>On Apr 17, 3:57 am, Pact.Capacity.Managem...@UK.FUJITSU.COM (Anthony >Steel) wrote: >> Good-day all, >> I was hoping someone may have a solution to a minor modelling issue we >> are encountering. >> >> Our aim is to produce "best-fit" analysis of response-variables against >> multiple regressors (for which PROC REG is currently used), but as we are >> dealing with sets of real positive numbers (>=0) we wish to restrict the >> range of possible parameter value outputs to also be in the same range >> (i.e. >=0). >> >> So in summary is there anyway of ensuring the multivariate linear >> regression analysis provides the "best-fit" that can be achieved with the >> calculated paramaters being greater than or equal to zero? > >

Paige replied

>I can't understand the reasoning here. Your predictor variables are >real positive numbers. Why then do you require the parameter estimates >to be also positive? Makes no sense to me. > >There could easily be an inverse (i.e. negative slope) relationship >between a specific predictor and your response, which would result in >a negative parameter estimate. > >Seems to me you need to rethink this.

Now I realize that I misinterpreted the original post, and I agree with Paige I don't see why you would want to have only positive parameter estimates.... I had thought you wanted only positive predicted values.

Peter

Peter L. Flom, PhD Statistical Consultant www DOT peterflomconsulting DOT com