Analysis of Association Between Diabetes, Stroke and Hypertension Using Joint Model, Mixed Effects Model and Cox Model

BackgroundDiabetes, stroke and hypertension are all common diseases, and they are known to be associated with each other. People with hypertension seem to be at a higher risk at diabetes and stroke. Previous studies have used Cox proportional hazard models to analyze the association between repeated measures and each disease independently. However, previous studies fail to take into account on repeated measurements of blood pressure and time-to-event data for two diseases into one model to analyze their association. Also, separate analyses would cause bias and invalid inference on parameter estimates, since those previous analyses fail to capture the latent relationship among clinical outcomes.MethodsWe used the joint model to evaluate the association between diabetes, stroke and hypertension on 5,000 UK Biobank participants. In joint model, we could link the longitudinal and time-to-event outcome through the random effects. We used linear mixed effect model and Cox proportional hazards model to do the model diagnostics for survival part and longitudinal part separately.ResultsIn the longitudinal sub-model, age at visit, body mass index (BMI), sex, ethnicity were all statistically significant (p<0.05). In the competing risks sub-model, BMI, sex and ethnicity were significantly associated with diabetes (p<0.0001), but not with stroke. Our results indicated that higher BMI and increased age were associated with an increased risk at diabetes, as well as a higher blood pressure. Male received higher risk at diabetes than female (p<0.0001). Non-white population received the higher risk than white. Discussion and Conclusion: Higher blood pressure contributed to higher risk at diabetes, but we didn't find significant evidence that higher blood pressure had an impact on stroke. One possible reason is that the participants had much lower occurrence rate at stroke. In the longitudinal data diagnostics, we found that the assumption of normal distribution of random effect may not be appropriate. In the survival data diagnostics, the Cox-Snell plot showed that Cox PH model is a good fit for modeling stroke, but not for diabetes. Therefore, further research on model diagnostics of joint model is necessary to take into account the latent association.