Skip to content
Prev 3258 / 7420 Next

Log transformed simple linear regression and Poisson regression

Dear Lin and others,
if I understood it well, You are now asking about variance decomposition
(how much variance in response data could be ascribed to individual
predictors or their groups).
If log transformation (or maybe sqrt to stabilize the variance?) works
well for you, I would do that and decompose the variance as usual, using
sum of squares. 
If the transformation does not work well and you need a GLM (poisson,
negative binomial, quasipoisson...), then I am not sure about solution.
I would use Nagelkerke's R (or some other Rsquared?), cross my fingers,
hope for its aditivity and use that as a measure of the variance
explained by the particular model.
(I hope someone corrects me if I am wrong).
I hope this helps.
Martin Weiser

lgj200306 p??e v ?t 02. 08. 2012 v 06:50 +0800: