A Structural Approach to Analyzing Occupational Cohorts: The Healthy-Worker Survivor Effect and an Application in the Trucking industry Cohort

摘要

Background: Occupational cohort studies are often challenged by the presence of the Healthy Worker Survivor Effect (HWSE). While standard methods of analysis may be biased in the presence of HWSE, g-estimation of structural failure time models has been proposed as an alternative for reducing bias due to time-varying confounding by a variable affected by prior exposure. Aim: We assess the effect of years worked in specific jobs in the trucking industry on mortality, using a Structural Cumulative Failure Time Model (SNCFTM), to adjust for time-varying confounding. We also apply a Structural Accelerated Failure Time Model (SNAFTM) and compare both g-estimation results to results from standard Cox proportional hazards models. Methods: We used g-estimation of a SNCFTM to assess the effect of major job titles in the trucking industry on total and ischemic heart disease (IHD) mortality, in 30, 758 males employed in the unionized U.S trucking industry in 1985. Exposure was defined based on detailed work records. We also applied g-estimation of a SNAFTM and standard Cox proportional hazards models. The time-varying health surrogate was time taken off work. Results: The Risk Ratio obtained from the SNCFTM for total mortality comparing the risk had everyone been exposed as a Long-Haul Driver for the duration of follow up to the risk had nobody been exposed was 1.42 (95% CI: 1.16 -1.77), and the RR for IHD mortality was 1.39 (95%: 1.01 - 2.09). SNAFTM and SNCFTM Hazard Ratio (HR) approximations were similar, but HRs from Cox proportional hazards models were lower. Conclusions: Effect estimates from the SNCFTM and SNAFTM indicate a higher all cause and IHD mortality risk for Long-Haul Drivers than Cox models. Thisdifference may partially reflect better control for the HWSE by the structural models. Structural failure time models may be useful in the analysis of occupational survival data and provide additional advantages compared to standard methods.