The index of concentration at the extremes for epidemiologic research: implications of categorization approach for observed health outcome associations

The Index of Concentration at the Extremes for race (ICErace) is increasingly used to quantify neighborhood-level racial segregation in epidemiologic studies. However, ICErace is commonly categorized using quantiles, potentially masking associations with health outcomes. We compared the performance of three categorization approaches for ICErace: quintiles, Jenks natural breaks optimization, and hierarchical clustering. Using Illinois zip code tabulation area (ZCTA) and census tract (CT) data from the 2018-2023 American Community Survey, we examined the impact of each approach on the resulting range of ICE values. Then, with generalized additive models, we compared each approach to a benchmark continuous ICErace model with spline terms to predict neighborhood-level prevalence of chronic obstructive pulmonary disease (COPD), healthcare utilization, and depression. At the ZCTA level, where ICErace values were highly right-skewed, the Jenks and hierarchical categories retained 85-95% of the benchmarks explanatory power. In contrast, quintile explanatory power was dependent on both the outcome and geographic level, ranging from 21% for healthcare utilization to 91% for depression at the ZCTA level. In fact, ZCTA quintiles masked a U-shaped association between ICErace and healthcare utilization observed with other approaches. Alternatively, CT quintiles outperformed other categorization methods for depression and COPD, highlighting the importance of considering health outcome and geographic level when selecting a categorization approach. Our results suggest that the ICE categorization approach is a consequential decision that can substantially alter research findings. With highly skewed data, data-driven approaches, including continuous modeling with spline terms, are recommended to preserve underlying relationships between ICErace and health outcomes.