DOE with Attribute Response

We has 2 options to do :

1. Try to convert the response to be discrete or continuous response. You have to define your response related the attribute resonse such as size of the defect, amount of defect, severity or etc. If none are applicable, most of the time we still have likert scale as the last choice.

2. Use attribute response. But it requires resource a lot.

My experience, I always use choice No. 1.

Jaran S.

Thanks for your response !

This DOE is about defective gas detectors , so my response is detected / not detected .......

Ovi

Hi,

          You have to take a transformations on the response data if it is of attribute type, depending on the distribution. May be in your case it may follows poission distribution. Take a transformation do the similar analysis as you will for continuous response.

      You can look into minitab for this. Are you can go through this website which have some questions and answers on DOE.

http://www.minitab.com/support/answers/index.asp?topic=DOE

hope this helps you

sridhar

Dear Ovidu,

You said that your response is gas detection right?.Some questions to probe.

1. Objective is to detect the gas leak right

2. Objective is to detect it fast  before a disaster occurs right

3. If this is the primary objective of your device the response should be the time in detection of gas leak by the instrument.

Hope this helps.

Shree Phadnis

Thanks , this is a really good idea, and something that definetly to take into account !

Ovi

Dear Ovi,

you are welcome.If you need any further help do not hesitate.

Shree Phadnia

I think that the suggestions given above are all good, but the technically correct tool to use here is logistic regression.  Minitab does a pretty good job of simplifying this tool and explaining how to use it.

The problem with a transformation approach is that you lose the power of the sample size.

I have had to perform DOE using attribute response for refrigeration systems leaks (defects). The response measured was leak / no leak. Since any leak would result in a defective product to the customer there was no allowance for  transforming the data top variable, including the application of a Likert scale.                                                    Transforming the data to a variable response may be applicable if the issue is how long can the gas (explosive or not) leak prior top failure of a unit or a disaster (explosion, toxic accumulation, etc.)                        If the real issue is leak / no leak I recommend that the data be examined as attribute - perform a regression analysis to determine the KPIV's and procedd from there to determine elimination / control of inputs.                                                                                                  Good luck on your DOE

Thank you !

Ovi

Hi Ovidiu Contras

If you have attribute data - what type you have ? Defective / Not Defective type or count of defects ? If it is defective / not defective kind of data do transformation of response using arcsin  SQRT p ( where p = proportion of defective ) & if it is defect count do transformation of response using SQRT (c) - where c is defect count.

Earlier one is binomial distribution & later one follows poisson distribution.

Let me know whether it's clear or not.

Thanks

Alek

can anyone help me,

is there any basic theory about DOE Attribute?

thank you very much

Diana,

   The paper "Analysis of Factorial Experiments with Defects or Defectives as the Response" by S. Gisgaard, and H.T. Fuller, Quality Engineering, 7(2), 429-443, 1994-95 contains a couple examples.

Where can I get a copy of the article "Analysis of Factorial Experiments with defects or Defects as the Response" by S. Gisgaard and H. T. Fuller?

Quality Engineering 7(2), p. 429-443 (1994-1995)

I've been trying to search for that on the web with no luck.  Is there a specific website you can give me to get the article? 

Qualtiy Engineering is a journal.  You should be able to get a copy from library.  If not, email me aldous.g.wongesso.ca your address.  I mail you a copy.

(1) Copy of the article can be obtained from

http://www.dekker.com

(2) Transformation from attribute data, e.g. defective rate (i.e. binomial variate) and defect count (i.e. poisson variate) to continuous variable, see BH2, Statistics for Experimenters by Box, Hunter and Hunter

(3) Quality Progress, March 2004, page 66 "3.4 per million" has another example of Attribute DOE

(4) Based on my experience, unless the range of the untransformed data (Y) is very large, (e.g. max Y >> min Y), conclusion from the transformed data is often the same as the conclusion from the untransformed data.

Aldous Wong  

  

Alek:

  Can you post a sample on how to transformation of attribute response?

  Tks! 

Dear Tony ,

Sorry for delay in response. I was on tour etc. Anyway , the way you can transform the data I 've explained in my original post on the subject above. So now I'm not explaining that & rather I'll answer specific to your question.. Say, in a DOE you have the response as a defect count in invoices.  Now for each set of run (you have to the choose the facors & level setting ) say you are taking 10 Invoices each & counting the no of defects as response. Now while dealing with attribute data , the assumtions of DOE are generally violated & hence you have to transform this data. Now in this example your data will follow poisson distribution. Hence you can use the transformation formulae what I have already posted in my earlier post on this subject.

Thanks

Hi, everybody!, do you all experienced with mixed (attribute and continue) response of DOE, especially in Taguchi method?? I really need more information for this kind of problem, any advices or suggestions or responses for this matter would be appreciated. Thanks B4..(Pratama)

Hi ,everybody

what's likert scale? where will i find the relative content ..

Thanks for you again .

A Likert scale, ironically, was right at the bottom of the page - how do you rate this article - a 1-5 scale.  Likert scales are discreet variables that are not continuous.

Hi There

If you can call me at 954-394-0773, I have some questions in regards of design review.

I am looking forward.

Best regards.
Mirza