- :
Impact of Number of Parts, Number of Operators and Number of Replicates on Measurement System Performance Metrics

Stephen Clarke

SABIC

Sugar Land, Texas 77478

Abstract

In conducting a Measurement System Analysis, the operational parameters of number of Parts, number of Operators, and number of replicate measurements is a question that is often discussed. This study endeavored to determine the effect of these parameters on the width of the 95% empirical prediction interval for the Intra-Class Correlation (ICC), and Gauge Reproducibility as a Percent of Tolerance (GRR%TOL). The empirical interval was chosen to account for the severe skewness observed in the ICC simulation results. In all cases (Good, Marginal and Poor measurement systems), the impact of the number of parts, operators and replicates on the width of the 95% prediction interval were negative, as expected. However, for the width of the prediction interval for ICC, the effect of the number of parts was much larger (as much as 8-fold) than the effect of the number of operators or replicates. For GRR%TOL the effects of parts, operators and replicates were similar in magnitude, but the effect of the number of parts was actually smaller than the operator or replicate effects. These results were compared with the simple regression on the resulting number of observations. ICC exhibited a larger R

^{2}for the three separate regression variables model compared to the number of observations regression. This supports the idea that the number of parts is considerably more influential in the width of the prediction interval for ICC. However, GRR%Tol exhibited a similar R^{2}to the Number of Observation regression, suggesting that for GRR%Tol, the impact of number of parts, operators and replicates is primarily a total number of observations effect. If one is relying on ICC, then the number of parts should be increased over increases in replication or operators. However, if one is using GRR%TOL, then increasing the number of operators is more effective in reducing the 95% prediction interval. Lastly, the severe skewness of the 95% prediction interval for ICC should be taken into consideration when determining the appropriate choice of measurement system metric. This skewness in ICC results from the use of variances (which follow a Chi-Square distribution), as compared to the use of standard deviation for - : Stephen E. Clarke
- : SABIC
- : Stephen E. Clarke
- : quality
- : intermediate
- : crofut@mac.com
- : 413-441-1491

Impact of Number of Parts, Number of Operators and Number of Replicates on Measurement System Performance Metrics