Review of Methods for Analyzing Recurrent Event Data: Evidence from Community-based KoGES Longitudinal Data

This study compares eight widely used models for analyzing recurrent event data in terms of risk interval, risk set, and baseline hazard function. It also presents a method for defining the likelihood function for each model, a guideline for model selection, and a measure for evaluating predictive accuracy. Applying the proposed guideline to the community-based (Anseong and Ansan cohorts) Korean Genome and Epidemiology Study data, which followed-up metabolic syndrome (MS) diagnosis status every two years from 2001 to 2018, revealed that the Prentice, Williams, and Peterson model with a counting process-type risk interval was the most appropriate. Estimation of baseline survival functions by event revealed similar intensities for transitions from 1→2 through 3→4 and from 4→5 through 6→7, while the intensity of transition 7→8 was the highest. The assumption of homogeneity among baseline hazard functions was rejected (p < 0.001). Factors significantly associated with time to MS diagnosis included cohort (Anseong vs. Ansan, odds ratio (OR) 1.11), increasing age (OR 1.01), female sex (OR 1.16), hypertension (OR 1.24), family history of chronic disease (OR 1.09), smoking (OR 1.18), and increasing BMI (OR 1.10).