Introduction

Statistical Methods in Pig Classification in the EU

 

Objective

The objective of the workshop was to reach a common level of understanding with respect to the present statistical methods used in the EU and to receive knowledge of and inspiration for use of new statistical techniques inside the frames of the EU regulation.

Today, most pig carcasses in the EU are graded by means of online measurements at the slaughter line. Normally, these measurements are carried out by use of a classification instrument which measures the lean meat content indirectly. Various technologically different instruments are used and naturally they have a different accuracy.

 

 

The main types of instruments:

 

 

Ruler:

The fat thickness is measured manually at one or two places on the split surface by means of a ruler.

 

 

Probe instruments:

Several instruments are based on a probe and the reflection of light. By insertion of the probe into a carcass it is possible to measure the thickness of the subcutaneous fat.

 

 

 

With manually operated instruments, the thickness is measured in the middle of the loin muscle at one or two places in longitudinal direction. Often, the thickness of the loin muscle is measured as well.

 

 

 

With an automatic instrument the insertion of several probes are made automatically at different places of the carcass.

 

 

Ultrasonic instr.:

An ultrasonic picture is the basis for a new type of measurements. The technique is used both in automatic instruments and in manually operated instruments.

 

 

The reference

 

 

 

In principle, the reference is the content of lean meat in the whole carcass. Practically, the lean meat content is determined by dissection of the main parts of one half carcass.

 

 

Types of calibration methods:

 

 

Multiple Linear Regression (MLR) and variants

 

Where only few on-line data is collected of each carcass – the fat thickness and perhaps the muscle thickness at one or two places – the methods are based on MLR. But there are some variants. Because of the expensive reference method, a double regression [1,2] is used to combine a few reference measurements and several less expensive measurements – a national method – correlated with the reference method. Another variant is a calibration dependent of the sex, which is only known indirectly [3].

 

Principal Component Regression (PCR)

 

When several on-line data is collected of each carcass – the actual case is the fat thickness at 7 places of the carcass and the muscle thickness in 3 places – the method is based on PCR. The full dissection method is used as references.

 

Partial Least Squares (PLS)

 

When a huge amount of on-line data is collected of each carcass – the actual case is up to 127 thickness measurements – the method is based on PLS. The full dissection method is used as reference.

 

EU regulation

 

In an attempt to harmonize the classification in the EU, the member states have agreed on a regulation, which states some minimum requirements concerning the calibration:

 

          The calibration has to be based on a sample of at least 120 representative carcasses – in case a national method is used, the sample size is chosen so the result is comparable with 120 regular reference measurements. Finally, the quality of the calibration is evaluated by means of RMSE, which has to be less than 2.5 percentage of meat.

 

Problems

 

Various problems are arisen like:

 

How do we choose a “representative” sample of 120 carcasses?

What does “RMSE” mean in case of PCR and PLS for instance?

Is it possible to use a cheaper national method in case of PCR and PLS? (A method parallel to double regression?)

 

Besides these questions, other problems arise like:

 

How do we determine the accuracy of destructive measurements?

How can we compare predictions with different accuracy?

How does the accuracy of the on-line measurements – and of the reference – influence on the accuracy of the calibration?

 

 

[1]      Causeur D, Dhorne T., 1998. Finite sample properties of a multivariate extension of double regression. Biometrics, 54, 1591-1601.

[2]      Engel, B. and P. Walstra, 1991. Increasing precision or reducing expense in regression by using information from concomitant variable. Biometrics, 47, 13-20.

[3]      Engel, B. and P. Walstra, 1993. Accounting for subpopulations in prediction of the proportion of lean meat of pig carcasses. Anim. Prod. 57, 147-152