The effect of the number of observations used for Fourier transform infrared model calibration for bovine milk fat composition on the estimated genetic parameters of the predicted data

Fourier transform infrared spectroscopy is a suitable method to determine bovine milk fat composition. However, the determination of fat composition by gas chromatography, required for calibration of the infrared prediction model, is expensive and labor intensive. It has recently been shown that the number of calibration samples is strongly related to the model's validation r2 (i.e., accuracy of prediction). However, the effect of the number of calibration samples used, and therefore validation r2, on the estimated genetic parameters of data predicted using the model needs to be established. To this end, 235 calibration data subsets of different sizes were sampled: n = 100, n = 250, n = 500, and n = 1,000 calibration samples. Subsequently, these data subsets were used to calibrate fat composition prediction models for 2 specific fatty acids: C16:0 and C18u (where u = unsaturated). Next, genetic parameters were estimated on predicted fat composition data for these fatty acids. Strong relationships between the number of calibration samples and validation r2, as well as strong genetic correlations were found. However, the use of n = 100 calibration samples resulted in a broad range of validation r2 values and genetic correlations. Subsequent increases of the number of calibration samples resulted in narrowing patterns for validation r2 as well as genetic correlations. The use of n = 1,000 calibration samples resulted in estimated genetic correlations varying within a range of 0.10 around the average, which seems acceptable. Genetic analyses for the human health-related fatty acids C14:0, C16:0, and C18u, and the ratio of saturated fatty acids to unsaturated fatty acids showed that replacing observations on fat composition determined by gas chromatography by predictions based on infrared spectra reduced the potential genetic gain to 98, 86, 96, and 99% for the 4 fatty acid traits, respectively, in dairy breeding schemes where progeny testing is practiced. We conclude that a relatively large number of calibration samples is required to be able to obtain genetic correlations that lie within a limited range. Considering that the routine recording of infrared spectra is relatively cheap and straightforward, we concluded that this methodology provides an excellent means for the dairy industry to genetically alter milk fat composition