Premium
Implementation of pattern‐mixture models in randomized clinical trials
Author(s) -
Bunouf P.,
Molenberghs G.
Publication year - 2016
Publication title -
pharmaceutical statistics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.421
H-Index - 38
eISSN - 1539-1612
pISSN - 1539-1604
DOI - 10.1002/pst.1780
Subject(s) - missing data , dropout (neural networks) , imputation (statistics) , computer science , econometrics , randomized controlled trial , statistics , artificial intelligence , data mining , machine learning , mathematics , medicine , surgery
Modern analysis of incomplete longitudinal outcomes involves formulating assumptions about the missingness mechanisms and then using a statistical method that produces valid inferences under this assumption. In this manuscript, we define missingness strategies for analyzing randomized clinical trials (RCTs) based on plausible clinical scenarios. Penalties for dropout are also introduced in an attempt to balance benefits against risks. Some missingness mechanisms are assumed to be non‐future dependent, which is a subclass of missing not at random. Non‐future dependent stipulates that missingness depends on the past and the present information but not on the future. Missingness strategies are implemented in the pattern‐mixture modeling framework using multiple imputation (MI), and it is shown how to estimate the marginal treatment effect. Next, we outline how MI can be used to investigate the impact of dropout strategies in subgroups of interest. Finally, we provide the reader with some points to consider when implementing pattern‐mixture modeling‐MI analyses in confirmatory RCTs. The data set that motivated our investigation comes from a placebo‐controlled RCT design to assess the effect on pain of a new compound. Copyright © 2016 John Wiley & Sons, Ltd.