From zech@physik.uni-siegen.de Wed Apr 12 16:21:19 2000 Date: Tue, 4 Apr 2000 10:48:25 +0200 From: "[iso-8859-1] \"Zech, Günter\"" To: 'Bob Cousins 310-825-1928' , Giulio D'Agostini Cc: [See mailing list in a separate file] Subject: AW: Conclusions of the CERN CLW Dear colleagues I think it is obvious that Bob's list presents his personal choice and was never agreed as the conclusions of the workshop. I am sure, Bob accepts that. The problems with the list are: i) Most of the statements are trivial (basic statistics) and cannot be subjected to a majority consensus. Those statements are mixed with personal preferences. ii) It contains irrelevant statements. Chi2 has nothing to do with error assignment or confidence intervals. Bob / Fred, how to you include the chi-squared in the error? iii) The selection of the points is obviously incomplete and biased and this is the main problem of the list. To demonstarte the last point, I produce a different list: 1) Classical confidence limits violate the likelihood principle. 2) They cannot be computed from the likelihood function. 3) They use information which is irrelevant for decisions. 4) They do not measure the precision of the measurement. 5) They are not compatible with the likelihood ratio test. 6) The unified approach produces disconnected confidence intervals for pdfs with tails. 7) Coverage requires to fix the ananlysis creteria (cuts) and the decision to publish independent of the data. 8) The proposed treatment of nuisance parameters by Cousins (take the best fit) can produce undercoverage. 9) In the Poisson case for a true value equal zero ther is complete overcoverage. 10) Poisson + background with uncertainty cannot be handled. 11) When there are physical bounds the confoidence limits of independent parameters can become correlated. 12) In the unified approach there is no prescription on how to combine limits in general and how to perform error prpagation. 13) An optimum way to combine measurements is to add the log-likelihood functions. 14) Systematic errors cannot be handled in the classical approach. All these points are valid. (Is there any point you cannot agree with?) Nevertheless presenting only this list to somebody who has not attenden the workshop gives a biased impression. I am quite unhappy about the way conclusions are drawn from the quite interesting workshop. I had hoped that people think more about the problems raised. Especially, one should encourage people to read the preceedings and papers CAREFULLY and to study the statistical literature and not concentrate on politics. Best regards, Gunter