iSixSigma

Normal Vs Non Normal data

Six Sigma – iSixSigma Forums Old Forums General Normal Vs Non Normal data

Viewing 16 posts - 1 through 16 (of 16 total)
  • Author
    Posts
  • #49128

    FL
    Participant

    Hi
    I have heard a common statement among the six sigma practitioners that “one should try to have a continuous metric and if the data is non normal then use transformation to make it normal”
    I am in the opinion that we have to be able to understand a process right at the transactional level. By transforming the data, the finer aspects are lost. I guess we transform the data to attempt calculating the Zshift.
    Awaiting your opinion
    Thanks
    FL

    0
    #167309

    BC
    Participant

    I would agree with your assessment.  And calculating Zshift isn’t all that important.  Normal data is nice if you need to do hypothesis testing also.
    The most important thing is to characterize your process:  understand what its distribution is, where it stands relative to the customer spec, and whether it’s stable over time. 
    BC 

    0
    #167313

    Mikel
    Member

    More six sigma rubbish !
    It takes 3200 data points to be certain that a distibution is normal out to 2.95 sigma … and by that time the process will have changed.
    Read Shewhart !  Control charts do not need any particular data distribution.
    A study of over 1000 different distributions by Wheeler “Normality and the Process Behavior Chart” (see page 88) showed that except for rare, extreme situations, 3 sigma limits work ! 
    Data transformations lose all meaning in the data … and of course require reverse transformations … a total waste of time.
    Use Shewhart charts.  Remember the AIM of control chart is to identify signals that should be investigated.  They are NOT probability charts as is commonly assumed.
     
     
     
     

    0
    #167315

    Six Sigma Shooter
    Member

    Exactly! (a loud applause from the Shewhart – Deming types!)

    0
    #167316

    Six Sigma Shooter
    Member

    By the way, you have just opened yourself up for the same old stuff from the probabilistic crowd.  We’ve been there, done that many times before on the forum.  Ah well, you can weather it well.

    0
    #167317

    SiggySig
    Member

    Stan, would you then not transform using process capability? Just use the distribution identifyer feature of Minitab?I ask because I ran a cycle time reduction project that had pretty non-normal data (Mean of 70 days, median of 43).

    0
    #167328

    Mikel
    Member

    Forget about transforms.  Most time related processes are non normal.
    Can you please describe your process in detail, what you are trying to do, and post some data ?
    By the way, forget about drawing normal distributions over the top of your histograms … this is totally meaningless.  Curve fitting achieves nothing.  The purpose of the histogram is to help gain an understanding of your process.  This can only be done using your raw data.

    0
    #167330

    monts
    Participant

    Use IMR Charts to analyse process behavior. They will through light on outliers and homogeneity of the data set. Trying to fit the data into a distribution is quite meaningless

    0
    #167331

    Fake ATI Alert
    Participant

    Stan
    You have  to  distinguish between  Parent Population  and  the Child Distribution

    0
    #167332

    Fake ATI Alert
    Participant

    Please  elaborate  the  role  of  the Monte Carlo Simulation  in  this  regard

    0
    #167340

    Mikel
    Member

    That’s easy, even for a guy who calls himself All Things Idiot (ATI).The child populations poops in their pants, the smell gives them
    away.

    0
    #167341

    Fake ATI Alert
    Participant

    You should  change  your  name  to “SATAN”
    It  suits  you  better

    0
    #167343

    Mikel
    Member

    Thanks for the advice.I think your name has hit your sweet spot – Don’t change a thing.

    0
    #167348

    BC
    Participant

    No, Satan is far more subtle.

    0
    #167667

    Nikki
    Participant

    In my 6+ years as a Six Sigma practitioner, I have used transformation for the purposes of obtaining expected process capability and go beyond just observed capability from the sample (there is a very nice transformation feature since the Minitab 14 release).  For hypothesis testing, when the data is non-normal parametric tests can be used instead of the usual regression, ANOVA, etc, which all assume/require normal data to be accurate.

    0
    #167678

    Robert Butler
    Participant

      The statement “For hypothesis testing, when the data is non-normal parametric tests can be used instead of the usual regression, ANOVA, etc, which all assume/require normal data to be accurate.” is boiler plate which, if followed by the beginner, will guard against their making egregious errors.  However, because it is boiler plate it creates barriers to analysis and limits to method usage which serve, in the long term, no useful purpose. 
      The following posts may be of some value in putting some of these limits in perspective.
    Issues surrounding questions of normality and ANOVA
     
    https://www.isixsigma.com/forum/showmessage.asp?messageID=110215
     
    Normality and regression
     
    https://www.isixsigma.com/forum/showmessage.asp?messageID=69370
     
    process capability for non-normal data
     
    https://www.isixsigma.com/forum/showmessage.asp?messageID=23506
     

    0
Viewing 16 posts - 1 through 16 (of 16 total)

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