DOE Screening

[#1]

I have a 5 Factor Placket Burman Screeening DOE. After the runs, it turned out that 2 of the runs are absolutely not good for the design and therefore should be removed before analysis. I do have results for the rest of the other runs. Could you help me with how to change my already generated design table so that I can get a correct analysis in the statistical analysis. Do I just click the undesired rows and click hide and exclude before I go on to evaluate design?

Thank you!

[#2 · Accepted solution]

First a question, when you wrote, "...2 of the runs are absolutely not good for the design..." what does this actually mean practically? No response? Nonsense response? Couldn't actually run the treatment combination?

I just hope you aren't throwing out responses because, well, they don't fit an urban legend/myth of what the response should really be? Sometimes the real gold in an experiment is in these responses. So if you dead set on eliminating the treatment combinations from the analysis the hide/exclude route is a good way to go. You might also want to put a note in the Column Properties around why you enabled hide/exclude certain rows.

[#3]

Thank you very much for your response. The target parameters cannot be measured in those two runs, which means there wont be any values in the repective DOE columns. You have answered my question and thank you for the other 'nuggets' of information!

[#4]

(A bit of serious humor on this general topic:)

Now that the problem is solved, I'd like to introduce a movie scene that we could be turned into a teachable moment for DOE screening methods.

In 2021, a movie titled The Tomorrow War premiered on Amazon Prime. About that movie:

Time travelers arrive from 2051 to deliver an urgent message: 30 years in the future mankind is losing a war against a deadly alien species. The only hope for survival is for soldiers and civilians to be transported to the future and join the fight. Determined to save the world for his daughter, Dan Forester teams up with a brilliant scientist and his estranged father to rewrite the planet’s fate.

Having watched the movie enough times to start thinking about other things at the same time, I noticed that at time 01:19:02 into the movie there is a discussion of experiments needed to identify an agent within the (female) invading aliens, that neutralizes a toxin designed to kill the creatures.

As the lab scenes unfold, we can see from the characters' reasoning and from computer displays that DOE is not being employed – though one can see how the dramatic requirements of the script might require this. The scientist on-screen speaks of thousands of factors, and of creating hundreds of variations of test sera:

...to make thousands of tests. With any luck, we should have a working female toxin in the morning."

Comments, or reverse engineering ideas, anyone?

[#5]

Hmmmm...if the scientists have a useable model, first principles, or empirical, then I'd head straight to the Prediction Profiler and simulation capabilities. Lacking a model...to borrow from another space related event..."Houston, we have a problem here."

[#6]

As is possible in Hollywood, over sensationalizing the problem.  One important question; Do the scientists want to understand the causal relationships between the treatments and the effects?  Or are they trying to "pick the winner" (this is my guess)?  If you are actually trying to understand causal relationships, this "requires" application of scientific method, which is an iterative process.  Start with hypotheses as to what treatments (e.g., toxins, amount/ratio, method of treatment, etc.) will have an effect on the survival of the species and why.  Identify what noise is in the system (e.g., environmental, age of alien, alien-to-alien). Design experiments to understand those relationships across changing noise variables using strategies like blocking or split-plots.  Likely start with screening and then proceed to modeling.  Oh, and it's not luck.

Or you could start with sampling, partitioning the components of variation into chunks of x's and determining leverage, vs. specific x's in an experiment.  There is no limit to the number of variables that can be evaluated using this approach.