Abstract:
|
Multiple imputation (MI) is a flexible method that allows us to make a variety of clinically meaningful assumptions about subjects whose observed data cannot be used for the planned estimand. Multiple imputation can follow the assumption that outcomes are missing at random, but can as easily make some other, perhaps more sceptical, assumption about what happens to subjects in a clinical trial who withdraw early or whose observed data does not serve the estimand of the trial. Also, MI can very naturally implement pattern mixture models, where assumptions differ by reason for discontinuation or by some other variable. Because MI analyses can have a clear clinical interpretation MI is now often used in clinical trials, but until recently MI has mostly been used with continuous outcomes. Researchers have now developed extensions of Rubin's classical MI that give similar flexibility for time-to-event and recurrent-event outcomes. This presentation summarises these extensions of MI and gives examples of their use.
|