Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Austin Nichols <austinnichols@gmail.com> |
To | "statalist@hsphsun2.harvard.edu" <statalist@hsphsun2.harvard.edu> |
Subject | Re: st: difficulty in understanding link function in glm |
Date | Thu, 19 Sep 2013 13:20:04 -0400 |
Reza Yousefi Nooraie <yousefr@mcmaster.ca>: Yes, you misunderstood. Using transformed y and the identity link assumes E(lny|X)=Xb but using y and the log link assumes ln[ E(y|X) ]=Xb and these are not the same thing. Only one makes sense when y can be zero or negative. See http://www.stata.com/meeting/boston10/boston10_nichols.pdf http://blog.stata.com/2011/08/22/use-poisson-rather-than-regress-tell-a-friend/ On Thu, Sep 19, 2013 at 1:02 PM, Reza Yousefi Nooraie <yousefr@mcmaster.ca> wrote: > Hi statalist members, > I have difficulty replicating the results using two alternative configurations of generalized linear model. > I'd like to predict a continuous dependent variable by a group of dichotomous and categorical variables. Since the dependent variable is highly skewed, I decided to log transform it. I assume that predicting a log transformed variable with a link(identity) function should give the same results as predicting the original variable with link(log) function. But not only the coefficients are different, but also their significance, and the overall log likelihood are different. > Perhaps I misunderstood the whole process. > Thanks for advice > Reza * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/faqs/resources/statalist-faq/ * http://www.ats.ucla.edu/stat/stata/