iSixSigma

DOE Confusion

Six Sigma – iSixSigma Forums Old Forums General DOE Confusion

This topic contains 12 replies, has 5 voices, and was last updated by  HF Chris 14 years, 1 month ago.

Viewing 13 posts - 1 through 13 (of 13 total)
  • Author
    Posts
  • #40746

    AT
    Participant

    I conducted a four factor two level DOE and found that only one factor has passed the alpha=0.10 significance test with P value less than 0.10. However, another factor shows the biggest main effect and appears strong even in the interaction plots. How should I proceed from here? Only consider the main efect of the first factor?

    0
    #127148

    Mikel
    Member

    What you have described is not possible. There is something wrong with your analysis.
    Post your data and you will get help.

    0
    #127149

    AT
    Participant

    Here are the trials with response replicates:
    A        B       C        D   Response replicates
    -1      1        -1       -1      8,6,9,8,10,8, 
    -1     -1         1        1      5,6,8,9,5,3,
    1       -1       -1        -1     5,8,6,7,8,6,
    1       1        1          1      1,7,4,5,1,7,
    1      1         -1         1     1,8,5,3,5,7,

    0
    #127150

    Mikel
    Member

    Your design is invalid. Where did you get it?

    0
    #127151

    Robert Butler
    Participant

    Stan is correct – as described you cannot look for any interactions at all. Your two ways are perfectly confounded with the mains or linear combinations of the mains. The best you can do is look at mains A,B,C,D.  On that same line – are the replicates actually replicates – that is did you run each one of the separate experiments 6 times for a total of 30 experiments or did you just take six consecutive samples from the results of a single run?  If it is the latter then you didn’t run replicates – you are taking repeat measures and that is an entirely different animal.

    0
    #127154

    Mike Walmsley
    Participant

    I agee with Bob and Stan. If you are not going to try to understand the tool and its proper use,do not use it. Purchase some decent software and use it. Minitab , Statistica , …  It has canned designs that should keep you out of trouble. I frequently use the alais matrix to assist in setting up and confirming the suitability of my designs. A good reference is listed elsewhere in the postings,yet will be repeated again here ” “Statistics for Experimenters” Box,Hunter & Hunter, Wiley Press.
    Don’t feel like we are beating on you. We have all been there at one time or another.
     

    0
    #127155

    HF Chris
    Participant

    Ramesh,
    I feel generous today so if you would like to understand where you went wrong post your sample in a way that makes sense to me. Post your data in matrix form. Is this a repeated measure design, between subject, discrete or continouous? discrete or continuous?

     
     
    Trial 1
    Trial 2
     Trail 3

    A
    Level 1
     
     
     

    A
    Level 2
     
     
     

    B
    Level 1
     
     
     

    B
    Level 2
     
     
     

    C
    Level 1
     
     
     

    C
    Level 2
     
     
     

    D
    Level 1
     
     
     

    D
    Level 2
     
     
     

    0
    #127156

    AT
    Participant

    Thanks for pointing out I am on the wrong track. The analysis is about purchase decision factors and each of the replicates are actually customer “likelihood to buy” ratings on a 1-10 scale. So I think the concern about repeat observations from same run is not applicable here as we are talking about different respondents. But I would really like to understand how you could figure out the problem ” two ways are perfectly confounded with the mains or linear combinations of the mains”. This will help me figure out if the same mistake is ever made again.

    0
    #127158

    Robert Butler
    Participant

      To figure out the confounding I used a computer to run diagnostics on your X matrix.  To do the same thing manually take your columns and multiply one column by another column and see what the resulting column looks like.  If it matches one of the other columns in your matrix then it is confounded with the variable associated with that column.  For example you have
    A        B       C        D 
    -1      1        -1       -1     
    -1     -1         1        1      
    1       -1       -1        -1    
    1       1        1          1     
    1      1         -1         1     
    Thus the AB interaction is just
    AB
    -1
     1
    -1
     1
     1
    which is identical to the D column thus the AB interaction is confounded with response D.

    0
    #127209

    AT
    Participant

    So the lesson to be learnt is that never do a less than orthogonal design. It is almost sure to have confounding which will make the results of the experiment completely invalid. Either do a full factorial or do a balanced fractional factorial as suggested by software or expert. Any exclusions/observations on this conclusion?

    0
    #127220

    HF Chris
    Participant

    No, the leassono f the day is to understand how to do an a priori; if you don’t understand do full 4 way interaction anova instead.
     
    chris

    0
    #127223

    Robert Butler
    Participant

    No, I’m afraid that isn’t the lesson learned. 
      Lesson #1 – if you are using some kind of software to generate a design make sure you know what you are asking and how the machine interprets your inputs.  Specifically – unless you are intrested in going into the very complicated world of super saturated designs a rough (very rough) rule of thumb is you will need at least one experiment for each variable of interest and one for the grand mean. Thus if you have 5 variables and 4 interactions of interest you wll need to have at least 10 experiments.  The fact that your design had only 5 experiments and 4 variables should have told you that, at best, you couldn’t possibly go after anything except main effects.
    Lesson #2 – Confounding will not make the results of an experiment completely invalid.  If you have a saturated design you will not have confounding between the main effects, by your choice of that kind of design, you have declared you are only interested in the main effects. Consequently,  the results of that kind of a design will be perfectly valid for the questions you wish to answer about main effects.
    Lesson #3 – By virtue of the fact that you can rarely set any given variable to exactly the same setting time after time even an orthogonal deisgn will be a touch non-orthogonal.  Designs, particularly the standard composite, factorial, and fractional factorial designs, are extremely robust to small departures from orthogonality.  In addition, if one of the horses die (that is one if one or more of the design points fails to run) the remaining design will no longer be orthogonal but you can still analyze the design for most of the effects of interest and still draw valid conclusions from your efforts.
    Lesson #4 – Learn the basics of experimental design – how to construct them and how to manipulate their results manually.  If you do this you will undestand how they are built and thus what a machine is doing for you. As a result you will be able to tell when the machine did what you told it to do as opposed to what you wanted it to do.

    0
    #127230

    HF Chris
    Participant

    Robert,
    A very nice response.
    Chris

    0
Viewing 13 posts - 1 through 13 (of 13 total)

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