OBJECTIVES: Trial based-economic evaluations are essential in deciding about the allocation of scarce health care resources. Although extensive efforts are carried out by researchers to prevent missing data in clinical trials, invariably, missing data cannot be completely prevented and is commonly inadequately handled which can lead to biased results and invalid conclusions. The aim of this scoping review was to identify and explore statistical methods for dealing with missing data in trial-based economic evaluations.

METHODS: We systematically searched for methodological studies assessing the performance of statistical methods for dealing with missing data in trial-based economic evaluations in MEDLINE and EMBASE (January 2000 to July 2018). From all the included studies, we extracted information on the advantages and disadvantages of the statistical methods as well as their relative performance.

RESULTS: Fourteen eligible studies were identified. Of them, two assessed statistical methods for dealing with missing effect data, three assessed statistical methods for dealing with missing cost data, and nine assessed statistical methods for dealing with both missing cost and effect data. The included studies suggest that naive methods, although easy to implement, are likely to yield biased estimates and do not account for imputation uncertainty; therefore their application is discouraged. Multiple imputation seems to be the best statistical method for dealing with missing data in trial-based economic evaluations because it reduces bias, improves efficiency and accounts for imputation uncertainty.

CONCLUSIONS: Current studies suggest that multiple imputation is most appropriate for dealing with missing data in trial-based economic evaluations. However, future studies of different variations of multiple imputation as well as how to combine bootstrapping and multiple imputation are required. Moreover, the way of dealing with multilevel missing data and the plausibility of simultaneous missing data mechanisms (e.g. missing at random and missing not at random) within one study need further exploration.