Hi,
a customer asked for sensory DOE. Concrete, she has two questions:
1. Creating a design with lets say 5 three-level discrete numeric factors estimating main and 2-factor interaction effects requires 6 runs minimum and 27 recommended. As there are 12 panelists the 16 runs should be enough for a good variance. She asked how to do this replication, my answer is to copy paste the design 12 times in the resulting data table. Any other idea (having 12 added columns for the panelists and then stack the table is another thought I had). At the end she wants to check the power for the replicated design (I suggested evaluate design then, however I'm not sure how accurate it works for such a setup). Any more straight forward way?
2. If they go with 16 runs, but one run they would like to exchange with a market product having special settings for the five factors. How can they adjust the design to have highest power but exchanging one of the runs with the given run? I can think of creating the table and manually check the power for each of the different resulting designs, but that's cumbersome. Any ideas?
Thanks,
Martin
A few thoughts on things to consider.
On point #1, for the design they could accomplish 16 runs for each of the 12 panelists. That is a total of 192 runs. Why not make a design for 192 runs with a random block that is of size 16? This would provide the correct number of runs, but it does mean that not every panelist will see the exact same combination of treatments, but that may not be a bad thing. After all, the design will not be balanced and by just repeating the same 16 runs, you are going to estimate some interactions (and possibly some main effects) better than others. By letting the Custom Designer choose the combinations it should give overall better and more uniform power for each of the interactions. When it comes to fitting the model, if each panelist is considered "equivalent" like an analytical instrument, you can remove the random block effect. If, however, the panelists are allowed to be a random effect then you will remove that known source of variability from one panelist to the next.
Point #2 is trickier.
For point #2, does each panelist need to see the market product with special settings? I assume they do. In that scenario, I would use a similar approach. I would create the design with 180 runs, panelists a random block of size 15. Then I would manually add the extra run of the market product for each panelist. Easiest way to do that in my opinion would be to put the market product in a separate table and join them. You could then evaluate the entire design. It does not guarantee the highest power using this approach, but the market product is a separate treatment combination, so it is not really part of the design anyhow.
Ideally I would like to create the table with the market products and augment it, but I can't seem to get a random block into that augmentation.
