Statistical analysis

In all, 13 569 women and 11 515 men provided information on birth weight at the baseline survey and were included in the analyses. Participants were categorized into four groups according to birth weight (<2500, 2500–3499, 3500–3999 and ≥ 4000 g), with the 2500–3499 g group used as the reference category. Entry time was considered as time since birth or age at enrollment, which were used to calculate incidence since birth and since baseline, respectively. Exit time was defined as age at death, age at diagnosis of T2DM or hypertension, or last follow‐up contact date, whichever came first. We did not observe significant heterogeneity between men and women in associations between birth weight and subsequent risk of metabolic disorders; thus, we combined the two databases in the analysis.

The potential curvilinear relationships of birth weight with risk of adult obesity, T2DM, and hypertension were evaluated using restricted cubic splines functions (RCS)33 using the 5th, 25th, 75th, and 95th percentiles as fixed knots and the 50th percentile as the reference. Age was included as a spline variable to minimize residual confounding. Logistic regression models (odds ratios [ORs] and 95% confidence intervals [CIs]) were used to estimate associations of birth weight with the prevalence of obesity, T2DM, and hypertension, whereas Cox proportional hazard models (hazard ratios [HRs] and 95% CIs) were used to estimate associations of birth weight with the incidence of the diseases since birth and since baseline. Potential confounders included in the models were age (as a continuous variable), sex (male/female), education (no formal education or elementary school, middle school, high school, and college or above, as dummy variables), per capita income (<5000, 5000–10 000, and > 10 000 RMB for women; <12 000, 12 000–24 000 and > 24 000 RMB for men, as dummy variables), cigarette smoking (never/ever), alcohol consumption (never/ever), regular exercise (never/ever), and having been breastfed (never/ever), which were collected at baseline.

All analyses were performed by using SAS version 9.1 (SAS Institute, Cary, NC, USA), and all tests of statistical significance were based on two‐sided probabilities.