de Vocht F, Kromhout H, Ferro G, Boffetta P, Burstyn I
Occupational and Environmental Health Research Group, School of Translational Medicine, Faculty of Medical and Human Sciences, The University of Manchester, Manchester M13 8GE, UK.
Occup Environ Med. 2009 Aug;66(8):502-8. doi: 10.1136/oem.2008.042606. Epub 2008 Dec 5.
Residual confounding can be present in epidemiological studies because information on confounding factors was not collected. A Bayesian framework, which has the advantage over frequentist methods that the uncertainty in the association between the confounding factor and exposure and disease can be reflected in the credible intervals of the risk parameter, is proposed to assess the magnitude and direction of this bias.
To illustrate this method, bias from smoking as an unmeasured confounder in a cohort study of lung cancer risk in the European asphalt industry was assessed. A Poisson disease model was specified to assess lung cancer risk associated with career average, cumulative and lagged bitumen fume exposure. Prior distributions for the exposure strata, as well as for other covariates, were specified as uninformative normal distributions. The priors on smoking habits were specified as Dirichlet distributions based on smoking prevalence estimates available for a sub-cohort and assumptions about precision of these estimates.
Median bias in this example was estimated at 13%, and suggested an attenuating effect on the original exposure-disease associations. Nonetheless, the results still implied an increased lung cancer risk, especially for average exposure.
This Bayesian framework provides a method to assess the bias from an unmeasured confounding factor taking into account the uncertainty surrounding the estimate and from random sampling error. Specifically for this example, the bias arising from unmeasured smoking history in this asphalt workers' cohort is unlikely to explain the increased lung cancer risk associated with average bitumen fume exposure found in the original study.
在流行病学研究中，由于未收集混杂因素的信息，可能会出现残余混杂。本文提出了一种贝叶斯框架，用于评估这种偏差的大小和方向。与频率论方法相比，贝叶斯框架的优势在于，混杂因素与暴露及疾病之间关联的不确定性能够反映在风险参数的可信区间内。
为说明该方法，我们评估了在欧洲沥青行业肺癌风险队列研究中，吸烟作为未测量混杂因素所导致的偏差。我们指定了一个泊松疾病模型，以评估与职业平均、累积和滞后沥青烟暴露相关的肺癌风险。暴露分层以及其他协变量的先验分布被指定为无信息正态分布。基于一个亚队列的吸烟流行率估计以及对这些估计精度的假设，将吸烟习惯的先验分布指定为狄利克雷分布。
在这个例子中，中位数偏差估计为13%，表明对原始暴露 - 疾病关联有衰减作用。尽管如此，结果仍然表明肺癌风险增加，尤其是平均暴露水平下。
这个贝叶斯框架提供了一种方法，用于评估未测量混杂因素导致的偏差，同时考虑到估计的不确定性和随机抽样误差。具体针对这个例子，在这个沥青工人队列中，未测量的吸烟史所产生的偏差不太可能解释原始研究中发现的与平均沥青烟暴露相关的肺癌风险增加。
