Bias Analysis to Guide New Data Collection

Bias analysis serves multiple objectives in epidemiologic data analysis. The objectives most often emphasized are quantification of uncertainty due to systematic errors and reduction in overconfidence by specifying hypotheses that compete with the causal hypothesis. A third objective is the utility of bias analysis to identify strategies for new data collection that will be productive in evaluating the validity of an association. The authors illustrate the value of this objective using two examples. The first example examines the value of comprehensive CYP2D6 genotyping in a study of tamoxifen resistance. Tamoxifen is metabolized primarily by CYP2D6 to more active forms. More than thirty polymorphisms in the CYP2D6 gene reduce its function. We genotyped the most prevalent CYP2D6 polymorphism and found a null association between genotype and breast cancer recurrence in a Danish population. One possibility is that incomplete genotyping of the multiple functional polymorphisms introduced non-differential misclassification and biased the association toward the null. We used bias analysis to evaluate the plausibility of this explanation and to guide a decision about devoting study resources toward more comprehensive genotyping of other polymorphisms in the CYP2D6 gene. The second example examines the association between vitamin K antagonist (VKA) therapy and the incidence of 24 site-specific cancers, using heart valve replacement as an instrumental variable. Earlier studies suggested a protective association between VKA anticoagulants and the incidence of cancer. We observed a null-centered distribution of associations, which may be due to non-differential misclassification of VKA therapy by the instrument. We used bias analysis to evaluate whether this misclassification was likely to explain the null-centered distribution of associations and to guide decisions about conducting a more expensive validation study. In the first example, the bias analysis showed that new data collection would be required to resolve the uncertainty, whereas the second example showed that new data collection was unlikely to be a productive use of scarce study resources.