I need help setting up an experiment for work. As a part enters production, certain dimensions get measured at different locations. The same dimensions are measured, however, they are done by different operators and machines. I have been presented with the task of analyzing the measurements from the different areas, and comparing the values. I would like to know if there are any other statistical tests that may be useful, rather than just visually comparing the numbers to see if they are close. what do you guys think?
Naomi- it is expected that each person has repeatable results. However, I need to somehow show that each person's repeatable data is in specifications. The whole purpose is because one person's measurement is used for the final records. To try and ease this persons workload we are trying to see if another operator's data is sufficient for final data.(repeatable and in spec, like the original operator). Any ideas?
In industrial quality assurance this is known as a gage repeatability and reproducibility study (also known as "Gage R&R") and is designed to determine the % contribution of sources of variability to an overall measurement system.
You set this up just like a designed experiment and analyze it with ANOVA. The rule of thumb is that you don't want your instruments and operators to account for more than 30% of the total variability in the measurement system (ideally they should account for less than 10%).
Here's a link to more info:
Most software packages like SPSS, Minitab, or Statgraphics or Statistica have the ability to run a Gage R&R, but you can also just use ANOVA.
John, I have taken an introductory class on Gage R & R's. However, can you explain to me how that will prove the purpose of the testing? There are only going to be two operators, each in different locations. Is it not statistically sufficient to just compare each measurement the operator produces, and note the variance? I have too many tools/measurements, and not enough time to complete a gage r and r for each tool that we plan on comparing. However, as I write this it does seems like a suitable and fitting idea if time was not an issue.
If the different operators will be using the same gage, then just do a t-test, but if not, because the two different operators will also be using two different gages, you won't be able to tell if any differences you see are operator or gage-related.
Ok, so that goes along with what I said up top. Since both operators are different, and are using different machines, there is nothing else I can do besides comparing their measurements visually and seeing if they are in spec. Is that correct?
You can do that (visual comparison), of course, but if you want a less subjective approach you should also run a t-test to see if their measurements are statistically different. If they are statistically different, though, you won't be able to determine the root cause of the difference (see my previous posts).
Even if both sets of measurements are "in spec," it's a dangerous assumption going forward that everything in the measurement process / system is "OK." You might get lucky this time, but problems may come up in the future. To address this thoroughly, if possible, I would go even further than my recommendation in the first paragraph and see if you can set up a Gage R&R study to examine the operator-to-operator and instrument-to-instrument variability and % contribution to total variation in the measurement system.
John, I have been thinking the same thing. However, instead of doing an R&R on every dimension I had to originally measure, I will only do the worst couple. I have about 6 dimensions to measure on one part, and there has been significant variance between one dimension(this will be the one I will start with). It would be logical to do an R&R on all dimensions(since they use diff. tools to measure) but this would be too time consuming for the technicians who would need to measure.
I agree that you don't need to do this on every dimension. Where I work we focus on "critical" dimensions, that is, any dimensions or specs that will or may affect product performance or safety.
I have taken your advice and completed an R&R. After running the data through minitab, the results do not seem as expected. I specifically think this is because we did not have an adequate number or samples to measure. In my previous R&R's we have usually had a sample size of 10. However in this experiment, we only have a sample size of 6. This is because we cannot afford to test in production parts and we are limited to scrapped parts.
Do you think my results are sufficient enough to draw conclusions from even though my sample size is smaller than usual?
No - I would question the basic validity of a Gage R&R study done on only 6 parts, especially on 6 "scrapped" parts (i.e., who knows what the heck they are supposed to be in terms of dimensions).
If you cannot afford to use in-production parts, then I would use some mock-ups that are similar to production, where the dimensions have been noted and verified, prior to running the R&R.
John, I think you have misunderstood my meaning of "scrap". The pieces we are using are very similar to the final product, but have failed for some other problem along the way. We have personally limited the amount to 6 because they are the only ones in spec for this measurement, and have failed for other reasons. Due to limited supply, I was just wondering how much this would affect results.
Yes, 6 seems rather "low" and I would have doubts about the precision and accuracy of the R&R estimates.
Advertise on Talk Stats