7th World Congress on Genetics Applied to Livestock Production, August 19-23, 2002, Montpellier, France Session 17. Estimation of additive and non-additive genetic parameters Communication N° 17-09 SIMULATION STUDY ON LINEAR MIXED MODELS WITH CONTAMINATED NORMAL DISTRIBUTION IN ANIMAL BREEDING I.G. Pereira 1 , H.N. Oliveira 2 and G.J.M. Rosa 3 1 Animal Science Department, University of Lavras - MG, Brazil 2 Department of Animal Breeding and Nutrition, São Paulo State University - SP, Brazil 3 Department of Biostatistics, São Paulo State University - SP, Brazil INTRODUCTION The methodology of linear mixed models is widely used in genetics and animal breeding. In general, the normal distribution is assumed both for residuals and random effects in these models, which makes inferences very sensitive to the presence of discrepant values in observations (Rogers and Tukey, 1972). An alternative in this sense refers to models with thick-tailed distributions, which has proved quite effective in robust estimation and of easy implementation in a Bayesian context (Rosa, 1999). However, the studies about the use of these distributions for estimation of (co)variance components and genetic parameters and prediction of genetic values of animals are few (Strandén and Gianola, 1999). This work compares estimates of variance components and predicted genetic values obtained by Gaussian and robust linear mixed models, in data simulated with different levels of discrepant values in observations. MATERIAL AND METHODS Fifty data files with 1,000 animals each, distributed in five generations of 200 animals without selection, were simulated. All the animals presented records for three hypothetical productive traits (FV, FV1 and FV2). For the 200 animals of the first generation, genetic and residual values were generated, from normal distributions with mean zero and variances equal to 36.0 and 64.0, respectively. The animals were randomly distributed into two levels of fixed effects and with probability determined by φ (0.20 or 0.40), they could be allocated to a contaminant or non-contaminant population. The values of the characteristics were obtained from the summation of the genetic value, the systematic (fixed) effect and the residual of each animal. The phenotypic value FV stood for the population with normal distribution with no contaminants and the other two, identified as FV1 and FV2 were generated in a similar manner to the former except that the residual value of each animal allocated to the group of contaminants was divided by the square root of a certain value τ (between 0.0625 and 0.25) before being summed to the others. For FV1, the value of τ was always greater than for FV2. For the animals of the other generations, the procedure used was similar, except that the genetic value of the animals was generated as the summation of the means of the parents’ genetic values plus the Mendelian sample variation. Thus, for each population three data sets were simulated. The first one had normal distribution without any contaminants, and the other two presented a proportion of contaminants given by φ, with residual variance equal to that of the original population divided by τ. The presence of