Comparison of models with a fixed or a random contemporary group effect for the genetic evaluation for litter size in pigs

This paper compares alternative models with respect to the accuracy of prediction of randomly chosen observations excluded from the genetic evaluation. REML variance components were estimated with a linear mixed model for number of piglets born alive of Swiss Landrace (SL) and Swiss Large White (LW) sows. The two alternative models assume a fixed herd* time period (h X t) effect and a random h X t effect, respectively. For two additional random models the variance ratios of the fixed model were used, but the ratio of residual to h X t variance varied. Repeatedly, a small set of randomly chosen observations were excluded from the data and predicted based on estimates obtained from the remaining data. The random models predicted excluded observations more accurately than the fixed model but there was nearly no difference between the models to predict observations from h X t with more records. Differences between the random models were small. Due to a better structure of the LW data, differences between models to predict observations were considerably smaller than for SL. For the analysed data a model with a random h X t effect is recommended. (C) 1997 Elsevier Science B.V.