iSixSigma

DOE Fractional Clarification

Six Sigma – iSixSigma Forums General Forums Methodology DOE Fractional Clarification

Viewing 5 posts - 1 through 5 (of 5 total)
  • Author
    Posts
  • #53593

    Doa
    Participant

    Hi To All,

    As I lumber through the many fine points of this powerful tool, I have the following question:

    When conducting a Level IV DOE, six factors, 16 runs, two levels, three replicates 48 runs or samples (in Minitab), is not blocking on replicates the point of replicates and one of the benefits of running replicates? Replicates increases our ability to detect significant differences and reduces Beta errors….a little affirmation or re-direction.

    Regards,
    Marty

    0
    #190788

    Robert Butler
    Participant

    You can block on replicates if you want to but that isn’t the point of replication. The reason you run replicates is to get a better estimate of internal error and you don’t need to block to do this.

    As an aside – in my experience full replication is a waste of time, money, and resources. I’ve built and analyzed hundreds of experimental designs and the only time I’ve ever run a full replication was when I was ordered to do so. In each of those cases I ran the analysis with the data points I’d recommended and then with all of the data points provided and in each case there was no difference in the findings.

    When running a design, the most efficient approach is to run the full design, replicate a couple of the design points (if you have a situation where you can run true center points then run the center and replicate it once) and then run the analysis and see what you see. If you have done your homework with respect to variable and variable range selection one of the following is very likely to happen:

    1. The results you get will be so impressive that the interest will shift from worrying about detecting some small difference to confirming the results of your analysis.
    2. You won’t see much of anything. Since you now have an estimate of internal error and estimates of mean differences you can use these estimate to address issues of power, and expected differences in means in the region of interest. In this case the chances are very good that everyone will want to stop and rethink the situation.
    3. You find either numerical trends that are not significant at your pre-chosen level of significance or you find one or two variables that appear to have some impact but not enough to account for a large portion of the observed variation. The usual procedure in this instance is to look at the directions of the non-significant trends and look at the levels of the variables that were significant and the direction of change in which they were significant, and re-think the design – most likely you will build a new design with different levels of the variables of interest.

    0
    #190790

    Doa
    Participant

    Good Morning Robert,

    I lucked out, you answered my post! This clears up my dilemma. I’ve spent much time on DOE…yet it is so important to have a firm and clear foundation.

    The key point for me concerning replications is a better estimate of internal error and better ability to detect signifcant diferences.

    Your sage advice always helps.

    Thanks again…

    Regards,
    Marty

    0
    #190791

    Robert Butler
    Participant

    I understand the sentiments behind your statement :

    “The key point for me concerning replications is a better estimate of internal error and better ability to detect significant differences.”

    And I also understand the source of this concern. However, the question you need to ask in these situations is the following:

    What is the point of an experimental design?

    The reason for running a design is NOT to get a great estimate of internal error so that you can identify statistically significant but physically meaningless differences. Again – I can only offer my experience but what I have seen time and again is precisely what I stated in the first post – if you find significant effects they will most likely be much larger than anyone imagined and, as a result, you will have spent a lot of time and money to generate unneeded precision with respect to the estimate of internal error.

    The easiest way to check this is to just do it. If you have prior data where you ran a complete replicate – take the initial design and a couple of randomly chosen replicate points – rerun the analysis and see if what you get differs substantially from what you found with the entire set of experiments.

    If you don’t have prior data and you are building a design – go ahead an plan for whatever number of replicates your machine says you need to detect the small differences someone has deemed to be important. Run the design, run a couple of points from the first rep and run an analysis. If there’s anything there they will very likely show up and I suspect no one will have any further interest in additional replication.

    0
    #191056

    MBBinWI
    Participant

    Robert: As always, excellent advice. The only thing that I would add is that oftentimes one or more variable is found to be non-significant. At this point, by reducing that variable from the analysis, you have effectively just added a replicate (to a balanced, two-level factorial anyway).

    The only caution – and this goes for any data gathering activity – is to ensure that your measurement system is adequate for the precision desired. I’ve had too many instances where novices plunged ahead with a DOE only to find in analysis that there was nothing significant. In most cases this turned out to be a measurement tool/method that did not provide needed precision, and replication would not have improved that situation.

    0
Viewing 5 posts - 1 through 5 (of 5 total)

You must be logged in to reply to this topic.