Augment Design not respecting Covariate Factors

[#1]

I am screening a reaction using a custom design generated 6 continuous factors and 2 covariate factors, with 1 continuous factor set to be hard to change, 3 whole plots, 24 runs. I have already run 8 reactions, and the hard to change factor (temperature) is easier to change than initially thought. So I want to change the design to one that has more power for predicting temperature effects whilst keeping the 8 completed runs.

I tried to do this by changing the temperature factor changes to be easy, deleting the 16 experiments I haven't run, and using Augment Design to add 16 new runs. This did not work however, as the generated runs did not honour the covariate factors. I'm using the covariate factors PC1 and PC2 to represent my solvent selection, and thus each PC1 value has a corresponding PC2 value. When augmenting the run, the PC1 and PC2 values would get mixed up. How do I get augment design to honour the covariate factors?

I've attached the JMP file, I am using JMP Pro 17.

[#2]

Hi @DecilePlatypus,

Welcome in the Community !

I think the easiest way to continue your DoE is by using the Augment Designs platform as you also noticed.

In your case, the covariates PC1 and PC2 are not independent, as they are linked to specific solvents. So only one covariate value is needed to identify the solvent used in the design.

You can change the Temperature Factor Change property from "Hard" to "Easy", and then add your 7 factors as X (6 continuous factors + PC1) and the Yield as Y in the Augment design window :

Then, you can check the option "Group new runs into separate block", and click on "Augment" (or specify a different type of design augmentation). 

Make sure all your factors main effects are entered in the Model panel (with the added "Block" effect if you have checked the option to group new runs into separate block), and define a suitable number of runs (including the 8 runs already done). In this example, I specified 20 runs in total :

After launching the augmented design generation and creating the table, I can add the columns PC2 and solvent name by updating the design table with the informations from the covariate datatable (with PC1 as the join variable).
Solvents chosen in runs are done in order to have the biggest difference/variability :

You can try several augmented designs and compare them thanks to Compare Designs platform to choose a suitable design for your needs. SOme additional suggestions :

• You may also start with less solvents in this screening stage (for example 3) and select them based on the maximum difference and chemical variability (for example Tributylamine, TMU and Methylcyclohexane).
• It may also be more helpful to use a categorical factor (solvent name) for the design creation to avoid any imbalance in the repartition of solvents used, and then add PCs in the table and use them for analysis.

Some previous discussions might help you :

Efficient DOE of one multi-level (3+) categorical variable and many continuous variables 
Re: 如果试验自变量为分类变量，响应为连续变量，是否可以使用正交设计进行筛选？ - JMP User Community

I attached your datatable with an added script for the augmented design generated.

Hope this first response will help you.

[#3]

Hi @Victor_G 

Thanks for the help. The problem with just augmenting with one column is that the resultant design will not explore as much space in PC2. Using your method only generates runs with either TMU or tributylamine, which maximises PC1 but leaves a lot to be desired in PC2.

[#4]

Hi @DecilePlatypus,

Yes, the best option here would be to augment the original runs using the solvent name as a categorical factor, to make sure the augmented design is not imbalanced regarding the solvent choice. 

Using 20runs in total in the augmented design, I have a more balanced design regarding the solvent choice in experiments :

And power for PC1 and PC2 is more homogeneous : 

So the best option could be to use name in the design generation, but use the PCs in the analysis (use update table with the covariate table to add the PCs in the augmented design table, with Name as join variable).

You can read the additional suggestions in my previous post, one was linked to the situation you described (imbalance in solvent choice in the experiments), and the other one is about the number of different solvents investigated.

Typically, removing dioxane from the list of solvents in the covariates could help. Here is the power analysis with 12 added runs (20 in total) but without considering dioxane in the covariate list :

I attached the datatable with the added scripts for the first (balanced) design and the second (without dioxane) as an example.