iSixSigma

Design of Experiment

Six Sigma – iSixSigma Forums Old Forums General Design of Experiment

Viewing 59 posts - 1 through 59 (of 59 total)
  • Author
    Posts
  • #44979

    Raghavendra Kalmadi
    Participant

    Hi, I would like to know whether it is necessary for doing a design of experiment before embarking on the improve phase  of a project and can DOE be applied for people centric causes of a problem.

    0
    #145417

    Eric Maass
    Participant

    No, it is not necessary to do DOE before embarking on the improve phase of a project – DOE is something you CAN use but not something you MUST use – there are other ways to go through analyze and improve, depending on the problem and situation.Yes, DOE can be applied for people-centric causes of a problem. I remember when a DOE was done with one factor being the experience (months of the job), another factor being the size of their hands (which tended to have some correlation with male/female, so not a completely politically correct study), and another factor being the type of training they received.  The first two factors were statistically significant.

    0
    #145440

    Cherukara
    Participant

    I believe it is not necessary as such. There are many ways and means for you to embark or use on the phases. Sometimes the organizational culture and resources may paly a part in deciding which tools to use.
    I remember reading some DOE applications in people centric problems in the internet. One of it is in the form of an analogy of a girl and 3 man. Who is she fond of?
    DOE is fun and very rewarding tool as such the use of it would give great result and satisfaction too. Just my opinion.
    Cheers! Dominic.
     

    0
    #145441

    Social experimentation
    Member

    Unless the DOE is performed in the environment to which you want to generalize your findings, you’ll have to deal with the problem of internal vs. external validity.

    0
    #145442

    Cherukara
    Participant

    Certainly : I agree with you. Very much so. Of course I think you are refering to the Control and Noise factors. Or is it something else?
    Would you appreciate DOE away from the spreadsheet style everyone is so familiar with?
    Cheers! Dominic.

    0
    #145451

    Francois
    Participant

    Who in their right mind is going to stop production and start experimenting ?
    DOE is for research labs.

    0
    #145467

    Robert Butler
    Participant

      DOE is not just for research labs. There’s no need to stop production to start expeimenting.  I’ve run dozens of designs under these conditions.  What is required is a lot of up front planning before making the runs.

    0
    #145479

    Social experimentation
    Member

    Dominic,
    I am not quite sure if you agree with me. But I defnitely agree with Robert Butler, who has the most experience in this area on this site (as proven by his many excellent contributions).
    Internal vs. external validity refers to the fact that the experimental situation can alter quite dramatically the way people act and react. In experiments you deal with a controlled environment. If you drop a stone over and over again, the stone won’t mind and do what physical laws will tell it to do. In an social experiment there are limitations to what you can and ethically should do to your “subjects”. Thus, you have to be careful to ensure that the experimental situation mimicks as best possible the situation that you want to generalize to. Therefore, Robert again is correct that the best way to run an experiment is to not stop production but make it part of the production in a way that your “subjects”, i.e. employees if possible don’t even realize that they are part of an experimental run.
    With this in mind, replications become an issue. If, for example, figuring out a new solution is the focus of the experiment, and you give the same person the same problem over and over again, the fact that the person has identified the solutions changes the dynamic. In some instances replications in the modern experimental sense creates its own measurement error.
    The difference between control and noise is part of the Taguchi design. You are already brave enough to conduct a DOE. Taguchi designs are even more complicated, so I am not quite sure if you want to go this route.
    The one thing, I would like to emphasize is randomization. Every single “people oriented” DOE that I have been involved with lacked randomization. For some reason, even trained and experienced black belts forget about randomization as soon as they deal with people.
    Finally, I can tell you from experience that the explained variance of DOEs in social settings is “dismal” in comparison with manufacturing experiments. For some odd reason, human behavior creates more variance than widgets, and most of that variance ends up in the residual. Consider yourself lucky if any variables are truly significant (you’d be the first GB/BB/MBB that I know who actually could utilize the information. In most cases, there is major enthusiasm first, then come the results, and the report gets quietly filed into the bottom of the drawer). In any case, good luck

    0
    #145482

    Mikel
    Member

    It’s still early, but I think we have a winner for the dumb post of the day.

