Statistical analyses

The data analyses for the present study were performed following a prospective research proposal (S2 Text). We performed pooled analyses of all the individual-level data. Multinomial (polytomous) logistic regression models with 5 categories of outcome for age at menopause (<40, 40–44, 45–49, 50–51, and 52 years and older) were used, with age 50–51 years at menopause as a reference group. The intensity, duration, cumulative dose, and age started smoking were analysed in combination with smoking status for current and former smokers. Never smokers were used as reference for all smoking measures. All statistical models were adjusted for BMI, education level, a combined variable for race/ethnicity/region, based on the established relations of these in the literature to the outcome. Age at menarche and parity were also potential confounders that could affect the association between smoking exposures and age at menopause. Thus, the models were additionally adjusted for both covariates but with only 14 studies included because the WHITEHALL and Southall And Brent Revisited (SABRE) studies did not collect age at menarche and Healthy Ageing of Women (HOW) study did not collect parity. Adjusted RRRs (i.e., the ratio of two relative risks, RRRs) [24] and 95% confidence intervals (CIs) for smoking measures and each category of age at menopause were estimated. The Bayesian information criterion (BIC) was used to compare the fit of the models that used each of the measures of smoking characteristic. A lower BIC implies a better fit.

Because the UK Biobank data contributed more than 60% of the total sample used in the cross-sectional analyses, we conducted a sensitivity analysis by excluding this study. Also, we analysed the associations of intensity, duration, age of starting smoking, and years since quitting with age at menopause by adjusting for all covariates, including the other smoking characteristics. Finally, in order to evaluate the heterogeneity among studies, in both cross-sectional and prospective analyses, we also performed study-specific regressions and random-effects meta-analysis for the associations between each level of the cigarette smoking measures and each category of menopausal age to estimate the magnitude of association.

Statistical analyses were performed by using SAS (version 9.4; SAS Institute Inc., Cary, NC) and Stata (version 14.0; Stata Corp., College Station, TX). In SAS, the SURVEYLOGISTIC procedure was used with the generalized logit link to adjust for the clustering of data within studies to obtain robust standard errors. The SGPANEL procedure was used to plot the associations. In STATA, the METAN command was used to perform meta-analysis. All statistical tests were based on the two-sided 5% level of significance.