Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Poisson residuals

From   Maarten Buis <>
Subject   Re: st: Poisson residuals
Date   Thu, 24 Nov 2011 10:33:33 +0100

On Thu, Nov 24, 2011 at 10:01 AM, Fabien Bertho wrote:
> I am trying to compute residuals after Poisson regressions.
> First, I used the following code:
> predict yhat
> gen poisson_residual= y - yhat
> Then, I tried the Stata command
> predict poisson_residual_stdp, stdp
> Finally, I tried this one
> predict poisson_residual_score, score
> The first code gives exactly the same values as the third one. And I don't understand the difference with the second computation.
> Please, what is the difference between both and and which is the right value?

With the -stdp- option you'll get the standard error of the linear
prediction, with the -score- option you'll get the first derivative of
the log likelihood with respect to xb, which is just by accident the
same as the raw residuals you computed in your first example. Your
first example just gives you the raw residuals, which are often not
very useful in non-linear models like -poisson-.So none of your
computations will give you the "right" numbers, it is even hard to
define what "right" is in this type of model. There is a whole cottage
industry inventing different types of residuals that are more or less
useful in such models. A fair set of those can be obtained by
estimating your -poisson- model using -glm- with the -family(poisson)
link(log)- options. You can see the different types of residuals and
their description by typing -help glm_postestimation-. There is also a
nice discussion in chapter 4 of

Hope this helps,

Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index