with multiple DUI arrests. Additionally, as a robustness check, we estimate a bivariate probit model on the dichotomous 5-year mortality outcome—the approach used in prior analyses of 24/7 sobriety and rearrest.² These 2-stage approaches allow inference of the local average treatment effect among those predicted to participate in the program when it is assigned.⁵ Analyses were conducted using Stata version 17/MP (Stata Corp) with 95% CIs based on county-clustered standard errors. RAND Corporation’s human subjects protection committee determined the study was exempt as data were collected for nonresearch purposes.

Results

The naive Cox model produces the most conservative estimate (HR, 0.88; 95% CI, 0.77-1.00) (Table). The preferred 2SRI-frailty model leverages a strong first stage to address selection and produces a substantially larger effect size than the naive Cox model (HR, 0.55; 95% CI, 0.40-0.75). The 2SRI-frailty-based survival curves show that 24/7 participants consistently experienced lower risk of mortality than nonparticipants (Figure). The 2SRI-frailty results are robust to alternative selection of index arrests, although the magnitudes vary (Table; eMethods in Supplement 1). Additionally, the bivariate probit model estimates a 55% lower probability of death for 24/7 sobriety participants within 5 years.

Table. Estimated Association of Individual-level 24/7 Sobriety Participation With Mortality Risk^a.

	Hazard ratio (95% CI)b		Bivariate probit model on dichotomous 5-y mortality, coefficient (95% CI)
	Cox hazard of time to mortality	Cox 2SRI gamma frailty hazard of time to mortality
Comparison of mortality models
24/7 Participation	0.88 (0.77-1.00)c	0.55 (0.40-0.75)d	−0.24 (−.41 to −.06)e
Implied reductionf	NA	NA	54.70%
χ2 Of instrument strength	NA	23.72	17.10
P value	NA	<.001	<.001
No.	60 661	60 661	60 661
	Hazard ratio (95% CI)
	Cox 2SRI to frailty sensitivity 1: last arrest for control group	Cox 2SRI to frailty sensitivity 2: first arrest for both groups	Cox 2SRI to frailty sensitivity 3: last arrest for both groups
Sensitivity results for preferred 2SRI-frailty model
24/7 Sobriety participation	0.47 (0.30-0.74)d	0.645 (0.39-1.07)c	0.55 (0.35-0.87)d
χ2 Of instrument strength	24.55	24.61	40.03
P value	<.001	<.001	<.001
No.g	60 659	60 657	60 657

Abbreviations: NA, not applicable; 2SRI, 2-stage residual inclusion.

^a Individual and criminal justice covariates include gender; race; age; indicators for prior violent, property, and drug arrests; number of prior arrests; number of prior driving under the influence convictions; prior prison; probation revocation; driving under the influence convictions at current arrest; sentenced to prison for current arrest; and intercepts for arrest year, judicial district, and arrested in a nearby county during the Sturgis Motorcycle Rally.

^b 95% CIs from clustered standard errors in parentheses.

^c P < .10.

^d P < .01.

^e P < .05.

^f Implied reduction: The bivariate probit coefficient implies a reduction (−0.009) relative to a 5-year mortality of 0.016, which translates to a 54.7% reduction.

^g Missingness in covariates accounts for change in sample.

Figure. Predicted Survival Functions for 24/7 Sobriety Participants and Nonparticipants Based on Cox 2-Stage Residual Inclusion-Frailty Model.

The dotted light blue line marks the 20th quarter (end of 5-year follow-up).

Discussion

These findings add a public health dimension to the growing evidence that 24/7 sobriety improves public safety by reducing rearrest.^2,6 To our knowledge, this is the first evidence from individual-level data that such programs may also improve health outcomes. While the associations with mortality appear robust, the magnitude depends on efforts to identify treated samples and methods to address selection. We also caution that our approach addresses individual-level observables and selection, but not other concerns, such as variation in county-level implementation. Nor can we assess the extent to which deterrence vs other mechanisms drive our findings. Additional research is needed to determine if such programs, whether stand alone or combined with other approaches (eg, substance use disorder treatment), generate similar findings.

Supplement 1.

eMethods. Creating the main analytic sample

eReferences.

Supplement 2.

Data sharing statement.