This paper warns that a common practice in health economic models can lead to mistaken answers. Model-based cost-effectiveness analysis (CEA) is often used in health technology assessment (HTA) to decide if a new treatment is worth adopting for a whole country or other target population. It is common to take a treatment effect from an international randomized controlled trial (RCT) and apply it to a country-specific baseline risk. The authors show that whether an input is a population average (a marginal estimate) or a subgroup/individual-specific number (a conditional estimate) matters a lot for the final cost-effectiveness result.

The paper explains two key ideas in plain terms. A marginal estimate is a population-averaged effect: what we expect on average across everyone. A conditional estimate is the effect for particular people or subgroups, after accounting for characteristics such as age or health status. The authors also explain collapsibility, which is about whether you can get the population average from subgroup effects by simple averaging. For example, the mean difference and the risk difference (RD) are directly collapsible. The relative risk (RR) is collapsible but the averaging weights depend on baseline risk. The odds ratio (OR) and hazard ratio (HR) are non-collapsible, so subgroup and population estimates can differ even when there is no interaction between treatment and other factors.

To study the implications for CEA, the researchers first set out the ideal modelling target for an HTA decision: a marginal cost-effectiveness estimand, meaning the expected effect in the target population. They derive that the most rigorous way to reach that target is an individual-level simulation. In such a simulation the model carries conditional inputs — baseline risk, treatment effect parameters, prognostic effects and any effect modifiers — predicts outcomes for individuals, and only then averages those outcomes over the target population. This late averaging yields the desired population-average (marginal) result.