[Bias from dependent errors in observational studies].

Measurement errors in two variables are dependent when the degree of error in one of them correlates with the degree of error in the other. When dependent error affects measured exposure and measured outcome, the estimated association between the two is likely to be falsely inflated. Such information bias is probably not uncommon in cross-sectional studies providing data on both variables from questionnaires. This often occurs in published studies, but there seems to be limited awareness of the problem. The basic source of dependent error is usually normal variation in certain personality traits, but it may also be in more transitional moods in the study population or inadequate measurement tools. The major precaution that should be taken in order to eliminate bias from dependent error is to break the bond between information on exposure and outcome by gathering data from two separate sources. We should also recognise that not all study designs and not all data are suited for establishing etiology.