iSixSigma

Correct Use of Hypothesis Testing

Six Sigma – iSixSigma Forums Old Forums General Correct Use of Hypothesis Testing

Viewing 15 posts - 1 through 15 (of 15 total)
  • Author
    Posts
  • #32316

    mcintosh
    Participant

    Hello,
    I am currently working on a project in an injection molding cell. My project is to get rid of a non value added operation that we are currently doing in the cell. This operation is when an operator is trimming excess rubber off of a part before they mold it in order to ensure they get a quality product. The customer is not paying us for this operation, therefore it is a waste and I have been tasked with eliminating it.
    My idea is to run a 1 day trial in the cell where I have an operator mold a part that is trimmed, and one that is not trimmed and keep repeating this cycle throughout the shift. When the operator gets a good part he marks down a 1, when he gets a bad part he marks down a 2.
    My null hypothesis would be that there is no difference between the two methods. My alternate hypothesis is that there is a significant difference. The alpha level that I will use is .05. Once I run the trial I plan on inputting the data into Minitab and running a One Way ANOVA on it to determine if there is a statistical difference or not.
    Could anyone provide me with some input as to whether or not you agree with this type of trial or if I am not doing something right? Any feedback would be greatly appreciated.

    0
    #86152

    mcleod
    Member

    I presume that you are trying to determine whether simply elimating the trimming step will add to the percentage of bad parts at final inspection. If this is the case, I would suggest you use the “2-proportions” test under basic stats to determing whether there is a difference between the percent good in the trimmed and untrimmed samples. Furthermore, your alternative hypothesis should be either greater than or less than in the options dialog box.Ask again if this isn’t what you are looking for.

    0
    #86160

    Erik L
    Participant

    Tom,
    Before you get started with the study, I would recommend that you look at the Power and Sample functionality of your S/W package and see what the recommendations would be for the appropriae sample size.
    Regards,
    Erik

    0
    #86162

    Zilgo
    Member

    As Erik said, check what the appropriate sample size would be and then as Scott said, use a proportions test.  ANOVA is not ideal in this case since you have discrete data.  Binomial (only choices are a 1 or 2) to be exact.
    It would be my suggestion to actually use a one proportion test.  You would test the proportion of failures from your sample (where you don’t do trimmings) against the historical proportion of failures (where you did do trimmings) and see if there is a difference in that case.  Comparing the two proportions from one sample opens itself to higher probability of poor data.  Presumably you have been keeping track of these failures (where you did do trimmings) for some time so that proportion should be representative of the process.

    0
    #86177

    captious
    Participant

    Another way is Chi-square.

    0
    #86201

    Michael Schlueter
    Participant

    Tom,
    Please consider not using hypothesis testing at all.
    I think your focus should be less on this rubber-removal step. Your colleagues do it as a precaution: not to get defects downstream the process. This fact can be either based on history (bad experiences in the past) or it can be just an assumption made.
    My focus would be on:

    excessive-rubber-follow-up problems
    creation-of-excessive-rubber process.
    Now I suspect, your line yield is a little less than 100%. I.e. you will probably be able to recognize dramatic changes in your yield anyway.
    Hypothesis testing will be clear on absolutely identical results and will be clear on absolutely different results. It will be ambiguous on little changes.
    Now, I suspect you are interested in dramatic improvement of your process downstream (the high-quality-product), i.e. much higher yield while the rubber-removal step has been eliminated. Unless your yield is very close to 100%, you probably do not want to keep it identicall at a lower level, do you ?
    I see you more in the need of two alternatives:

    improve the robustness of your process (DOE a la Taguchi) with continous data
    AFD – analysis.
    AFD (Anticipatory Failure Determination) is an analytical thinking tool. You can use it to identify that mechanism, which generates your excessive rubber, unless it is rather obvious to see. Because there are these question:

    why does your process create excessive rubber at all (it is waste, lost money, lost time, lost reputation etc.) ?
    why can’t they do better ?
    how to do it better ?
    Best regards,
    Michael Schlueter

    0
    #86225

    mcintosh
    Participant

    Michael,
    I have to focus on this rubber – removal step in the process. This is the project that I have been given by senior management. Right now this process is costing us approximately $60,000 a year in Indirect Labor. We did not quote this process originally to the customer, so therefore we are eating all of this cost.
    You are correct, my colleagues are doing it as a precaution to not get defects in the downstream process. However, it has become such a band-aid that now it is has become the norm. No one has challenged the fact that we may not have to do it at all.
    I appreciate your feedback. However, I still feel that I need to focus on this before I start attacking the overall end result (defective) in the work cell. Thanks for keeping me on my toes!

