Custom DOE With Covariate Factors

[#1]

I get the following JMP alert when trying to build a DOE with covariate factors, "Insufficient number of covariate runs to fit all model terms.". I think I understand what this means but am confused as to why I am getting the alert. I have two covariate factors. There are 17 values to select from in the candidate set. The covariate factors are hard to change. I have three easy to change process factors. Two of the process factors are continuous and one is a two-level categorical. I am designing for a model with linear, interaction and quadratic terms. I am asking for 10 whole plots with a total of 30 runs. I have attempted using fewer and more whole plots and get the same alert. There is very little correlation between the two covariate factors in the candidate set and it has plenty of levels to support the non-linear terms. Has anyone experienced getting this alert and know the issue causing it to occur? Thanks.

[#2]

Confusion Alert!

A Covariate Factor has a special purpose, which might not suit your needs. A covariate factor has pre-determined levels. That generally means that randomization is done, so leave Changes set to the default Easy setting.The number of runs in the covariate data set, not the number of distinct levels, determines the maximum number of runs, not the minimum. So if I provided a covariate data set with 15 rows, I can make a design with 15 runs or less.

The number of levels required is determined the terms in the model. If you have only first-order terms, then only two levels are necessary. If I have second-order terms, then only three levels are necessary. In fact, any levels other than the minimum level, center level, and maximum level are sub-optimal but a covariate factor is used when I cannot control the levels but I can determine them ahead of time and I want to match them optimally in the treatments of the other factors.

So, maybe don't try to make the covarate factor hard to change.

[#3]

I can easily imagine a scenario where covariates would be hard to change.  One of the most common uses for covariate factors would be raw material lot properties, and raw material lots are often hard to change in a process.  In these cases, the only thing that should be limited by the number of covariate rows is the number of whole plots, not the number of total runs.  I think @H2OSUP has identified a real bug in Custom Design.

[#4]

Hello @cwillden. This is exactly my situation.

[#5 · Accepted solution]

Hi @H2OSUP , I was able to reproduce your problem and I think Custom Design wants to have at least as many covariate rows as runs in your design and you really only need as many rows as whole plots.  You can work around the issue by concatenating your covariates table to itself to double the number of covariate rows to 34.  The risk is selecting the same set of covariate levels more than once for the whole plots, but in my test that did not happen.  This looks like a bug.

[#6]

Hey @cwillden, I think you are on to the issue. I believe it has to do with the number of parameters compared to the number of rows in the candidate set. I tried a model with fewer parameters than rows and it created a design with 30 runs. So, the total number of runs exceeded the number of rows in my candidate set but the number of parameters was less than the number of rows. I also believe it is a bug.

[#7]

I tried your solution of concatenating my covariates table to itself and it worked. It did select one row twice but that is not a problem for my scenario. Thanks!