JMP DoE analysis with repeated run in either block

[#1]

Hello

I have a question in relation to DoE analysis in JMP. Lets say, in one of the blocks, one run did not go well and it is repeated in the second block. Could anyone provide any potential pointers in this scenario how to adjust the entire run for the further DoE analysis? Which block parking the repeated run would make more sense for DoE analysis in JMP and what would be pro and cons if its is parked in either of the blocks?

[#2]

I understand the situation and your question but there is not enough detail for me to give a definitive answer.

Did you try to analyze the data yet? The designed experiment is just a way to get the best data for the regression analysis. It is usually not fatal to the analysis if you lose a run or you cannot achieve an exact level for a factor. You are using regression analysis, which can be used with any data, even if it was not collected with an experiment. Of course, there are requirements and assumptions of regression, but you should try it and see what happens.

[#3]

Thank you very much to everyone for their input. We are into generating the complete dataset before we do JMP analysis.  However, it was great to know various options to approach the issue-will be in touch!

[#4]

Hi @ProbitStarfish1 ,

(Interesting user name )

I was just responding and noticed @Mark_Bailey had got there before me. I agree with Mark.

My understanding is that you have a run where the response result is questionable. But you have repeated the same factors settings in a run in the second block. That is fortunate!

My advice would be to analyse with the "dodgy" run. And then analyse with the dodgy run "excluded". 

How does the model change? Is it a big difference?

This can then inform your decision as to whether you should exclude the dodgy run for your ultimate analysis.

Obviously, missing runs from a design is never ideal. It will have "unbalanced" your design to some extent. And if this is a DSD, you will no longer be able to use the Fit DSD platform because the foldover structure will have been spoiled (unless it was a centre point). Like Mark says, we would need to know more to know the exact implications.

I hope this helps,

Phil

[#5]

You have already received some good ideas from very experienced practitioners.  I have some additional thoughts:

1. First, why did it not go well?  What did you learn?  Were the conditions that created this "event" understood?

2. Can you get an estimate of the block effect?  Do you have other "replicated" runs in the second block that were identical in the first block?  If so, you may be able to account for the block effect in the missing data point.

3. Here are some things to try: Use the mean of the data to replace the missing data point and run the analysis, then try your predicted value for that treatment (this assumes you predicted results). Lastly, Remove the highest order effect from the saturated model and regress on the remaining data. Go to the red triangle options for the response >Save Columns>Prediction Formula.  This will use the remaining degrees of freedom to predict what the missing value would be.  Use that and perform the analysis.  

In any case, don't accept one attempt replacement technique.  Do multiple replacements and look at the results.  If the results of the analysis are similar, you are probably OK, if they are different, then you'll have to think about collecting more data.

---