Statistical analysis

We categorized weight change between the baseline and 10-year surveys into seven categories: weight loss (>10, >5–10, >2.5–5 kg), stable weight (within ±2.5 kg), and weight gain (>2.5–5, >5–10, >10 kg) based on previous studies^14,15 and ensuring enough number of events in each weight change category. Cox proportional hazards regression with age as the time metric was used to estimate hazard ratios (HRs) and 95% confidence intervals (95% CIs) of mortality related to weight change, using the stable weight group as the reference. The models were adjusted for race/ethnicity as a strata variable, and age, BMI at 10-year follow up and years between the baseline and 10-year surveys as covariates. BMI was selected for inclusion as it was related to subsequent mortality and to weight change. The interval between the two surveys was included to account for the autocorrelation of measures over time. For a majority (75.7%), the interval was between 8 and 12 years, since the goal had been 10 years. Smoking was an important confounding factor related to both BMI change and mortality. Thus some models were further adjusted for smoking at 10-year follow-up based on a comprehensive smoking model that was developed for lung cancer study in the MEC.¹⁶ The model explicitly included average number of cigarettes; average number of cigarettes squared; indicator variables for former and current smokers; number of years smoked (time-dependent); number of years since quitting (time-dependent); and interactions of race/ethnicity with the following variables: average number of cigarettes, average number of cigarettes squared, smoking status, and number of years smoked. We also considered education and alcohol consumption as covariates but did not include them in the final models because adjustment did not alter the associations. The proportionality assumption was tested by Schoenfeld residuals and was found to be valid. Since the associations between weight change and all-cause mortality were similar between men and women, some of these analyses were performed in men and women combined, with adjustment for sex as a strata variable.

We evaluated models for all-cause mortality and for cause-specific mortality (cardiovascular disease and cancer). We performed subgroup analyses to examine whether the associations between weight change and mortality varied by race/ethnicity, age group (45-54; 55-64; 65-75 years at baseline), baseline BMI (<25; 25–<30; ≥30 kg/m²), and smoking status (never, former, current). We also examined the associations by years between the baseline and follow-up surveys (<10, 10-12, >12 years), but did not present the results because there was little substantive difference. In sensitivity analyses, we excluded the 1,070 deaths that occurred within the first 2 years after the 10-year follow-up survey, and excluded the participants with baseline BMI <20 kg/m² (n=3,591) who might have had an underlying illness. We also examined weight change per year (by dividing by the number of years between questionnaires) and percent weight change relative to the baseline weight. Tests for heterogeneity across subgroups were based on the Wald statistics for cross-product terms. Tests for heterogeneity between cardiovascular disease and cancer mortality were based on competing risk methodology.¹⁷ We also examined a nonlinear relation between weight change and mortality nonparametrically with restricted cubic splines.¹⁸ All analyses were conducted by using SAS statistical software version 9.4 (SAS Institute, Inc., Cary, North Carolina).