    0
    #145484

    Anonymous
    Guest

    Someone who ‘tweaks’ every run :-)

    0
    #145494

    Social experimentation
    Member

    Stan,
    Why don’t you post at least one post whereby which anyone could actually gage your level of knowledge. All of your posts are about dumb etc. … Doesn’t sound very knowledeable either, does it, old boy?

    0
    #145498

    Marlon Brando
    Participant

    I  agree  fully  with  you

    0
    #145503

    Mikel
    Member

    Social,
    You are just the latest Johnny Come Lately here. Do a search to see if I actually help people
    Your observation is like going to open house at Sandia Labs and concluding that the folks there sit around and drink punch and eat cake all the time.
    Show your own worth before starting to take swings at others.

    0
    #145505

    Mikel
    Member

    BTW, Francois’ post was dumb.

    0
    #145507

    Social experimentation
    Member

    Stan,
    You have either become so jaded in your little, isolated engineering cube that you have become socially totally illiterate, or you have regressed to the social dumbness of a stone. In either case, grab something for lunch and take a deep breath. In with the good, out with the bad. It looks like you’re exploding any minute … Not a good day little old Stanly, is it :-)))).

    0
    #145517

    Mikel
    Member

    Very astute observation. I would have been afraid of Dominic not knowing this based on his post.

    0
    #145519

    Social experimentation
    Member

    Congratulations, Stan, you just won AGAIN a gold medal in the Internet Special Olympics. Pad yourself on the back :-).

    0
    #145520

    Marlon Brando
    Participant

    Andy
    I  believe  that  you  are  a  real  expert  in  DOE.Can  you  please  submit  a  simple  direct  example ,or  just  guide  me (us)  to  any  article  or  link.I’m   really  confused  and  need  some  explanation.Have  read  some  examples  in  books  ,but  not  yet  understood  the  concept.Is  it  related  to  the  famous  equation
    Y = F(X)??

    0
    #145521

    Mikel
    Member

    Again, very astute answering Robert’s post and addressing it to Dominic. You just learn about this internal/external stuff lately and trying to impress someone.
    Your observations about Taguchi, randomization and social setting are dead wrong. Taguchi designs are not the least bit complicated and Taguchi doesn’t really worry about randomization – I wonder why? Experiments in social settings are carried out daily by some of the most progressive marketing organizations out there.
    Social – get out of the textbook experience and textbook advice and go get some real experience.

    0
    #145522

    Mikel
    Member

    That would be pat yourself, not pad yourself.

    0
    #145525

    Social experimentation
    Member

    At least you’re finally talking about what miffs you off.
    The randomization and Taguchi issue is clearly a product of your mind. I don’t know where you come up with that. The Taguchi designs have not been used in the social sciences, so that shows how limited your knowledge is. And yes, marketing researchers use DOEs and yes, I was part of the researcher group that first developed a simulation study based on the Jacoby matrix (field experiment -> laboratory experiment -> simulation) … GfK Nuermberg … so lol. This is an unpublished dissertation that has become proprietary knowledge of the company. I doubt that you have that level of credentials :-).
    The randomization is just an observation that I have had given that Black Belts don’t follow instructions very well (but with MBBs like you, I finally understand why some of them are the way they are: assumptions, assumptions, assumptions … lots of posturing and little to no substance).
    I am looking forward to further outbursts of your ignorance and inability to read posts and quote correctly.
    As we are on the topic: “You just lean” … learned????

    0
    #145526

    Anonymous
    Guest

    Marlon,
    Many DOE  experts visit this site … for instance, Robert Butler, and many others. I mention Robert because he answers most of the DOE questions.
    In mu opinion, one of the best ways to understand the concept is to find a simple example.
    If you contact me directly I’ll ‘walk’ you through a simple example with a ‘real’ interaction’ you can try on your office desk.
    Personally, I don’t like Y = F(X) because it misses two of the most important concepts of robust design.
    My email is a.urquhart(-)ntlworld.com
    Cheers,
    Andy

    0
    #145528

    Mikel
    Member

    Wow! You participated in a double secret DOE in a social setting that I am not good enough to learn about. Do I need a special decoder ring or something?
    And all this since you posted that DOE never worked in a social setting! Boy you work fast – I am certainly no match for you especially since you already have the decoder ring.
    And Taguchi designs have not been used in social sciences? What BS, you think you know that much about the entire universe of social sciences?

