Optimising early selection using longitudinal data

This study analysed the use of longitudinal data, i.e. repeat ed assessment of the same individuals at different ages, in the context of early selecti on. Autoregressive relationships, banded correlations and unstructured ('un smoothed') matrices were used to model the additive genetic covariance matr ix (Go) for 10 total height measurements of a Pinus radiata open-pollinated progeny test. We examined the effects on response to selection of inferred covariance structure, mass versus combined selection, one or multiple asse ssments, and two breeding-delay intervals. End results are expressed as pre dicted average gain per year. The patterns of predicted response to selecti on vary widely between inferred covariance structures. Considering the auto regressive model (based on logarithm of age ratios between assessments) as an example, the effect of combining information from relatives on response to selection is more important (16% to 41% extra gain) than using extra mea surements (2% to 25%), when predicting individual breeding values, although the economics of extra gain vs extra assessment costs must be carefully an alysed. It is expected that using multiple assessments could be advisable f or datasets with lower genetic autocorrelations. An approximate comparison across covariance models showed the autoregressive model to exhibit the bes t ability to produce 'correct' selections as well as the highest predicted response to selection.