The effect of misclassification on the estimation of association: a review

Misclassification, the erroneous measurement of one or several categorical variables, is a major concern in many scientific fields and particularly in psychiatric research. Even in rather simple scenarios, unless the misclassification probabilities are very small, a major bias can arise in estimating the degree of association assessed with common measures like the risk ratio and the odds ratio. Only in very special cases--for example, if misclassification takes place solely in one of two binary variables and is independent of the other variable ('non-differential misclassification')--is it guaranteed that the estimates are biased towards the null value (which is 1 for the risk ratio and the odds ratio). Furthermore, misclassification, if ignored, usually leads to confidence intervals that are too narrow. This paper reviews consequences of misclassification. A numerical example demonstrates the problem's magnitude for the estimation of the risk ratio in the easy case where misclassification takes place in the exposure variable, but not in the outcome. Moreover, uncertainty about misclassification can broaden the confidence intervals dramatically. The best way to overcome misclassification is to avoid it by design, but some statistical methods are useful for reducing bias if misclassification cannot be avoided.