Evaluation of statistical modeling approaches for epidemiologic studies of low-dose radiation health effects. (3rd April 2022)

• Record Type: Journal Article
• Authors: Golden, Ashley P.; Cohen, Sarah S.; Chen, Heidi; Ellis, Elizabeth D.; Boice, John D.
• Is Part Of: International journal of radiation biology. Volume 98:Number 4(2022)
• Page Start: 572
• Page End: 579
• Publication Date: 2022-04-03
• Subjects: Epidemiology -- models of radiation responses -- statistical models -- Cox proportional hazards model -- Poisson regression model; Radiation -- Physiological effect -- Periodicals; Radiobiology -- Periodicals; 571.45
• DOI: 10.1080/09553002.2018.1554924
• Languages: English
• ISSNs: 0955-3002

Abstract

Abstract: Purpose: A substantial body of epidemiologic literature addresses risks associated with occupational radiation exposure but comparing results between studies is often difficult as different statistical models are commonly used. It is unclear whether different methods produce similar results for estimates of radiation risk when applied to the same data. The goal of this study was to compare the radiation risk estimates for leukemia other than chronic lymphocytic leukemia (non-CLL) and ischemic heart disease (IHD) produced by both Cox and Poisson regression models for time-dependent dose-response analyses of occupational exposure. Materials and methods: For brevity, this methods paper presents the results from one cohort, the Nuclear Power Plant workers (NPP), though the evaluation considered five cohorts of varying size and exposure as part of the Million Worker Study. Cox Proportional Hazards models, with age as the underlying timescale for hazard, were conducted using three computer software programs: SAS, R, and Epicure. Doses lagged 2 years for non-CLL and 10 years for ischemic heart disease were treated as time-dependent exposures at the annual level and were examined both in categories and as a continuous term. Hazard ratios (HR) and 95% confidence intervals (CI) were reported for each model in SAS and R, while the Peanuts program of Epicure was utilized to produce Excess Relative Risk (ERR) estimates and 95% CI. All models were adjusted for gender and year of birth. Four piece-wise exponential Poisson models (log-linear regression for rate) were developed with varying cutpoints for age strata from very fine to broad categories using both R and the Amfit program in Epicure for ERR estimates. Results: Comparable estimates of risk (both RR and ERR) were observed from Cox and Poisson models, regardless of software utilized, as long as appropriately narrow categories of age were utilized to control the confounding of age in Poisson models. The ERR estimates produced in Epicure tended to agree very closely with the HR or RR estimates, and the statistical software program used had no impact to risk estimates for the same model. Conclusions: As computational power is no longer the burden today as it has been in the past, the results of this evaluation support the use of the Cox proportional hazards or the ungrouped Poisson approach to analyzing time-dependent dose-response relationships to ensure that maximum control over the confounding of age is achieved in studies of mortality for cohorts occupationally exposed to radiation.