The impact of the low number of records when selecting for uniformity: a simulation study of birth weight in guinea pigs

Abstract

Selection for uniformity in livestock species is desirable because it is associated with welfare and robustness. To address variability, it is essential to have more than one record per animal; therefore, the number of records per individual is crucial. The guinea pig, characterised by its low litter size, is a species of notable economic and nutritional importance, especially in the Andean region of South America. The objective of this study was to evaluate, through simulation, the impact of low record numbers on the performance of heteroscedastic models used in selection for uniformity, using guinea pig birth weight as an example. Different data structures that modulated litter size (LS), number of litters (NL), number of breeding animals (BAs), and number of generations of records were evaluated. The analysis of simulated data was performed under Frequentist and Bayesian statistical approaches. In general, the model was able to adequately estimate the variance components and predict genetic values, although its performance depended on the data structure. Under the Frequentist approach, the litter variance affecting the residual variance was biased downward in LS = 2 and LS = 3 (between -69 and -96%), although the bias decreased in LS = 5 and LS = 7 (between -10 and -44%). The BAs did not influence the bias. At NL = 2, the bias ranged from -68 to -76%, and at NL = 6, from -35 to -69%. The Bayesian method failed to converge in the scenario with two generations and NL = 2. The accuracy of the genetic values for both the trait and the residual variance was identical regardless of the statistical approach used. An increase in BAs did not lead to favourable changes in accuracy, but higher LS and NL tended to improve the accuracy of the genetic values of the trait and the residual variance. We conclude that selection for uniformity under a low number of repeated records, such as the limited number of within-litter birth weight data in guinea pigs, is feasible. However, it is recommended to carefully monitor the data structure to predict the expected response and, if possible, to use populations with the highest possible litter sizes.