How to fix a broken experiment done in 3 blocks?

MJZ82

I have a DOE where I've blocked runs by month - I can run 4 runs/month, for a total of 12 runs over 3 months. 

Is there a way to salvage my design if 1 of the runs in a block fails due to outside circumstances unrelated to the DOE factors?

For example, run number 4 in my 1st block of experiments had to be stopped half way through due to a raw material supplier issue, so that block only has 3 runs completed. Is there a way to re-tool the DOE so the 1st block has only 3 runs, the second block has 4 runs, and the third block has 4 runs, or should I just start over (would set me back a month or so)?

What if I can run a 3 run block, a 5 run block, then a 4 run block instead? Just curious what people would do in this situation, and how do I leverage data from those 3 completed runs without having to start from scratch- should I look at paring down the number of factors, and just do blocks of 3 runs with less factors etc...? 

statman

I wouldn't say the experiment is "broken". There is still opportunity to learn with the remaining data.

It is extremely difficult to provide specific useful advice with the amount of context provided. Can you provide the experiment data table? It seems that you allocated 4 treatments to each block, what are the factors associated with the block (e.g., what is the noise)? Have you done any predictions? What is the predicted rank order of model effects? What is the predicted size of the block effect? What is the resolution of the design structure? etc.

There are several options to "replace" the missing data, but there are restrictions on how useful those would be given the design structure. Here are some replacement options:

1. Use the mean of the response from the remains 11 treatments (or use the mean from the 3 treatments within that block). This tends to reduce the impact of that treatment.

2. Use your predicted values (this assumes you did predictions before running the experiment).

3. Regress on the remaining 11 treatments leaving 1 DF out of the intended model (usually the highest order effect or the effect predicted to be the smallest). Save the prediction formula and the predicted value for the missing treatment will be completed in the JMP table. Note: You also might want to do a quick look at the size of the block effect and adjust values to compensate for block effect before doing that prediction formula.

4. Do all of the above, analyze the results and determine how much does that one treatment impact results. If there is general agreement, you might conclude that lost treatment isn't a big deal. If there is disagreement, then you will likely need more data. So the question is do you re-run that one missing treatment? Should you run in the same design space (or in the projected design space)? If you re-run it, you introduce a fourth block effect, so should you also re-run another treatment to "calibrate" block effect?