Variability in prevalence and incidence of dementia and their contributing risk factors by algorithmic classification: Results from the Health and Retirement Study

Abstract

AbstractBackgroundStudies that use observational population level data with research classification for dementia have to contend with error in classification. Errors might be particularly pronounced in studies examining racial and ethnic disparities in dementia where classification can be confounded by certain risk factors (e.g. Socioeconomic Status; SES).MethodWe use data from the health and retirement study to examine differences in dementia prevalence (in 2000 [n = 6,860] and 2016; [n = 6,892]) and 2‐year incidence (2000‐2002 and 2016‐2018) over two time periods between non‐Hispanic White (NH‐W) and NH‐Black older adults (ages 70+‐years) across 4 algorithmic classifications of dementia in HRS (Langa‐Weir, LASSO, Hurd, and Expert). We then use Oaxaca‐Blinder decomposition techniques to examine the contribution of lifecourse factors (demographic, childhood SES and health, adult SES, and later life health behaviors, and health characteristics) to these racial differences in prevalence and incidence over the two time periods.ResultAll four algorithmic classifications point to consistent differences (favoring NH‐W) in prevalence (Figure 1) and, to a lesser extent, incidence (Figure 2) of dementia. Additionally, dementia prevalence but not incidence decreased, particularly among NH‐B, between 2000 and 2016. Despite these consistent findings, the estimated rates of dementia prevalence and incidence differed depending on the algorithmic choice with the Lange‐Weir classification consistently underestimating rates for NH‐W. Nearly 3/4th of the differences in dementia prevalence between NH‐B and NH‐W were explained by advantageous (among NH‐W) adult SES factors, and later life health behavior, and health characteristics; with SES status consuming the largest share. The contributions of these factors were less notable for differences in dementia incidence.ConclusionAssessments of racial disparities in dementia with observational data, where cognitive and physical health and function measures are used to generate research classifications, can lead to varying estimates by choice of algorithm. Sensitivity to algorithmic classification is more notable for dementia prevalence estimates, particularly given SES confounding, than for incidence. Later life health and health behaviors are also notable drivers of racial differences in dementia prevalence. More work is required to calibrate classification techniques to reduce bias and enhance the precision in estimates of magnitude and risks for dementia disparities.