In plant breeding, multilocation trials form the major means for the comparative evaluation of cultivars. Based on such trials, recommendations may be given to farmers. Commonly, cultivars that did best on average are recommended for all locations. However, the intended growing region can be ecologically heterogeneous and cultivar-location interactions can be substantial. As a result, the cultivar with the largest mean is not generally the best in all locations of the region. When information is available on the response of a cultivar to varying environmental conditions, cultivar-location interactions can be (partly) predicted, thus allowing for more specific recommendations. This article discusses a regression-based approach for predicting cultivar performances using covariate information on locations such as average rainfall and soil type. Special emphasis is given to the selection of covariates useful for prediction. The mean squared error of prediction is used as a selection criterion.

Key Words
Covariate selection; Cultivar-location interaction; Genotype-environment interaction; Mean squared error of prediction; Mixed model.

Hans-Peter Piepho is Senior Statistician, Biometrics, Institut für Nutzpflanzenkunde, Universität Kassel, 37213 Witzenhausen, Germany. Jean-Baptiste Denis is Senior Statistician, Unité de Biométrie, INRA, 78026 Versailles, France. Fred A. van Eeuwijk is Associate Professor, Dept. of Agricultural, Environmental, and Systems Technology, Subdepartment Mathematics, Wageningen Agricultural University, 6703 HA Wageningen, The Netherlands.