Multivariate time-to-event models for studies of recurrent childhood diseases

Abstract

Standard Poisson regression analyses of infectious disease incidence may not be optimal when covariates change over time. When a longitudinal study is of sufficient duration so that multiple disease episodes per person may occur, within-subject correlation may be present and require special statistical consideration. Seasonality of disease incidence is a common feature of field studies of respiratory and diarrhoeal diseases in children. Accurate analysis may require keeping the children aligned with respect to the same baseline hazard function. These data features often are ignored in the analysis of such studies. Methods for accounting for multiple events are discussed. We propose the use of a counting process model to retain the alignment with respect to season. This model is supplemented by a simple bootstrap procedure to account for the within-child correlation. The bootstrap technique is illustrated with data from a randomized trial of the effects of vitamin A supplementation on childhood morbidity. Standard errors for the main effect of vitamin A are larger using the bootstrap, indicating the presence of positive correlation. Time-to-event models can be useful in the analysis of studies of childhood morbidity, especially for situations with seasonality, waning of effect, or multiple events. The bootstrap provides an appropriate, straightforward method for handling within-child correlation in such settings.