S01 - Session O3 - Flexible linear models for complex data in horticultural tree breeding

Authors: Craig Hardner *, Joanne Defaveri

Response to selection and improved confidence in the commercial value of new germplasm is dependent on objective and accurate prediction of its performance. Nevertheless, many of the traits that influence the commercial value of new germplasm are complex. Here we review opportunities for implementing linear mixed models to obtain the best (most accurate) and unbiased prediction of these traits. Firstly, we discuss how multiple sampling of a single tree is pseudo-replication and cannot be considered replication of the genetic treatment, hence will lead to incorrect estimates of heritability and tests of significance. Secondly, many traits exhibit a trend with tree age, and not accounting for the correlation in errors and trends in variance may bias predictions. Thirdly, traits may exhibit genotype-by-environment interaction. For these traits, not accounting for this interaction may overestimate heritability and deliver inaccurate predictions of genetic worth. In addition, while multi-environment models (and with multiple measurements) may be complex, approaches that reduce the complexity of these models but maintain the correct covariance structures can be used. Fourthly, many field trials may have spatial trends and correlations in data, and not accounting for these again can result in biased predictions. We present how these issues may be dealt with in a linear mixed model framework to assist with the implementation of these approaches in horticultural tree breeding.