Examining inequalities: cautions and considerations when shifting intersectional MAIHDA from a descriptive to a causal inference framework

Intersectional MAIHDA (multilevel analysis of individual heterogeneity and discriminatory accuracy) is a method for analysing intersectional inequalities that is increasingly popular in health-related research. The method is characterised by its unusual application of a standard multilevel modelling framework, which allows users to describe inequalities in an outcome across multiple intersecting characteristics simultaneously. It does this by nesting individuals within the relevant combination of intersecting characteristics, or strata. Intersectional MAIHDA and its closely related approaches (including multicategorical MAIHDA and Eco-Intersectional Multilevel models) were initially conceived as descriptive methods, yet some recent applications stray into causal interpretations. There are also several notable risks for analyses situated in the grey area between description and causal inference, specifically where strata can plausibly be inferred as important for both selection models (which may be causal), and descriptive analytic models (which are intentionally not). In this study, we consider three key issues that warrant attention: (1) how to interpret the coefficients of stratum-defining variables, (2) the potential for inadvertent collider adjustment and generalisability bias, illustrated using simulated data, and (3) the inclusion of 'control' variables. We provide recommendations for producing more robust descriptive research using MAIHDA.