Interaction effects are only a piece of the puzzle: reintroducing MAIHDA as a powerful tool to advance understanding of intersectional inequities

This article is linked to ‘’(https://doi.org/10.1093/aje/kwaf233). As intersectionality becomes a central framework in epidemiology for examining population health patterns and inequities, the multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA) approach has gained prominence as the “gold standard” method.1-3 However, uncertainty about how to interpret its resulting estimates, as well as misunderstandings about what the approach is designed to accomplish, persist. As such, we appreciate the opportunity to engage with two papers recently published in the journal (Merchant et al.’s application of MAIHDA to study changes in suicidality among US adolescents pre- to post-2020 and the associated commentary by Al-kassab-Córdova) as a way to address common sources of confusion. Al-kassab-Córdova argues that Merchant et al.’s results do not substantively reflect intersectional inequities, given the magnitudes of and observed reduction in the nonadditive between-stratum variance (5.1% pre-2020 to 1.1% post-2020)—obtained by 1 minus the proportional change in variance (PCV). However, the PCV is a variance decomposition metric that reflects how much of the between-stratum variance (ie, differences in the predicted prevalence of suicidality across 40 intersectional strata) is explained by the additive main effects of race/ethnicity, sex/gender, and sexual orientation—not the magnitude of any inequities themselves. It is entirely possible for both additivity (ie, the degree to which stratum differences follow main-effect gradients) and absolute inequities (ie, the size of the differences in predicted prevalence across strata) to increase simultaneously over time; importantly, such a scenario is highly likely in the context of Merchant et al.’s study. Consider the myriad system-level crises that occurred in 2020: the COVID-19 pandemic and associated social/economic disruptions, heightened racialized violence, and an intensifying wave of anti-LGBTQ policies.4,5 Irrespective of statistical interaction, all of these system-level crises plausibly increased levels of suicidality across multiple axes of marginalization in parallel (eg, increasing suicidality among all or most strata inclusive of girls, sexual minority youth, and youth of color), thereby strengthening the predictive power of the main effects while simultaneously widening the absolute gaps between multiply marginalized groups and their more advantaged counterparts. And indeed, Merchant et al. observed precisely this pattern: the prevalence of suicidality increased for nearly all groups, and often most sharply for multiply marginalized youth. Moreover, and again irrespective of statistical interaction, the authors observed sizable between-stratum heterogeneity (given by a variance partition coefficient of ≈11%), and substantial differences in predicted probabilities (from <10% among heterosexual Black boys to >50% among bisexual multiracial/Other girls), which collectively reveal meaningful intersectional inequities. Together, these findings illustrate why interpreting MAIHDA results requires evaluating the full suite of estimates in concert, rather than relying on any single statistic (see Table 1 for details).