Recent studies use sibling fixed effects to estimate the influence of the family environment on children, a practice we call the ‘discordant family design’. These studies suffer from a disconnect between the use of within-family variation, on the one hand, and relevant theories which mostly refer to variation between families on the other. In addition, reverse causality, within-family confounding, selection into identification, and measurement error complicate their interpretation further. We discuss three applied examples—the effects of parenting, family income, and neighbourhood context—and provide some general guidance. To avoid misinterpretation, researchers should have a strong grasp of the variance that enters into estimation, and not just the potential confounders a given strategy is designed to deal with.