Validation strategy can result in an overoptimistic view of the ability of milk infrared spectra to predict methane emission of dairy cattle

Because of the environmental impact of methane (CH ₄ ), it is of great interest to reduce CH ₄ emission of dairy cattle and selective breeding might contribute to this. However, this approach requires a rapid and inexpensive measurement technique that can be used to quantify CH ₄ emission for a large number of individual dairy cows. Milk infrared (IR) spectroscopy has been proposed as a predictor for CH ₄ emission. In this study, we investigated the feasibility of milk IR spectra to predict breath sensor–measured CH ₄ of 801 dairy cows on 10 commercial farms. To evaluate the prediction equation, we used random and block cross validation. Using random cross validation, we found a validation coefficient of determination (R ² val) of 0.49, which suggests that milk IR spectra are informative in predicting CH ₄ emission. However, based on block cross validation, with farms as blocks, a negligible R ² val of 0.01 was obtained, indicating that milk IR spectra cannot be used to predict CH ₄ emission. Random cross validation thus results in an overoptimistic view of the ability of milk IR spectra to predict CH ₄ emission of dairy cows. The difference between the validation strategies could be due to the confounding of farm and date of milk IR analysis, which introduces a correlation between batch effects on the IR analyses and farm-average CH ₄ . Breath sensor–measured CH ₄ is strongly influenced by farm-specific conditions, which magnifies the problem. Milk IR wavenumbers from water absorption regions, which are generally considered uninformative, showed moderate accuracy (R ² val = 0.25) when based on random cross validation, but not when based on block cross validation (R ² val = 0.03). These results indicate, therefore, that in the current study, random cross validation results in an overoptimistic view on the ability of milk IR spectra to predict CH ₄ emission. We suggest prediction based on wavenumbers from water absorption regions as a negative control to identify potential dependence structures in the data.