Alternative analytics strategies to complement meta-regression analyses in animal nutrition: an example exploring milk yield and composition

• Grant number: 25/23356-2
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: February 01, 2026
• End date: May 31, 2026
• Field of knowledge: Agronomical Sciences - Animal Husbandry - Animal Nutrition and Feeding
• Principal Investigator: Vinícius Carneiro de Souza
• Grantee: Beatriz Kinue Shiraishi
• Supervisor: Robin White
• Host Institution: Faculdade de Zootecnia e Engenharia de Alimentos (FZEA). Universidade de São Paulo (USP). Pirassununga , SP, Brazil
• Institution abroad: Virginia Polytechnic Institute and State University, United States
• Associated to the scholarship: 25/05103-0 - Effect of hepatic urea recycling on nitrogen use efficiency in ruminants: a meta-analysis, BP.IC

Abstract

A number of variable selection strategies have been employed in recent meta-analyses in the animal nutrition space, including backward elimination, all-possible-models, and forward selection, among others. In each of these methods, the modeler is responsible for identifying sets of potential explanatory variables. When considering only main effects, this task is fairly easy to complete and justify; however, given the need to more comprehensively model interactions within animal nutrition, complications arise. There are a few guidelines, outside the research question to be investigated, regarding the number of interactions that should be included in animal nutrition models or the appropriate level of interaction (2-way, 3-way) that should be tested. This project will evaluate 2 alternative analytical approaches, namely, recursive feature elimination (RFE) and additive Bayesian networking (ABN), in parallel to the traditional mixed-model meta-analysis as a mean to identify potential interactions that should be included in animal nutrition models and the appropriate level of interaction that should be tested. Recursive feature elimination is an approach (based on the random forest machine learning method) to select features with high relative importance and fit a model while dropping features with low relative importance. Additive Bayesian networking is a method to determine an optimal directed acyclic graph and a multivariate approach using machine-learning, and is well adapted to study messy, highly correlated data. To evaluate how these alternative data analytics approaches might complement traditional meta analyses, our objective will be to explore the strengths and limitations of linear-mixed effect regression, RFE, and ABN in identifying relationships among diet, rumen, animal,and milk performance variables. An existing database composed of one hundred ninetyfour (194) studies representing 705 individual treatments is being updated to include published studies between 2021 and 2025 and will be used in this project. Incorporation of a more holistic, and data-driven strategy for variable selection such as RFE and/or ABN, capable of representing interaction-like effects, has the potential to expand confidence in existing meta-analyses to support the decisions made by the modeler in retaining or omitting relationships, either in initial variable selection or during stepwise variable selection procedures. The objective of this work is to explore the strengths and limitations of mixed-models with backward elimination, RFE, and ABN in identifying relationships among diet, rumen, and milk performance variables. We hypothesize that, based on this analysis, future meta-analyses could benefit from incorporating RFE and/or ABN as part of the exploratory data analysis process or as a complementary tool to traditional mixed-model approaches.