OBJECTIVES: Matching adjusted indirect comparison (MAIC) is a useful tool to compare the relative effectiveness of treatments across clinical trials and inform health technology assessments. Despite its increasing popularity, there is a need to further understand the appropriate ways to apply this method. Prior studies indicate that matching the individual-patient-data (IPD) and the and aggregate-level data (AD) at the treatment arm level, rather than at the trial level, produced more precise estimates, and that simultaneously matching on mean and variance of baseline characteristics rendered similar results than when matching on the mean only. Here we investigate how these observations hold when 1) the sample size of the IPD changes and 2) the relationship between baseline characteristics and outcomes is non-linear.

METHODS: We present a simulation study where, in each instantiation, two trial populations (one for IPD and one for AD) are simulated based on different assumptions related to the baseline covariates, type of outcome and population size. MAIC is then applied using different specifications and the corresponding bias and mean square error (MSE) is computed and compared across different specifications.

RESULTS: The simulations indicated that the utility of matching on variances in addition to means depended on the linearity of the relationship between the covariates and the outcomes. The bias of the arm-to-arm versus trial-to-trial matching was impacted by sample size. When sample sizes were large enough, randomization led to well-balanced trial arms, negating the advantages of arm-to-arm adjustment. At smaller sample sizes, the trial-to-trial method could result in more stable estimates of the weights, while at moderate sample sizes, the methods trade off stability of weights for improved matching of the arms.

CONCLUSIONS: This study characterizes how the accuracy of MAIC depend on the properties of the population sample, the outcome type and the operational specifications of the method.