Methods to Analyse Cost Data of Patients Who Withdraw in a Clinical Trial Setting

Background: Missing data resulting from premature study withdrawal are a common problem in the analysis of longitudinal data in clinical trials. To date, this subject has received little attention in the context of economic evaluations and with regard to the analysis of cost data. Objectives: To (i) demonstrate the impact of patients who drop out during the study on the outcomes of an economic evaluation, and (ii) to compare the mean and variation in costs after applying five different methods to deal with incomplete data: multiple imputation, complete cases analysis, linear extrapolation, predicted mean and hot decking. Study design: The study was performed using cost data collected in two randomised clinical trials comparing patients with chronic obstructive pulmonary disease receiving either tiotropium bromide or ipratropium bromide. The overall dropout rate was 17%, with the daily costs of the dropouts approximately 4 times higher than the costs of the completers. Methods: Multiple imputation is a principled method that deals with missing observations by replacing each missing observation with a set of multiple plausible values. The variance between the resulting multiple datasets is combined with the variance between the datasets to take account of the extra uncertainty that results from missing data. The outcomes after multiple imputation were compared with the results of four naive methods to deal with missing observations: complete cases analysis, linear extrapolation, predicted mean and hot decking. All costs were expressed in 2001 euros. Results: In the tiotropium bromide group, mean (standard error) costs varied from Conclusion: This study showed that the method of dealing with the data of the dropouts had a large impact on the outcomes of an economic evaluation. Information about the rate of patient withdrawal and the way data of dropouts are treated is of vital importance in assessing the results of economic evaluations and should always be reported. Multiple imputation is a principled method that can be used to deal with the data of these patients.