# Statistical design of experiments revealed some surprises -- what next?

I recently became the manager of a complex piece of laboratory equipment. It has many knobs and dials to turn, and the different knobs and dials affect the significantly affect the responses I try to measure. Since there was no theory of how exactly each knob affects the response, I turned to statistical design of experiments figure out how to set my knobs. Using the same exact input each time, I use different settings for the various knobs and am trying to find the optimal (in my case, the maximal) setting for the knobs given this identical, unvarying input.

I initially expected that many settings would have an optimum level somewhere within the design space I was analyzing. Thus, I decided to use a definitive screening design to design my runs. I did exactly this, measuring the same sample nineteen different ways, with each of the nineteen runs having one of the nine knobs set to either a "high", "medium", or "low" setting. This was straightforward, using the the DefScreen() function from the daewr package in R. My (normalized) responses as well as the (scaled and centered) experimental conditions are displayed below.

    >>> scaled_data %>% dput
structure(list(method_num = 1:19, y = c(0.90085050161025, 0.00973935039938332,
0.046907811292605, 0.0477583584054755, 0.000734906456458273,
1, 0.101817637536817, 0.140301184447718, 0, 0.421559489775223,
0.636661367259339, 0.00451827441824333, 0.0714993448043543, 0.0605393859587476,
0.115095370854818, 0.222299049967782, 0.176960666274614, 0.0455311915787348,
0.203887872677827), A = c(0, 0, 1, -1, -1, 1, -1, 1, 1, -1, -1,
1, 1, -1, -1, 1, -1, 1, 0), B = c(1, -1, 0, 0, 1, -1, -1, 1,
-1, 1, -1, 1, 1, -1, -1, 1, 1, -1, 0), C = c(1, -1, 1, -1, 0,
0, 1, -1, 1, -1, -1, 1, -1, 1, -1, 1, 1, -1, 0), D = c(1, -1,
-1, 1, -1, 1, 0, 0, -1, 1, -1, 1, -1, 1, 1, -1, -1, 1, 0), E = c(1,
-1, 1, -1, 1, -1, 1, -1, 0, 0, 1, -1, 1, -1, 1, -1, -1, 1, 0),
F = c(1, -1, -1, 1, -1, 1, -1, 1, 1, -1, 0, 0, 1, -1, 1,
-1, 1, -1, 0), G = c(1, -1, -1, 1, 1, -1, -1, 1, 1, -1, 1,
-1, 0, 0, -1, 1, -1, 1, 0), H = c(1, -1, 1, -1, -1, 1, -1,
1, -1, 1, 1, -1, -1, 1, 0, 0, 1, -1, 0), J = c(1, -1, -1,
1, -1, 1, 1, -1, -1, 1, 1, -1, 1, -1, -1, 1, 0, 0, 0)), class = "data.frame", row.names = c(NA,
-19L), .Names = c("method_num", "y", "A", "B", "C", "D", "E",
"F", "G", "H", "J"))


Some ad-hoc model-fitting of this data revealed a few surprises:

• First and foremost, no quadratic terms seemed significant.
• Three (possibly four) of the variables of my original nine seemed the most important.
• One model that fits my data well is y ~ H*J + B + B:G + B:G:H
• Even with those interactions included, there isn't clear evidence for any distinct optima within my design space.

A graph of my data vs. model predictions is below.

My main question is, what should I do next? Practically speaking, it isn't much work or time for me to do further runs, and so agonizing over the "best" way to proceed from the point of view of statistics not necessarily worth tons of time...yet I am curious about statistics and design of experiments in general. I figured the CV community would be a good place to ask for opinions and guidance on "official" ways to proceed. So, suppose for the sake of argument that further experimentation was very expensive and time-consuming. Some specific questions I have include:

1. The modeling I have done so far suggests (to me) that my design space didn't span any strong (local) optima. Thus, I need to extrapolate. How should I choose to extrapolate my original design space?

2. Should I use D-optimal design methods to augment my initial design?

3. How (if at all) should I use my model from my runs so far to inform future runs?

4. Can I just decide that the four variables I've identified as significant are all that matters, and re-do a definitive screening design centered at the point I identified as optimal in my first round of experiments?

5. What am I missing that I should be thinking about? My goal is to design an efficient set of experiments that will get me to the "optimum" setting for the knobs and dials on my experimental machine.

• You didnt define what you mean by "optimal setting of the knobs" Can you explain? Mar 23, 2017 at 17:13
• Sorry, I'll re-word that paragraph...Thanks for giving this a look. Mar 26, 2017 at 22:41
• Well, you definitely jumped into the deep end of the pool. The difficulty with experimental designs and the general approach is that there are so many "it depends" type answers. I have no official answers, but speaking as an applied statistician (engineer), I would recommend picking up Statistics For Experimenters by Box, Hunter and Hunter. You can get a very good condition used 1st edition for about \$6. It will take you through all the basics, including path of steepest ascent, which would be a way to start extrapolating, as mentioned by kjetil b halvorsen. Mar 5 at 11:20
• Also, I'd suggest starting with 2-factor experiments for learning purposes, rather than doing 4-factor experiments. It's so much easier to visualize things in 2-D and, when adding a response, 3-D. I've found visualizing designs and results as much as possible really helped me to advance my own understanding of experimental design. Mar 5 at 11:23