Longitudinal data analysis is frequently complicated by drop-out. In this paper we consider several methods for dealing with drop-out afflicted data. Along with a general comparison, particular attention is paid to the consequences of model misspecification. The purpose of our approach is two-fold. We first deliberate the form of the drop-out model and compare two alternatives. Furthermore, the extent to which each method is dependent on its core assumptions is assessed through scenarios where one or more such assumptions are compromised. Second, the extent to which we can identify adequacy of model fit is investigated via recently developed diagnostics. These twin targets are pursued via simulation scenarios and application to a schizophrenia trial of over 500 patients with near 50 per cent drop-out.