In this paper we report on the results of a recent dissection experiment in The Netherlands where prediction formulae for the percentage lean meat in pig carcasses with the Hennessy Grading Probe (HGP) and a vision system (from now on referred to as Vision) were determined. Predictions with the HGP were based on one fat and one muscle depth measurement only, while predictions with Vision were based on as many as 115 direct and derived measurements. The data from this dissection experiment were used to illustrate the statistical calculations involved in relation to the number of carcass measurements. Prediction with instruments that gather a large number of measurements per carcass is not covered by the present European Community (EC) regulations. Therefore the calculations were conducted according to new regulations for statistical methodology in pig carcass grading that are expected to be adopted by the EC in the near future. The calculations included consideration of 3 subpopulations (females, entire males and castrated males). The Vision data were also used to show that ordinary regression after selection of a subset of carcass measurements severely under estimates the accuracy of prediction: instruments and associated prediction formulae are seemingly much more accurate than they truly are. When standard regression methods are used for instruments that gather a large number of measurements, there is a considerable risk that measurement instruments will be selected for the wrong reasons. Accuracy of approved instruments may not even comply with the EC-regulations, with poor consequences for harmonization within the EC.
- squares regression