Because there was a high proportion of missing data for some exposures, we performed secondary analyses in which we repeated the main analyses after multiple imputation of missing data (eTable). We assumed data were missing at random (dependent on values of other nonmissing variables but not on values of missing variables). We did multiple imputation by chained equations (MICE) using IVEWARE V0.1 (University of Michigan, Ann Arbor, MI) within SAS software.34,35 We did 10 imputations of missing data, and included all early-life exposures, race/ethnicity, birth decade, race/ethnicity-by-birth-decade interaction terms, childhood family income, participant’s education level, maternal death prior to baseline, and age at menarche in the imputation regression models. Even though mother’s vital status and participant’s education level were not confounders in our polytomous logistic regression analyses, they were included in the imputation regression models because they influenced whether early-life data were missing. We used PROC MIANALYZE in SAS V9.2 to summarize polytomous logistic regression results across all 10 imputation datasets. In analyses using imputed missing data, we could mutually adjust for related early-life factors to account for the possibility of additional confounding without reducing sample size.
