Rabe C, Gefeller O
Department of Medical Informatics, Biometry and Epidemiology, Friedrich Alexander University Erlangen-Nuremberg, 91054 Erlangen, Germany.
Methods Inf Med. 2006;45(4):404-8.
Different approaches to partition the attributable risk into exposure-specific components are methodologically evaluated.
Two methods of partitioning the attributable risk in a multifactorial situation have been suggested. One is based on a solution adopted from game theory, the Shapley value, whereas the other recently suggested approach uses a heuristically motivated proportional weighting scheme. These two concepts are reviewed and compared in a situation with three exposure factors. A hypothetical numerical example is discussed illustrating differences in the case of complex interaction structures.
The two methods are found to differ in two critical features that affect the outcome of partitioning: i) including or ignoring the full interaction structure between exposure factors involved in the partitioning, ii) using an equal or proportional weighting scheme for the marginal excess risks of the exposures. As a result, not only the individual partial attributable risks for the exposure factors may be quantitatively different between the methods, but also their ranking depends on the partitioning approach.
The epidemiologic properties of the partitioning procedure based on the Shapley value are known and fit to the needs of epidemiologic applications. The alternative approach recently suggested can lead to considerably different results. As long as its epidemiologic properties are not fully understood, the traditional partitioning method should be given preference in practical applications.
对将归因风险划分为特定暴露成分的不同方法进行方法学评估。
提出了两种在多因素情况下划分归因风险的方法。一种基于博弈论中的一种解法，即夏普利值，而另一种最近提出的方法使用启发式动机的比例加权方案。在有三个暴露因素的情况下，对这两种概念进行了回顾和比较。讨论了一个假设的数值例子，说明在复杂相互作用结构情况下的差异。
发现这两种方法在影响划分结果的两个关键特征上存在差异：i）在划分中包含或忽略所涉及暴露因素之间的完整相互作用结构，ii）对暴露的边际超额风险使用相等或成比例的加权方案。结果，不仅两种方法之间暴露因素的个体部分归因风险在数量上可能不同，而且它们的排序也取决于划分方法。
基于夏普利值的划分程序的流行病学特性是已知的，并且符合流行病学应用的需求。最近提出的替代方法可能会导致相当不同的结果。只要其流行病学特性尚未完全理解，在实际应用中应优先使用传统的划分方法。
