Once the subjects to be included in a study have been selected, the focus of research shifts to the collection of information about the subjects. There are two critical types of data that must be captured in any comparative study. First, for each subject we need to know the exposure status. Second, we need to record the value of the outcome variable. Errors of measurements that arise for either the exposure or the outcome can result in a biased estimate of the treatment effect.
Information bias is sometimes used also to describe the implications of measurement error in covariates. We explained in Chapters 6 and 7 how statistical adjustment techniques are based on the idea of conditioning on covariates. In theory, conditioning on these covariates can reduce or eliminate confounding bias. However, sometimes the covariates are measured with error. In the final section of this chapter, we briefly discuss how using a fallible covariate can affect the adjusted empirical effect.
Like selection bias and confounding, the concept of information bias has evolved in the absence of an explicit causal model. As a result, the essential nature of the problem has not been clearly articulated. Information bias is conceived as a discrepancy between the effect based on actual (distorted) data and the “true” effect measured without any distortion. But the true effect is simply the statistical association that would have been measured without error. Classical measurement theory ...