Department of Epidemiology, University of California, Los Angeles, CA 90095-1772, USA.
Int J Epidemiol. 2009 Dec;38(6):1662-73. doi: 10.1093/ije/dyp278. Epub 2009 Sep 9.
I present some extensions of Bayesian methods to situations in which biases are of concern. First, a basic misclassification problem is illustrated using data from a study of sudden infant death syndrome. Bayesian analyses are then given. These analyses can be conducted directly, or by converting actual-data records to incomplete records and prior distributions to complete-data records, then applying missing-data techniques to the augmented data set. The analyses can easily incorporate any complete ('validation' or second-stage) data that might be available, as well as adjustments for confounding and selection bias. The approach illustrates how conventional analyses depend on implicit certainty that bias parameters are null and how these implausible assumptions can be replaced by plausible priors for bias parameters.
我提出了一些贝叶斯方法的扩展，以解决存在偏差的情况。首先，使用婴儿猝死综合征研究的数据说明了一个基本的分类错误问题。然后给出了贝叶斯分析。这些分析可以直接进行，也可以通过将实际数据记录转换为不完整记录，将先验分布转换为完整数据记录，然后将缺失数据技术应用于扩充数据集。这些分析可以轻松地纳入任何可能可用的完整（“验证”或第二阶段）数据，以及对混杂和选择偏差的调整。该方法说明了传统分析如何依赖于对偏差参数为零的隐含确定性，以及如何用合理的偏差参数先验来替代这些不合理的假设。
