Time-to-Event Modeling of Hypertension Reveals the Nonexistence of True Controls

Abstract

AbstractGiven a lifetime risk of ~90% by the ninth decade of life, it is unknown if there are true controls for hypertension in epidemiological and genetic studies. Here, we use the Bayesian framework to compare logistic and time-to-event approaches to modeling hypertension. Using a proportional hazards model, we explored nonparametric and parametric models of the baseline hazard function, accounting for interval censoring. In the Howard University Family Study (HUFS), a population-based study of African Americans from Washington, D.C., the median age at hypertension was 48 years, baseline hazard rates increased with age until 55 years, and the probability of being free of hypertension at 85 years of age was 12.2%. In the nationwide NHANES study, the median age at hypertension was 42 years in African Americans, in contrast to 57 years in European Americans and 56 years in Mexican Americans. Baseline hazard rates increased with age until 58 years in African Americans, comparable to 60 years in European Americans and 58 years in Mexican Americans. The probability of being free of hypertension at 85 years of age was 8.4% in African Americans, in contrast to 21.4% in European Americans and 20.6% in Mexican Americans. In all four groups, baseline hazard rates decreased but did not reach zero, consistent with the nonexistence of controls. Model fits were comparable for a proportional hazards model based on gamma-distributed hazard rates under a correlated prior process and a logistic model adjusted for age and age2. Last, using an agnostic model screening approach of 38 potential covariates, we identified and replicated in all groups a model that included chloride, low-density lipoprotein cholesterol, uric acid, and weight. With age and age2, the variance explained by these covariates was 40.9% in African Americans, 34.8% in European Americans, and 28.3% in Mexican Americans. Taken together, modeling of the baseline hazard function of hypertension suggests that there are no true controls and that controls in logistic regression are cases with a late age of onset. These findings shed considerable insights into the design of genetic and epidemiological studies of hypertension with implications for ethnic health disparities.