iSixSigma

Validating a measurement as a root cause

Six Sigma – iSixSigma Forums Old Forums General Validating a measurement as a root cause

Viewing 6 posts - 1 through 6 (of 6 total)
  • Author
    Posts
  • #38969

    Nwajei
    Participant

    Anyone,
    How can you validate that “not having a measurement system” is the root cause of a defect or performance metric? Seems impossible to do since validation is only acheived if you can measure (the measurement?). 
    The why because can only extend to…because we cannot define the parameter…why? ….because we do not know how …why? ….(not we’re getting away from a measurable factor).
    Without getting to specifics (yet), I would appreciate some insight from those BBs or MBBs who have experienced a similar dilemma.
    Thanks in advance.
    Frank
     

    0
    #117459

    walden
    Participant

    Frank,
    I’m relatively new to the Six Sigma world however I know a little about stats.  The key question you should be asking is what kind of variable are you dealing with: nominal, rank order, string (numerical)? Based on the answer to that, you can assign numbers to nominal and rank variables to separate the items into like groups.  Mean responses (number of changes) can then be measured and monitored. The key to determining if a measurement is valid is to run descriptive stats on the like groups to determine likeness and differences (homogeneity and heterogeneity). If this is the direction you were going, let me know and I can elaborate. If I’m off-base then please clarify my error.
    Chris
     

    0
    #117461

    Nwajei
    Participant

    Chris,
    thank you for the response.
    It’s not a question of validity of the measurement system.  Here’s an elaboration.  If I determined that the root cause of a defect was the incapability to identify it as a defect (yes/no or go/no-go) and to validate that, in fact, it was a cause then I would need a statistical method to associate the cause to it’s effect.  The effect, if I am right, would be the yield of good units or the throughput would increase in statistical and practical significance.  It’s intuitively obvious that it does, since by simple definition and application has increased the yield but is that good enough as validation? I don’t think it is because here’s the kicker, anyone can then challenge me with a question like…”how sure are you that your measurement system (which is a screening mechanism) is overly generous (yield more that it really should or passing units that shouldn’t pass) when you never had a base for “good” to begin with?”  What I do have is emperical data from “good” units behind which I based my definition and established the measurement criteria, How can I associate that data (which is not normal) to an way to validate that the criteria is statistically valid (if it can be done at all)? 
    I can be more specific.
     
    Thanks again….Frank 

    0
    #117573

    walden
    Participant

    Frank,
         I see the answer as part of your question. Prior to validating misses, review your criterion selection.  Are you familiar with signal detection theory? http://psych.hanover.edu/Krantz/STD/STDbasic.html. If you are too conservative in your selection process you will reject items that were defective but you will also reject items that were not defective.  Being too generous would have the opposite affect. The question is what is the cost to the customer or the company if you miss defects, how many defects? Using signal detection theory, which does add a subjective variable, you set the criterion based on severity of wrong decision, good or bad.  If you create a new life saving medicine, is it better to save 10 lives with the possibility of killing one?
       Back to the stats question of cause and effect, can you prove cause and effect on something that you have not manipulated? NO. Most stats are based on deviations from the normal distribution.  The results are usually correlations not cause and effect unless in a very very controlled environment.  And final results are probabilities if you have the same sample every time which is not possible.  What you are proving is that your criterion is isolating defects better than chance alone thus proving it valuable.  Compare your defect identification results with a no criterion base and see what the difference is. I hope this helps.  There are always debates on the validity of f and p values but if you can prove that your study is better than chance alone and show based on Signal detection theory were you d-prime should lie I think this should help your justification.

    0
    #117697

    Nwajei
    Participant

    Chris,
    You have hit the cord when saying “you set the criterion based on severity of wrong decision, good or bad.”.  Kind of like evalutating a risk reward ratio.  In my particular instance, the important question is “what is the risk of being wrong considering the continengy plans that would mitigate the effects of being wrong”
    I’m not familiar with the signal detection theory and I’m not sure how to use it but I will research to see if there is a simple application I can learn from.
    Appreciate your input.
    Kind regards,
    Frank
     

    0
    #117699

    walden
    Participant

    Frank,
    Glad I could help. Without boring you with details, Signal detection theory originated from peoples’ yes or no selection in their ability to hear a sound over the noise in their particular environment.  The question was asked, why do some people selectively (by choice) say they hear the noise whether it was present or not.  Anyway, your data can be simplified by entering said product into the critical and parallel paths. How much time and money (measurable)do you lose trying to install a bad part, having to fix the bad part, or having to send it back and order a new part (product pull)? What is the wait time for receiving a new product?  If the part is not used to build a product in your company, the same questions are asked to the customer. The hit or miss criterion in SDT is a probability solution, but you can just use the principle as guidance for false hit, correct hit, false rejection, or correct rejection.
    Chris

     
     

    0
Viewing 6 posts - 1 through 6 (of 6 total)

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