Evaluation of Temporal Trends in Racial and Ethnic Disparities in Sleep Duration Among US Adults, 2004-2018

This cross-sectional study describes the temporal trends in racial and ethnic disparities in sleep duration during a 15-year period among US adults with sleep data available from the National Health Interview Survey.

complemented with demographic and socioeconomic characteristics.

National Health Interview Survey Response Rates and Non-Response Bias Mitigation
The NHIS assesses response rates at several different levels and reports two different sets of response rates, "Conditional" and "Final," each reported at the level of "family" and "adult." The difference between Conditional and Final rates is whether the household nonresponse is taken into account ("Final") or not ("Conditional"). [1] The following are the response rates (both annual, and the pooled mean) for NHIS from 2004-2018: school/general equivalency diploma, or less than high school); 2 categories for insurance coverage status at the time of the interview (insured and uninsured); 2 categories for marital status (married or living with partner, and not); 3 categories for employment status (with a job or working, not in labor force, and unemployed); 2 categories for flu vaccine in past 12 months (yes and no). We also included self-reported chronic conditions, including hypertension, diabetes, prior stroke or myocardial infarction, cancer, and emphysema or chronic bronchitis.

Statistical Analysis
To estimate the annual prevalence of short and long sleep duration prevalence rate for each racial/ethnic group we used separate multivariable multinomial logistic regression models. In these models, categorical sleep duration (short, recommended, and long) was the dependent variable, and age, a dummy variable for each region, and a dichotomous indicator for each year of interview were the independent variables. Age and region were centered on their overall mean for the study sample. The coefficients for each year, when combined with the intercept, then represented the logit of the annual rates of each sleep duration category adjusted for age and region. The results were used to generate estimated annual prevalence, using the inverse logit of each year effect as the annual prevalence and applying the method of parametric bootstrapping to calculate the standard error (SE) and the confidence interval (CI) for the transformed coefficients. [2] To measure the racial and ethnic differences in short and long sleep duration, we subtracted the annual prevalence among White people from the annual prevalence among each of the other 3 groups for that year, also constructing SE for the differences. Then, to estimate trends over the study period, we used weighted linear regression models where the dependent variable was the adjusted annual sleep duration disturbance prevalence or difference, and the independent variable was time in years. To account for varying precision of each estimated prevalence or the difference over time, each observation was weighted by the inverse square of the SE.
To evaluate the association between race and ethnicity and each of these sleep duration disturbances by age, we used a similar approach as above. We used multinomial regression model with categorical sleep duration as the dependent variable and age groups as the independent variables. Age was categorized by 5-year groups for this analysis. To measure the racial and ethnic differences in short and long with age, we subtracted the age group prevalence among White people from the age group prevalence among each of the other 3 groups for that year, also constructing SE for the differences.
To estimate the annual low-income prevalence by race and ethnicity we used the mean annual estimate obtained by separate multinomial logistic regressions using a similar approach as above, but with each of the multiply imputed low-income variables as the dependent variable and an indicator for each year as the independent variables. Similarly, we used the mean prevalence estimate of each sleep duration outcome (short or long) obtained from separate multinomial regressions using each of the income groups. [3] Lastly, we performed a sensitivity analysis to assess if the observed disparities in short sleep duration were explained by differences in self-reported sleep duration bias. In this sensitivity analysis, considering that the publicly available NHIS data includes sleep duration in integers, we artificially created hourly subdeciles by randomly and equally distributing individuals into 10 bins based on their NHIS sleep duration value (e.g., X number of individuals that reported 6 hours were randomly assigned to equally distributed deciles of an hour between 6 and 7 hours; each bin containing with X/10 number of individuals). We then added 73 minutes and 54 minutes to White and Black individuals' sleep duration, respectively, based on the mean sleep duration overestimation of each race group when compared with polysomnography.
[7] We then performed the same multinomial logistic regressions as above to assess the annual prevalence of short, recommended, and long sleep duration for each race; and the differences between both groups.   +3.17 (-0.11 to 6.46) 0.06 --+2.12 (-0.13 to +4.37) 0.07 --Data source is the National Health Interview Survey from years 2004 to 2018. Short-sleep duration was defined as fewer than 7 (<7) hours of sleep in a 24-hour period. For change in prevalence and change in difference: a positive sign (+) means the prevalence (or its difference with White individuals) increased and a negative sign (-) means it decreased. Prevalence estimates were adjusted by age and region. See details of this sensitivity analysis in eMethods. Abbreviations: CI, confidence interval