Aplicação de modelos robustos para a predição de valores genéticos em bovinos de corte

Abstract

Extreme values can distort the result of a genetic assessment. Similarly, deleting these values can hide relevant changes in a herd. Prediction of genetic values in a population of individuals should have a higher level of accuracy when the available phenotypic and pedigree information matches reliable data. However, factors such as the potential effect of unknown injuries, disease, differential treatment or even data entry errors are variables that are not considered in statistical models, but are capable of compromising data quality to the extent that they significantly influence performance of an individual or group of individuals, generating extreme values that may skew estimates of genetic parameters. Mixed statistical models are the most used for predicting genetic values, but are sensitive to data with extreme values and need to edit or discard these data to mitigate the distortion of results. Therefore, the objective of this work is to demonstrate that the implementation of a robust model can reduce the influence of this data with extreme values and improve the prediction result without discarding data. For this, an algorithm was developed that calculates the mixed model equations, identifies the relationship between the extreme values and the accuracy of the prediction and introduces, when necessary, a weighting variable capable of reducing the deviation of each observation from the mean. your sample unit. The results showed that it was possible to improve the accuracy of the estimates, reducing, in some cases, the influence of extreme values by up to 90 percent, according to the calculated standard deviation, without discarding them from the model. Thus, in the face of data sets with extreme values, the robust prediction model presented more accurate results compared to the mixed model. In both characteristics evaluated, there were reductions between 55 and 79 percent in the interval between the highest and the lowest estimated value.