Effects of data structure on variance of prediction error and accuracy of genetic evaluation.

Abstract

Several features of data structure were studied to determine their effects on variance of prediction error and accuracy of evaluation. Assigning 50 sires with progeny to a portion of 10, 25, or 50 contemporary groups according to a sire model with and without additive genetic relationships, or assigning 50 individuals with their own record to one of 2, 5, or 10 contemporary groups according to an animal model, established the designs. Additive genetic relationships were based on stimulated pedigree files. Low, medium, and high heritabilities (.10, .25, and .40, respectively) were considered. The inverse of coefficient matrices gave variances of prediction error. Populations derived from the sire model (n = 8,100) consisted solely of progeny-tested individuals. For them, number of progeny had a quadratic (P < .001) association with variance of prediction error (R2 = 56 to 82%), which selection index theory underestimated when there were < 100 progeny. Number of direct connections (sires of contemporaries of progeny) together with progeny numbers explained variance of prediction error (R2 = 76 to 90%) better than either variable alone. With no direct connections, variance of prediction error was maximum unless a relative with at least one direct connection itself existed. Populations derived from the animal model (n = 900) consisted of animals with designs representing a progeny test, performance test, or a combination of both (34, 41, and 25% of the total, respectively). For performance-tested animals (without progeny), number of genetic connections was not highly correlated with variance of prediction error (r = -.10, across h2), but relatives prevented zero accuracies when contemporary groups consisted of one animal. Even when animals had no relatives, more than five members per contemporary group gave little additional increase in accuracy. For other than a progeny test, designs were complex, being described by many variables that were confounded.