Weight factors in a DOE (Minitab feature)

Question

Hello,

we have a company with Minitab users who want to add weights on some factors in a DOE in JMP (as Minitab seems to permit). Does it make sense, is that a fair comparison if you weight some factors, what the purpose ?

Can we reproduce that on JMP ? I have in mind to create a formula "weight = [3X1 + X2 + ... + Xn" (weighting X1 by 3 fold on normalized factors as instance) and use it as weight. Any thoughts, recomendation ? Thanks for the help.

Emmanuel

Phil_Kay · Accepted Answer

Hi @Emmanuel_Romeu ,

We certainly can weight factors or effects in Custom Design in JMP.

There are 2 options that could help:

1. Weighted A-Optimal. This allows you to put different weighting on each effect in the model. We recommend that you weight on different orders of magnitude. For example, weight your most important effects 10 and the others 1. You should see that the heavily weighted effects will have increased power and decreased correlation with other effects.

2. You can select "If Possible" option for estimation of less important effects. This allows you to create a design with fewer runs than effects in the model  - a supersaturated design.

Brad's recent Discovery presentation covers A-optimal.

I am surprised that Minitab has this sophistication. I would be interested to know more about their capability.

Regards,

Phil

DonMcCormack · Accepted Answer

Based on what they have on their website they're talking about weighted regression. Use the Weight role in the Fit Model dialog box. The trickiest part is coming up with the column of weights. If the value for a row corresponds to an average measurement, the inverse of the variance of the raw data is commonly used as a weight.

Weighted regression helps to ameliorate violation of the assumption of homogeneous (response) variance across the design space. A-optimality affects the design matrix and the precision of the parameter estimates. That is to say weighted regression is about variance in the Ys and A-optimality on variance in the Xs.