Using Response Optimizer to find the optimum parameter setting - Minitab

[tejaswani]

Hi I had a problem with minitab. I want to use Response Optimizer to find the optimum parameter setting and I want to use one Text variable in it and a text result or an absolute value should be present in the result instead of an integer value. So, help me in solving this problem.

[Statistical Steven]

In Reply to Parent Post by tejaswani

Hi I had a problem with minitab. I want to use Response Optimizer to find the optimum parameter setting and I want to use one Text variable in it and a text result or an absolute value should be present in the result instead of an integer value. So, help me in solving this problem.

In general the response optimizer (based on Response Surface Methodology) requires all independent factors to be continuous. If you have categorical variables, you can usually optimize for each level of the categorical variable then build a covariate model. I do not think you can do it in Minitab directly, but I am not a Minitab expert.

[Allattar]

Not sure I fully understand the problem.
Here goes anyway, if the output is categorical, then you cannot analyse for this in the ANOVA that the analyse factorial design tool is using.
You can analyse for the categorical output using one of the logistic regression tools in the regression menu. There is a problem here in that this model is not stored for the response optimiser and wont be included. The response optimiser only knows about responses analysed with the analyse tools in DOE.

You can store the calculated event probabilities from the logistic regression back in the worksheet, and I suppose you could then put those event probabilities in the analysis, but, I cant see it being a particularly valid analysis as your mixing probabilities and responses. You'll end up with problems where probabilities are close to 1 or 0 with any fit from an Anova or regression.
(not the best explanation but it will do)

The other issue it could be here is that the response optimiser does not work for General full factorial designs.

2 level factorial designs, mixture, response surface, all fit a regression model, and that model is used by the optimiser.

In a general full factorial you are only finding the average response of the level, hence as your only dealing with average of the level even for continous factors the response optimiser, and indeed any of the tools that rely on a fitted regression are not available.

So if your using a general full factorial design for categorical inputs you cannot use the response optimiser. A note is you can use categorical inputs for a 2 level factorial design, but it does fit a regression using coded values of -1 and +1 and it will allow the use of the optimiser. 2 levels is easy, you can fit a straight line to 2 levels.

General full factorial designs though cannot aassume a linear fit through different categorical inputs.

Hope that covers at least some of what you where asking

[Juan Dude]

Also not sure I fully understand the problem but I ran into the following problem once while doing a DOE so here it goes.

If what you are saying is you want to do a DOE with attribute data instead of variable data then you have to "convert" the attribute into variable, i.e. pass = 1, fail = 0 and then you will be able to do your DOE analysis.

Hope it helps.

[Allattar]

The problem you get with that solution Juan is that the residual error may not be normally distributed.

The analysis is run using ANOVA, and fitted with a regression. These tools are fitting assuming that the average of a point, or the fitted line describes a mean, with individual measurements scattered around that with a normal distribution.

Translating results into 1 or 0 will not give you that shape, hence the mean, or line is unlikely to positioned correctly. Leading to a result that does not accurately inform you of the true result.

Also another one to beware of when using pass/fail data is do make sure you have enough samples tested at each run of the DOE.

Lets say when combination of factors (A)has a 10 percent chance of failure.
Also another combination has (B) 80% chance of failure.
If you test only one sample on each run its possible that you see a failure for run A but no failures for run B.
You will need, using a power and sample size on a 2 proportion test, 9 samples in each run for 90% power of seeing a difference between run A and B assuming A has a prob of .1 and B of .8.

This of course requires more samples should the difference between any two runs be smaller.

Your looking at around 500 samples in each run to see a difference between 2% failure and 6%.
(assuming 2 sided rather than one sided for alternative)

Now saying all that you should be able to get away with using a count of failures as a continous variable if there are enough observed failures to be able to aproximate a normal distribution with the proportion. Typically you want to make sure you are seeing more than 5 defective items in each run or it certainly wont be suitable to use a normal approximation.