Incomplete data due to premature withdrawal (dropout) constitute a serious problem in prospective economic evaluations that has received only little attention to date. The aim of this simulation study was to investigate how standard methods for dealing with incomplete data perform when applied to cost data with various distributions and various types of dropout. Selected methods included the product-limit estimator of Lin et al. the expectation maximisation (EM-) algorithm, several types of multiple imputation (MI) and various simple methods like complete case analysis and mean imputation. Almost all methods were unbiased in the case of dropout completely at random (DCAR), but only the product-limit estimator, the EM-algorithm and the MI approaches provided adequate estimates of the standard error (SE). The best estimates of the mean and SE for dropout at random (DAR) were provided by the bootstrap EM-algorithm, MI regression and MI Monte Carlo Markov chain. These methods were able to deal with skewed cost data in combination with DAR and only became biased when costs also included the costs of expensive events. None of the methods were able to deal adequately with informative dropout. In conclusion, the EM-algorithm with bootstrap, MI regression and MI MCMC are robust to the multivariate normal assumption and are the preferred methods for the analysis of incomplete cost data when the assumption of DCAR is not justified.
(c) 2005 John Wiley & Sons, Ltd.