AbstractIn cowpea breeding, multi‐environment trials are conducted to select lines with high yield. The occurrence of genetic and/or statistical imbalance is common in these experiments, in addition to the possibility of (co)variance between genetic and non‐genetic effects. We explore the restricted maximum likelihood/best linear unbiased prediction features to select the model with the most appropriated covariance structure and compare the results with the traditional model (homogenous variances and no covariances). Then, 17 inbred lines and three cultivars were evaluated in six experiments during two crop years in the semiarid zone of Northeast Brazil. The trait evaluated was the 100‐grain weight. We selected the best model considering the Akaike Information Criterion. The model with diagonal structure for the residual effects and heterogeneous compound symmetry for the genetic effects had the best fit. The predicted genetic gain of lines selected in this model was 1.18% higher compared to the traditional model. Modeling different (co)variance structures for genetic and non‐genetic effects is an efficient approach in selecting superior genotypes in multi‐environment trials in cowpea breeding.