OBJECTIVES: Missing data are a common occurrence and could have a significant effect on cost-effectiveness conclusions. Past studies recommended multiple imputation (MI) at the aggregate level to impute total costs or health utilities. However, costs and health utilities are products of multiple individual variables. Therefore, our objective was to compare the performance of MI at the aggregated level vs the alternative imputation at the individual response level.

METHODS: We simulated the missing data based on observed complete data of 318 patients from the Delphi-trial (Dementia: life- and person-centered help). Healthcare resource uses and corresponding unit costs were used to calculate the cost. SF-6D-derived health utilities were used to estimate QALYs. 300 reputations of nine different missing data scenarios were generated that differed according to the proportion of missing data (10%,20%,40%) and the pattern of missing data. MI by Chained Equation was used to impute missing values. Mean differences, standard deviation and range (25% to 75% quantile) in the percentage change from true incremental ratios were used to assess the precision.

RESULTS: For all scenarios, the individual response level MI was more precise than the aggregated level, represented by lower deviations of incremental costs (up to 9% vs. 27%, 15% vs. 33% and 27% vs. 65%) and incremental QALYs (up to 15% vs. 19%, 15% vs. 38% and 25% vs. 77%) for the different proportion of missing data. Compared with imputing health utilities, imputing missing SF-6D items tended to generate lower QALYs.

CONCLUSIONS: MI at the individual response level is recommended when there is less than 20% missing aggregated data, despite the fact that the imputation could still change the incremental cost effectiveness ratio up to 50%. Imputing aggregated outcomes could lead a substantial deviation from true incremental ratios if more than 10% of the data is missing, which could alter decision making.