    0
    #86564

    Devesh Mathur
    Participant

    Hi,
    I think that your null (Ho) and alternate (Ha) hypotheses are correct.  But after doing ANOVA, you have to be careful about drawings inferences about Ho and Ha.  “p” value should be very close to 1 in order to accept Ho.  Else, the defect level will be high.
    You can do a DOE to find out optimum level of Xs in order to ensure that excess rubber is not produced at all.
    Best regards,
    Devesh

    0
    #86565

    Devashish
    Participant

    Hi Devesh,
    Was looking for you.I tried your no.Pl let me know your contact details.
    Thanks

    0
    #86567

    Devesh Mathur
    Participant

    Hi Debashish,
    My no. is 9845540879.  I am in Wipro now. Please give your no. and where you are.
    Best regards,
    Devesh

    0
    #86573

    faceman888
    Participant

    Is this ‘flash’ rubber?  If it is flash rubber then I understand that you might not be able to eliminate it without capital (I work in rubber too).  I would consider Michael’s recommendations though and see if it fits your project’s timetable and capital restraints (sometimes ‘flash rubber’ is inherent in the design of equipment and therefore beyond the scope of a  lot of DMAIC implementations).  If you have to go as you are currently planning:
    1) Use either a Chi-square test or a test of two proportions (assuming that your are evaluating two methods not 3 or more).
    2) I would do some power analysis up front.
    3) If you have 3 or more methods I would recommend GLM, it should handle the binomial response better than traditional ANOVA with a transform. (See Design and Analysis of Experiments, Douglas C. Montgomery, 5th Ed., Chapter 14).
    Good Luck, flash is tough, been there.

    0
    #86577

    Ron
    Member

    Look at the whole process not just the one operation.
    What is the reason there is flash on the part to begin with? EliminaTE THAT flash and you eliminate your problem.
     
    To your question a Chi Square would be appropriate, however if you rate your results from 1 to 10 regarding acceptability  you can use ANOVA.
     

    0
    #86601

    jediblackbelt
    Participant

    Tom –
    I may be completely in left field but I read this as you are wanting to see if there is a difference in appearance if you do an operation or don’t do an operation.  If you are interested in another method that is quick have you tried a Tukey Tail Count test?  It is an extremely quick test that I use for a quick “is one method better than another.”  You can have the operator run 10 parts of each method and then rank them from best to worst.  The person ranking should not know which process is which.  Then count the ends until you change from one state to another.  If you end up with a count of 7 then you can say that the “better” process outranks the other process with 95% confidence.  Check out Keki Bhote’s book “World Class Quality” for information.
    Example:
    Rank from best to worst
    11112212211121212222
    Your tail count is 8.  You have 4 #1’s and 4 #2’s on the ends.  You can now say with 95% confidence that process 1 is better than process 2. 
     

    0
    #86744

    J Andell
    Participant

    I lean toward agreement with Michael. Even if your management made a specific request, a first-rate Black Belt can recognize and attack the underlying issue. To paraphrase Deming: a non-Black Belt cannot always recognize a Black Belt problem when he/she sees one.
     
    I think the approach should be:

    Establish (using some of the hypothesis testing tools that have been discussed) just how much trim-able material the customer can tolerate. Be sure your decision is based on the customer. Taguchi did some great work on developing appropriate repsonse metrics; you may want to look into that approach.
    Evaluate what process parameter(s) cause you to have material that needs trimming. That could come from flow charts & fish-bone diagrams, or from design of experiments (if you have a good response metric).
     
    Even if this exceeds what your management says they want, there are major costs associated with a) a trimming process, and b) not giving customers what they require. You need to estimate those costs, so management knows how your team’s work is helping their bottom line.

    0
    #86782

    Mike B.
    Participant

    Tom,
    I don’t know anything about your process but a very simple tool that might help is a concentration diagram.  (Forgive me if I’m telling you something you already know.)  All you do is draw a picture of the part and then have the operator put symbols for the various defects on the drawing in the spot they occur.  After a while, if they concentrate in a particular spot, it might give you a clue as to what is happening.  For example, does the mold not close on one side or corner, does it happen with all products/raw materials?  A very simple tool but it works well.

    0
Viewing 15 posts - 1 through 15 (of 15 total)

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