Searching for causal networks in experimental data: a swine production application

Efficient agricultural production systems require integrated management of complex physiological mechanisms. Recent developments in network methodologies can enable meaningful directional insight into the inner workings of such complex systems. Motivated by a designed experiment in swine production, we explore potential causal biological relationships between physiological outcomes in high-performing gilts and sows using structural equation models implemented in a mixed modeling framework. Data consisted of short- and long-term reproductive outcomes for 200 sows and 440 gilts arranged in a randomized block design and randomly assigned to nutritional treatments during late gestation. We implement structure- learning algorithms adapted to a hierarchical Bayesian framework to search for and quantify causal links between physiological traits separately for gilts and sows, while recognizing the multilevel architecture of the data given by the experimental design. Using a modified Jackknife resampling approach, we evaluate stability of the learned network structures and make power considerations for network inference. Results indicate distinctly heterogeneous networks for gilts and sows, consistent with differences in their physiological mechanisms. These findings have practical implications for differential management of gilts and sows to improve efficiency of swine production systems. In addition, these finding motivate further methodological extensions to structural equation models to enable specification of heterogeneous networks.