A Bayesian multilevel modeling approach to identifying determinants of diabetes in Bangladesh using BDHS 2022 data

Diabetes mellitus is an emerging national public health problem in Bangladesh, and the prevalence rates are rising at an alarming rate due to demographic and lifestyle changes. Although many studies have identified diabetes risk factors in Bangladesh, few have used advanced statistical methods to estimate risk probabilities. This study aimed to identify and estimate the most significant determinants of diabetes among Bangladeshi adults using three MCMC-based Bayesian model specifications: a standard Bayesian logistic regression and two multilevel Bayesian logistic regression models with informative and weakly informative priors applied to nationally representative BDHS 2022 data. Informative priors were constructed on the log-odds scale from published BDHS 2017–18 adjusted odds ratios, and the survey design was accounted for through sampling weights and enumeration area–level random intercepts. Each model was estimated using four independent MCMC chains, with convergence confirmed by the Gelman–Rubin statistic (R̂ < 1.01) and high effective sample sizes for all parameters. Model comparison using Leave-One-Out Cross-Validation (LOO) and the Widely Applicable Information Criterion (WAIC) identified the multilevel Bayesian logistic regression with informative priors as the best-fitting model (LOOIC = 10,819.63; WAIC = 10817.67), confirming the importance of accounting for the clustered survey structure. Age was the strongest predictor, with adults aged 65 years and above having more than three times the odds of diabetes compared to those aged 18–34 years (OR: 3.52, 95% CrI: 3.10–3.99). Obesity (OR: 1.96, 95% CrI 1.68–2.30), hypertension (OR: 1.59, 95% CrI 1.46–1.72), and higher wealth index (OR: 2.23, 95% CrI 1.94–2.56 for the richest quintile) emerged as major correlates, while urban residents had higher odds than rural residents (OR: 1.12, 95% CrI 1.01–1.25). Notable regional variation was observed, with the Dhaka division recording the highest odds of diabetes relative to all other divisions. As the data are cross-sectional, these estimates describe associations rather than causal effects, but they offer a robust probabilistic evidence base for targeting high-risk populations through evidence-based screening and intervention programmes in Bangladesh.