The use of regression analyses in a cohort mortality study of welders

Using an internal nonexposed comparison group, and two types of regression analysis, we have reanalyzed the data from a cohort mortality study of 3247 welders and 5432 nonwelders who worked in Western Washington during the period 1950-1976. Previous analyses relied primarily on the standardized mortality ratio (SMR) for lung cancer as a summary statistic to estimate relative risk; welders were compared to the U.S. population, stratified by age, sex, race, and calendar time. In the present study, nonwelders from the same union were used as the reference group; an internal referent group is often preferable to the U.S. population, because it is likely to have shared similar lifestyles (e.g. smoking habits) as the exposed. Cox regression was used to estimate relative risk. In addition, the data was analyzed via logistic regression in order to compare Cox to logistic regression. A regression analysis of these data enabled us to analyze exposure as a continuous variable and simultaneously adjust for potentially confounding variables without stratification. Interactions terms could also be easily tested. The previous analysis found as SMR for lung cancer of 1.32 (p = 0.06). Cox regression with an internal comparison group yielded similar results. Issues discussed include the differences between logistic and Cox regression, the appropriate variable for time in Cox regression, and the definition of risk set in Cox regression. Finally, a newly available program is discussed which selects random samples of the risk sets used in Cox regression, in order to do a case-control analysis.