Workup bias in prediction research.

Studies often suggest that accepted clinical predictors actually have little predictive strength. One explanation for some such results is the presence of workup bias. To explore the effects of workup bias in prediction research, the authors modeled the effects of workup bias on the ability of early clinical findings to predict intracerebral hemorrhage in patients with stroke. In a simulated biased sample, workup bias resulted in distorted operating characteristics for those clinical findings influencing application of the "gold standard" and for other related findings. Sensitivity was increased, but both specificity and likelihood ratios were decreased in the biased sample. Workup bias can spuriously decrease predictive abilities for accepted clinical findings when such findings guide application of the "gold standard." Investigators should be aware of the potential effects of workup bias, search for clues to its occurrence, and interpret study results carefully when it is present.