Predicting the reference –

Introduction to the handbook


ITP – Le Rheu - France

The reference for pig classification is defined as a lean meat proportion. As this criteria is destructive and very expensive to obtain it has to be predicted on slaughterline. The EC regulations contain some requirements on how to predict the reference.

Nevertheless, member states have chosen in the past different ways for sampling and for parameters estimation. Moreover, two issues have appeared : the reduction of experimental costs and the management of large sets of highly correlated variables. These difficulties have been solved by more complicated statistical methods.

In order to help meat scientists dealing with pig classification methods in the (present and future) EU it has been decided to write some guidelines in a handbook. The first official version (draft 3.1) has been distributed on February 2000 to the member states delegates at a meeting of the Pigmeat Management Committee. Contributions come from statisticians or meat scientists who are national expert for pig classification. All are involved in EUPIGCLASS.

The aim of WP2 is to improve the first version of the statistical handbook, going further into details, achieving a large consensus and illustrating with examples. We can distinguish two parts : the statistical methods and sampling, because most of the sampling considerations are general even if each method can have some specificities.

Some important issues in sampling are : how to choose a representative sample (management of sub-populations), how to select on predictors, which size ?

All the statistical methods derived from classical linear regression which was the initial method in the EC regulation. They can be splitted into two groups :

  1. Methods for reducing experimental cost
  2. Methods for managing large sets of highly correlated variables.

To cut the cost of one experimental trial “double regression” has been first introduced in pig classification by Engel & Walstra (1991). To reduce the costs of testing different instruments Daumas & Dhorne (1997) have introduced “regression with surrogate predictors”. These methods fulfill the present EC requirements and estimation of accuracy is available.

In parallel the Danes have first introduced PCR (Principal Component Regression) to test the danish robot Classification Centre (10-20 variables) and then PLS (Partial Least Squares) to test the danish robot Autofom (> 100 variables). But no models are behind these methods and the parameters estimation is under discussion.