Professional Documents
Culture Documents
Background
0011 0010 1010 1101 0001 0100 1011
Each measurement made by an instrument or measuring device consists of the true, unknown level of the characteristic or item measured plus an error of measurement. In practice it is important to know whether or not the variance in errors of measurement of an instrument, or the imprecision of measurement, is suitably small as compared to the variance of the characteristic or product measured, or the product variability.
Background
0011 0010 1010 1101 0001 0100 1011
For efficiency of the measuring process, the variance in errors of measurement should be several or many times smaller than the variability of the characteristic measured or product variance. High measurement error causes the power of most statistical tests to decrease unless compensated for by larger sample sizes
Power = 1 - Beta Beta is the Type II error
Device 1 = B1 + Xi + Ei1 where B1 = bias for device 1 Xi = true value Ei1 = random errors for device 1
Device 2 = B2 + Xi + Ei2
Regression Approach
0011 0010 1010 1101 0001 0100 1011
Y = b0 +b1*X Device 1 = Intercept + Slope * Device 2 Intercept value should be zero, if not, it indicates bias of the two devices Slope term should be around 1 indicating Similarity Mean Square Error estimate Precision R-Square term estimates some measure of strength
Inverse regression can pose several problems when trying to resolve the asymmetry problem
How to handle the case when there are multiple measuring devices?
Pairwise comparisons multiple testing error problem
Some Solutions
0011 0010 1010 1101 0001 0100 1011
Grubbs model
Tan and Iglewicz modify the standard regression approach based on Mandels work of Errors in Variables
Grubbs Model
0011 0010 1010 1101 0001 0100 1011
Involves calculating the Sums and Differences of the two devices The differences will be used to estimate bias
standard paired t-test with n-1 degrees of freedom
Performing a correlation analysis analysis on the sums and differences is used to estimate precision of the two devices
follow students t distribution with n-2 degrees of freedom
one for bias of the two measurement devices one for precision equivalency
Multivariate approach instead of independent tests Differences are regressed on the Sums
Difference = Intercept + Slope * Sums
Model F-test uses the UNCORRECTED SUMS OF SQUARES to get the correct number for the df
instead of the familiar corrected sums of squares the overall Type I error (Alpha) rate is exact
If the Model F indicates significance, then tests for the Bias and Precision are just the individual F-tests
the overall test is generally more powerful it can reject the equivalence assumption of the two devices even though each individual test does not
The Precision test is exact The Bias test is exact only when the precision between the two devices is equal
use the paired difference t-test otherwise more powerful Uses UMVU (uniform minimum variance unbiased) estimate of the variance
Type I Error Rate is exact Overall test could reject even though individual test do not
power of test
Statistical modeling
usual array of diagnostics
residuals
Regression approach
0011 0010 1010 1101 0001 0100 1011
Regression can still be used, but an adjustment to the model must be made In Simple Linear Regression, the results (Beta Hat) are achieved by minimizing the sum of squared residuals in the direction of the dependent variable The correction is to achieve Beta hat by minimizing the sum of squared residuals in the direction of -Lambda/Beta hat
Regression approach
0011 0010 1010 1101 0001 0100 1011
Lambda is determined by performing the standard Gauge RxR studies or taking repeated values
Regression approach
0011 0010 1010 1101 0001 0100 1011
The approach still uses regression and Lambda value helps solve the asymmetry problem
as lambda approaches infinity, implies X values approach 0, so Beta hat is Sxy/Sxx, which is where X is the independent variable as lambda approaches 0, implies Y values approach 0, so Beta hat is Syy/Sxy, which is where Y is the independent variable
Regression approach
0011 0010 1010 1101 0001 0100 1011
Approach handles precision but not bias Uses Polar coordinates for confidence intervals
Slope = TAN (Theta) Intercept = Tau/COS (Theta)
Example
0011 0010 1010 1101 0001 0100 1011
Device1 5.00 5.17 5.17 5.00 8.50 5.67 8.00 9.00 8.50 11.17 9.00
Device2 4.73 4.83 4.63 4.37 7.03 4.50 7.03 7.93 7.50 9.57 7.70
Sum 9.73 10.00 9.80 9.37 15.53 10.17 15.03 16.93 16.00 20.73 16.70
Diff 0.27 0.33 0.53 0.63 1.47 1.17 0.97 1.07 1.00 1.60 1.30
R-Square = 98.24% SQRT MSE = 0.31 Device 2 = -0.28 + 1.2 * Device 1 95% Confidence Intervals
Intercept (-1.07,0.52) Slope (1.07,1.31)
Conclusion
No Bias, but not similar
Regress Diff on Sums (Uncorrected SS) Model F = 73.22 ==> p-value = 0.0000027 Therefore, devices are different relative to their bias and precision Individual Precision Test
F = 17.5 ==> p-value 0.0024
Paired t-test has a T value = 6.98 and pvalue less then 0.00001
Correct Conclusion
0011 0010 1010 1101 0001 0100 1011
Potential Problems
0011 0010 1010 1101 0001 0100 1011
Methods do not account for a difference in Gain, or slope of devices Devices might measure equally well or poor at the low and high ends of the scale, but the relationship is not constant
collect data at one end of the data range power of the test could be compromised
Multivariate methods lead to fuller choice of sub-hypothesis and can be used regardless of the number of measurement devices
Independent variable is averaged each part across all the measurement devices Dependent variable is calculated by the differences of each value from that mean
Generally have to fit a Full model and a Reduced model (intercepts only) Then compare the two models
usually through some matrix manipulation
Technique can be performed by most software packages that can perform MANOVA techniques
Authors Opinion
0011 0010 1010 1101 0001 0100 1011
As the title says, this is just my opinion and not based on any concrete proof
such as simulation studies
My preferred method of analysis would be the Multivariate approach using Blackwood and Bradleys Regression with the Uncorrected Sums of Squares
this procedure seems to have a more powerful test in finding differences eliminates the possibility of getting a negative variance, which Grubbs method could get
Authors Opinion
0011 0010 1010 1101 0001 0100 1011
With the Multivariate approach, there is a natural extension to testing more than two measuring devices
Of course, there is no reason to try both the multivariate approach and Grubbs approach since they are easily computed using standard data analysis techniques
Final Comments
0011 0010 1010 1101 0001 0100 1011
These methods are not to replace Gage RxR studies, but to evaluate two devices against each other
Each device should be tested for bias and repeatability and linearity as desired
corrective action should be taken as needed
Final Comments
0011 0010 1010 1101 0001 0100 1011
The test for Bias is only a test for agreement between the two devices, not a bias against a standard
both devices could be grossly off from the standard (but in the same direction and amount)
If there is a claim that one device is superior to another (better precision), these methods could prove the validity of the claim and provide the precision estimates
Grubbs, F.E. (1973). Errors of Measurement, Precision, Accuracy and the Statistical Comparison of Measuring Instruments , Technometrics Vol. 15 pp. 53-66 Bradley, E.L. and Blackwood, L.G (1991). An Omnibus Test for Comparing Two Measuring Devices, Journal of Quality Technology, Vol. 23 pp. 12-16 Tan, C.Y. and Iglewicz, B. (1999). MeasurementMethods Comparisons and Linear Statistical Relationship, Technometrics, Vol. 41 pp. 192-201
Christensen, R. and Blackwood, L.G (1993). Tests for Precision and Accuracy of Multiple Measuring Devices, Technometrics, Vol. 35 pp. 411-420
Bedrick, E.J. (2001). An Efficient Scores Test for Comparing Several Measuring Devices, Journal of Quality Technology, Vol. 33 pp. 96-102
brian.novatny@us.michelin.com