A comparative analysis of hypertension risk in urban informal settlement in Kenya using hierarchical and marginalized models

Background: Cardiovascular disease (CVD) is the leading cause of death globally and accounts for approximately 13% of all deaths in Kenya. Despite this burden, few statistical models accounting for the clustered structure of longitudinal public health data have been applied to CVD risk assessment in Kenya. Conventional regression approaches that ignore within-cluster correlation yield underestimated standard errors and invalid inferences. Methods: We analysed longitudinal binary data on hypertension from the SCALE UP study (2012–2014) and the Nairobi Urban Health and Demographic Surveillance System (NUHDSS, 2002–2015) among adults aged 35 years or older in two urban informal settlements in Nairobi, Kenya. Two complementary modelling frameworks were compared: (i) a hierarchical generalized linear mixed model (GLMM), providing subject-specific estimates via individual-level random effects and (ii) the marginally specified logistic-normal model of Heagerty, which parameterises the marginal mean directly while retaining a latent normal random-effects structure to characterise within-subject dependence. Results: Hypertension prevalence declined from 75.7% at baseline ( N = 622 ) to 60.4% at endline ( N = 414 ; p < 0.001 ). Both models identified the same six significant predictors at the 5% level: time, dehydration, blood glucose, body swelling, waist circumference, and hip circumference. The GLMM random-effects variance was σ ˆ 2 = 3.322 , indicating substantial between-subject heterogeneity. The marginal model yielded a conditional-scale variance of 4.358, reflecting the additional variability required to induce the specified marginal logistic form. The marginal model consistently produced narrower confidence intervals and attenuated odds ratios relative to the GLMM, consistent with a population-averaged interpretation. Conclusion: Both models identified central adiposity, measured by waist circumference (marginal OR = 1.57, 95% CI: 1.18–2.09), as a key modifiable risk factor for hypertension, while greater hip circumference was inversely associated (OR = 0.61, 95% CI: 0.46–0.80). The marginal model provides more stable population-averaged estimates suited to public health planning, whereas the GLMM captures individual-level heterogeneity useful for clinical risk stratification. A dual-model approach is recommended for comprehensive CVD risk modelling in similar resource-limited settings.