Statistical Analysis

All cognitively normal individuals at baseline were considered at risk for incident MCI. MCI onset was assigned at the midpoint between the last assessment as cognitively normal and the first-ever assessment as MCI. Persons who developed dementia without an intervening MCI diagnosis were presumed to have had an undetected MCI phase. We computed the follow-up duration from baseline to MCI onset for incident cases and through the date of the last follow-up for participants who remained cognitively normal. Subjects who refused participation, could not be contacted, or died during follow-up were censored at their last evaluation.

We computed the rate of weight change (kg/decade) from the maximum weight in midlife through follow-up, including late life weights as a time-dependent variable. We investigated the association of weight change with incident MCI using Cox proportional hazards model and reported hazard ratios (HR) and 95% confidence intervals (CI); in separate models we included maximum weight in midlife or late life weights. We examined confounding by APOE ε4 allele, diabetes mellitus, hypertension, stroke, and cigarette smoking with each variable added separately, and effect modification by age, sex, and APOE ε4 allele by including interaction terms of these covariates with rate of weight change.

Multivariable models included: Model 1: sex (where applicable), education and APOE ε4 genotype (ε4 carrier vs non-carrier); Model 2: included model 1 variables and potential confounders: alcohol problem (yes vs. no), depressive symptoms (BDI score ≥16), use of statins, diabetes mellitus, hypertension, coronary heart disease, cigarette smoking (never vs. former or current) and stroke.

We also examined the effect of including height in the models. In a separate model, we included both maximum weight in midlife and late life weights, but not weight change, in the same model and examined the associations with MCI. We repeated the analyses using rate of change in BMI (unit/decade). All hypothesis testing was conducted assuming an alpha = 0.05 significance level and a two-sided alternative hypothesis. All statistical analyses were performed using SAS version 9.3 (SAS Institute, Cary, North Carolina).