217
Improving REML estimates of genetic parameters through penalties on correlation matrices
Penalized REML estimation can substantially reduce sampling variation in estimates of covariance matrices, and yield estimates of genetic parameters closer to population values than unpenalized analyses. A number of suitable penalties based on prior distributions of correlation matrices suggested in the Bayesian literature are described, and an extensive simulation study is presented demonstrating their efficacy. Results show that reductions in `loss' in estimates of the genetic covariance matrix well over 50% are readily obtained for multivariate analyses of small samples, in particular when more than a few traits are considered. Default settings for a mild degree penalization are proposed, which achieve a substantial proportion of possible improvements whilst safe-guarding against over-penalization. These make such penalized analyses suitable for routine use without increasing computational requirements, and are envisaged to become standard procedure in future.
Keywords:
Estimation of genetic parameters
Penalized REML
Priors for correlations
