Juan Merlo
This essay traces the two-decade development of multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA). It begins at a course in London in 1999, where I first saw a clear contextual association coexist with an intraclass correlation of only 1%. This paradox exposed two paradigms in epidemiology: one on group averages, the other on the heterogeneity around them. In this Journal, beginning with a 2003 editorial, the heterogeneity paradigm crystallised into a framework of three components: the specific contextual effect, the general contextual effect and discriminatory accuracy. The three must be read together: a large difference between contextual means can coexist with almost no ability to classify individuals, which measures Rose's prevention paradox.MAIHDA is a conceptual framework rather than a statistical technique, agnostic to context definition, statistical method and social theory. It analyses unidimensional contexts, such as geographical areas or healthcare institutions, and multidimensional contexts, such as intersectional strata and other cross-classifications. Intersectionality is just one of the theories MAIHDA can carry. None of the three components requires a random-effects model, and for describing inequalities, simple-means MAIHDA is canonical, avoiding untenable exchangeability assumptions and shrinkage.MAIHDA extends to causal inference: in a weighted pseudo-population it maps the heterogeneity of causal effects and how far that heterogeneity is contextually organised. As an alternative to mean-centric risk-factor epidemiology, MAIHDA harmonises population and precision epidemiology, and makes the mapping of health inequalities a civic function.