82942_en-US

liliconcom

Augmenting DOEs with changed factor ranges

 

 

Hi,

 

a pharma customer asked how to change best factor ranges when augmenting a design. Sometimes this is necessary as the analysis provided the hint that you should look in a slightly different region. This different region could be shifted or enlarged as these images below show.

 

Is there any best practice or thngs one should take care of?

 

I know that you can change the ranges in the augentation dialog but not sure about the possible effect on power. From a simple example I didn't see a big issue, but want to be sure.

 

Thanks for your thoughts,

Martin

Re: Augmenting DOEs with changed factor ranges

I agree with all @markbailey suggests. Can I throw in one other thought? Maybe in the second subsequent design, throw in one or two treatment combinations from the first design as a form of poor person's replication? One thing I see sometimes is 'Oh we have a new bag of materials that weren't used in the first design since we used all of the old stuff up.' What if that new bag is NOT materially identical? This would give you a fighting chance at seeing if there's some kind of offset in average or variance of response from 'first experiment' to 'second experiment'. And what about the measurement system's potential changes from 'first experiment' to 'second experiment'. Calibrations, new operators, etc. can muddy the second experiment.

Re: Augmenting DOEs with changed factor ranges

For one thing, I see a tendency to narrow the range, even if it is shifted, after the first round of experimentation. "I know where the good stuuf is!" Don't do it. That change will definitely compromise estimation and power. You didn't specifically mention this consideration but I encounter it so often in other situations that I just wanted to point it out.

 

A wider factor range means that non-linear effects are likely. Be sure to consider quadratic terms in the new model.

 

I am not sure how important power is at a second stage of experimentation. It might be, for sure, but prediction might be more important at this point. Pay attention to prediction variance as a profile and as a fraction of the design space.

 

Consider changing the optimality criterion if prediction is now more important that in the first experiment.