PNS211 ASSESSING POPULATION-ADJUSTED INDIRECT COMPARISON METHODS (MATCHING-ADJUSTED INDIRECT COMPARISON AND SIMULATED TREATMENT COMPARISON) USING MONTE-CARLO SIMULATIONS

Abstract

Traditional methods for network meta-analysis and indirect comparisons rely on aggregate data, but use of individual patient data (IPD) to perform population-adjusted indirect comparisons (PAIC) is becoming increasingly common in submissions to health technology assessment bodies. When IPD is available in a study along with summary statistics from another study, PAIC can be conducted if effect modifiers exist. Here, we performed a simulation study to determine how two proposed methods, Matching-Adjusted Indirect Comparison (MAIC) and Simulated Treatment Comparison (STC), perform under different types of scenarios and outcomes. We simulated both continuous and binary outcomes of two treatments having three effect modifiers (binary, count, and continuous) along with a non-effect modifier for 200, 400, 600, and 1000 subjects in IPD. The simulations were performed for three different scenarios: three effect modifiers (correctly specified model), two effect modifiers (mis-specified), and three effect modifiers with a non-effect modifier (over-specified). The mean-squared error (MSE) or root mean-squared error (RMSE) was used as a test measure for comparing MAIC and STC. For binary outcomes when the model is correctly specified, the MSEs for MAIC are much smaller than those for STC (e.g. 1.85 vs. 26.48 for 200 subjects) for all numbers of subjects examined. For continuous outcomes with a correctly specified model, the RMSEs for MAIC are also smaller than those for STC for all numbers of subjects. When the model is mis-specified, the results for MAIC and STC can vary considerably from the correctly specified results, depending on the strength of the missing effect modifier. For binary outcomes, use of MAIC is strongly preferred over STC. For continuous outcomes, using either MAIC or STC shows improvements vs. a naive approach. In using either method, model mis-specification can lead to biased results.