    0
    #145531

    Social Experiments
    Member

    Stan, good talking to you.

    0
    #145532

    Case250
    Participant

     
    What grade are you in?

    0
    #145535

    Mikel
    Member

    Hi Case.
    Mr. Social is confused and I’m just trying to help him out.
    First he give lectures on bad amnners and then exhibits them.
    He says that DOE doesn’t work in a social setting but then cites a double secret study that following some Jacoby nonsense that can’t be verified.
    Truth is I know my manners are bad with some. And truth is they are at theire worst when dealing with some blowhard that can’t even answer a simple question.
    The original poster asked if it was necessary to do a DOE as part of Improve (no) and if DOE could be applied to people centric projects (yes). All Social’s answer was about was trying to answer everything but that and giving really bad advice.
    What grade am I in? – I got kicked out for bad behavior. I beat up the star quarterback for being a pompous blowhard.

    0
    #145538

    Marlon Brando
    Participant

    Andy
    I  appreciate  greatly  your prompt response.I  will  contact  you later for  further  explanation.Thanks

    0
    #145539

    Social
    Member

    Stan,
    While you were wasting your’s and everybody else’s time grandstanding and impressing everyone with your bad manners, I just sent out another peer-reviewed paper for publication. Maybe you want to condense your tremendous knowledge and take on the true challenges. Right now you appear like a whiny little girl who has lost her bridges …yikes, not a pretty thought :-).

    0
    #145540

    Savage
    Participant

    I think it’s “britches” – or “breeches” depending on your tastes.  Not bridges.
     
     

    0
    #145548

    Francois
    Participant

    Stan gets my vote for the dumber than dumb award. 

    0
    #145549

    Francois
    Participant

    Perhaps pea brain Stan can explain how to do DOE without interrupting production ?

    0
    #145553

    Robert Butler
    Participant

      It’s done like this.  You pick your variables(let’s say conveyor speed, and temp) and identify their settings and you choose the physical properties of interest (let’s say viscosity of raw material type A) and you set up the design.
    Experiment  Speed  Temp  Viscosity
        1             Low     Low    High
        2             High     Low     Low
        3             Low     High     Low
        4             High     High     High
      You randomize the run order, you get your people together, give them walkie-talkies and lay out the plan for the days efforts which will go something like this:
     Let’s say you are going to run #3 first.  The guy/gal at the speed control station calls in and tells you the speed is now at the low setting, the guy/gal at the temp control station calls in and tells you when the temp is at the high setting, then you call the guy/gal at the raw material site and tell them to start feeding the low viscosity lot material.  They call back and tell you when that lot of raw material has entered the system.
      You and several others keep track of where that lot is in the system and you start taking samples at the far end before, during, and after you think the material has passed through.  You let the system go back to the “normal settings”, you wait for a wash out period, you take additional samples for control purposes and then you repeat the same process all over again for the next experiment. 
      At the end of the day you take your pile of tagged and marked samples (by run and by time) you test them, record their results, and run your analysis.

    0
    #145555

    Mikel
    Member

    What BS

    0
    #145558

    Francois
    Participant

    Robert,
    Poor fellow, you’ve obviously never worked in production.

    0
    #145559

    Ang
    Participant

    You sad little fellow Stan,
    I can see it now, meeting with Factory Manager and R&D Manager.  BB says:
    “I’ve just done a 4 week course !!!!     Look I have a certificate and everything.   And just look at the color of my belt !
    I’m going to take over production and do some experiments.  “
    Get real little man.
     

