OBJECTIVES: This study aimed to evaluate the impact of multilevel modelling as compared to ordinary least squares regression on the joint uncertainty surrounding costs and effects in cost-effectiveness analysis (CEA) using data with a multilevel structure.

METHODS: Thousand datasets were simulated with varying correlation coefficients between costs and effects and Intracluster Correlation Coefficients (ICCs) for costs and effects. Ordinary least squares regression (OLS) and multilevel modelling (MLM) were used to estimate cost and effect differences. The performance of the methods was assessed according to confidence interval coverage, standard errors (SEs) of the cost and effect differences, and the mean change in SEs. Joint uncertainty around costs and effects was estimated using bootstrapping and was summarized in cost-effectiveness acceptability curves (CEACs) for both methods.

RESULTS: MLM resulted in smaller standard errors of cost and effect differences as well as narrower confidence intervals. The higher the ICC, the more precise MLM was as compared to OLS (e.g. SE of MLM versus OLS 8% smaller at an ICC of 0.05 and 30% smaller at an ICC of 0.30). At ICCs of 0.20 and more, the difference in the probability that the intervention was cost-effective in comparison with control between OLS and MLM increased, with MLM showing higher probabilities of cost-effectiveness. Results were similar across datasets with varying correlation coefficients between costs and effects.

CONCLUSIONS: Overall, multilevel modelling (MLM) was more precise than Ordinary Least Squares regression (OLS), especially for higher ICCs. However, this did not alter the conclusion about the cost-effectiveness of an intervention as compared to control as shown by the CEA curves. This was a simulation study and further research should investigate whether results are similar in empirical datasets.