iSixSigma

MSA Attribute study to 3rd edition

Six Sigma – iSixSigma Forums Old Forums General MSA Attribute study to 3rd edition

Viewing 10 posts - 1 through 10 (of 10 total)
  • Author
    Posts
  • #34835

    boettler
    Member

    Thanks Matt for the basic guidelines for conducting an attribute study.
    I have read over page 137 a number of times and still can not figure out how to calculate bias, and repeatability taking the information you have detailed Matt and trying to calculate the bias. Especially after reading on how to caluclate the bias as page 137? I am really confused. I have a deadline of March 10 to calculate a number of attribute studies as per the 3rd edition.
    Help!!

    0
    #96561

    DaveG
    Participant

    What is it you don’t understand?

    0
    #96562

    boettler
    Member

    Thanks Dave for replying.
    The concept presented in the study as present by Matt and Joy I understand but how does one caluculate the BIAS from the data presented. I understand the study presented to be 20 samples appraisers 2 number of times each appraiser puts the samples through the attribute gage is once, find out the number of times the appraisers aggree between the tewo of them and then find the % where the appraisers aggree with the expert.  Cab I calculate bais? How?
    I have read over MSA 3rd edition a few times starting page 125 but I seem to be baffled.
     

    0
    #96577

    Gabriel
    Participant

    This method works only to evaluate agreement between operators and between operators and the standard. It does not compares the decision (good bad) against “how far from the limit the part was”.
    The “bias” would be the difference between the average “break point” (where the desicion changes from good to bad) and the actual specification limit, so the previous method does not gives you the bais.
    Read the analytical method, page 135 and on.

    0
    #96579

    DaveG
    Participant

    What Gabriel said.

    0
    #97002

    boettler
    Member

    I have got the 3 appraisers, the 8 samples length has been built where by the min is right at the bottom end of the tolerance spec ie 10000 mm and max is 1006.00 all others are nicely divided between 1000 and the 1006 mm length, therfore the tolerance band is 6 mm.
    According to the MSA 3rd edition each sample is to be checked 20 times, for a total of 160 measurments per appraiser since each  one that was checked to the tolerance band I made one osammple the min. sopec length alowed which is the one sample that the appraosers will question and another sample at the top end of the spec here again it is questionable all the other samples range in length which lay nicely within the 6 mm tolerance in the length of the samplethat meet the 6 mm tolerance. I have a total of 480 attribute results. How do I go about calculating the bias for the study. the example in the book i find very confusing. Is the bias reported to be between the appraisers or the bias of the entire study end results?
    The repeatablility, reproducibilty has been done using all 480 results. When I set up the study all 8 samples lengths are within the 6 mm tolerance length band. 
    The three appraisers I had one appraiser indicate 2 samples out and all others in . One sample at the min. end of the tolerance band and one at the max end, the appraiser indicated not ok when indeed the samples was in spec.
    appraiser two results they indicated one sample as being not ok this one was recorded as a not ok, when infact it was ok. this one was found on the min end of the spec of the tolerance band.
    Appraiser three had 2 questionable results indicated as “No ok”, here both the results they reported was for sample 1 of 8 that was on the min end of the tolerance band.
    help!!!!!!!!!!!

    0
    #97003

    faceman
    Participant

    I have never looked at the 3rd edition MSA.  I left that field when 2nd was the standard.  I can share this.
    Bias is typically the difference between the ‘true’ value of a part and what is indicated by the gage.  This is typically handle by a calibration study not the R&R.  If you used the same gage to establish the dimensions of the part (low end of spec, upper end of spec, and the rest distributed throughout the spec range then you built the bias into those parts anyway).  Usually a lab environment is better for bias.  Also, this isn’t a hard and fast rule, but some would suggest that you should select pieces that represent you actual process range not the spec range.  It can affect you % to process values.  Some R&R programs allow you to enter the historic process variation to address this.
    The ‘bias’ between appraisers that you mention should be captured in your reproducibility.
    See question three on this FAQ page.
    http://www.aiag.org/publications/quality/msa_faq.asp
    also do a ‘find’ on bias, it is mentioned several times on this page.
    Good luck, hope this helps.

    0
    #97004

    boettler
    Member

    Thank you for the information, the attachement you have added to your e-mail, for some reason I can not open it, the command has been cancelled. according to the MSA manual 3rd edition for attribute studies,one can calculate the bias for attribute study pages a 125- 140. Once I am able to open your attachement hopefully, it will clarify the quaqtions i have and how to do it. Like I said the example in the MSa Manuel 3rd edition I find confusing.

