Evaluating interventions with differential attrition: the importance of nonresponse mechanisms and use of follow-up data.

Evaluations of psychological interventions are often criticized because of differential attrition, which is cited as a severe threat to validity. The present study shows that differential attrition is not a problem unless the mechanism causing the attrition is inaccessible (unavailable for analysis). With a simulation study, we show that conclusions about program effects (a) are unbiased when there is no differential attrition, even with usual complete cases analysis; (b) may be severely biased when based on usual complete cases analyses and there is differential attrition; (c) are unbiased when based on the expectation-maximization (EM) algorithm, even when there is differential attrition, as long as the attrition mechanism is accessible; and (d) are biased, even with the EM algorithm, when the attrition mechanism is inaccessible. Following Little and Rubin (1987), we advocate the collection of new data from a random sample of subjects with initially missing data. On the basis of these data, we propose a simple correction to the EM algorithm estimates. In our study, the correction produced unbiased estimates of program effects parameters, even with an inaccessible attrition mechanism and substantial differential attrition.