iSixSigma

How to treat and apply transformed data.

Six Sigma – iSixSigma Forums Old Forums General How to treat and apply transformed data.

Viewing 4 posts - 1 through 4 (of 4 total)
  • Author
    Posts
  • #33952

    Nwajei
    Participant

    Anyone with the stats mind set,
    I took data from samples of two machines tested each for performance “before” and “after” an engineering upgrade (two factors… machine and test)  The y response variable data is not normal and therefore I applied a Box Cox transform to get in normalized (to all the data combined).  The probability chart (in Minitab) shows normality.  However, performing an anderson darling test of the xformed data gives me a P-value less than .05.  Does this invalidate the xform? or could the low p value come from the possible differentiations of the y responses between machines and/or the “test”. I noticed that when I scope down to say, one machine or one test, the histogram depicts better normality and the Anderson Darling test rises above .05.  Is this the approach sufficient to ensure the data I’m using is sound to do hypothesis testing (I am comparing the before and after scenario to see that the engineering upgrade made a significant improvement to each machine)?  Any clarification would be appreciated.

    0
    #92929

    Philip Green
    Participant

    Frank
    I don’t fully understand the situation you have.  However it seems that one set of dat is normally distributed and the question is then whether a simialr set of dat is normally distributed.  My suggestion is that you use a non-parametric test (Mann-Whitney) to determine whether the data are from the same distribution.

    0
    #92965

    Nwajei
    Participant

    thanks for your input Phil.  I think I can offer some clarity.  If I want to compare data from a population of machines that have been upgraded (refer to as Upgrades) to a population of machines before the upgrade (refer to as Baseline) and I have a performance measurement (my Y response variable) which is not normally distributed (as is measures % failures), the transformed data of both combined doesn’t past the normality test, however if the transformed data between upgrade and baseline is close enough to call it normal but have different lambda values, can I still use it?  I don’t believe so since I’m comparing apples to apples.
    BTW I have used the Mann Withney test and believe that will lead to the eventual outcome.  I appreciate your thoughts. 

    0
    #93309

    WW.
    Member

    Hi Frank
    May I post some idea to your problem . I think that , The normality of y respone is not request for testing  base on your data collection ( Two factors ) Why not you use ANOVA ( Two way ) then test Residual of respone to validate the statistical assumsion ( NID(0,1) ) . The respone is not normallity , It may be come from the effect of one of two factors that you are interrested . The requriment of Normallity is used for subgroup only and Before use ANOVA , your should test both factors have equal Variation… That is my idea . If you have any suggession please let me know…Good luck..
    WW.

    0
Viewing 4 posts - 1 through 4 (of 4 total)

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