Definitive Screening Design question

[#1]

Should you only run a Definitive Screening Design if you know or think you're very close to the process optimum? What happens if you run a DSD, but you're outside the optimum factor space? Would you have wasted your time, and have been better off doing a fractional-factorial 1st to screen items, then doing the full factorial to get curvature afterward? 

[#2]

Effective screening relies on several key principles. The design methods count on these principles holding. It sounds like you are past the screening stay: you know the important factors and suspect the best settings. You can optimize with a Custom Design for the RSM model. Just be sure to use a factor range around the expected best level that is wide enough to elicit a strong effect (large change) in the response. Otherwise, you will have large standard errors in the model coefficients and suffer poor prediction performance.

[#3]

Here are my thoughts:

1. The purpose of DSD's is to screen a large number of factors taking into account potential curvature and 2nd order effects. So the answer to your 1st question is no.

2. There is no "right" answer to your additional questions.  Much depends on where you are in the knowledge continuum (e.g., how specific are your hypotheses, how much do you understand noise), the response variables you are modeling (e.g., continuous, ordinal or nominal), the type of factors you are manipulating (e.g., continuous, categorical), your sense of urgency, your budget, etc.

3. I am a firm believer in sequential investigation.  The purpose of the first experiment is to design a better experiment.  Go big and bold to start (e.g., lots of factors at bold levels with bold blocks of noise), understand noise and how to  handle it, reduce the number of factors, perhaps test for curvature with center points, develop a useable model.

[#4]

But if you're not near the optimum, you won't see any curvature right? This is why in RSM you typically start w/ a full-factorial + center points to test for curvature, & if you don't find curvature you use the model of steepest ascent to "climb the mountain" in the direction of a higher response until your response declines. If you do a DSD in a factor space not near the optimum, I feel like you've wasted runs since you then have to move in the direction of steepest ascent again & re-run another model right?

That's why I'm wondering DSDs advantage is if you're not sure you're near an optimum, and why you would use a DSD if you're not sure you are in fact near an optimum? 

[#5]

DSD's are first and foremost screening designs. So as both @Mark_Bailey and @statman commented, typically your primary focus is on finding the most influential factors in the bolder factor space, to borrow a phrase from @statman . A metaphor I used often with my colleagues is, "With screening designs were just trying to find the right church. With RSM designs were trying to find the right seat in the right pew, to hear the choir best." DSD's are a useful method when relatively early in the learning process...not towards the end, where the more typical RSM oriented designs will come in...like either the classics like Box-Behnken or central composite designs or more modern and efficient approaches in the optimal experimental design tool kit.

[#6]

So are you're saying DSD's can be used as regular screening designs for an unknown system, just as a fractional factorial might be used, but their advantage lies in detecting curvature. So, while they may have only a few more runs than a typical Res IV fractional factorial, they could potentially save you runs if you're already near the optimum? Whereas w/ the fractional factorial, you'd then have to move to the full-factorial, and then RSM? 

[#7]

You are on the right track...but not quite. Yes, DSDs can (and I'd go even further and say in many cases, should be used as regular screening designs). Do the math for number of runs calculations and you see pretty quickly that as the number of factors increases beyond 5 or so...DSDs actually have FEWER runs than a typical RES IV fractional factorial design. And you can get a hint of 2 factor interactions and curvature. So there's lots to like when comparing DSDs with RES IV two level fractional factorial designs. Were it me...in many instances I start with the DSD and then as @statman discusses, move to a more RSM approach leveraging what you learned from the DSD. And I'd generally stay away from the classic full factorial and RSM oriented designs (like central composite) and use the more modern optimal experimental design approaches.

[#8]

Thanks, I've been hesitant to use a DSD b/c I don't know much about them, & was wondering what their advantage is - but it seems they're Res IV designs already, can lead to fewer runs w/ a high # of experiments, & can detect curvature (though w/ some alias issues). 

What are more "Optimal experimental design approaches" in place of RSM? Do you have any examples? 

[#9]

When I refer to 'optimal experimental design approaches', I'm referring to the expansive body of techniques codified in JMP's Custom Design platform. A recommended reference is Goos and Jones, "Optimal Experimental Design: A Case Study Approach". I'd also search the on demand JMP webinars for this topic which has been covered in many instances.