PRM196 - Indirect Treatment Comparisons With Guided Matching-Based Adjustment: A Hybrid Of The Stc And Maic Techniques

Abstract

Matching adjusted indirect comparisons (MAIC) and simulated treatment comparisons (STC) produce comparisons between treatments in different trials after adjusting for imbalances between their populations. These methods overcome issues like disjointed evidence networks and heterogeneity in network meta-analyses, and can produce comparative evidence where it may otherwise be impossible. While MAIC and STC are conceptually similar, they differ in the way they adjust for population differences. The former uses weights to match on all available baseline characteristics and the latter predicts outcomes in the comparator’s population using equations. Matching on all baseline characteristics is intuitively appealing, but when there are large differences between populations or numerous, possibly correlated, variables to match, the derived weights can have an unbalanced distribution, giving a relatively small subset of patients the majority of the weight. This can lead to spurious findings and loss of precision. In STC, predictive equations can be very informative as they identify determinants of the outcome, which represent the variables that that may confound comparisons. The equations may also be useful beyond the STC for use in economic modelling. Prediction of the adjusted results for non-linear outcomes (like time-to-event endpoints) requires additional analytical steps, however. A hybrid approach that leverages the strengths of the two approaches can overcome these challenges. A predictive equation is developed as in STC to identify potential confounders among the all available variables. Balancing weights are then derived to match on these confounding variables, and used to derive adjusted outcomes via reweighting instead of prediction. The robustness of results to omitted baseline characteristics can be assessed via sensitivity analysis. Despite the reduced matching list, there could remain substantial loss of effective sample size. In such cases, tone can explore whether some precision can be gained via prediction by reverting to an STC approach.