Simple-Means MAIHDA puts inequality estimates back on the observed probability scale
Health inequalities are often described through model-based estimates that smooth away the very differences researchers set out to measure. This paper argues for a different standard: estimate each stratum directly from the observed individuals within it, keep the probabilities on the probability scale, and let the uncertainty remain visible rather than compressed by modeling assumptions.
Simple-Means MAIHDA, or S-MAIHDA, does exactly that. Each stratum prevalence comes straight from the observed counts, with no distributional assumption imposed on the data. The method then organizes inequality into three linked components: specific contextual effects, general contextual effects, and discriminatory accuracy of the context. In practice, this means S-MAIHDA does not flatten inequality into a single model coefficient. It separates the raw prevalence, the composition-standardized prevalence, the area-related absolute risk difference, the sociodemographic risk difference shared across areas, and the modification of that difference by the area itself.
The distinction matters at the bedside and in public health planning. If a model shrinks sparse strata toward the overall mean, a small but clinically meaningful concentration of disadvantage can disappear into statistical convenience.
Why random-effects MAIHDA can blur the pattern it is meant to measure
The usual implementation of MAIHDA relies on random effects, and that introduces shrinkage. Small strata are pulled toward the overall mean, which is useful when the goal is prediction, but less useful when the goal is faithful description of inequality. In settings where policy decisions depend on where disadvantage is concentrated, attenuation can be costly. A neighborhood, social group, or joint area-sociodemographic stratum may look less extreme than it truly is, and the resulting map of inequality becomes smoother than the lived reality.
S-MAIHDA treats the observed stratum proportion as the canonical estimate. Its confidence intervals express the uncertainty honestly, rather than letting model structure dictate the size of the difference. Every quantity is computed from a simple table of counts by stratum, which means the individual records do not need to be shared for the analysis to proceed. The method is implemented in a Stata command, with free code also available in Python and R.
A probability-scale description that preserves structure
The method operationalizes the three MAIHDA components on the probability scale. The specific contextual effects are expressed as raw and composition-standardized prevalences, plus an exact decomposition of each observed stratum prevalence into a reference level, the absolute risk difference of the area, the absolute risk difference of the sociodemographic stratum common to all areas, and the area-driven modification of that difference. The general contextual effects include the variance partition coefficient, the standardized contextual area under the ROC curve, and the contextual structuring of between-stratum inequality.
The discriminatory accuracy is described with the contextual area under the ROC curve and with sensitivity and specificity at the population prevalence used as a potential intervention threshold. Those are not decorative statistical extras. They answer a direct clinical and policy question: how well does context separate one individual from another when the goal is to target action, not merely to summarize variation.
Malmö data show two very different inequality structures
The method was illustrated with 2006 register data from Malmö, Sweden, covering 43,291 individuals and 300 area-sociodemographic strata. Two outcomes behaved in strikingly different ways. Psychotropic medication use appeared almost purely sociodemographic, remaining stable across areas with weak contextual structuring. Its variance partition coefficient was about 4%, its standardized contextual area under the ROC curve was about 63%, and contextual modification of inequalities was close to 0%.
The second outcome moved in the opposite direction. Choice of a private general practitioner was strongly geographical, with a variance partition coefficient of about 11%, a standardized contextual area under the ROC curve of about 72%, and contextual modification of inequalities of about 17%. In wealthier areas, sociodemographic differences were reshaped and amplified. The local environment did not merely sit beside social position; it changed how social position translated into observed behavior.
Psychotropic medication use looked socially patterned but geographically steady. Private primary care use was geographically structured and socially amplified.
What the comparison with RE-MAIHDA changes in practice
Random-effects MAIHDA attenuated the inequalities in both outcomes. That is the central warning in the paper. When shrinkage pulls sparse strata toward the mean, the output may be cleaner, but it is no longer the direct description of what was observed. For researchers mapping inequity across neighborhoods, social categories, and their intersections, the consequence is straightforward: a smoother model can understate the concentration of risk or access in exactly the strata that matter most for intervention.
S-MAIHDA offers a stricter standard. It does not ask the data to conform to a distribution that may not fit the clinical or social process generating the disparity. Instead, it estimates what was actually observed, then quantifies how much context structures those differences. For cross-sectional inequality work, that is a stronger fit to the analytic task. The method is especially relevant when the question is not whether a difference exists in some abstract model, but how large it is, where it sits, and whether area-level context changes its size.
A more faithful description of inequality comes from the data themselves
The paper frames S-MAIHDA as the canonical description of inequality because its estimates are the observed data themselves. The compare diagnostic quantifies how far random-effects MAIHDA departs from that description. In fields where policy targets depend on recognizing small strata with concentrated disadvantage, that diagnostic step is not optional. It reveals when a model is summarizing variation in a way that may conceal the pattern under review.
The strongest contribution here is not a new abstract metric. It is a refusal to let convenience replace observation. By preserving raw prevalences, composition-standardized prevalences, and the exact decomposition of inequality components on the probability scale, S-MAIHDA keeps the structure of health inequality visible without recourse to shrinkage-driven smoothing. The method still needs to be judged in other settings, outcomes, and datasets, but the immediate comparison is concrete: when the task is to describe inequality without distortion, the observed stratum counts deserve priority over model-driven averages.