|So, Yoon-Sup -|
Submitted to: Crop Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: November 10, 2009
Publication Date: November 10, 2009
Citation: So, Y., Edwards, J.W. 2011. Predictive ability assessment of linear mixed models in multienvironment trials in corn (Zea mays L). Crop Science. 51:542-552. Interpretive Summary: Identifying plant varieties (commercial and experimental) that provide superior performance for growers across a wide range of growing conditions is a very expensive task. Varieties must be evaluated in many environments and performance data summarized in order to identify not only the best varieties, but also those with the most stable performance. Several statistical methods of analyzing crop performance data were compared in order to determine the best comparisons among varieties can be obtained by assuming each test-site is equally informative. We compared 25 different models with unbalanced data sets to identify the best method for comparing varieties. We found that different environments in corn yield tests can be quite unequally informative and that modeling the differences in data quality among environments can provide better comparisons between cultivars. Recommendations were made for analysis of yield trial data that will benefit seed companies and producers trying to choose the best breeding lines and culitvars.
Technical Abstract: Prediction of future performance of cultivars is an important objective of multi environment trials (MET). To achieve this goal, predictive ability of statistical models is often examined using cross validation that partition a data set into modeling and validation data. In the present paper, a series of linear mixed models with varying degrees of heterogeneous genotypic variance, correlation and error variance structure were compared for their ability to predict performance in an untested environment in 51 unbalanced data sets from the Iowa Crop Performance Test for corn. Simulation studies were conducted to investigate the relationship between variance component estimates in data sets with small sample size due to unbalancedness and prediction ability of the models. In most cases there was no substantial improvement predictive precision among models that included heterogeneity of genotypic variance-covariance components, but the best prediction model included heterogeneous environment-specific error variances in 63% of data sets analyzed. The largest differences in predictive precision among models appeared to be due to poor estimation of genotypic covariance components due to a small number of common hybrids across two years in a data set. Simulation confirmed the observation from cross validation. We also found that BIC chose models similar to models chosen by cross-validation, suggesting that BIC could be a reasonable approach to choosing the best model for predicting performance in unobserved environments.