Impact of missing data on the mixed model estimation of plant breeding parameters from multienvironment yield trials

AGUATE, F.M.; IBAÑEZ, M.A.; BALZARINI, M.G.

Florencia

Conferencia; XXVIIth International Biometric Conference; 2014

International Biometric Society

The success of a genetic improvement program is reflected by its capacity to produce superior genotypes (G) for a variety of environments (E). To identify superior genotypes in multienviroment trials (MET) it is necessary to estimate genotypic yield performance and its stability across environments. Additionally, genetic correlation among environments is used to indicate the relative size of genotype-by-environment interaction (G×E) effects that are not due to heterogeneity of genotypic variance among environments and are, therefore, of practical interest to define breeding strategies. Genotypes evaluated in MET vary across environments, as some genotypes are introduced and others discarded from year to year, resulting in incomplete databases for multi-year analysis. The Linear Mixed Model (LMM) framework constitutes a valuable tool to obtain genetic parameters from incomplete data. Our goal was to evaluate the impact of different levels of missing data on predictors of genotype yield performance (BLUP), stability variances for each genotype (SVg), and the genetic correlation among environments (rg). Evaluations were performed under complete databases and increasing levels of incompleteness (10-to-70%, at 5% intervals, of missing data at random and not at random). Databases were simulated under different scenarios regarding the combination of the following setting parameters: dataset dimension, G×E and G relative magnitude. Variance components and BLUPs were obtained from SAS Proc Mixed, using REML. To derive stability variances, the covariance structure of the interaction effects, at each environment, was modeled with a diagonal matrix with no structure over the principal diagonal (type=UN(1)). We found that the number of E was the factor of highest impact on the robustness of breeding parameters under incompleteness. Not at random missing data schemes are more likely to have a genotype missed in higher number of E than missing at random, and consequently a higher impact of incompleteness was observed for all parameters. Stability variance was overestimated for G tested in few E, mainly in small METs: in 3-environment datasets, SVg started increasing rapidly over 20% of incompleteness, in 6-environment METs the increase started at 40% of missing data and when 25 environments were evaluated, the break point for this parameter was at 50% of incompleteness. The genetic correlation among environments decreased when the level of incompleteness increased and this effect was higher when the relative genetic variance was 5% rather than 10%, when G×E variance was 30% rather than 5%, and for 3-environment datasets. Increasing levels of inconsistency were observed for the rank of the 25% top genotypes, created from the BLUPs of G effects, with more missing data. However the break point was over 30% of missing data for all scenarios, and the selected G remained the same in a real dataset of high dimension (89 G × 25 E).