What analysis can be done on attribute inspection sampling results or range of values data?

Six Sigma – iSixSigma Forums General Forums Methodology What analysis can be done on attribute inspection sampling results or range of values data?

Viewing 3 posts - 1 through 3 (of 3 total)
  • Author
  • #53496


    :) I have a database of results for inspection sampling in what I consider a strange form and wonder what analysis can be made. Data is mostly variable, but has been treated as an attribute where pass / fail results are recorded for batch lots. Despite the number of samples measured, only the min and max value of the values are recorded and a pass fail decision made. (batch sizes and sample sizes are known)

    Sampling provides a decision to accept a lot for shipment, but is there any other use for the data?

    What does a history of batch acceptance tell me about the process?

    Can a run chart, capability, DPM, or anything at all be surmized from such a collection of results? I am accustomed to continuous variable data and I am unable to see a good direction in this case.

    Would it make sense to treat all the different product types made as equal and analyze the entire data set as attribute pass / fail data to try and show an ability to “pass” any given lot?


    Robert Butler

    You have a pass/fail rating for each batch/lot (are we assuming batches and lots are the same thing or are they different and if they are different are the pass/fail ratings for lots within a batch or just for a batch or some kind of mixture of both?)

    If you have pass fail ratings that are consistent (either batch or lots) and if you have sample sizes and if you know that the specifications for pass/fail haven’t changed over the time period covered by your data for all of the products then there are a few things you could do with the data but I don’t see that they would be of much value.

    1. You could look at probability of pass/fail as a function of sample size, product type, and time.
    2. You could generate two point estimates of lot variability over time and examine it for trending both in a temporal and in a pass/fail sense to see if there are any connections between variance estimates and time or pass/fail (one would assume that, if the two point estimates really reflected the lot-to-lot variation that there would be a connection between them and pass/fail ratings).
    3. You could make the large assumption that the midpoint between the min and the max values is representative of the lot mean and you could assume the two point estimates were representative of the lot variation and if you knew the specs for each product had not changed over time then you could use this to get a crude estimate of process capability.

    The problem with all of the above is that each of the possibilities makes an inordinate number of assumption and, as a result, any analysis you might run and any conclusion you might draw has a high probability of being nothing more than poorly written (and very expensive) science fiction.

    There are times when you are given data which has little or nothing to offer with respect to the issues of interest. About the best you can do under these circumstances is thank everyone for their effort, put the data in a round file, and get on with putting together a sampling plan that will provide you with the data you need.

    Based on your post I would say the data you have falls in this category.



    Your eloquent summary concurs with my suspicions and it is reassuring to have this forum to get such input. To quote one of my new collegues; “it is what it is.”

    My best effort to make some sense of the data so far has produced some cumulative defect (ratios) history, which when graphed with reject events, at least depicts on one page a picture indicating somewhat a performance summary, leaving prediction to the imagination.

    Formally constructing a prediction for Pass vs Fail result is what I am seeking to do now, and I AM NOT SURE THE BEST ROUTE. If there are 3 rejected batches and 97 accepted batches over a 5 year period where the 3 rejects were batch # 2, #5 and #6 in time sequentially does this mean that there is a 3% chance of batch failure today? How do I include the time factor to indicate better performance of late? Production units between failure, time between failure? This discussion thread is helping me formualte an attack angle. Any ideas for the best presentation of such information? Can I get a real number for liklihood of failure and a confidence interval? Each successive accpetance should improve the confidence interval.

    If I explain the weaknesses and the reasons surrounding a current estimate of capability (variation and mean estimates) I hope to make a good case for the value of changes needed to improve data collection. I expect to need to pleed a case for the $$$ value of knowing capability by finding current weaknesses that exist because it is not known, and plan to establish the causation link to the $$$ opportunity. The opposition to increased data collection; though needed, is real for several stake holders but I consider myself lucky to have the opportunity.

    Thank you Robert.

Viewing 3 posts - 1 through 3 (of 3 total)

You must be logged in to reply to this topic.