Constraining parameter estimates on regression models

Question

Hi all,

I have a customer that would like to constrain coefficients to be positive only.  I searched the community and found the suggestion to use the Parameter Bounds option in the Nonlinear platform (https://community.jmp.com/message/223014#223014 ). Just wondering if anybody has any other suggestions.

Thanks,

Jerry

Jerry Cooper

JMP Systems Engineer - Gulf Coast Region

JMP Statistical Discovery from SAS

SAS Institute, Inc.

Tel: +1-512-840-6255

E-mail: jerry.cooper@jmp.com<mailto:jerry.cooper@jmp.com>

Byron_JMP · Answer

depending on what they are modeling, like if its count data, the generalized linear model with the poisson distribution function will keep the response positive.

There is a fish.jmp data set that this approach works with, it keeps you from building models where you catch less than 0 fish.

jerry_cooper · Answer

Thanks, Pete. The data we used for this yesterday was proprietary, so I wasn’t able to keep it. I’d love to see your data and the approach you took if you can share it.

jerry_cooper · Answer

Thanks, Jian. I agree with the concern, I definitely had the same reaction at first. But, since, in this case, in reality the factors can only have an additive effect on the response, I justified it to myself by saying, “All models are wrong… let’s just not be so obvious about it.”  Actually, when I asked what the goal was, their response was “for the model to reflect reality.” Since it wasn’t “predictability” I thought the request had merit.

When we used the Nonlinear platform approach on their data, it seemed to work OK, but it was cumbersome (lots of factors!) and, as was mentioned in the thread on the community site,  there were no model diagnostics to evaluate the results.