A0326
Title: A multi-state modeling with Poisson regression utilizing grouped data in a radiation epidemiological study
Authors: Munechika Misumi - Radiation Effects Research Foundation (Japan) [presenting]
Hiromi Sugiyama - Radiation Effects Research Foundation (Japan)
Abstract: By treating the event occurrence hazard as a piecewise constant function, we can perform survival time analysis using Poisson regression. This approach enables flexible modeling of the baseline hazard and accommodates multiple time-dependent covariates within the generalized linear/nonlinear model framework. In radiation epidemiology, particularly for the analysis of long-term follow-up data from epidemiological cohorts, Poisson regression has become a standard methodology. To implement this regression, we must stratify the time-to-event dataset to create grouped data, a complexity that has hindered the execution of more flexible survival analyses in the field. Furthermore, traditional Poisson regression assumes independent and non-informative censoring, an assumption that is invalidated when subjects experiencing a competing event are censored. In response to these challenges, we propose the use of a multi-state model based on Poisson regression with grouped data to analyze long-term follow-up data. As a motivating example, we use data from the Life Span Study (LSS) of Japanese atomic bomb survivors. The LSS dataset comprises over 120,000 survivors from Hiroshima and Nagasaki, with a follow-up length exceeding 50 years. The analysis provided a comprehensive interpretation of the relationship between radiation exposure and incidence of rectal cancer and adenoma, which are multiple events in a carcinogenesis pathway.
