Abstract:
|
Random-coefficient pattern-mixture models (RCPMMs) have been proposed for longitudinal data when dropout is thought to be nonignorable. An RCPMM is a random-effects model with summaries of dropout time included among the regressors. The basis of every RCPMM is extrapolation. We review RCPMMs, describe various extrapolation strategies, and show how analyses may be simplified through multiple imputation. Using simulated and real data, we show that alternative RCPMMs that fit equally well may lead to very different estimates for parameters of interest. We also show that minor model misspecification can introduce biases that are quite large relative to standard errors, even in fairly small samples. For many scientific applications, where the form of the population model and nature of the dropout are unknown, interval estimates from any single RCPMM may suffer from undercoverage, because uncertainty about model specification is not taken into account.
|