Comparison of statistical methods for recurrent event analysis using pediatrics asthma data

Abstract

Summary When the same type of event is experienced by a subject more than once it is called recurrent event, which possess two important characteristics, namely “within‐subject correlation” and “time‐varying covariate.” As a result, the traditional statistical methods do not work well on recurrent event data. Over the past few decades, many alternatives methods have been proposed for the analysis of recurrent event data. In this article, the six most prominent methods for recurrent event analysis have been compared on pediatric asthma data. Three variance corrected models (viz “Anderson and Gill [AG] model,” “Prentice, William, and Peterson‐Counting Process [PWP‐CP] model,” and “Prentice, William, and Peterson‐Gap Time [PWP‐GT] model”) and three corresponding frailty variants (AG‐frailty, PWP‐CP‐frailty, and PWP‐GT‐frailty) were compared using three mathematical criterion (AIC, BIC, and log‐likelihood) and one graphical criteria (Cox‐Snell goodness of fit, visual test). All model comparison indices showed the PWP‐GT model as the most appropriate model on asthma data over other models. By using PWP‐GT model, seven predictors of asthma exacerbation (viz “abdominal pain at previous visit,” “Z5 (%) at previous visit,” “diagnosis of asthma at previous visit,” “calendar month of exacerbation,” “history of maternal asthma,” “monthly per capita income,” and “emotional stress”) were identified. The PWP‐GT model was identified as the most appropriate model over other models on pediatrics asthma data.

Previous psychiatric hospitalizations as risk factors for single and multiple future alcohol‐related hospitalizations in patients with alcohol use disorders

“…We then assessed the time to multiple AUD rehospitalizations using the Prentice, Williams and Peterson gap-time model (PWP-GT). PWP-GT models assume that recurrent events within the individual are related: individuals are not at risk for the n th AUD hospitalization until they experience their (n-1) th [22,23]. As PWP-GT models require a large number of study subjects for every failure time [22,23], based on the number of subjects experiencing multiple AUD hospitalizations, we set the maximum number of rehospitalizations at three.…”

“…PWP-GT models assume that recurrent events within the individual are related: individuals are not at risk for the n th AUD hospitalization until they experience their (n-1) th [22,23]. As PWP-GT models require a large number of study subjects for every failure time [22,23], based on the number of subjects experiencing multiple AUD hospitalizations, we set the maximum number of rehospitalizations at three. Both Cox and PWP-GT models were then fitted with and without prescription covariates.…”

Adoption of the Website and Mobile App of a Preventive Health Program Across Neighborhoods With Different Socioeconomic Conditions in the Netherlands: Longitudinal Study

“…As reflected in Figure 1 , the adoption of the website and the mobile app were sequential events, with each participant being at risk for only 1 of these events at a time. The modeling approach that handles this structure best is an extension of the classical Cox model [ 29 ], namely the variants of the Prentice, Williams, and Peterson (PWP) model [ 30 - 32 ]. To answer the research question of this study related to differences in the rate of adoption between means of delivery, the PWP Gap–Time (PWP-GT) model is most appropriate, which estimates the effects of the following event since the time from the previous event [ 30 , 32 ].…”

“…The modeling approach that handles this structure best is an extension of the classical Cox model [ 29 ], namely the variants of the Prentice, Williams, and Peterson (PWP) model [ 30 - 32 ]. To answer the research question of this study related to differences in the rate of adoption between means of delivery, the PWP Gap–Time (PWP-GT) model is most appropriate, which estimates the effects of the following event since the time from the previous event [ 30 , 32 ]. This is achieved using time-dependent strata, where the hazard function is allowed to vary from event to event [ 33 ].…”