GLM - Offset term

[AB Ball]

Hi,
I am running a GLM - Poisson Family Identity Link model to show the impact of various predictor values on low birth weight. I am trying to include an offset term to take into account the total number of births. I have the following code:
glm Nlbw PM25cent IncomeScoreMulti NBAME NWhite, family(poisson) link(log) offset(Nbirths)

The output gets stuck looping on iterations. Please can somebody help?

[Carlo Lazzaro]

AB:
welcome to this forum.
Does the same happen when you go -poisson- instread of -glm-?

[AB Ball]

Hello Carlo, thank you for your reply. Tt works by replacing the GLM to Poisson. I am a stats & Stata novice and just trying to learn as I go on. Would omitting the GLM still work in the same way?

[Paul Dickman]

Are you using an identity link (as you write in the text description) or a log link (as in the code)? Using an identity link can lead to convergence problems because you are modelling the number of LBW (which cannot be negative) as a function of the linear predictor (which can be negative).

If you are using a log link, a common reason for non-convergence is that your model is trying to estimate rate ratios that are zero or infinity. This is more common when using cross-classified data (categorical predictors) whereas you look to be using continuous predictors. It might, however, be an issue. Does a model with fewer predictors converge? If so, try and identify which predictor is causing problems.

You could also have a look at the last values of the parameter estimates and their variances to try and identify which ones are problematic.

matrix list e(b)
matrix list e(V)

May I ask why you're using Poisson regression when it appears your outcome is a proportion (for which a binomial error structure would be standard)?

[AB Ball]

Hello Paul,

Thank you for your reply.

Sorry that was my error in the text - I am using the log link. All my variables are continuous. Thanks for the advice - I will try running the model one variable at a time.

I have a discrete number of LBWs which are positive skewed which is why I thought the Poisson regression would be appropriate. My model is looking at the number of LBW babies per small geography area.

[Carlo Lazzaro]

AB:
regardless -glm- or -poisson- choice, you should check your model for overdispersion.