Evaluating Social Determinants of Health–Based Alternatives to Race-Based Cognitive Normative Models

BACKGROUND AND OBJECTIVES: Race and ethnicity are proxy measures of sociocultural factors that influence cognitive test performance. Our objective was to compare different regression-based cognitive normative models adjusting for demographics and different combinations of easily accessible/commonly used social determinants of health (SDoH) factors, which may help describe cognitive performance variability historically captured by ethnoracial differences. METHODS: We performed cross-sectional analyses on data from Black and White participants without mild cognitive impairment/dementia in the Atherosclerosis Risk in Communities Study who attended visit 5 in 2011-2013. Participants underwent a battery of 11 cognitive tests (3 domains: memory, executive function, language). We fit 6 separate normative models for each cognitive test, all including age and education, with different combinations of race, the Wide Range of Achievement Test (education quality proxy), and area deprivation index (neighborhood deprivation) associated with current residence. We compared model fits and calculated concordances/discordances between models using z-scores derived from each normative model and a z-score <-1.5 threshold for impairment. RESULTS: Participants (n = 2,392) had a mean age of 74.4 years, 60.4% were female, and 17.1% were of self-reported Black race. The "Full" model with race alongside demographic and SDoH measures consistently outperformed other nested submodels (likelihood ratios ≥ 100) for all domains/tests except Delayed Word Recall. Models with education quality alone ("WRAT") generally outperformed models with neighborhood deprivation ("ADI") or race ("Race") alone for memory and language tests while "Race" models performed better for executive function tests. Adding neighborhood deprivation to education quality ("WRAT + ADI") did not improve models vs using "WRAT" alone. Across all domains/tests, the concordance compared with the "Full" model was lower for "Education" and "ADI" models than for other nested models. Although numbers were small, there was greater discordance among Black (range = 8.2%-23.2%) compared with White (range = 2.2%-3.4%) participants, particularly for Boston Naming Test and executive function tests. DISCUSSION: Education quality outperformed neighborhood disadvantage as an additional/alternative SDoH measure in normative models and may be useful to collect in cognitive aging studies. While performance varied across cognitive domains and tests, routinely reported SDoH variables (education level, education quality, late-life neighborhood deprivation) did not fully account for observed ethnoracial variability; future work should evaluate SDoH across the lifespan in more ethnoracially diverse populations.