Mallow's Cp Statistic as criterion for regression

I am building a multiple regression model using stepwise procedures and using the minimum Mallow's CP statistic as criterion for adding variables to the model. Instead of getting any values of Mallow's CP that are near p (which is what it should be if the model is a good fit), I am getting very negative values for this statistic. Here is a little information about my data: selecting best predictor variables from 128 possible variables, and each variable has n=1020 observations. Am I doing something wrong, or are negative values ok? If anybody has any experience in this area, can you please help me out?



New Member
In multiple regression if Mallow's CP is close to P then we can say that fit is good.But you are getting negative values for CP.Cp is calculated as :
(SSEp/S^2)-N+2P,so if number of observations are more than 2p then you may get negative CP.


TS Contributor
An Information Crtietrion would be more appropriate (AIC,BIC). Mallow;s Cp is close to such a criterion

AIC=-2logL+2p, where p=#(parameters),that includes the predictors AND the estmated variance of the error term. The smaller the AIC value the better.It takes real values