Statistical techniques for analyzing data from prevention trials: Treatment of no-shows using Rubin's causal model.

Abstract

Noncompliance is a common challenge in the analysis and interpretation of prevention trials. The authors describe new formulations of the problem based on D. B. Rubin's (1974, 1978) causal model. The formulations help clarify assumptions underlying estimation procedures and yield more efficient methods of estimation. The authors apply the methods to a trial of a job training intervention in which nearly half the participants randomly assigned to the intervention failed to attend the job training seminars. An interesting feature is the presence of covariates measured prior to treatment randomization. Versions of the model that condition on these covariates suggest positive results for the intervention in a high-risk group but no evidence of gains in a low-risk group. Prevention trials that assess the efficacy of interventions apply the interventions to groups of participants and then compare the distributions of relevant outcomes across groups. If statistically and substantively significant differences are found, the central issue is whether these differences can be attributed to causal effects of the interventions rather than to confounding factors. Randomized treatment assignment is a key design tool for limiting the undermining effects of confounding factors. The potential for bias in the assignment process is removed, and randomization balances the distribution of confounding factors across groups on average. Chance imbalances can be