Defining and Evaluating Bias in Longitudinal Trials with NMAR Missing Data
View Presentation View Presentation
*Ronald W. Helms, Rho Inc. and UNC 


Although it is well known that not missing at random (NMAR) missing data from longitudinal clinical trials inflict biases on estimators and consequently lead to tests of corrupted hypotheses, the model equation, , the definition of the primary parameter , and the definition of all depend critically on E[Y], which is not necessarily straightforward. To illustrate, one basic question is: what is the expected value of a NMAR missing element of Y? This presentation gives a more complete introduction to the problem, elaborates on the basic question, and presents a potentially satisfactory definition of E[Y] in this context that leads directly to definitions of ß and Bias. These definitions provide a basis for evaluating and comparing biases of specific statistical analyses and/or imputation methods for NMAR incomplete data from longitudinal clinical trials. We illustrate the methods of evaluating and comparing biases and the effects of testing corresponding corrupted hypotheses via analyses of realistic datasets from longitudinal analgesic studies.