A hybrid likelihood algorithm for risk modelling

Abstract

The risk of radiation-induced cancer is assessed through the follow-up of large cohorts, such as atomic bomb survivors or underground miners who have been occupationally exposed to radon and its decay products. The models relate to the dose, age and time dependence of the excess tumour rates, and they contain parameters that are estimated in terms of maximum likelihood computations. The computations are performed with the software package EPI-CURE, which contains the two main options of person-by person regression or of Poisson regression with grouped data. The Poisson regression is most frequently employed, but there are certain models that require an excessive number of cells when grouped data are used. One example involves computations that account explicitly for the temporal distribution of continuous exposures, as they occur with underground miners. In past work such models had to be approximated, but it is shown here that they can be treated explicitly in a suitably reformulated person-by person computation of the likelihood. The algorithm uses the familiar partitioning of the log-likelihood into two terms, L1 and L0. The first term, L1, represents the contribution of the 'events' (tumours). It needs to be evaluated in the usual way, but constitutes no computational problem. The second term, L0, represents the event-free periods of observation. It is, in its usual form, unmanageable for large cohorts. However, it can be reduced to a simple form, in which the number of computational steps is independent of cohort size. The method requires less computing time and computer memory, but more importantly it leads to more stable numerical results by obviating the need for grouping the data. The algorithm may be most relevant to radiation risk modelling, but it can facilitate the modelling of failure-time data in general.