It is generally believed that non-differential misclassification will lead to a bias toward the null-value. However, using one graphical and one numerical example, we show that in situations where underestimation more than overestimation is the problem, non-differential misclassification may lead to a bias away from the null-value for intermediate categories of exposure variables. We show that a true threshold level for an exposure may, subsequently, appear as a dose-response relationship. Underestimation more than overestimation is likely to occur in studies in which self-reports are used to obtain data on activities which respondents voluntarily engage in and which are socially not acceptable or known to be potentially dangerous. Researchers should be aware that both a shown deleterious effect at a low level of exposure and a dose-response relationship may in fact be spurious and due to underestimation of exposure at higher levels.