Abstract #7

# 7
Automated model selection: Part II (exercises).
V. L. Daley*1, T. J. Hackmann2, M. D. Hanigan3, 1National Animal Nutrition Program (NANP), University of Kentucky, Lexington, KY, 2University of California, Davis, CA, 3Virginia Tech, Blacksburg, VA.

Automated model selection (AMS) can be applied in different research areas for the selection of the best fitting models. The objective of this exercise is to apply the AMS approach as a tool for the selection of prediction models. A hypothetical example will be used to help the audience better understand and apply AMS. This exercise will use RStudio program, which can be freely downloaded from the internet. As an example, the investigator will develop empirical models to predict the dry matter intake of lactating dairy cows using AMS and parallel computation in R (MuMIn). A meta-analytical data set from the National Animal Nutrition Program (NANP, https://animalnutrition.org) is available for this exercise. First, data quality and range will be checked to identify and remove outliers, and data will be visualized (ggplot2). A global mixed model (lme4) will be fitted to the data using all potential predictor variables from the data set. A set of candidate models will be generated using combinations of the fixed terms of the global model. Then, Akaike’s information criterion corrected for small sample size (AICc) is calculated and used to rank the models. The candidate models are collected in a data set named “all models” for future use. The models with the lowest AICc values are collected in another data set named “best models.” The estimated parameters of best candidate models are automatically collected. Only candidate models with a variance inflation factor (VIF) less than 10 are kept in the “best models” data set. Evaluation of the biological coherence and ANOVA to compare the models are performed. The best candidate models will also be evaluated using the root mean squared error of prediction (RMSE) and concordance correlation coefficient (CCC). The best 3 models are automatically collected in a table (sjPlot). During the exercise, additional instructors will be available to help the participants. The attendees will be able to undertake AMS to select the best models and apply those models in research.

Key Words: multi-model inference, mixed models, Akaike information criterion