Presentation (CTI)

OBJECTIVES: The modern era of targeted therapies and narrowed indications may be a boon for patients but poses a challenge for indirect comparison of health technologies: clinical trials of targeted therapies can struggle to recruit many patients; and competitors often set a new standard of care before trials conclude. In newer indications, there are few competing treatments, and the method of Matching-Adjusted Indirect Comparison (MAIC) is commonly applied. MAIC analyses may attempt to match many characteristics, particularly if there is uncertainty about which factors are important confounders of the treatment effect, and the method can fail under these conditions. Moreover, weighting methods like MAIC reduce the effective sample size (ESS) available for comparison, which in practice tends to favor the current standard of care if precision is too low to demonstrate a significant improvement under the novel treatment. This limits adoption of new treatments which could benefit patients. The objective of this research is to determine whether regularization of MAIC is a viable solution to these challenges.
METHODS: The MAIC method of Signorovitch (2010) - based on a logistic parameterization of the propensity score - was used as the foundation of the regularized MAIC. Logistic parameters were penalized using an L1 (lasso), L2 (ridge), or combined (elastic net) penalty term, and weights were computed from the regularized parameter estimates.
RESULTS: Statistical simulation with 100 patients per cohort and 10 variables to match demonstrated a bias-variance tradeoff, resulting in smaller errors for the regularized weights and markedly better ESS compared to the default method. Moreover, when imbalances between the two cohorts were large, the regularized method had a solution even if the default method had none.
CONCLUSIONS: Regularized MAIC should be considered a viable alternative to default MAIC, particularly when ESS is limited or when default MAIC has no solution.