The effects of missing data due to study dropout on longitudinal analysis inference using outcome-dependent sampling.