Objective

To test for any long-term effects on the death rates of domestic assault suspects due to arresting them versus warning them at the scene.

Methods

The Milwaukee Domestic Violence Experiment (MilDVE) employed a randomized experimental design with over 98 % treatment as assigned. In 1987–88, 1,200 cases with 1,128 suspects were randomly assigned to arrest or a warning in a 2:1 ratio. Arrested suspects were generally handcuffed and taken to a police station for about 3 to 12 h. Warned suspects were left at liberty at the scene after police read aloud a scripted statement. Death records were obtained in 2012–13 from the Wisconsin Office of Vital Statistics and the Social Security Death Index, with the support of the Milwaukee Police Department.

Results

In the first presenting case in which the 1,128 were identified as suspects, they were randomly assigned to arrest in 756 cases and to a warning in 372. No clear difference in death rates from all causes combined (d = 0.04) was ever evident between the groups, or for five of the six specific categories of cause of death. However, a clear difference in homicide victimizations of the suspects emerged between those arrested and those warned. At 23 years after enrolment, suspects assigned to arrest were almost three times more likely to have died of homicide (at 2.25 % of suspects) than suspects assigned to a warning (at 0.81 %), a small to moderate effect size (d = 0.39) with marginal significance (two-tailed p = 0.096; relative risk ratio = 2.79:1; 90 % CI = 1.0007 to 7.7696). Cox regressions controlling for suspects’ stakes in conformity (employment and marriage) show that homicide victimization for arrested suspects is three times that of warned suspects (p = 0.07), although no interactions are yet significant. Logistic regression with more covariates increases arrest effects on homicide to 3.2 times more than warnings (p = 0.06).

Conclusions

Suspects randomly assigned to arrest died from homicide at a consistently higher rate than controls over a two-decade period, but the difference was not statistically discernible until the 22nd year after assignment. Long-term follow-up of randomized experiments is essential for detecting mortality differences that substantially affect cost–benefit analyses of criminal justice practices.