    0
    #145562

    K.O. Livin
    Participant

    Stan,
    You need to read the post again. I haven’t had to jump in and flag you when Reigle and Harris Moonies were back for quite a while. All you have to do is read the dribble on dropping the stone and the peer review publication. This smacks of the Harris camp. They have been quiet for some time but the interaction of getting expelled from ASU and getting kicked around last week on the forum I am sure turned over whatever rock Reigle was hiding under.
    K.O. Livin

    0
    #145575

    Mikel
    Member

    My note to Francois was spot on. Look at his note to Robert Butler.

    0
    #145576

    Mikel
    Member

    Morning Franny,
    Let’s start with how stupid the stipulation of not interupting production is. Whoever says that is the criteria is wrong. There are occasions where we would not interupt to learn, but they are the exception.
    You want to put a name on how to do DOE’s without interupting production? It goes by the name of EVOP and Simplex to name a few.

    0
    #145577

    Mikel
    Member

    Just checked and my bridge is still there.

    0
    #145580

    Robert Butler
    Participant

      I’m not offering theory here Francois, I’m offering first hand experience in the production environment. The post concerning running a design while permitting production to continue is a quick summation of how I’ve dealt with that problem time and again. (Not that it matters but we are talking “time and again” as in “time and again” over a 25+ year career as an industrial statistician) 
      Under these circumstances what is extremely difficult is the identification of the acceptable ranges of the variables.  Essentially what one is doing is the “tweaking” mentioned by Andy U, the big difference being that you are doing “tweaking” in a very organized manner.  In order for this to be of any value you have to have accurate information concerning historical minima and maxima within the production environment.  Far too often when this data is requested it is found to be nowhere in evidence. 
      When this occurs, I shift the focus from thinking about a design to gathering accurate information on the existing process in order to find out where you are as opposed to where you think you are. 

    0
    #145582

    Brit
    Participant

    Not sure why you guys are so bent on thinking you can’t run a DOE on a producing process.  I have done it in steel mills (2), an alloy facility, a galsses manufacturer, and chemical plants (3).  It’s a little tougher for a continuous process, but it can be done without interrupting production. You simply have to be careful on the extent to which you make changes.  This is done all the time.    

    0
    #145593

    Jered Horn
    Participant

    …and you can add rubber and plastics mixing, rubber and plastics extruding, rubber and plastics calendering, rubber and plastics molding, steel forging, grinding, automated paint lines, robotic welding, multiple types of finishing operations, fiber optic glass manufacturing…All first hand experience, btw.
    I think the list of manufacturing processes where you can NOT run a DOE (without interrupting production) would be shorter.
    Sounds like some people need some practice thinking outside the box.

    0
    #145657

    thandi
    Participant

    Send the SS jackass’s ANOVA to the R&D labs.
    Use ANOM.

    0
    #146136

    Jonathon Andell
    Participant

    I’m neither a Taguchi “worshipper” nor a Taguchi “basher,” so I have the opportunity to p*** off both camps with what I’m about to say.I have yet to observe a major difference between the success of a Taguchi vs. a classical matrix in and of itself. I have seen vast differences in the success of both kinds of DOE, related to how well the leader prepares the team before designing and executing the experiment. A well-prepared experiment is highly robust; no choice of matrix can rescue a badly prepared experiment.I have beaten my brains in trying to randomize easily 3 dozen DOE’s, and I have yet to detect a single instance where the practice has revealed some heretofore hidden source of time-dependent variation. I think it’s because of the preparation issue. Nowadays, I randomize when it’s convenient, but I don’t knock myself out (partially because the brain-beating has started taking its toll, but don’t tell Rush Limbaugh).I have done some inner-outer array runs. They don’t create analytical complexity, but they significantly increase the baby-sitting needed to run the experiment correctly. They are worth doing in some situations, but definitely not every situation.The purists can identify statistical flaws with some of Taguchi’s arrays. Personally, if they get me the information I seek in the fewest runs, I am OK using them. I also am OK choosing another matrix when necessary.Now that I’ve had my say, feel free to attack my competence and integrity.

