When Missing Does Not Mean Lost [OED]

Background: Missing data due to subject dropout from study is common in late phase clinical trials. It is possible for missing to be equivalent to lost and has no relevance to time course of the data. On the other hand, circumstances exist where compelling reasons could be formed to argue that dropout is related to underlying unobserved data and thus informative. In this case, treating missing as simply lost may be inefficient and can also lead to biased estimation of time course of the data. This may be the case with some common methods used in practice to deal with missing data, including the “last observation carried forward” (LOCF), or mixed-effect modeling carried out in ordinary manners. The topic of informative missingness in the context of longitudinal data has received considerable attention in the statistical literature.

Methods: A brief overview will be given on different statistical approaches and the philosophical difficulty of unverifiable assumptions. Motivations will be given on the need of appropriate dropout handling under different missing data mechanisms using the hazard function to model dropout, assessing goodness of fit, and jointly developing parametrical model on both observed longitudinal data and dropout. Model development can be relatively easily implemented in NONMEM. Using the selection model approach, data described in [1] are explored with more refined dropout models to evaluate the influence of dropout model on the ability of the joint model to predict observed longitudinal data patterns.

Results and discussion: The joint model approach provides a framework to test whether missing means lost, and how to assess the consequences. Modeling informative dropout does require assumptions unverifiable from current data, such as that the underline longitudinal data patterns are known. Dropout diagnostic plots are possible but not easy. The Weibull distribution can be a better alternative to describe the dropout, compared with the constant hazard model. Differences in dropout model will affect predicted average longitudinal data trends, usually done with simulation. The influence of dropout model on estimation of longitudinal data trends appears minor, although further investigations are still needed.

[1] C. Hu and M. Sale, A joint model for nonlinear longitudinal data with informative dropout, Journal of Pharmacokinetics and Pharmacodynamics, 30(1): 83-103, 2003.