Methods/Statistics
Extending Intersectional Methods to Population Health Research Victoria Fisher* Victoria Fisher Kiana Ramos Nicole Alkhouri Nadia N. Abuelezam
Background: Intersectionality theory has been used to describe the interconnectedness of identities within social, institutional, and structural systems of power. Recently introduced methods aim to produce better intersectional analyses (beyond multiple interactive terms), including multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA). MAIHDA relies on categorical strata as both fixed and random effects but has been met with concerns of collinearity and issues with interpretation. We propose an extension of MAIHDA to population-level infectious disease data, which allows for the use of continuous fixed effects to address potential collinearity.
Methods: Concerns about MAIHDA come from the reduction in random effects variance after the strata are added to the model as fixed effects. We propose using continuous measures as fixed effects that correspond to the categorical strata. This allows for greater within-group variation and the possible application to population-level research. We apply our proposed methods to U.S. county-level COVID-19 mortality data (CDC Wonder) via three multilevel models: random effects only, with continuous fixed effects, and with the strata as fixed effects (3044 counties were included in our analysis).
Results: The results from our three models suggest that using continuous count fixed effects provided the best fit. Continuous fixed effects reduced variance accounted for by the strata from 19% (null model) to 14%, while categorical fixed effects reduced the variance accounted for to 1.2%, suggesting that the use of categorical fixed effects may be sufficient for modeling the data on their own, but would not capture the interactive effects integral to intersectional analyses.
Conclusions: We illustrate the use of a novel application of MAIHDA methods in a population-level analysis, where strata were created from continuous variables. This method may be beneficial when individual-level data is not accessible.