    0
    #146144

    Robert Butler
    Participant

      I think the problem here is that you are looking in the wrong direction for the wrong thing which, based on my understanding of what you wrote you couldn’t see anyway. 
    You said, ” I have beaten my brains in trying to randomize easily 3 dozen DOE’s, and I have yet to detect a single instance where the practice has revealed some heretofore hidden source of time-dependent variation.” 
    As I see them, these are your possibilities:
    1. You randomized your design – by definition you won’t see any time trends – the effect of any lurking variable will just show up as an increase in the residual variability.
    2. You didn’t randomize your design and a lurking variable “hitched a ride” with one of your controlled variables.  The lurking variable is perfectly aliased with the effect of your controlled variable – you won’t see anything here either. Depending on how it varied with respect to your controlled variable the lurking variable could either make your controlled variable highly significant, highly insignificant, or not do anything at all.  What you WILL see is, when you try to use your resulting equations for improvement/control the variables won’t do what you said they would do to your process when you change them in the directions indicated by your correlation equation.
    3. You didn’t randomize your design and your lurking variable moved around a bit and hitched a ride first with one controlled variable and then another.  Same as #2 above only now we are impacting (or not) the significance of more than one controlled variable.
      If you ran a complete design (randomized or not) in some period of time and if you then replicated that design at some later period of time your best bet for checking a time trend would be to difference the experiments and their replicates and see if you have either a constant offset or an offset in the same direction (i.e. all of the experiments in the first run are numerically lower/higher than all of the experiments in the second run).
      At the risk of hijacking this thread/thought I’d like to offer the following:  For whatever reason, when a statistician says the word “randomize” it seem everyone immediately assumes he/she means complete randomization and thus:
     1 Anything less won’t/can’t be tolerated
     2. Anything less is a complete waste of time
     3 If this assumed complete randomization can’t be achieved then there really isn’t any point in bothering with randomization at all.
     Neither 1,2, nor 3 have any connection with the facts of experimentation.  Offhand I couldn’t tell you the percentage of designs I’ve run where I had only partial randomization vs. total randomization but I suspect partial would far exceed total. 
      The key to all of this is, as you noted, the preparation and execution.  In those cases where it simply isn’t possible to randomize in any direction (and I’ve had a few like this) or randomization for many of the believed to be important variables is extremely difficult the way you introduce “randomization” is to run the design and then pick some subset of the design and run it again (the formal name is partial block replication) and analyze everything as though it was a single design.
    …as far as p**** me off is concerned….you’re not even close!  :-)

    0
    #146414

    Jonathon Andell
    Participant

    I think we are 99% in agreement, if not more. I want to clarify: the 3 dozen DOE’s to which I refer WERE randomized, and none revealed a time-dependent lurking X. I used to insist fanatically on randomizing rigorously, and now I have a mellower outlook.If randomizing is straightforward, there’s no harm and possibly a benefit, but fanatical insistence on rigorous randomization can be impractical. I think your post supports that belief.In my experience, both in conducting my own DOE’s and in observing others’, I have seen very few instances where randomizing actually exposed a lurking X. I have seen a bundle of experiments fail because of poor preparation. Again, I think we are in agreement there.

