Pattern-mixture models adjusting for non-ignorable dropout with administrative censoring in longitudinal studies

In longitudinal studies and clinical trials, estimation and comparison of the rates of change over time in a continuous response variable are often of primary interest. Due to various reasons, patients often drop out before the study terminates. This results in no data being available after the time of dropout. Also, in many studies, the patients are recruited at different calendar times but the follow-up is ended at a common calendar time. This is called administrative censoring in that patients who do not drop out will have different lengths of follow-up time when the study ends.; If the dropout mechanism is related to the unobserved features of the underlying disease process, e.g., the unobserved true initial value, rate of change, or true value of the response, the standard methods which ignore the missing data mechanism are biased in estimating the rates of change. In the literature, many proposed approaches for analysis of data with non-ignorable dropout require complicated computations which can not be implemented with standard statistical software. In addition, most of these methods do not allow for administrative censoring.; We present several pattern-mixture models to adjust for non-ignorable dropout, while also accommodating administrative censoring, based on the two-stage linear random effects model. The average rates of change conditional on the pattern of dropout and censoring pattern are estimated using SAS Proc Mixed. The rates then are averaged over the dropout patterns to estimate group mean rates of change, and standard errors are calculated using the delta method. We illustrate these models and compare them with the usual maximum likelihood approach assuming ignorable dropout, and the Schluchter selection model (1992) using data from a multi-center randomized clinical trial, the Modification of Diet in Renal Disease (MDRD) study. Simulations under various scenarios where the dropout mechanism is ignorable and non-ignorable with and without administrative censoring are employed to evaluate the performance of these models.