Accuracy of Single- and Multiple-Trait REML Evaluation of Data Including Non-Random Missing Records

We examined the accuracy of single- and multiple-trait REML procedures by studying estimates of within-individual genetic correlations between an ordered categorical trait and a continuous trait. The traits were derived from simulated bivariate, normally distributed data including selectively deleted records. Ten thousand data sets were generated for each partially factorial combination of two levels of genetic correlation (0.3 and 0.6), and environmental correlation (0.3 and 0.6), and three levels of narrow-sense individual heritability (0.05, 0.15 and 0.25) and mortality (0, 10, 30 and 50%). All data sets consisted of data on 200 unrelated parents, each with 20 halfsib progenies. The accuracy of the evaluations was illustrated in terms of average bias and variation of derived correlation estimates. The average bias values generated by multiple-trait REML were generally low. In contrast, single-trait REML was sensitive to selective deletion of records and systematically underestimated the genetic correlations. For both methods, especially at low heritabilities, the magnitude of the variation was generally high, showing that there is a substantial probability of obtaining seriously misleading genetic correlation estimates if the analysis is based on a single experiment and data include non-random missing records.