Work

Comprehensive Demographic Correction Improves Sensitivity and Reduces Bias in Cognitive Assessment

Year: 2026

Type: preprint

Abstract: Background Scores on neuropsychological assessments are typically corrected for the influences of age, education, and gender (AEG). However, other demographic factors, such as crystallized ability and race/ethnicity, independently affect test performance. As a result, standard scores systematically over- or under-classify impairment in patients whose demographic profile differs from that of the reference population. Methods We developed a Comprehensive (C-) model scoring algorithm that added vocabulary, age², race/ethnicity, Latino background, a coarse socioeconomic status proxy, computer use, and daily prescription medications to the standard AEG predictor pool. The model was developed using data from 1,914 community-dwelling adults assessed with the California Cognitive Assessment Battery (CCAB; Woods et al., 2024). For each of 118 individual cognitive measures, stability-selection LASSO identified robust predictors in 300 random 80/20 splits retained at ≥ 80% frequency and then estimated mean coefficients and confidence intervals in 1,000 bootstrap OLS samples. Cross-sample frozen-coefficient validation was used to evaluate scoring model generalization in two subgroups: Group 1 (n = 1,033, older, first enrolled cohort) and Group 2 (n = 881, a recently recruited younger cohort). Results Stability selection retained a mean of 2.81 predictors per measure (range 1-6). Compared to the AEG model, the C-model approximately doubled variance explained (r² = 0.50 vs 0.25; mean across cognitive domains r² = 0.32 vs 0.18) and outperformed AEG in 98.8% of individual measures with non-trivial demographic signal. Racial disparities in MCI classification (the bottom-7th-percentile) were substantially reduced: Black-vs-White ratios fell from 5.6 (AEG) to 1.8 (C). Conversely, sensitivity was improved in individuals with elevated premorbid function: MCI classification ratios in low-vs-high vocabulary quartiles fell from 11.3 to 2.1. AIC favored the C-model in 88.1% of measures (mean ΔAIC = −167), ruling out overfitting. Frozen-coefficient validation preserved the C-model’s r² advantage in every cognitive domain. Conclusions By correcting scores for race, premorbid cognitive functioning (vocabulary), and other demographic predictors, the C-model explains substantially more variance than the AEG model, reduces racial bias, and increases sensitivity to cognitive decline in high-functioning participants. C and AEG models can be used in parallel: model concordance increases diagnostic confidence, while disagreement carries diagnostic information. Highlights We developed a Comprehensive (C-) model for scoring California Cognitive Assessment Battery (CCAB) tests that supplements standard age + education + gender (AEG) demographic corrections with additional predictors including vocabulary, age², race/ethnicity, Latino background, socioeconomic status (SES), computer use, and daily medications. To avoid model overfitting, significant predictors were identified with stability-selection LASSO, resulting in a mean C-model retention of 2.81 (range 1-6) predictors per individual test score. In 1,914 community-dwelling adults assessed with CCAB, the C-model approximately doubled explained variance for overall performance (OMNI) scores when compared with the AEG model (r² = 0.50 vs 0.25), and accounted for more variance than the AEG model in 98.8% of measures with non-trivial demographic signal. Cross-sample validation with two demographically distinct cohorts showed that the C-model’s r² advantage was preserved in every cognitive domain (Δr² variation < ±0.020 across fits), supporting generalizability. The C-model reduced demographic classification disparities in classifying mild cognitive impairment (MCI, bottom 7% of participants) and normalized MCI detection across different levels of premorbid cognitive reserve: Black-vs-White MCI-classification ratios fell from 5.6 (AEG) to 1.8 (C), and low-vs-high vocabulary performance ratios fell from 11.3 to 2.1. C-model scores are orthogonal to vocabulary; comparisons with premorbid crystallized intelligence are incorporated directly in the scoring algorithm, eliminating the need for post-hoc comparisons. The C-model is best used in parallel with AEG scoring: concordance between models increases diagnostic confidence, while disagreement provides additional diagnostic information.

Source: medRxiv

DOI: 10.64898/2026.06.27.26356750