Sleep disparities in the United States: Comparison of logistic and linear regression with stratification by race

Analyses of risk factors associated with poor sleep/deprivation often use nationally representative surveys of the United States such as the National Household and Nutrition Examination Survey (NHANES). Outcomes are dichotomized as <6, <7, or <8 h of sleep and modeled with logistic regression where race or ethnicity is treated as an independent variable. Converting a continuous variable (sleep hours) to a categorical compromises statistical power. Treating race as a confounder fails to uncover how sleep disparities affect minorities.

This analysis of NHANES from 2005 to 2008 of White and Black participants compares interpretations from logistic regression models using ≤6 h of self-reported sleep to linear regression models using number of sleep minutes as the outcome. The analysis includes bivariate and multivariable models of risk factors associated with poor sleep including race, markers of low socioeconomic status (SES), sleep difficulty measures, self-reported health, and clinical comorbidities (obesity, hypertension, diabetes). All models were generated for the complete sample and stratified by race.

Linear regression models produced quantifiable, clinically meaningful results such as women slept ~20 additional minutes than men for both Black and White strata or were OR=0.63 times as likely to sleep ≤6 h. Markers of low SES (education, poverty) and self-reported health were associated with sleep deprivation for Whites, but not for Blacks in both linear and logistic regression.

Stratified analyses by race using the amount of sleep as a continuous outcome in linear regression is more rigorous and informative than logistic regression for sleep research using US representative surveys.