OBJECTIVES: MAIC is a popular method of population-adjusted indirect treatment comparison. It uses a propensity score approach, which re-weights patient characteristics in the index trial to match the baseline characteristics of a target population. The primary challenge with MAICs is that reweighting may produce significantly smaller effective sample sizes. RF is a non-parametric ensemble technique that averages outcomes from multiple decision trees and can be used to weight patient characteristics based on the number of times any pair of subjects end up in the same terminal nodes. Our hypothesis was that the RF approach estimates the weights more accurately versus the propensity score approach, thus improving the results of MAICs.

METHODS: Data was simulated for a two-study comparison (AB and AC) involving three treatment levels. The MAIC included five covariates, an effect modifier (age), and three prognostic variables (time since diagnosis, smoking status, race, and gender). Weights were estimated to match the effect-modifier distributions between the two trials. MAICs using the RF and propensity score approach were applied over 1,000 iterations.

RESULTS: With a sample size of 300 patients, MAICs using the RF approach resulted in significantly higher accuracy/lower mean absolute error (MAE) i.e., the absolute difference between the point estimates of the log odds ratio of treatment C vs. B and their true value averaged over 1,000 iterations (MAE=1.29 vs. 1.72, p<0.001). RF also resulted in a smaller reduction in the effective sample size (73.4% vs. 86.8%, p <0.001).

CONCLUSIONS: For MAICs, using simulation studies, RF MAIC appears to improve some outcomes of the matching process, including higher accuracy and loss of sample size. Further studies, including actual patient-level data and simulation studies, should be conducted to explore results with varied sample sizes, sparse data, and varied number of covariates.