OBJECTIVES: Matching-adjusted indirect comparisons (MAICs) are a popular method of population-adjusted indirect treatment comparison used to support health technology assessment submissions. MAICs rely on a propensity score approach that rescales the weight of patients in the index trial to a target population. We sought to explore the impact of approaches used to rescale weights on the results of MAICs.

METHODS: Data were simulated for two single-arm studies from Weibull distributions varying three experimental factors: binary vs continuous covariates, level of imbalance between trials, and strength of effect treatment effects. Each patient from the target study population was assigned a weight of 1. The index study population was adjusted and assigned weights based on a method-of-moments logistic regression. Three approaches were considered for rescaling the MAIC weights: (1) raw weights; (2) effective sample size (ESS); (3) maximum rescaled weights equal to 1 (M1). A weighted Cox regression with treatment as regressor was fit and bias, root mean square error (RMSE), and coverage probability were estimated for each scaling technique over 2,000 iterations.

RESULTS: The weighted samples sizes using raw and ESS approaches were negatively correlated with the level of imbalance between the index and target trials ranging from a 3%–80% reduction. The M1 weighted sample size was up to 75% smaller than the raw and ESS methods, suggesting that MAICs rely on assigning weights substantially larger than 1 to patients to achieve a balanced population. The bias and RMSE from all population-adjusted analyses were 60% smaller than the naïve comparisons; however, a further decrease in bias between 2%–10% was observed using the M1 approach.

CONCLUSIONS: The weighted sample size of the adjusted individual patient data had a marginal impact on MAIC results. Further simulation studies should be conducted to identify the optimal scaling strategy.