iSixSigma

Interaction and confounding

Six Sigma – iSixSigma Forums Old Forums General Interaction and confounding

Viewing 12 posts - 1 through 12 (of 12 total)
  • Author
    Posts
  • #37189

    ROSS
    Member

    Hi:
        I have two questions as belowing:
        Interaction:
        When we do a two-way anova want to know affection of factor-A  factor-B. From the result of Minitab, we can find factor-A and the interaction of A and B is significant, and we can improve factor-A directly, but what should we do about the significant interaction?
       Confounding:
         When we do DOE analysis, sometime we will face condounding, what should we do about confounding in a DOE analysis.
       Tks!
     

    0
    #108920

    mjones
    Participant

    Tony-
    I’m glad to help, but this is the third basic question you have asked today. What is the deal man? I have to wonder what you’re into. Are you in GB or BB training? Do you have an instructor or MBB? Have you talked to them? 
    Are you actually doing ANOVA and DOE? If so, you may getting into some kind of dangerous issues, like concluding differences that do not exist or ignoring real, valid differences. But it’s your career. To your questions…
    If a factor is significant, you must decide how to address it. If the significant factor is an interaction, you must decide how to address it as well. Sometimes, significant interactions can actually help in your control and improvement of the process — gets to the issue of process knowledge and how to manage and control it. Deciding “how to address it” can mean using the information to improve the process, but it could mean deciding to ‘lock down’ the variable and use other factors to influence/manage the process; and there are other alternatives as well. Mostly, DOE and ANOVA give us information to make good decisions; they do not make decisions for us.
    Confounding? You do what you must do. If statistically significant factors are confounded, you will have to rerun the experiment, perhaps folding the design to resolve the confounding to determine, typically, if the true effect is a main effect or an interaction effect.
    And, if this is not enough, you have to be concerned about 3-way interactions if the process is such that 3-way interactions are possible/likely.
    By the way, as a caution: you’d best make sure your MSA is done well or you can’t trust your ANOVA or DOE.

    0
    #108934

    ROSS
    Member

    Hi mjones:
        Thank for your instruction very nuch, and I am a BB, but we get training from consultant firms and we have no MBB, so I may ask you some basic concept.
        For interaction, do you mean Minitab can not do any further analysis when we doing anove or DOE? What can we do is go to real process and find solution with our process knowledge? But we always go to statistic for lack of process knowledge.
       For the confounding, do you mean we need more experiments on it to avoid confounding?
       Look forward to your answer! Tks!
     
     
     

    0
    #108997

    ROSS
    Member

    Hi:
       Can sombody who has experience on it help me?

    0
    #108998

    Ken Feldman
    Participant

    Confounding can be a function of the resolution of your experiment.  If cost becomes a problem and you decide to try and minimize the factorial design then you will start to encounter confounding.  You generally won’t worry much about confounding of higher order interactions.  For example, a Resolution IV experiment will have confounding of three way interactions.  Assume a 2 level, 4 factor experiment.  A full factorial would give you 16 runs.  A half fraction would give you 8 runs and be a Res IV experiment.  But you would lose the ability to discriminate 3 way interactions.  This would usually not matter to you so you would go with the 8 and save a few bucks.

    0
    #108999

    ROSS
    Member

    Darth:
    Thanks for your feedback first!
    Do you mean we need not do anything about confounding for it is nature when we do the factorial design for cost and experiment situation reason?
    And can you have some explain about interaction?
     

    0
    #109001

    Tim F
    Member

    One intuitive way to think about interactions is to use cooking analogies. Suppose you have a recipe for cookies, but they don’t taste very good to you. You decide to adjust the amount of ingredients to find a better combination. Perhaps you adjust flour, sugar, milk, butter, and eggs, using a some factorial design. You discover that on average, the cookies are better with more sugar. That is a main effect.Suppose the recipe has both white sugar and brown sugar. You make some more cookies and discover that increasing EITHER the white sugar OR the brown sugar helps, but increasing both makes it just too sweet. That is a two factor interaction.Suppose the recipe has sugar, corn syrup, and water. If you add more sugar, you might decrease the corn syrup to balance the sweetness. However, this takes fluid out of the recipe, so you need more water. This would be a three factor interaction. Tim F

    0
    #109003

    mjones
    Participant

    Hi Tony-
     
    Darth answered much of your question already. But to expand a bit.
     
    You ask: For interaction, do you mean Minitab can not do any further analysis when we doing anove or DOE? Yes, that is precisely what I meant. You use your process knowledge to understand what is happening and why; make sense of it all and take the right action.
     
    You ask: What can we do is go to real process and find solution with our process knowledge? But we always go to statistic for lack of process knowledge. Yes again. As stated, use your knowledge to do the right things. Statistics are not a substitute for process knowledge. Statistics are tools/methods to support process knowledge by showing you what effects are, or are not significantly different from random noise. There is no magic here. Just a bunch of clues and a lot of hard practical work.
     
    You ask: For the confounding, do you mean we need more experiments on it to avoid confounding? Yes once again. (Note, this was addressed substantially by Darth.) Basically, to remove the confounding you must have a larger sample. There are several alternatives including folding the design, adding replicates, and/or reducing factors.
     
    You have confounding when you have a fractional factorial experiment. But the purpose of this type of experiment is for screening, or determining which factors are and which are not significant. These designs are not intended to completely understand all factors at all levels. You need to run additional, more detailed designs to get you to that level. Again, I strongly suggest you get with your training consultant to better understand these things; and/or do more research on your own. This is not a subject to easily ‘learn’ from a web forum.
     
    Good luck.

    0
    #109006

    ROSS
    Member

    Thanks MJones Tim F Darth mjones
        Now I get it. Tks!

    0
    #109012

    Dog Sxxt
    Participant

    When AxB interactive is significant, you shall find out optimum level for A and B by running more experiments.
    Do not need to worry confounding for a full factorial experiment (with a balanced orthogonal matrix).  Suggest to use Taguchi method if you are not very sure how to avoid confounding effects.

    0
    #109020

    Ken Feldman
    Participant

    Confounding is a result of fractional designs.  The key is whether the confounding causes you to not be able to fully understand your outcomes.  For example, if in your design, you have a three way interaction confounded with your Main Effect and the three way interaction is not significant, you might not worry about it.
    A simple way to understand confounding might be if you and your pet dog Killer both went on a diet.  You then got on a scale with Killer in your arms and the scale said weight was lost.  Was the lost weight due to you or Killer or both?   You can’t really tell since there is confounding or mixing of the results.  There is lots of info out on the Web so try doing a few searches and you will find all you need to know.

    0
    #109021

    Mrroeb
    Participant

    A confound is anything other than your independent variable(s) that is having an effect on your dependent variable.

    0
Viewing 12 posts - 1 through 12 (of 12 total)

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