Descrição do crescimento de frangos via inferência bayesiana

Abstract

Broiler growth is a dynamic biological process, influenced by genetic, sexual, nutritional, and environmental factors, whose description is essential for management, nutrition, and production decisions. Nonlinear regression models are widely used for this purpose, as they summarize the growth trajectory into parameters with direct biological interpretation. Traditionally, these models have been estimated by classical inference; under the Bayesian approach, broiler growth remained little explored, particularly regarding the use of formal tests for parameter comparison and the treatment of truncated data. The central contribution of this thesis is to apply Bayesian inference to this context, integrating parameter estimation, uncertainty quantification, and the formal comparison among groups in poultry growth curves. In the Bayesian approach, the information from the data is combined with prior knowledge about the parameters, yielding posterior distributions that express the uncertainty of the estimates and allow parameters to be compared among groups, such as males and females, breeds, or strains, through the posterior distributions of the differences or formal measures of evidence. This feature is particularly useful in the presence of truncated data, a common situation in broilers, in which the discontinuation of weighings at commercial age limits the information about the late growth phase. The thesis comprises two articles. In both, the posterior distribution was obtained by sampling, using the Hamiltonian Monte Carlo (HMC) algorithm with the No-U-Turn Sampler (NUTS), implemented in the brms package of the R software. In the first, the Logistic, Gompertz, von Bertalanffy, and Brody models were fitted to the body weight data of males and females of the black variety of the Utrerana breed, raised in a free-range system; model selection was based on the LOO-IC, WAIC, and ELPD criteria, and sexual dimorphism was assessed using the Savage Dickey method, which quantifies evidence both for and against the equality of parameters. In the second, the Gompertz model was fitted to the growth of the Cobb 500 and Ross 308 strains from truncated data, adopting informative prior distributions, anchored in biological references, for the parameters subject to greater uncertainty due to truncation, especially the asymptotic weight; comparisons between sexes and between strains were based on the posterior distributions of the differences, with highest posterior density (HPD) intervals and analysis of the critical points of the curve. The Bayesian nonlinear models proved adequate for describing the growth of the birds. In the Utrerana breed, the Gompertz and von Bertalanffy models showed the best performance, and sexual dimorphism was expressed only in the asymptotic weight, which was higher in males. In Cobb and Ross, the Gompertz model identified sexual dimorphism in the asymptotic weight in both strains, with no consistent evidence of differences between sexes in the age at maximum weight gain or in the relative growth rate. The comparison between strains, based on the critical points, revealed no differences between Cobb and Ross when posterior uncertainty was taken into account.