Abstract / Summary
Hypertension prevalence is rising in rural Ghana, but the comparative behaviour of frequentist and Bayesian inferential frameworks for ordinal blood pressure outcomes in this setting remains underexamined. This study applied both frequentist (ordered probit) and Bayesian ordered probit models to identify factors associated with blood pressure status and to compare the two frameworks on coefficient agreement, uncertainty quantification, and model fit. This was a cross-sectional analysis of 2,010 adults aged 40–60 years in the Kassena-Nankana districts. Blood pressure was classified into four ordinal categories (Normal, Elevated, Stage 1, Stage 2 hypertension) per ACC/AHA 2017 thresholds. A directed acyclic graph informed a minimal adjustment set of age, sex, marital status, occupation, smoking, alcohol use, pesticide exposure, body mass index, and moderate-to-vigorous physical activity. Ordered probit models were fitted under frequentist and Bayesian (Hamiltonian Monte Carlo, weakly informative priors, four chains) frameworks on the complete-case analytic sample ( n = 1,979). The proportional-odds assumption was assessed with the Brant test. Model fit was compared using AIC (frequentist) and Pareto-smoothed importance-sampling leave-one-out cross-validation (Bayesian), placed on a commensurable deviance scale. The distribution of blood pressure category was Normal 46.1%, Elevated 12.4%, Stage 1 hypertension 19.6%, and Stage 2 hypertension 21.9% (combined Stage 1/Stage 2: 41.5%). Older age (probit β = 0.032 per year, 95% CI 0.023–0.041; Bayesian mean 0.032, 95% CrI 0.023–0.041), male sex (probit β = 0.279, 95% CI 0.141–0.417; Bayesian mean 0.279, 95% CrI 0.137–0.420), and higher BMI (probit β = 0.054, 95% CI 0.039–0.069; Bayesian mean 0.054, 95% CrI 0.039–0.069) were consistently associated with higher blood pressure category in both frameworks. The Brant test indicated that the sex association was not constant across severity thresholds (χ2 = 10.78, df = 2, p = 0.005), while all other covariates were consistent with proportional odds. Married status (probit β = -0.129, 95% CI -0.240 to -0.018) and pesticide exposure (probit β = -0.122, 95% CI -0.225 to -0.019) were each inversely associated with blood pressure category in both frameworks. Current smoking showed a negative but non-significant association (probit β = -0.119, 95% CI -0.264 to 0.025, p = 0.106). Frequentist and Bayesian coefficients agreed to within 0.001–0.002 across all predictors, and model fit was comparable on a commensurable scale (frequentist AIC = 4,936.5; Bayesian − 2×elpd_loo = 4,937.3; difference = 0.8). Older age, male sex, and higher BMI were robust factors associated with blood pressure category in this population, though the sex association varies in magnitude across severity thresholds. The inverse associations with married status and pesticide exposure are unexpected and are best interpreted as hypothesis-generating given the cross-sectional design. Frequentist and Bayesian ordered probit models produced materially equivalent inference once the adjustment set and data quality were addressed, indicating that framework choice is secondary to rigorous data preparation in this setting.