An Application of Semiparametric Structured Additive Model to Cancer Data

Abstract

In many epidemiological studies where times to event data are clustered, introducing frailties in the Cox model can account for heterogeneity induced by such clustering. Analyses were carried out using data collected on a sample of cancer patients from University of Ilorin Teaching Hospital, using Full Bayesian inference based on Markov Chain Monte Carlo (MCMC) simulation technique. The approach allows the estimation of very complex and realistic models.  The results showed that sex and age were significant risk factors associated with death from some selected cancers. The risk of dying from the selected cancers was observed to progressively increase as age of patient’s increases. Using Deviance Information Criterion (DIC) for model comparison, it was observed that the flexible semi parametric additive P-spines model, which allows for nonlinearity due to metrical covariate age, was better than the model that introduced metrical age linearly as fixed effect. It was also found that models that accounted for heterogeneity induced by clustering observations were more adequate than those that ignored it. On effect of interaction between sex and age on the death due to cancer, the model that contained interaction between sex and age when metrical age was modeled nonlinearly was observed to be better than those that modeled metrical or categorized age as linear effect.