Confounding is one of the most common and important biases in primary care research. This article explains the genesis and effects of two common misconceptions of confounding: 1) Confounding can be assessed with a statistical test. 2) All covariates should be included in a multivariate model to control confounding. Assessment of confounding by testing the statistical significance of baseline differences or the significance of a potential confounding factor in a multivariate model can produce underestimates or overestimates of the true association between an exposure and an outcome. Inclusion of all covariates in a multivariate model may lead to controlling for variables that are not, in fact, confounders. This may produce underestimates or overestimates of the effect in question, as well as artificially widened confidence intervals. Both of these misconceptions can lead to profound misinterpretation of research results. To prevent problems resulting from these misunderstandings, researchers should consider drawing causal models prior to conducting the research and use the change-in-estimate criterion, rather than a statistical test, to detect confounding.