0

I'm using R to run some generalized linear models (GLM) and I want to interpret the estimates.

If I use a GLM assuming Gaussian distribution with identity link, interpretation is easy: if the estimate for factor $x$ is $y$ then an increase of 1 of factor $x$ (if all other factors remain constant) will be estimated to result in a change in the response variable by $y$ units.

Now, if I run a GLM assuming Poisson distribution with log link, in the example above, will the response variable be expected to change by $y$ units, or by $\exp(y)$ units?

note: I'm getting the estimates from summary(model object)

Maarten Buis
  • 19,189
  • 29
  • 59
Ben G
  • 1
  • Other threads to review include: [1](http://stats.stackexchange.com/q/24821/), & [2](http://stats.stackexchange.com/q/23672/). – gung - Reinstate Monica Mar 06 '14 at 15:58
  • Yes, those should be very helpful. It's a good idea to look on the site for the answer to your question - then if something doesn't add up for you, you can post a follow-up question. Also, had you included "Poisson" in your title, i.e. been a bit more spcific, the site would have suggested some relevant links. – D L Dahly Mar 06 '14 at 16:11

1 Answers1

1

For a unit change in $x$ you expect the response to change by a factor of $\exp(y)$. So say $\exp(y)=1.5$, then a unit change in $x$ is associated with a $(1.5-1)\times 100\% = 50\%$ increase in the response.

Maarten Buis
  • 19,189
  • 29
  • 59