The design of this study included oversampling of those at higher risk for SDB and women with markedly higher levels of BMI to increase the precision of the risk estimates. Because of this sampling strategy, numeric sampling weights were developed for the analysis so that the estimates could be inferred to the general population. A comprehensive presentation of this sampling strategy has been presented elsewhere,20–27 including the use of the NHANES III laboratory data as the standard33 to adjust both the men and women in terms of sociodemographics to be representative of the national population.
Mean (standard deviation [SD]) and proportions of the demographic characteristics were calculated for the entire population, as well as stratified according to insomnia and objective sleep duration status. Multinomial logistic regression models were used to assess the independent association of objective sleep duration with normal sleep, fully remitted, partially remitted, and persistent insomnia, while controlling for potential confounding factors. Objective sleep duration was entered in the regression models as a continuous variable of hours of sleep. We calculated the odds ratios (OR) and the 95% confidence intervals (95% CI) from the regression models to estimate the risk of fully remitted, partially remitted, and persistent insomnia associated with objective shorter sleep duration, simultaneously adjusting for covariates. The covariates we adjusted for included major confounding factors expected to affect this relationship, i.e., age, race, gender, obesity, SDB, physical health problems, mental health problems, caffeine, cigarettes, alcohol consumption, and sampling weight. From the fully adjusted multinomial regression model, we also calculated the OR and the 95% CI to estimate the risk of fully remitted, partially remitted, and persistent insomnia associated with each covariate. Analysis of covariance (ANCOVA) was used to examine mean differences on MMPI-2 scores between the 4 study groups, while controlling for potential confounding factors. Bonferroni corrections were applied to control for type I errors when performing post hoc multiple comparisons. All analyses were conducted with SPSS version 17.0 for Windows.