A comparison of mixed-model analyses of the Iowa Crop performance test for corn
Multienvironment trials generally have highly unbalanced data structures in which a particular cultivar is only observed in a subset of all environments for which data are available. A very common approach to reporting data from such unbalanced data is to subset the data into balanced subsets and restrict comparisons within balanced subsets. Such an approach results in much information being ignored. We undertook an empirical study of 65 individual data sets from the Iowa Crop Performance Test for corn (Zea mays L.) to compare eight different mixed linear models to determine which features in the data need to be considered in developing approaches to make use of all available information. We used a model selection approach to identify the best model based on the presence or absence of heterogeneity of error variances among environments, heterogeneity of genotypic variances among environments, and heterogeneity of genotypic correlations between pairs of environments. The trait analyzed was grain yield. We found evidence of heterogeneity of error variances among locations in 58 of 65 data sets for two model selection criteria. Heterogeneity of genotypic variances and correlations between pairs of environments was found in about half of the sets we analyzed. A general recommendation for model selection cannot be made from this analysis. In general, we found that heterogeneity of variances and correlations was prominent in many data sets. Identification of the best statistical model for a particular data set may be dependent on application of a model selection approach.