    0
    #146430

    Robert Butler
    Participant

      No, I don’t think we are.  The reason for randomizing is to avoid the impact of a lurking variable – specifically avoid its being aliased with one of the variables you are controlling.  If you randomize you won’t see the effect of a lurking variable and there isn’t any way to plot or analyze your randomized runs to make it appear. 
      As I said, if it is there and if you have randomized, then its effect will show up in the residual error but it will not impact your conclusions concerning the value or lack thereof of the variables you did control. If you want to test for the presence of a lurking variable/time effect you will need to run genuine replicates of the same experiment during the time you are running your design. 
      For example, let’s say you were running a 2**3 design, you could completely randomize your design, you could run 3 experiments per day, and there was a significant lurking variable that increased your response over time.  On day 1 you ran 3 experiments.  On day 2 you ran two more experiments from the design and you replicated experiment #1 from the first day.  On day 3 you ran two more from the design and replicated #1 from the first day and on day 4 you ran the final point and replicated experiment #1 from the first day. To see the impact of the lurking variable you would plot the results of experiment #1 by day.
      I cringe every time I hear the phrase “If randomizing is straightforward” …because it is usually a prelude to justifying no randomization.  I can only offer personal experience in this regard but in my experience, failure to live with the “inconvenience” of the effort needed to inject randomness into a study and hence to randomize has been the source of more failed/botched studies than I like to think about.
      A very common study killer (which I’ve seen more times than I like to think about) is the following:
      We didn’t randomize
     and
     some distance into the study:
      the
    key test instrument/facility/operator
    failed/burned down/resigned  
     and the
    new instrument/facility/operator
    is
    different/doesn’t do it quite like we want/is unfamiliar with our methods.
    …and there we are – a significant “lurking” variable – the money/time is spent, we can’t easily go back and repeat anything, and now we have to face the fact that any significant effect our analysis is attributing to one or more variables may not have anything to do with the variables at all. 
     The amount of time/energy that will be wasted attempting to “fix” this failure to randomize will be out of all proportion to the cost we would have incurred had we spent the time and money to build in randomization to begin with. 
    The other big killer:
     We didn’t randomize and an unknown (but much later discovered) lurking variable was aliased with a study variable.  That study variable “test” out to be very significant.  We set up the correlations – confirm the settings of the significant variable – start running the process – everything looks great – the lurking variable shifts the other way – the process falls apart – we adjust the “critical variable” to compensate and, because it wasn’t the critical variable – nothing happens …and all of our effort and what we thought we knew about the process are for naught.
     
    As an industrial statistician I spend a lot of time with my engineers and the people involved in the process in order to understand the system so I can provide ways to randomize which won’t require the complete rebuilding of the production line but which will insure the integrity of the effort. I suppose my efforts to always find a way to randomize could be viewed as fanatical but in my experience it is far easier than many would have you believe.  
     
    For me, the real fanaticism seems to come from managers/engineers who view randomization as some kind of impossible dream. Time and again I’ve been told by the senior engineers that “everyone knows”  it’s impossible, you “can’t do that kind of randomization” and time and again when I’ve gone out on the line and explained my needs to the operators I find they are able to give me a half a dozen easy ways to do exactly what I need for “that kind of randomization.”
     

    0
    #146539

    Jonathon Andell
    Participant

    The classes I took on DOE suggested that a “lurking time dependent factor” could be exposed by plotting either the results or the residuals in the order in which the randomized experiment was run. Presumably a non-random pattern would emerge, and that pattern would lend insight into what the variable might be.If you are saying that this outcome does not happen, then I am learning something new. I certainly can say that, even in my rigorously randomized experiments, I never saw a single non-random pattern emerge in this manner.As for straightforward randomization, I have dealt with experimental X’s that truly required an entire day to convert from plus to minus, or vice versa, along with variables that could be switched easily. If the blocking options did not accomodate the number of factors that were slow or complex to change, then I randomized within those variables.I still feel comfotable saying that my experiments were successful, because my capping runs constantly verified what the data suggested they would do. Those who failed to confirm on capping runs frequently randomized, but they failed to do their preparation in the other ways that are important.Where I think you and I agree, is that we don’t jump into an experiment without a lot of up-front prep. We agree that it’s better to invest in that prep than to try to “rectify” after the data is in hand.

    0
    #146542

    Anonymous
    Guest

    Jonothon,
    I’m not aware of ‘capping runs’ in the literature of ‘classical statistical’ method, unless it’s been added recently. I mean .. does Minitab provide a means of ‘self-testing’ or a capping run? I’m not aware it does.
    Perhaps someone will enlighten me.
    Cheers,
    Andy

    0
    #146547

    Jonathon Andell
    Participant

    I don’t know whether capping runs appear in classical literature (I am quite confident you won’t find the term in anything Shakespear wrote), but a mentor told me that it’s always a good idea. After analyzing data, it’s advisable to run the process at settings we expect to produce “bad” and “good” outputs, to validate that the DOE did indeed find the right X’s and the right settings. It’s really the acid test of whether the DOE got the job done.

