nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Urgent - Why I get other values than -1,0,1 in my design?

martindemel

Hi,

I create a 4 factor design, 3 continuous with lower and upper Level, one categorical with 3 Levels (it's the example from Brad and Peter in their book on page 70ff (a robust and optimal process Experiment). When I generate a D- or I-Optimal design I would expect only -1, 0 and 1 as well the three Levels for the categorical factor to appear in the design. But I get everytime at different positions a numeric value like below in the design. I cannot explain why this happens. Can you help me find an Explanation or Show me a way to get the same design as in the book?

Here is a screenshot of the design I created with the script attached:

I Need this for tomorrow, therefore this urgency. I haven't noticed this before, but it probably will raise questions on customer side I cannot answer right now.

Thanks in advance!

Martin

PS: I checked in JMP Pro 11, JMP Pro 12.1, JMP Early Adaptor 13, always with the same unexpected results (besides the decimal is at different positions everytime I generate the design, though that's what I expect)

Skript_DOE_issue.jsl

Accepted answers

Ryan_Lekivetz

Hi Martin,

This behavior is due to the coordinate exchange algorithm not being able to find a better choice than the decimal value. Almost all the time that you see something like this, changing the value to the nearest integer will result in a slightly less efficient design (but usually not by much).

I would say there's nothing wrong at all with rounding to the closest integer - you can go to evaluate to confirm that not much has changed, and at the end of the day, you have control over the design that's going to be run.

Even when I tried letting it run for awhile, I still had the one decimal place for I-optimal.

Cheers,

Ryan

All comments

Bill_Worley

Martin,

I don't have a true answer for you I can only tell you that I was able to replicate your result with a D-optimal design, but not the I-Optimal design.

I did an RSM design and if I left the interaction terms as "Necessary" then I got the same result as you. If I made the interaction terms "If Possible" I did not see the unexpected value for the D-Optimal design.

This does not explain the unusual value, but hopefully you will be able to recreate the design you want to see now in D-Optimal.

Best,

Bill

martindemel

Hi Bill,

thanks for the first comment. I also see this behavior but not consistent. Sometimes I still get a decimal even with "If Possible" set.

I could "cheat" and set those values in the table to the nearest integer as looking to the effect measures in evaluate design then they only change slightly, but it still feels as I'm doing something bad.

Martin

MikeD_Anderson

It has to do with the D-optimal and I-optimal algorithms. Brad actually had this happen during a DOE seminar he did for a customer in my area. The key to this is to remember that the custom designer is actually calculating the best design, not using a predefined one like full or fractional factorial. What is happening is, based on the constraints and mode you're trying to fit, the algorithm is finding an optimal set of conditions where 0.32 is better than just 0.

One of the solutions to just change it to 0 and then use the design evaluation tool to see if the 0.32 vs 0 really makes a difference in the power calculations and design evaluation criteria (at the bottom of the design evaluation report). This is the one that Brad used in the seminar I was in.

The other possible solution is to just up the fitting time and see if it converges to 0 over time.

Best,

Ryan_Lekivetz

Hi Martin,

Even when I tried letting it run for awhile, I still had the one decimal place for I-optimal.

Cheers,

Ryan

martindemel

Hi Michael and Ryan,

ok, this makes sense then. I will check for the power difference and the evaluation criteria. Now I feel confident enough to explain what's happens when it comes to that question.

Thanks a lot!

Martin