In mental health research, it is common for there to be numerous variables measured, many of which may have some missing values. When sample sizes are modest, existing strategies that address missing data using multivariate statistical models can easily have too many parameters for the available data to estimate. This R01 proposal-revises an earlier application, seeking support to evaluate a newly-developed factor-based multivariate statistical method for imputation of missing values, to extend the method to accommodate longitudinal factor-analysis models, to integrate software programs into a user-friendly Windows-based interface, and to snake the programs available to researchers over the world-wide web. Two applied problems motivate the present work, one from a suicide prevention study by Rotheram-Borus, Piacentini, Van Rossem, et al. (1996), where because of scattered missing values only 107 of 140 subjects for whom data were collected were available for model fitting, and another from an investigation of medication use and unmet need among children who live in foster care in Los Angeles County, where scattered missing values resulted in only 216 out of 302 subjects being available for model fitting (Zima, Bussing, Crecelius, Kaufman, Belin 1999). The proposed statistical methods offer the prospect of including all subjects in these analyses, of incorporating information from partially observed cases, and of accurately reflecting uncertainty. Ideas from factor analysis are central to the strategy.
