A sensitivity analysis to separate bias due to confounding from bias due to predicting misclassification by a variable that does both

Variables that predict misclassification of exposure, outcome, or a confounder cannot be controlled by techniques that adjust for predictors of risk. They must be controlled by external adjustments. We confronted an analysis in which a variable predicted misclassification of the exposure and of a con founder. The same variable confounded the exposure-outcome relation. The analysis focused on the relation between less than-definitive therapy and breast cancer mortality in the 5 years after diagnosis. Receipt of less-than-definitive prognostic evaluation predicted misclassification of definitive therapy (the exposure) and stage (a confounder). Prognostic evaluation also confounded the therapy-breast cancer mortality relation, We used a sensitivity analysis to separate the misclassification biases from the confounding bias. The relative hazard associated. with less-than definitive therapy in the original multivariable model equaled 1.75 (95% confidence interval = 1.02-3.00), The median estimate in 2,500 repetitions of the sensitivity analysis was a relative hazard of 1.64, and 90% of the estimates fell between 1.47 and 1.83. The sensitivity analysis suggests that less-than definitive therapy confers an excess relative hazard of breast cancer mortality in the 5 years after diagnosis. The original analysis, which adjusted for confounding by prognostic evaluation but not its misclassification biases, overestimated the relative hazard.