Genetic analysis of growth curve in Moghani Sheep using Bayesian and restricted maximum likelihood

Abstract

Abstract This study was conducted to predict the genetic (co)variance components of growth curve parameters of Moghani sheep breed using the following information: birth weight (N = 7278), 3-mo-old weight (N = 5881), 6-mo-old weight (N = 5013), 9-mo-old weigh (N = 2819], and 12-mo-old weight (N = 2883). The growth parameters (A: maturity weight, B: growth rate, and K: maturity rate) were calculated using Gompertz, Logistic, Brody, and Von Bertalanffy nonlinear models via NLIN procedure of SAS software. The aforementioned models were compared using Akaike information criterion, root mean square error, adjusted co-efficient of determination. Also, both Bayesian (using MTGSAM) and RMEL (using WOMBAT) paradigms were adapted to predict the genetic (co)variance components of growth parameters (A, B, K) due to the best fitted growth models. It was turned out that Von Bertalanffy best fitted to the data in this study. The year of birth and lamb gender had a significant effect on maturity rate (P < 0.01). Also it turned out that within the growth parameter, with increasing (co)variance matrix complexity, the Bayesian paradigm fitted well to the data than the restricted maximum likelihood (REML) one. However, for simple animal model and across all growth parameters, REML outperformed Bayesian. In this way, the h2a predicted (0.15 ± 0.05), (0.11±.05), and (0.04 ± 0.03) for A, B, and K parameters, respectively. Practically, in terms of breeding plan, we could see that genetic improvement of growth parameters in this study is not a tractable strategy to follow up and improvement of the management and environment should be thoroughly considered. In terms of paradigm comparison, REML’s bias correction bears up an advantageous approach as far as we are concerned with small sample size. To this end, REML predictions are fairly accurate but the mode of posterior distributions could be overestimated. Finally, the differences between REML and Bayesian estimates were found for all parameter data in this study. We conclude that simulation studies are necessary in order to trade off these parading in the complex random effects scenarios of genetic individual model.