Data structure designs for breeding value estimation of performance-tested boars using mixed-model methodology were compared. Computer models were based on estimates of parameters from the literature and from results of a survey of test station managers. Results were compared using accuracy (the correlation of true and estimated breeding values) and prediction error variance (PEV). The single-trait animal model included a fixed effect due to station-season, a random effect due to breeding value for ADG or backfat, and a random error term. Family size, number of families per test, and relationships among animals within and across tests were varied. Prediction error variance decreased faster for small families than for large ones as number of families increased, but increasing numbers of animals per pen was most important, especially if test size was optimized. With no other genetic ties, full-sibs were much more accurately evaluated than half-sibs. Designs that included sire ties among families within a station-season resulted in increased PEV. Increasing the number of full-sibs and(or) increasing the number of families per test would help to optimize PEV and correct this problem. Tying station-seasons with the relationship matrix improved the average accuracy of predicted breeding values. Placing full-sibs in different stations resulted in the greatest accuracy of evaluation, but a large number of half-sib (sire) ties resulted in comparable accuracies. Half-cousin ties did not improve accuracy of evaluation but could result in significant genetic progress by increasing the selection differential.