Institute of Transport Economics, Gaustadalléen 21, NO-0349 Oslo, Norway.
Accid Anal Prev. 2011 Jan;43(1):253-64. doi: 10.1016/j.aap.2010.08.018. Epub 2010 Oct 8.
This paper discusses the application of operational criteria of causality to multivariate statistical models developed to identify sources of systematic variation in accident counts, in particular the effects of variables representing safety treatments. Nine criteria of causality serving as the basis for the discussion have been developed. The criteria resemble criteria that have been widely used in epidemiology. To assess whether the coefficients estimated in a multivariate accident prediction model represent causal relationships or are non-causal statistical associations, all criteria of causality are relevant, but the most important criterion is how well a model controls for potentially confounding factors. Examples are given to show how the criteria of causality can be applied to multivariate accident prediction models in order to assess the relationships included in these models. It will often be the case that some of the relationships included in a model can reasonably be treated as causal, whereas for others such an interpretation is less supported. The criteria of causality are indicative only and cannot provide a basis for stringent logical proof of causality.
本文讨论了将因果关系操作标准应用于多元统计模型的情况，这些模型旨在识别事故计数中系统变化的来源，特别是代表安全处理的变量的影响。讨论的基础是制定了九条因果关系标准。这些标准类似于在流行病学中广泛使用的标准。为了评估多元事故预测模型中估计的系数是否代表因果关系还是非因果统计关联，所有因果关系标准都是相关的，但最重要的标准是模型对潜在混杂因素的控制程度。举例说明了如何将因果关系标准应用于多元事故预测模型，以评估这些模型中包含的关系。通常情况下，模型中包含的一些关系可以合理地视为因果关系，而对于其他关系，则不太支持这种解释。因果关系标准只是指示性的，不能为因果关系的严格逻辑证明提供依据。
