Multiple linear regression was used to determine the effect of weight change on RDI change, adjusting for age, baseline weight, weight change squared, the mean of baseline and follow-up RDI, and sex. Multiple linear regression was also used to test for sex–by–weight change interaction. The square of weight change, the cube of weight change, and a spline term at weight change of 0 were included to look for nonlinearity in the relationship between weight change and RDI change. After sex stratification, the quadratic term created a better model fit than the cubic weight change or spline at weight change of 0. Regression diagnostics were performed to test model assumptions and assess for adequacy of fit. The averages of the baseline and follow-up RDI were used to avoid bias due to regression to the mean.21 Additional multiple linear regression models explored the effect of controlling for race and other measures of body habitus, including BMI, neck circumference, and waist-hip circumference ratio, and tested for interactions among weight change and age, sex, ethnicity, and starting weight in the associations with change in RDI. Because of significant sex interactions, all subsequent models were stratified by sex. The χ2 test was used to assess differences between weight change categories in the distribution of RDI change categories and in incidence of regression and progression of RDI. Multiple linear regression and the χ2 test were performed again using change in BMI and change in neck circumference as alternative measures of weight change. Sex-stratified multinomial logistic regression was used to determine the effect of weight change categories on RDI change using stable weight and stable RDI as the reference groups, with age, race, baseline weight, and the average of baseline and follow-up RDI as covariates. The statistical significance (2-tailed; P<.05) of the multiple linear regression coefficients was assessed by t tests, and the coefficients in the multinomial logistic regression were determined by the Wald χ2 statistic.19 Models were reassessed, excluding data with high outlier influence scores, with no substantial change in results. Statistical analyses were conducted using SAS statistical software, version 8.01 (SAS Institute Inc, Cary, NC), except for the multinomial logistic regressions, which were conducted using Stata statistical software, version 7.0 (Stata Corp, College Station, Tex).