    0
    #97014

    Gabriel
    Participant

    Rob,
    I posted theis message just before yours:
    “This method works only to evaluate agreement between operators and between operators and the standard. It does not compares the decision (good bad) against “how far from the limit the part was”.
    The “bias” would be the difference between the average “break point” (where the desicion changes from good to bad) and the actual specification limit, so the previous method does not gives you the bais.
    Read the analytical method, page 135 and on.”
    You are not doing what I told you, and you are making conceptual mistakes too. That message is still valid, but since it was not enough to make you do what you have to do to get what you want, here you have more details. Let’s see if now I can convince you (if I don’t, then good luck and it was ice to meet you).
    – You are still using the wrong method for the bias. The method you are using is only useful to assess agreement within operator, opertor-to-operator, and operator-to standard. But no bias gan be guessed because the decision is recorded just “agree” or “disagree” without measuring the size of the dissagreement.
    – The result you got was not only logical and expectable: It is the best result you could have gotten. You must anderstand hat there exist not such a “part just in the limit of the spec”. The only thing is that you don’t know “exactly” the size of the part, and as far as you can measure it you don’t detect a difference against the specification limit. It is logical that, if the attribute measuring system works Ok, parts that are very very close to the limit of the specification will be sometimes accepted and sometimes rejected. You cannot expect a part at 999.999 to be rejected allways and a part at 1000.001 to be accepted allways. That’s not realistic as you would need to have a measurment system which is free of any kind of variation (which does not exist). In a perfect attribute measurement system, any part that is an infinitesimal out of tolerance would be accepted with probability 0 and a part that is an infinitesimal in tolerance would be accepted with probability 1, where a part “exactly” at the limit would not exist (in a continous distribution you have probabilityes that an individual can be within a certain range, but the probability that an individual has exactly one value is allways zero). In a real masurement system, you will have three zones. The zone where the part is so bad that it will be accepted with probability virtually 0, the “grey zone” where the parts will be sometimes accepted and sometimes rejected, and the zone where the part is good and so far from the specification limit that it will be accepted with probability virtually 1.
    Try to imagine a graph where you plot the probability to accept a part against the size of the part (or don’t imagine, see page 140). In your example, the plot of a perfect attribute measurement system would start flat at zero up to a size of 1000, at this point it will have a step to 1 and it would keep flat at 1 up to 1006, and at this point it would step down to zero again and remain flat at zero.
    The plot of your real-life less-than-perfect measurement system, however, wil behave different. The plot would also start flat at zero. But somewhere close to 1000 it will begin to slope up with an increasing slope. At some point it will reach its maximum slope and, while still going up, it will begin to reduce its slope until somewere at the right of the poit where it departed from zero it will rach 1 and keep flat. The zone from the point where it departs from zero to the point where it reached 1 is the grey zone. Ideally, this grey zone should be as narrow as possible and centerd on 1000. The width of the grey zone is the repeatability and the offset from the center of the grey zone to the specification limit is the bias. Of course all the same could be said for the other specification limit.
    So what you should do is to take parts arround and very very close to 1000, starting from at least one part that is allways rejected to at least one part that is allways accepted, and with at least 6 parts in the middle that are accepted an incrasing percent of the times as the size goes up (that’s at least 8 parts). The increment from part to part must be as constant as practical. Just as an example:
    Size           Accepted (on 20 trials)
    999.86           0
    999.88            0
    999.90          1
    999.92           3
    999.94           6
    999.96          9
    999.98          11
    1000.00        14
    1000.02        17
    1000.04         18
    1000.05        19
    1000.06        20
    1000.08        20
    With this info you can plot the probability to accept against the size graph, and measure the witht of the grey zone and the offset from 1000.
    Then you would redo the process for 1006.
    Follow the indications on page 135 and on (had I told you this before?)
    By the way, bias is an instrument offset, so the test should be done in laboratory conditions, and only by one operator (a well trained one). If the bias becomes “operator dependent” then that will be found in the r&R (agreement) study as a lack of agreement in parts that are close to the limits.

    0
    #97030

    boettler
    Member

    Thank you Gabriel.
    I undertstand what you mean, it will become more clear once i read over from page 135, it is the Manager who seems to be on the path as I described previously.
    Thanks for the help 

    0
Viewing 10 posts - 1 through 10 (of 10 total)

The forum ‘General’ is closed to new topics and replies.