Impact of missing values on variance component estimates in multienvironment trials

AGUATE, F.; BALZARINI, M.

Marrakech

Congreso; 61st World Statistics Congress; 2017

Multienvironment trials (METs) are conducted to evaluate genotypes (G) across locations (L). Datasets collecting field testing along years (Y) are often incomplete because several G are not tested in some environments (E). Provided enough connectivity among trials through shared G, variance components can be estimated from likelihood-based methods (REML) under a single linear mixed model pooling several years of MET. However, the relative bias of estimates can depend on dataset dimension, missing data proportion and the G drop-out mechanism. The aim of this study was to quantify the impact of missing values in variance estimates for data collecting from 2 to 8 years of METs. A simulation-based study varying data dimensions (numbers of G, L and Y) and the proportion of missing data was performed. The G replacement mechanism was based on G-BLUP to dropout cultivars from year to year, generating missing values as function of retained G. In 2-year analyses, the G variance was more than 10% overestimated, with 40% or more missing values. When adding years, improvements in variance estimates where notable: 8-year analyses, produced lesser than 5% biased estimates and lower standard errors, even with highly unbalanced datasets. An increase in the number of L and G was beneficial. The genotype-mean repeatability was negatively affected by missing values, but benefit from long-term data. Results set the boundaries of dimensions in MET analyses to obtain better variance component estimates. REML procedures need 4 or more years of unbalanced MET data to obtain good estimates. In a process of selection, analyses of 2 or 3 years are not good enough for estimations of G and GxY variances when dealing with missing values.