• :

    There are many interlaboratory comparisons where the data are discrepant and the reported within-study uncertainty estimates cannot be relied upon.  The classical random effects model, which explains this phenomenon by additional noise with constant heterogeneity variance, may not be adequate when the smallest reported values correspond to the cases which are most deviant from the bulk of data. Two augmented random effects models are suggested. In the first version the reported  uncertainties are considered as lower bounds to the true uncertainties, the second proposes to split the data into different classes with the same heterogeneity variance only within each cluster. The model choice is to be made on the basis of the classical or restricted likelihood. The properties of  corresponding maximum likelihood estimators are discussed. The assessment of the final uncertainty of these estimators is also reviewed. The non informative, reference  prior provides the Bayes  procedures for the model averaging which are recommended for practical use. The results are illustrated by the  collaborative determination of two physical constants and by some  heterogeneous data sets from International Key Comparisons.

  • : Andrew L. Rukhin
  • : National Institute of Standards and Technology
  • : Andrew L. Rukhin
  • : statistics
  • : advanced/theoretical
  • : andrew.rukhin@nist.gov
Selecting a Model for Collaborative Studies: Augmented Random Effects Paradigm