611 – Missing Data Imputations
A Simulation Study to Compare Multiple Imputation Methods Under Missing Not-at-Random Assumption
Jianjun (David) Li
Pfizer, Inc.
Lingfeng Yang
BMS
Because missing values are neither observable nor depend on observed values, data missing not at random (MNAR) poses unique challenges in data analysis. A simulation study that compares four multiple imputation (MI) methods under MNAR was conducted to address regulatory concerns of missing data in a clinical trial. The four MI methods for comparison are imputation with the mean of observed placebo values, imputation with the change since last visit of observed placebo values, imputation with observed means in each arm, and last z-value carried forward. A variety of scenarios of missing data proportions in drug and placebo arms was considered to evaluate these methods in terms of power and type I error rate. The simulation study shows that (1) imputation with the change since last visit of observed placebo values performs best among the four MI methods; (2) intuitively conservative method (e.g., imputation with observed placebo mean) does not necessarily protect against false positive findings that favor active drug better than imputation with the change since last visit; (3) analysis results are not sensitive to the number of imputations if the number of imputations is 10 or