Is intersectionality selective? The role of collider bias

We examine whether observed intersectional inequalities may arise from selection processes rather than underlying social mechanisms. Intersectionality is commonly operationalized using interaction terms in regression models or approaches such as Multilevel Analysis of Individual Heterogeneity and Discriminatory Accuracy (MAIHDA). An implicit assumption is often made that the analyzed sample reflects the population of interest, so that inference from the sample can be extended to that target population. Using theory illustrated by directed acyclic graphs, and empirical evaluation through Monte Carlo simulations, we demonstrate that conditioning on a selection variable influenced by social characteristics and the outcome induces collider bias. That is, it occurs when selection into the analytic sample depends jointly on the intersectional variables and the outcome of interest. In a data-generating process where the outcome is independent of intersectional variables, restricting analysis to a selected sample produces apparent intersectional inequality. These spurious findings emerge regardless of the method used, reflecting the fact that quantitative methods applied to data that have undergone selection processes can produce misleading conclusions. Collider bias may generate, amplify, attenuate, or reverse apparent intersectional inequalities. Importantly, if the data-generating mechanism induces collider bias, it propagates through all downstream quantities that depend on the observed joint distribution, not just interaction components evaluated here. Therefore, general descriptions of inequality, and not just the interaction components, will also be biased relative to the target population. When drawing conclusions about a target population from an analytic sample, accounting for selection is essential for valid estimation and interpretation of intersectional patterns.