NMIJ-BIPM workshop, May 16, 2005

Transcription

1 . p.1/26 Confronting the Linkage Problem for Multiple Key and Supplementary Comparisons Nell Sedransk and Andrew L. Rukhin Statistical Engineering Division National Institute of Standards and Technology Building 820, Gaithersburg, MD USA NMIJ-BIPM workshop, May 16, 2005

2 Linkage. p.2/26 l i l i l i L 1 L 2 l i

3 Linkage problem. p.3/26 In a linkage study only indirect comparisons possible. Linking labs measure common artifacts Delahay and Witt (2002): update measurements of the labs participating only in the EUROMET project 345 to obtain the best estimate of what would have been the result from such a laboratory had it actually participated in CCEM-K4. Goal: the table of pairwise laboratory contrasts (degrees of equivalences) with associated uncertainties.

4 Figure. p.4/26

5 Model. p.5/26 Raw data, laboratories, artifacts, sample sizes lab for artifact effect of laboratory effect of artifact Assume independent. Variances Initially no Type B error in this model (will be included later) not necessarily equal., mutually

6 Contrast Estimation. p.6/26 Laboratory effects and artifact effects identifiable. Goal: to estimate contrasts (well-defined) are not and to give a confidence intervals When,, Behrens-Fisher problem so only approximate confidence intervals. Good estimators: depend on sufficient statistics, have the small uncertainty (mean squared error).

7 . p.7/26 Sufficient statistics Sufficient statistics for unknown parameters sample means sample variances

8 . p.8/26 Sampling distributions independent of estimates the variance Sample sizes matter!

9 Graph. p.9/26 1st Pilot Lab 2nd Pilot Lab k-th Pilot Lab

10 Graph. p.10/26 Laboratories: linking (pilot) measure several of artifacts (or participate in several out of non-linking (supplementary) Set of all laboratories which measure artifact different studies) (or is empty) For labs if all linking labs measuring both artifacts and for some

11 Connectedness condition. p.11/26 The graph has vertices=laboratories, edges = common artifact (equivalence under ), the total number of labs, = the total number of vertices in. In the graph the non-linking labs are connected to the linking labs measuring the same artifact

12 Completeness condition. p.12/26 gheorem Assume: the graph is connected, i.e. for any two labs and there exists a sequence (path) such that each linking laboratory measures exactly two artifacts a and each non-linking lab measures just one. Then complete sufficient statistic if and only if. is the number of linking laboratories measuring two of given artifacts.

13 Complete sufficient statistics. p.13/26 Simple case 1: measures artifacts measures and and, measures and. Simple case 2: all measures artifact saturated designs Whenever two different linking labs measure the same measure artifact two artifacts, say, and, completeness is lost

14 Saturated design When, any estimator is a UMVUE of the parametric function If for each,. p.14/26

15 Welch-Satterthwaite approximation. p.15/26 can be unbiasedly estimated via Unique unbiased estimator for contrasts

16 Simple case revisited The Welch-Satterthwaite confidence interval the critical freedom point of a -distribution with degrees of Three artifacts, and The first circulates through labs the second through labs with the last through labs Two linking labs and. admits a UMVUE.. p.16/26

17 Simple case revisited, the UMVUE is. p.17/26 An unbiased estimator of the variance: The Welch-Satterthwaite confidence interval for point of a -distribution with the degrees of freedom.

18 Example. p.18/26 i (a 1 ) i (a ) 3 L 1 M (a, a ) (a, a ) L 3

19 Simple case revisited. p.19/26 the UMVUE is (linkage through the labs and An unbiased variance estimator ) The additional linkage through the lab increases the variance!

20 Type B uncertainty I. p.20/26 Model 1. The sample means for as before; independent realizations of and of known (Type B error in the uncertainty budget) independent of Linear model, one of variance components is known Uncertainty budget stated accurately For known non-zero, ISO GUM (1993) recommends it must be subtracted from the data: replaced by

21 Type B uncertainty II. p.21/26 If Type B uncertainty is expressed as an estimate e of the variance of the Type B error together with associated degrees of freedom Model 2 as before independent realizations mean zero variances are not known independent of and of

22 Type B uncertainty II. p.22/26 random with the distribution and known Type B error is expressed through Degrees of freedom in W.-S intervals: the denominator includes an additional term replace

23 Type B as a prior distribution. p.23/26 For lab Type B uncertainty for lab think of as the parameter of the prior inverse gamma-density, (assume this distribution does not depend on Effective degrees of freedom = the scale parameter can be also found Usable formulas: ) -confidence interval for

24 . p.24/26 Contrast Estimation:Incomplete Sufficient Statistic generic path passing through labs in the graph such that

25 Contrast Estimation Proposition Each path leads to an unbiased estimator of, and to an unbiased estimator of. p.25/26

26 Contrast Estimation For any other path different from, Any convex combination,, is an unbiased estimator. The vector of estimated best weights is is an estimate of the covariance matrix formed by all paths., of the vector. p.26/26