Statistical analysis

Based on 1.25 million participants, we estimated the comprehensive age-, sex-, region-, and year-specific average BMI levels using a temporal–spatial hierarchical Bayesian model,¹⁹ taking advantage of information on regions and age over time for missing data. Cross-validation was used to examine the predictive performance of the model by comparing the estimated BMI levels with their observed counterparts. The details of distribution of high BMI estimation are in the Supplementary Materials.

Age-sex-specific RRs measuring the effects of BMI level on the outcome were synthesized from published meta-analysis, and where available, pooled analyses of prospective observational studies, which were consistence with GBD 2019 study. More information on RR estimation had been previously provided by the Global Health Data Exchange via a web tool.¹¹ The details of assessment of risk-outcome pairs and their RRs were in the methods of Supplementary Materials, and RRs used in this study were showed in Supplementary Table S1. The theoretical minimum risk exposure level was established with the most recent pooled analysis of prospective observational studies,²⁰ defined as a uniform distribution for BMI between 20 and 25 kg/m². We estimated the population attributable fraction (PAF) for cause-specific deaths attributable to high BMI using the following equation:

where RRoas(x) is the RR as a function of BMI at x, for cause o, age group a, and sex s, with the lowest level of observed BMI as l and the highest as m. Paspt(x) is the distribution of BMI at x for age group a, sex s, province p, and year t.

YLLs were obtained by multiplying the number of deaths by the standard life expectancy at each age. Deaths and YLLs attributable to high BMI were calculated by multiplying the deaths and YLLs by age-, sex-, region-, and cause-specific PAF. Mortality rates and YLL rates were calculated by dividing the number of attributable deaths and YLLs by the corresponding population, respectively. We applied the direct standardization to adjust for differences among the populations, with the standard population of the sixth national population census in 2010 (Supplementary Table S3).

We analysed the decomposition of changes in mortality attributable to high BMI by socioeconomic regions during 2005–2018. Methods developed by Das Gupta were used to conduct a decomposition of changes in mortality attributable to high BMI due to population growth, population ageing, risk-deleted mortality rate (the expected mortality that would be observed if high BMI was removed), and changes in exposure to high BMI.²¹ We also calculated the percentiles 2.5 and 97.5 of the 1000 Monte Carlo runs for each quantity of interest as 95% uncertainty limits (UI), which were able to propagate uncertainty in the final analyses.

The analyses were performed using R version 4.2.1 (R Foundation for Statistical Computing, Vienna, Austria) and SAS version 9.4 (SAS Institute Inc., Cary, NC, USA).