Abstract #T171
Section: Production, Management and the Environment
Session: Production, Management & the Environment II
Format: Poster
Day/Time: Tuesday 7:30 AM–9:30 AM
Location: Exhibit Hall B
Session: Production, Management & the Environment II
Format: Poster
Day/Time: Tuesday 7:30 AM–9:30 AM
Location: Exhibit Hall B
# T171
Factor screening for prediction of retention-pay offs of dairy cows using standardized regression coefficients, random forests, and the method of elementary effects.
A. Beyi*1, A. De Vries1, 1University of Florida, Gainesville, FL.
Key Words: elementary effect, sensitivity analysis
Factor screening for prediction of retention-pay offs of dairy cows using standardized regression coefficients, random forests, and the method of elementary effects.
A. Beyi*1, A. De Vries1, 1University of Florida, Gainesville, FL.
A dairy cow retention pay-off (RPO) of individual cows may be used to support culling decisions. Sensitivity analyses regarding herd inputs (factors) that affect RPOs are generally carried out with a one-factor-at-a-time design and typically include only the change in the average RPO. Our objective was therefore to carry out a formal, systematic sensitivity analysis of herd factors as recommended in the literature. RPOs were generated using a stochastic dynamic programming model (DairyVIP) using 15 herd inputs. Each combination of herd input variables resulted in 2,304 RPO for cow categories, a combination of lactation number, month in milk, and relative milk yield. We used Standardized Regression Coefficients (SRC), Random Forests (RF), and Morris’ method of Elementary Effects (EE) to investigate the importance of the 15 factors. Data for SRC and RF were calculated using independent Monte Carlo variations of the 15 herd variables within preset ranges (n = 320 runs with DairyVIP). Data for EE were calculated with the “economical design” of the original Morris study using 4 levels per factor and 20 replicates (n = 320 runs). The 5 output variables were average RPO and change in herd profit, as well as Spearman’s rank correlation coefficient, Root Mean Squared Error, Mean Absolute Error compared with the set of RPOs using default herd inputs. Results for the 5 rescaled variables were averaged. All 3 methods ranked heifer price as the most important factor. The methods did not agree on the rankings of most of the other factors. Replacement heifer price, milk price, and price per body weight of replaced cow were the 3 most important variables based on RF and EE. Using EE, the ranking of the 7 most important herd inputs was heifer price, milk price, price per live body weight, calf price, relative milk production, relative body weight, feed price. We concluded that the 3 methods may result in different rankings of herd input factors. The EE method is recommended in the literature and warrants wider use in sensitivity analysis in dairy sciences.
Key Words: elementary effect, sensitivity analysis