Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Fabien Bertho <fabien.bertho@sciences-po.org> |
To | statalist@hsphsun2.harvard.edu |
Subject | Re: st: Interpretation GLM coefficients |
Date | Thu, 19 Jan 2012 14:15:30 +0100 (CET) |
Hi David, Yes, it is very helpful. Thank you very much. I would have one more question concerning the - margins - command. If I use this post-estimation command, can I interpret all coefficients as marginal effects? For instance, - A one percent change in x2, generate a b2% change in x1 - The movement of x4 from 0 to 1 produces a [exp(b4)-1]% change in x1 And, what is the difference if I include the - atmeans - option or not? Thank you again for your help. Fabien > ---------------------------------------- > From: David Hoaglin <dchoaglin@gmail.com> > Sent: Thu Jan 19 13:45:30 CET 2012 > To: <statalist@hsphsun2.harvard.edu> > Subject: Re: st: Interpretation GLM coefficients > > > Hi, Fabien. > > Since you are using the log link, the relation between the mean of x1 > and the linear predictor is > > mean(x1) = exp(b1 + b2*ln_x2 + b3*x3 + b4*x4), which you can rewrite as > mean(x1) = exp(b1) * [exp(b2)^ln_x2] * [exp(b3)^x3] * [exp(b4)^x4]. > > Interpretations often use "incidence rate ratios." > > A change of 1 unit in ln_x2 (not a change of 1 percent in x2) would > change the mean of x1 by a factor of exp(b2). > > A change of 1 unit in x3 would change the mean of x1 by a factor of exp(b3). > > And changing x4 from 0 to 1 would change the mean of x1 by a factor of exp(b4). > > For each variable, the interpretation should include the explanation > (as applies also in OLS) that the changes are adjusting for the > contributions of the other variables in the model. Because the > predictor variables generally change together in the data, it is an > oversimplification and often misleading to talk in terms of holding > the other variables fixed. (Unfortunately, many textbooks do not > explain this, but it is an inherent feature of regression-like > models.) It may be possible to make predictions that change one > variable and hold others fixed, as long as you stay within the region > covered by the data. Similarly, a change of 1 unit should not take > you (far) outside that region. > > I hope this discussion is helpful. > > David Hoaglin > > On Thu, Jan 19, 2012 at 4:39 AM, Fabien Bertho > <fabien.bertho@sciences-po.org> wrote: > > Hello, > > > > I am running glm Poisson regressions and I have some problems in interpreting the coefficients. I am running the following regression: > > > > - glm x1 ln_x2 x3 x4, robust cluster(distance) family(poisson) link(log) irls - > > > > Where x1 is an integer, x2 and x3 are continuous variables and x4 is a dummy variable. > > > > Is the interpretation of glm coefficients different from OLS coefficients? My attempt: > > > > => A one percent change in x2, generate a b2 unit change in x1 > > => A one unit change in x3, generate a b3 unit change in x1 > > => The movement of x4 from 0 to 1 produces a b4 unit change in x1 > > > > Is this correct? > > > > Thank you. > > > > Best Regards, > > > > Fabien BERTHO > * > * For searches and help try: > * http://www.stata.com/help.cgi?search > * http://www.stata.com/support/statalist/faq > * http://www.ats.ucla.edu/stat/stata/ -------------------------------------------------------------------------- Tous les courriers électroniques émis depuis la messagerie de Sciences Po doivent respecter des conditions d'usages. Pour les consulter rendez-vous sur http://www.ressources-numeriques.sciences-po.fr/confidentialite_courriel.htm * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/