New methods for detecting and modelling heterogeneity in survival responses

schedule le mardi 27 mars 2018 de 10h45 à 11h45

Organisé par : Castillo, Fischer, Giulini, Gribkova, Levrard, Roquain, Sangnier

Intervenant : Olivier Bouaziz (Universty Paris Descartes)
Lieu : UPMC, salle 15-16.413

Sujet : New methods for detecting and modelling heterogeneity in survival responses

Résumé :
In survival analysis it is quite common that heterogeneity between patients results in various survival response distributions. This heterogeneity can be controlled through known covariates (such as date of birth, age at diagnosis, gender, treatment, co-exposure, BMI, etc.) using regression-type models such as the Cox proportional hazard model or by performing stratified analyses. Other types of heterogeneous dataset arise when the hazard rate changes over the calendar time in a cohort study and specific models like age-period-cohort have been extensively studied to take into account this kind of heterogeneity. In the present talk, we present two new approaches to deal with this kind of heterogeneity.

The first one considers survival heterogeneity as a breakpoint model in an ordered sequence of survival responses. These responses might be ordered according to any numerical covariate like the date of diagnosis. In such a model, we aim at estimating the hazard rates in each homogenous region using a Cox model and at accurately providing the number and location of the breakpoints. A constrained Hidden Markov Model (HMM) is implemented which performs a full change-point analysis in a segment-based model providing linear estimates of the parameters (using the EM algorithm) and a full specification of the posterior distribution of change points.

The second method specifically models age, period and cohort effects (with the relation period=age+cohort) through a bi-dimensional hazard estimation. Since the number of parameters can be quite large compared to the sample size, an L0 penalized likelihood method is implemented to avoid overfitting issues. The method is based on the adaptive ridge method from Frommlet and Nuel 2016. It allows to detect the homogeneous hazard rate regions in the age/period, age/cohort and cohort/period planes and can be seen as an extension of classical age-period-cohort models allowing for interactions between the three effects.