Division of Environmental Health Sciences, School of Public Health, University of California, Berkeley, California, USA.
Occup Environ Med. 2013 Oct;70(10):736-42. doi: 10.1136/oemed-2012-101332. Epub 2013 Jul 19.
The healthy worker survivor effect is a bias that occurs in occupational studies when less healthy workers are more likely to reduce their workplace exposures. When variables on the pathway from health status to exposure are measured, g-methods can avoid this bias. However, studies in which follow-up ends at employment termination have additional potential for selection bias. This paper examines the structure of the healthy worker survivor effect, compares results with and without censoring at employment termination, and addresses how to prevent bias when such censoring occurs.
G-estimation of structural accelerated failure time models was applied in the United Autoworkers-General Motors cohort study to examine relationships between metalworking fluid exposure and cause-specific mortality. Subjects were followed from hire through 1994, regardless of employment status. To answer the central question, g-estimation analysis was repeated after truncating at employment termination and censoring outcomes that occurred thereafter, with adjustment for censoring by inverse probability weighting.
Using full follow-up time, HRs were estimated for all-cause mortality (1.09), ischaemic heart disease death (1.19), and death from any cancer (1.09), comparing 5 years of metalworking fluid exposure to no exposure. For all three outcomes, the HR estimates based on data censored at termination of employment were below 1 (respectively, 0.92, 0.97, 0.79).
In this application, g-estimation together with weighting did not prevent selection bias due to employment termination. However, the bias might be avoided in studies with measured health-related variables on the pathway from health status to employment termination.
健康工人幸存者偏差是一种在职业研究中出现的偏差，当身体状况较差的工人更有可能减少其工作场所暴露时，就会出现这种偏差。当测量健康状况到暴露之间的路径上的变量时，g 方法可以避免这种偏差。然而，当随访在就业终止时结束时，这些研究还存在额外的选择偏差的可能性。本文研究了健康工人幸存者偏差的结构，比较了在没有和有就业终止时的结果，并讨论了当发生这种删失时如何防止偏差。
在联合汽车工人-通用汽车队列研究中，应用 g 估计结构加速失效时间模型来研究金属加工液暴露与特定原因死亡率之间的关系。研究对象从入职开始随访到 1994 年，无论其就业状况如何。为了回答核心问题，在对就业终止进行截断并对随后发生的结果进行删失，并通过逆概率加权进行删失调整后，重复 g 估计分析。
在全随访时间中，与无金属加工液暴露相比，所有原因死亡率（1.09）、缺血性心脏病死亡率（1.19）和任何癌症死亡率（1.09）的 HR 估计值为 5 年的金属加工液暴露。对于所有三个结果，基于就业终止时的数据删失的 HR 估计值均低于 1（分别为 0.92、0.97 和 0.79）。
在这项应用中，g 估计与加权并没有防止由于就业终止而导致的选择偏差。然而，在具有测量健康相关变量的研究中，可能会避免健康状况到就业终止之间的路径上的偏差。