    0
    #146548

    Robert Butler
    Participant

    Ah Ha!  What we have here is a failure to communicate!  :-)  If we go back to your original statement “I have beaten my brains in trying to randomize easily 3 dozen DOE’s, and I have yet to detect a single instance where the practice has revealed some heretofore hidden source of time-dependent variation.”  I took this to mean you were trying to see time trends and lurking variables by plotting the raw results of the experiments and all of my subsequent discussion was based on this premise. However, your 4 November post suggests you were actually talking about a residual analysis.  So, let’s start again.
      Let’s assume you randomize and:
      a. Your residual plots against predicted and your residual plots against time don’t show any trends. 
      b. Your residual plots against time show a trend of some type.
      c. Your residual plots against predicted show a trend of some type. 
      In the case of a) your analysis would suggest that for the time and conditions over which the experiment was run there were no lurking variables and there was no impact of time.
      In the case of b) time has an effect but, because you randomized and ran a design, you can still talk about the effects of the variables in your study independent of the effect of time.
      In the case of c) there was some lurking variable which had an effect but, again because you randomized and ran a design, you can still talk about the effects of the variables in your study independent of the effect of that lurking variable.
      If we assume you didn’t randomize and there was a lurking variable/time effect present then we have the following possibilities:
    1. The lurking variable/time effect was not confounded with any of the variables in your design.  In this case when you run your residual analysis the lurking variable/time effect will make its presence known as a trend in one of your residual plots and you can discuss the effects of the variables in your analysis independent of this lurking variable/time trend.
    2. The lurking variable/time effect was confounded with one of the variables in your design. It will not show up as a pattern in any residual plot, you won’t know that it is actually part of one of and you won’t know it was there until things start to go wrong.
      The fact that you have taken the time to do a lot of randomization and that your subsequent residual analysis has not indicated the presence of some additional lurking variable or time trend is great but the fact that your precautions didn’t result in the discovery of a lurking effect doesn’t justify not randomizing.
      Other than this I concur – we are in very good agreement with respect to the issues surrounding up front preparation.  As for “capping runs” I assume you mean confirmation runs.  This is, as far as I know, standard practice.  You would be taking a big risk to just apply a regression equation to a process without first confirming that it is a reasonable mathematical description of the way your process behaves – This goes back to the heart of the saying – “all equations are wrong and some are useful.”
     

    0
    #146554

    Jonathon Andell
    Participant

    I had a feeling we were approaching a state of violent agreement :-) I have read some of your posts on prior discussion threads, and I was confident that you and I have similar viewpoints on DOE. You described three hypothetical outcomes after a randomized experiment: “a. Your residual plots against predicted and your residual plots against time don’t show any trends. “b. Your residual plots against time show a trend of some type. “c. Your residual plots against predicted show a trend of some type.”My experience has been almost entirely outcome “a.” Clearly, my very next experiment could yield either outcome “b” or outcome “c,” so I remain willing to randomize.I hope you didn’t regard my posts as taking an anti-randomization stance. I just feel that one can blow things out of proportion, and add significant time and cost to an experiment, with an excessively dogmatic insistence on randomization. With a little practical flexibility one can get 90% of the benefit of randomization, and often trim the experiment’s cost by 50% or more.
    We also agree regarding confirmation runs, our best protection against the “gotcha” outcome. When I teach the benefits of confirmation runs, I quote the universal law of throwing a Frisbee: “Never precede an action with a statement more predictive than, ‘watch this!'” Until the capping run is finished, we just don’t know where the durn thing is going to fly. (Badly prepared experiments are at greatest risk of surprises during capping runs, as you already know.)

    0
    #146555

    Cynic
    Participant

    Stan,
    are you off of your medications?  Try answering the questions without the attitude.  Nobody needs it.

    0
    #146556

    Cynic
    Participant

    Stan,
    as usual, your response is full of arrogance.  And in this case it seems from the perspective of someone who has never lived in the real world.  There are cases for interrupting production immediately (as with empowered operators who can stop a line to 5-why and fix a problem) and there are absolutely cases when a criteria for DOE is that production is not interrupted.
    What a stupid response your last post is – I hate to sound so negative but I have a repulsion to your attitude.

