Purely prognostic variables may modify marginal treatment effects for non-collapsible effect measures

In evidence synthesis, effect measure modifiers are described as variables that induce treatment effect heterogeneity at the individual level, through treatment-covariate interactions in an outcome model. As such, effect modification is defined with respect to a conditional measure. However, marginal effect estimates are required for population-level decisions in health technology assessment. For non-collapsible measures, purely prognostic variables that do not predict response to treatment at the individual level may modify marginal treatment effects at the population level. This has important implications for recommended practices. Firstly, unadjusted indirect comparisons of marginal effects may be biased in the absence of individual-level treatment effect heterogeneity. Secondly, covariate adjustment may be necessary to account for cross-study imbalances in purely prognostic variables. Popular summary effect measures in evidence synthesis such as odds ratios and hazard ratios are non-collapsible. When the relevant target estimand is a marginal effect, the use of collapsible measures would reduce dependence on model-based covariate adjustment for transportability and facilitate the selection of effect measure modifiers.