Modelling individual uncertainty and population variation in phenotypical traits of livestock

Characterising between-animal variation and its population-level consequences is key to effective management and selective breeding of livestock. The aim may be to predict trait development (e.g. performance) from early growth, or to estimate unobserved traits (e.g. maximum growth or maturity parameters). A usual sequence of steps is as follows. 1) To develop a mathematical model of chosen animal-level traits. 2) To estimate individual parameters across a sample of animals. 3) To estimate a population distribution of parameters. 4) To generate a population distribution of chosen traits by simulating the model across distribution 3. The challenge is in the parameter estimation given usual data limitations. We use a Bayesian inference methodology to estimate the population distribution of predicted traits. The approach i) accounts for individual-level uncertainty in parameters (step 2) due to their correlation and data limitations, e.g. short growth span or infrequent records, and ii) does not invoke distributional assumptions and estimation of variance-covariance parameters (step 3). We present results derived from individual data with typical limitations. Results include distributions of growth parameters within breeds and across species (pigs, chicken, rabbits); they extend the literature by showing the extent of uncertainty and variation in parameters and by comparing variation not only across species but against that within breeds. We show distributions of protein and lipid growth parameters and metabolic heat production (HP) estimated across animals and species and predicted distributions of dynamic body composition. Literature body composition estimates usually condition on input of average HP data; by estimating both jointly, their individual-level correlation is included and no metabolic data is needed. HP estimates distributed about 0.7MJ/Kg/d in pigs in line with literature, and body fat content variation was much larger than that of body protein. We suggest this approach has general application in model parameterisation and prediction of trait development in populations using limited individual data. This project has received funding from the European Union’s Horizon 2020 research and innovation programme, grant agreement No 633531