The design of family studies to estimate the value of an intraclass correlation coefficient p is considered when ni individuals are to be selected from each of k families, i = 1, 2, ..., k. In particular, the accuracy of a balance design (ni = n, i = 1, 2, ..., k) for estimating p is compared with the accuracy of an unbalanced "natural" design, in which the ni are sampled at random from family size distributions that tend to occur in practice. It is found for two different estimators of p that the balanced design is usually preferable, but only to a small degree if the number of families sampled is greater than 50.