Models are proposed as a basis for unbiased and biased estimation from sample survey data of a regression relation. Unbiased estimates of the variances and covariances of estimates of the regression coefficients are obtained for certain cases but do not appear to exist for all cases. It is suggested that past surveys, conducted mostly to obtain efficiently estimates of means and totals, be analyzed to yield estimates of regression relations, to obtain information needed in designing future surveys, and to test the adequacy of the proposed models.