Survival analysis in longitudinal studies for recurrent events: Applications and challenges

Abstract

Background and objectiveRecurrent event data analysis is most commonly used in biomedical research. However, the researchers dealing with recurrent events in survival analysis have ignored the assumption that the recurrent events are correlated. There are methods available that takes into account dependency between recurrent events. The main objective of this study was to demonstrate the recurrent event models using upper respiratory infection (URI) among Indian infants. MethodsThe frequency of URI among a birth cohort of 210 babies was evaluated monthly with nasopharyngeal swabbing. Data on 11 potential risk factors were noted. The extended Cox models, such as Andersen-Gill counting process (CP), Prentice-Williams-Peterson (PWP-CP), PWP–Gap time model, Marginal Model and Cox frailty model were applied. The better model was assessed based on Loglikehood test statistics. ResultsOf the four models, PWP-CP model had lower log likelihood value. The URI incidence rate was estimated to be 2.27 (95%CI: 1.70–3.03)times significantly higher in the month of July–October and 1.43 (95%CI: 1.19–1.71) times higher in the month of November–February as compared to March–June (p < 0.001). Nasopharyngeal colonization with S. pneumoniae was also another important risk factor (HR = 1.18, 95%CI:1.01–1.39, p = 0.03). ConclusionIn the current study, PWP-CP model was found to be better model as compared to other models. Also biologically appropriate as subsequent events depend on the first event of URI. Hence, the choice of an appropriate method for analyzing the recurrent event data should not be decided only on statistical basis but also based on the research question.