How do I find the best design for 20+ factorsHi all,
I am very new to the whole DoE, so forgive me if the question is very naïve. I am trying to work out the best type of model to apply to my experiment. I have 20+ factors to combine and my aim is to find the best combination of these collecting four timepoins.
Actually, I need to find the best combination with the lowest number of factors.
I'll try to explain this better. The way this experiment goes, I can expect each one of these factors will have a certain positive impact on the sample. So this would give me a final best combination that most likely contains all of the 20+factors. And this is what I am trying to avoid, I need the best response with the lowest number of factors in it.
Thank you in advance to anyone who can help
Regards
Ilaria
DOERe: How do I find the best design for 20+ factorsI'm sure that folks will need a few more details to be able to offer some guidance. If you are just starting off with this topic, I would strongly recommend taking a look at our 'Statistical Thinking for Industrial Problem Solving' online course (for which there is no charge, and for which you don't need JMP installed locally).
Even at this early stage, it's probably worth saying that one of the golden rules of DoE is that 'If you want to know the effect of a factor, you need to vary that factor'. So I'm just a little concerned by your "And this is what I am trying to avoid . . ." comment (but I may have mis-understood).
Re: How do I find the best design for 20+ factorsMy thoughts:
Realize advice on what "tool" or design to select is dependent on a number of factors and the particular situation. It sounds to me as though you want a test not an experiment. A test is used to "pick a winner". An experiment is used to get understanding of the causal structure. If you truly want to understand what is going on and find the factors that have the greatest impact (and are most useful for prediction), this usually takes several iterations. Always start with hypotheses as to how and why each factor contributes to the variation in the responses. In general, when dealing with >15 factors, I am biased to a directed sampling plan (component of variation study) to compare groups of x's (components like within sample, sample-to-sample, etc.) and evaluate the measurement component. Sampling is an efficient way to get insight into a process with lots of variables. Select factors from the group that exhibits the greatest leverage from the CoV study for initial screening experiments. Ultimately developing a model (model development can be done several ways, but again I am biased to the Taylor series approach). Principles of Scarcity, hierarchy and heredity are useful guides as you iterate.
