We review the conditions under which cross-sectional family data (e.g., data on twin pairs or adoptees and their adoptive and biological relatives) are informative about direction of causation. When two correlated traits have rather different modes of inheritance (e.g., family resemblance is determined largely by family background for one trait and by genetic factors for the other trait), cross-sectional family data will allow tests of strong unidirectional causal hypotheses (A and B are correlated "because of the causal influence of A on B" versus "because of the causal influence of B on A") and, under some conditions, also of the hypothesis of reciprocal causation. Possible sources of errors of inference are considered. Power analyses are reported which suggest that multiple indicator variables will be needed to ensure adequate power of rejecting false models in the presence of realistic levels of measurement error. These methods may prove useful in cases where conventional methods to establish causality, by intervention, by prospective study, or by measurement of instrumental variables, are unfeasible economically, ethically or practically.