    0
    #146564

    Anonymous
    Guest

    My point precisely ..
    “Though this be madness yet there is method in it.”
    Cheers,
    Andy
    PS: There is more to Taguchi than meets the eye!

    0
    #146576

    Robert Butler
    Participant

      One of the big problems with posts is that all you have in front of you is the typed word and with that as your only source of feedback/information/communication it is far too easy to infer that which isn’t there.  My take on this exchange was that you and I were discussing the pros and cons of the issues surrounding randomizing both from the standpoint of theory and the standpoint of necessity.  I thought from the beginning we had more to agree about than disagree.
      The only reason for offering an extended discourse on randomization is because I’ve had to contend with this problem many times. I hoped by laying out all of the issues concerning the risks involved our discussion would provide some clarity to what can be a confusing and irritating issue. To that end I also tried to make it clear that in most of my work partial randomization is far more likely than complete randomization and I tried to indicate (partial block replication) how to get randomization even when it is impossible to change anything while running your study.  Again my only worry is that it is very easy to view randomization as a pest (which it is) and then decide not to expend any effort to try to inject it into the work.  
      About the only other puzzle to resolve in looking over the posts to this topic this morning are those from Cynic.  I can’t decide if he/she is actually referring to some of Stan’s earlier posts or if he/she is referring to what I wrote and confusing posters.  If the reference is to what I’ve posted in the past few days – please accept my apologies- I wasn’t trying to be rude and I didn’t think I was giving offence.

    0
    #146577

    Jonathon Andell
    Participant

    I’m sure Cynic can speak for himself/herself. However, my read is that the only thing offending Cynic seems to be one respondent’s selection of offensive and insulting tones, and you are not that respondent. Even when you thought we disagreed, your posts were polite and stuck to the facts of the situation.I have seen situations when interrupting the process was unnecessary as you said. However, my mentors always advocated setting the plus and minus levels of the X’s as far apart as safe. Sometimes those settings put the resultant product at risk of being unacceptable. For those cases, it should be acceptable to interrupt production.An example was a plastics injection molding place, where we were working to reduce an out-of-round condition. The “traditional” (that is, theory X in the extreme) management had shut down production while they tried the dartboard approach. SOmehow their shut-down was OK with them, because they expected that at any moment the parts would magically start coming out perfect. However, another day of shut-down for a systematic DOE was an unacceptable cost to them. After blowing two weeks of down time, their bosses imposed a DOE on them, and we solved the problem in a day. By the way, we did a semi-randomization, due to the long times it took to stabilize at new temperature settings.It was fun to do the capping run. We operated at the “bad” setting, and the dysfunctional managers used the bad data to attack me. Then we switched to the good settings and turned out the best parts they had ever produced.

    0
    #146586

    Mikel
    Member

    Robert I am sure they were referring to me.
    As I have pointed out before you, Andy, Barry and others are on the high road and trying to educate. I personally don’t think a lot of the people you help are worth it – but that’s my opinion and I admire the resilience you display with trying to teach.
    Your exchange with Jonathan was good, just like what I’ve seen with Andy and Barry. You don’t have to agree to have a good exchange of ideas. I believe Jonathan is another one worth having good exchanges with.
    Keep up the nice help and if anyone goes after you for the help you provide, you can rest assured I’ll divert their attention – somebody has got to be the bad guy.
    By the way, I threw out the idea of Taguchi not really worrying about randomization – just look at the layout of his designs column one is for the hardest to change, column two for the second hardest – you don’t do that if you intend to go randomize. Why do you think that is the case?

    0
    #155429

    Allthingsidioto
    Participant

    I guess that  you  have  to  find  figures (results) for  each  experiment in order  to  optimize the  given  factors (what  to  increase  or  decrease),then  you  may  out  to  study  the  interaction and  to  find  the  appropriate  confounding.You  may  want  to use (+) sign for  high  and (-) for  low,you  may  wish  also  to  extend your  experiments  to 8 or  even  to  16,to  come  out  with  the  proper  conclusion.
    Thanks  for  the  enlightenment
    Just  my  opinion,regards 

    0
Viewing 59 posts - 1 through 59 (of 59 total)

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