A biometrical-genetical model for human variation adequately predicts observed correlations for measured intelligence. The estimation procedure is outlined.IN a recent review Jinks and Eaves (1974) stressed the close agreement between observed correlations between relatives for IQ and their expectations on the basis of a simple model. Their model involved five parameters: additive and dominant components of gene action (DR and HR); the marital correlation (a); the correlation (A) between spouses' additive genetical deviations; the component (Er) reflecting the environmental covariation of parent and offspring. The contribution of the environmental influences specific to individuals (E1) is a sixth parameter whose magnitude is fixed by the other five by the restraint that the total variance is unity.Jencks' (1973) application of path coefficients to the analysis of intelligence was partly questioned because of what was believed to be an upper limit upon the value of the path between the genotypes of parent and offspring. No such limit exists in fact so his conclusions cannot be discounted on this basis. More critical, however, was Jinks and Eaves' re-analysis of published correlations (Burt, 1966; Jencks, 1973) which demonstrated that any signficant heterogeneity of heritability estimates obtained from different degrees of relationship can be removed if the contribution of dominance is precisely specified and a weighted least squares procedure is adopted.The correlations analysed have already been tabulated (Jinks and Eaves, 1974), but details of the model and estimation procedure were necessarily omitted and are considered here. The model (table I) is that of Fisher (1918) for the correlations between relatives for a population in equilibrium under assortative mating with additions to specify the contribution of certain plausible environmental components. The model is a modification of that in the review because u has been reparameterised in terms of A and DR.This restriction is necessarily imposed if all the assumptions in Fisher's model are to be tested adequately. It will be seen that although the restraint alters the estimates somewhat the fundamental interpretation remains unaltered.In view of the fact that biometrical genetical models are usually linear and this model is substantially non-linear it is appropriate to outline the estimation procedure.The model for our vector of observed correlations x may be written x=f(4)-i-8where f() represents a vector of functions of the parameters , and g is a vector of deviations. We require to minimise the weighted sum of squared deviations(1)