Estimation of Spatial Trend and Automatic Model Selection in Augmented Designs

Zakaria Kehel, D. Z. Habash, S. A. Gezan, S. J. Welham, Nachit Miloudi. (1/11/2010). Estimation of Spatial Trend and Automatic Model Selection in Augmented Designs. Agronomy Journal, 102 (6), pp. 1542-1552.

Abstract

In plant breeding, large numbers of test entries are tested annually and across many environments, making field designs that can reduce costs and time highly advantageous. In the ICARDA durum breeding program the augmented design (AD) is used extensively. In this study, two series of multi-environment agronomic trials that used an AD for a set of replicated check lines and unreplicated test lines were analyzed using restricted (or residual) maximum likelihood (REML) estimation with mixed models, allowing for spatial correlation in the experimental layout. The characteristics of a subset of the trials were used to simulate the precision of estimates of variance parameters and test line effects when the test lines were fitted as fixed or random, and to establish the performance of Akaike's information criterion (AIC) and the Bayesian information criterion (BIC) for model selection in this context. With test lines fitted as random, estimates of variance parameters were less biased and more precise, the information criteria chose the correct model more often, and use of the information criteria led to a model that gave good estimates of test line means. Estimation and model selection performed less well with the test lines fitted as fixed effects, but a hybrid method improved both estimation and model selection in this case. Statistical models that improve trial accuracy and reduce costs are important tools when testing a large number of breeding lines over multi-environment trials.