OBJECTIVES: Matching-adjusted indirect comparison (MAIC) is a common method of population-adjusted indirect treatment comparison between two studies. It uses a propensity score (PS) weighting approach which is sensitive to poor effect modifier (EM) overlap and small sample size of the index trial as reweighting often leads to significantly smaller effective sample sizes. G-computation is a marginalization method that can achieve more accurate estimates than MAIC when EM overlap is poor. Random forest (RF) is a non-parametric ensemble technique that averages outcomes from multiple decision trees and can weight patient characteristics based how many times any pair of subjects ends up in the same terminal nodes. This study aimed to evaluate and compare the convergence and fitting of RF and G-computation in sample sizes <100.

METHODS: Data were simulated for an anchored two-study comparison (AB and AC) with three treatment levels. The MAIC included five covariates: two EMs (age, time since diagnosis) and three prognostic variables (smoking status, race, sex). Weights were estimated to match the between-trial EM distributions. MAICs using RF, G-computation, and PS approaches were applied over 1,000 iterations.

RESULTS: MAIC with RF converged in all iterations (sample size = 50), while MAICs with PS and G-computation converged in zero and 34 out of 1,000 iterations, respectively. Mean absolute error (MAE, i.e., absolute difference between the point estimates of the log odds ratio of treatment C vs B) was significantly lower using the RF approach and their true value averaged over the converging G-computation iterations (MAE=0.94 vs. 1.42, p<0.001).

CONCLUSIONS: This simulated study demonstrated that RF was an alternative, highly accurate MAIC method when there is a small sample size and poor imbalance. Additional simulation and patient-level data studies should be conducted to explore results with varied sample sizes, sparse data, and number of covariates.