Empirical Comparison of Non-model- and Model-based Statistical Methods for Cancer Mortality in Occupational Cohorts

Objective: The aim of this study was to compare effect measures from two non-model and three model-based statistical methods in assessing the association between lung cancer/multiple myeloma mortality and exposure to oncogenic viruses. Data from the ongoing Cancer Risk in Workers Exposed to Oncogenic Viruses (CRIWETOV) project for members in a local Union Pension Fund belonging to the United Food &Commercial Workers (UFCW) international union was utilized. The workers were followed–up for mortality from January 1, 1972 to December 31, 2003. This cohort was comprised of workers in poultry slaughtering/processing plants, and non-poultry workers. Effect estimates and 95% confidence intervals were calculated using direct and indirect standardizations (non-model based statistical methods), and Poisson, Cox proportional hazards, and binary/multiple logistic regression models (model based methods). Percentile bootstraps based on the case resampling bootstrapping method were used in generating confidence intervals for the non-model based methods. The entire cohort and subgroups of poultry and non-poultry separately had higher risks of mortality from both malignant diseases (statistically significant for lung cancer) compared to the United States' general population, but slightly lower (statistically not significant) risks among poultry compared to non-poultry workers. Results of comparative effect measures from the various statistical methods were similar under limiting conditions, with a very slight difference in variability and precision. Although these methods may lead to similar general conclusions, they require different underlying assumptions/specifications that the data must meet for their applications to be valid.