Effects of interaction, confounding and observational error on attributable risk estimation.

Abstract

The properties of attributable risk are discussed for situations in which there are several risk factors that are possibly interacting or confounded. Conditions are identified when the attributable risk among the exposed is constant, when the marginal attributable risk estimate is valid, and when the public health effects of separate risks are additive. Such conditions reflect, in various ways, the interaction and confounding of the different risk factors involved. For diseases with more than two risk factors, these conditions are sufficient but not necessary; thus it is possible to have additive public health effects of two risk factors even though they are confounded and interactive. In contrast, when there are exactly two binary risk factors, the conditions are sufficient and necessary. It is shown that bias in attributable risk through misclassification of exposure arises primarily through insensitivity errors. Particularly with zero false negative rates and equal false positive rates for cases and controls, the attributable risk estimate is unbiased; however, a larger standard error pertains to the estimate based on misclassified data.