Bayesian variable selection for analyzing longitudinal substance abuse treatment data with informative censoring
*Susan M Paddock, RAND Corporation
Keywords: variable selection, conditional linear model, informative censoring, non-ignorable non-response, Bayesian methods, pattern-mixture model
Measuring the process of care in substance abuse treatment requires analyzing repeated client assessments at critical time points during treatment tenure. These assessments are frequently censored due to early departure from treatment. Informative censoring is often characterized by the last observed assessment time. However, if missing assessments for those who remain in treatment are attributable to logistical reasons rather than to the underlying treatment process being measured, then length of stay might better characterize censoring than would time of measurement. In this talk, I will describe how to incorporate Bayesian variable selection into the Conditional Linear Model to assess whether time of measurement or length of stay better characterizes informative censoring while incorporating uncertainty about the effect of censoring on treatment process change into the analysis.