Estimating Incidence Rate using poisson regression

[Catharina Alberts]

Hi All,

Up to now I have benefitted a lot from this forum. One question I can’t get my head around is the following:
I have a longitudinal dataset in a long-format (the long-format is because I have variables that change every visit a patient comes is). I want to estimate the IR (by1.000) using poisson regression.

 
poisson hpvinc i.agec i.smk, exposure(pdomo) irr

 
xtset id
 
   
xi: xtgee hpvinc i.agec i.smk, exposure(pdomo) eform family(poisson) 

gen 
pdomo=(pdo/30.5)/1000

Because the data is in long-format, would it be better to use GEE to take into account clustering of the individuals records within one id?
Both approaches give almost the same answer, with some subtle differences behind the comma for the IR as well as the 95%CI.
Thanks for your time,
Catharina

[Carlo Lazzaro]

Catharina:
welcome to this forum.
If you have panel data, why not considering -xtpoisson-?

[Catharina Alberts]

Dear Carlos,

Thanks for your reply!

-1- Would you mind explaining why you suggest to use xtpoisson?

-2- The reason I am using xtgee, is because my outcome is HPV. Our approach is inspired on Xue et al. his/her paper, where (s)he suggest to use exchangeable correlation structure to get one point-estimate for all different types of HPV (rather than estimating it for all the different types of HPV individually or using 'anyHPV' as outcome).

So my syntax actually looks as follows:

xtgee hpvinc  i.agec i.hpvtype , exposure(pdomo) eform family(poisson) corr(exc)

I have long-formatted dataset in two dimension: time and hpv-type.

Please note my background is in epidemiology (rather than statistics).

Any thought would be welcome,

Thanks,
Nienke

[Carlo Lazzaro]

Catharina:
thanks for the link.
If a previous authoritative article in your reserach field used -xtgee-, follow that road, as it will be easy to justify in your paper.

[Catharina Alberts]

Thanks for your quick reply:

So just to recap, if I would not had to take into acount the multiple HPV-types and I had to choose between the two syntax lines (I mentioned in my first post) to execute a Poisson regression the xtgee would be the preferred one? (I also want to report incidence rate for the individual HPV types, and I am trying to understand whether using xtgee would be correct).

[Carlo Lazzaro]

1. Catharina:
2. - my original take was in the light of the seemingly panel structure of your dataset;
3. - if you use -poisson- with panel data, you should invoke clustered standard errors, as the within-panel observations are not independent;
4. - your -xtgee- assumes panel data and allows you to model within-group correlation structure more precisely. I would consider this last code the way to go.