An Extension of Unanchored Matching-Adjusted Indirect Comparison to Verify the Results of the Comparisons Between Poorly Overlapping Studies

OBJECTIVES: Matching-adjusted indirect comparison (MAIC) is a method of choice for unanchored comparisons. However, when the overlap between trial populations is poor for at least one prognostic factor or effect modifier, the effective sample size (ESS) is small. Therefore, poorly overlapping variables are often omitted, leading to ‘incomplete models’ likely to provide biased results. We present a procedure for investigating the impact of poorly overlapping parameters on the outcomes of the MAIC.

METHODS: The implications of omitting variables were analyzed in simulated scenarios, altering the sample sizes (20-200 patients) and levels of overlap (-2.2–0 z-scores) for binary or continuous variables. We also proposed an extension of MAIC which simulates outcomes for individual patients based on different values of the poorly overlapping variable and assumed parameters characterizing the influence of this variable on health outcomes. This allows for investigating the bias of incomplete MAIC models and re-including problematic predictors/effect modifiers into the inference.

RESULTS: The inclusion of poorly overlapping predictors/effect modifiers, compared with incomplete models without these variables, resulted in reduced ESS and broader confidence intervals, which was exacerbated in studies with small samples. Poor overlap and limited samples were also associated with higher variability in the treatment effects. Consequently, the results of incomplete models were more stable and precise, although more biased. The extension of MAIC provided the minimum and maximum values of the measures of association between poorly overlapping variables and health outcomes required to change the conclusions of incomplete models.

CONCLUSIONS: The proposed extension allows for validating the results of MAIC excluding poorly overlapped variables and prevents erroneous conclusions due to violations of the inherent assumptions of unanchored MAIC. The method is the most useful for small-sample studies.