Quantitative methods for descriptive intersectional analysis with binary health outcomes.

Area of Science:

• Epidemiology
• Health Disparities Research
• Biostatistics

Background:

• Intersectionality acknowledges how multiple social identities interact within power structures to influence health.
• Epidemiological studies increasingly use intersectionality to describe and understand health disparities, but lack methodological guidance.
• Evaluating various statistical methods is crucial for accurate intersectional health disparities research.

Purpose of the Study:

• To advance methods for intersectional estimation of binary health outcomes in descriptive epidemiology.
• To evaluate the accuracy of seven intersectional data analysis methods compared to non-intersectional regression.
• To assess variable selection performance among decision tree methods.

Main Methods:

• Compared cross-classification, regression with interactions, Multilevel Analysis of Individual Heterogeneity (MAIHDA), and four decision tree methods (CART, CTree, CHAID, random forest).
• Evaluated accuracy of estimated intersection-specific outcome prevalence using simulated data across 192 intersections.
• Utilized National Health and Nutrition Examination Study data for illustrative example analyses.

Main Results:

• At larger sample sizes, all methods except CART outperformed non-intersectional main effects regression.
• In smaller samples, MAIHDA showed the highest accuracy, while random forest, cross-classification, and saturated regression were least accurate.
• CTree and CHAID demonstrated moderate performance; CART performed poorly in estimation and variable selection.

Conclusions:

• MAIHDA is recommended as an unbiased method for accurate estimation of high-dimensional intersections, especially with smaller sample sizes.
• Larger sample sizes are critical for the performance of other evaluated intersectional methods.
• Results advocate for the adoption of intersectional approaches in descriptive epidemiological research.