This is an early access version, the complete PDF, HTML, and XML versions will be available soon.
Dry Matter Intake Prediction Models: Evaluation Across Energy-Corrected Milk and Lactation-Stage Classes in Holstein Cows
Department of Animal Science, Agriculture Faculty, Çukurova University, 01330 Adana, Türkiye
*
Author to whom correspondence should be addressed.
Animals 2026, 16(12), 1824; https://doi.org/10.3390/ani16121824 (registering DOI)
Submission received: 19 May 2026
/
Revised: 9 June 2026
/
Accepted: 11 June 2026
/
Published: 12 June 2026
(This article belongs to the Section Animal Nutrition)
Simple Summary
Dry matter intake (DMI) equations are widely used in dairy nutrition, but their directional bias may depend on the production context rather than on equation structure alone. We evaluated the bias of five DMI prediction models across energy-corrected milk (ECM) and lactation-stage classes in a literature-derived database of lactating dairy cows. Model bias differed among equations and changed across ECM and lactation-stage classes, with the lactation stage showing a stronger effect. These results suggest that no single DMI model is uniformly optimal across all situations and that context-specific model selection may improve the practical use of DMI predictions in dairy nutrition.
Abstract
Accurate prediction of dry matter intake (DMI) is essential for ration formulation, nutrient supply, and evaluation of production efficiency in lactating dairy cows. Several DMI prediction models are currently used, but most comparative studies have emphasized overall accuracy rather than whether model bias changes across biologically relevant production contexts. The objective of this study was to evaluate the context-dependent bias of widely used DMI prediction models in lactating dairy cows across classes of energy-corrected milk (ECM) and lactation stage. A literature-derived database was assembled from 135 studies consisting of 436 treatments from 6985 Holstein cows, reporting observed DMI and the variables required to implement five prediction models and evaluate their prediction error (PE): NRC2001, the Cornell Net Carbohydrate and Protein System (CNCPS), NASEM2021, Agroscope2021, and GfE2023. PE was calculated as predicted DMI minus observed DMI, such that positive values indicated overprediction and negative values indicated underprediction. Observations were classified according to ECM and days in milk (DIM). Mixed models were fitted separately for the ECM class and the lactation-stage class, with the study fitted as a random effect. PE differed among models, and the pattern of bias depended on both the ECM and the lactation-stage classes. The interaction between the ECM class and the model was significant, indicating that productive level modified model bias. The interaction between lactation-stage class and model was also significant and more pronounced, indicating marked changes in model bias across lactation stages. Across classes, NASEM2021 generally remained closest to zero, whereas GfE2023 and CNCPS showed more negative PE values in most contexts. Agroscope2021 showed a more context-sensitive pattern, and NRC2001 remained comparatively moderate across several classes. These findings indicate that the evaluation of DMI prediction models based only on global mean bias may conceal an important biological structure in PE. Context-specific evaluation, particularly across the lactation stage, may provide a more informative basis for selecting DMI prediction models for research and practical ration formulation.
