Setting Ranges For DOE Factors

[#1]

Hi,

I am pretty new to using JMP and statistics in my designs, but am thoroughly enjoying learning about it!

My question is about how to choose ranges for your factors. Say I am designing a process with a factor that can run between 1-100. I also "expect" the optimal range to be around the 15-20 mark. (Assume the desirability is maximized at that range, as in quadratic increasing from 0-20 then decreasing beyond that).

My concern is that when I design a DOE, I get a scatter of low, middle, high conditions, (0, 50, 100), but even the 50 in this case is expected to be very high and most likely undesirable.

• Is there a reason to tighten the range so that the center point is close to expectation or will JMP still provide the optimal condition regardless? (And I would therefore just be giving up explored range without gaining anything)
  • Ex. Range = 0-30 vs. Range = 0-100
• Is there a time and place when and when not to tighten or widen ranges?
  • Ex. Definitive Screen (Where factors are being explored for significance) vs. optimization DOEs

Some reading has led me to believe you want the largest ranges possible so that the model can effectively "fall off" on both sides of an optimum, but I wonder if you can still be too wide...

[#2]

Good question, @MetaLizard62080 .

It's not really possible to give a definitive answer on this. It depends on many things, not least your objectives. In this case it sounds like you are quite confident about the location of the optimum factor setting, so I am wondering why you need to do the experiment.

But here are some things to thing about...

Having broad ranges will generally increase the power of your experiment to detect important effects; making big changes to the system should result in big differences in the measured responses. Setting the ranges too narrow can mean that you don't see a big enough "signal", and it gets lost in the "noise" of the experiment.

Setting the ranges too broadly could mean some of the runs result in total failure of the system and no measurable response data. That is not ideal, of course. But it could still be useful information.

In my experience, it is more common for people to set ranges too narrowly than too broadly. So I generally encourage people to be bold with setting their factor ranges. Scientists often think they understand the system better than they really do. Also, they tend to underestimate the noise. 

Bear in mind that you can always augment with more runs to look at broader or narrower ranges in the next phase of your experiment.

I hope this helps,

Phil

[#3]

I agree completely with Phil and his experience regarding level setting.  Keep in mind your prior knowledge regarding what may be the best levels for factor may have been developed using OFAT.  You may want to expose interactions and non-linear effects before selecting optimum levels.