Comparing proportions positive

[Saskia Bogers]

Hi world!

Can anyone advice me on a model to compare proportions between groups?

I've been doing some analyses in STATA to compare the number of diagnostic tests ordered by doctors from 2 groups (attended education and didn't attend education). Every record in my dataset is a doctor in one quarter-vear period (ie one doctor will have multiple observations), and contains a variable with the number of tests ordered and the number of tests that were reactive (positive). I compared the number of tests between groups using a Poisson model as I'm working with count data. I've also compared the number of positive tests this way, but it would be more informative to compare proportions positive.

I therefore created a variable that shows the proportions positive (#pos/#ordered). However, I'm unsure which model to use here. I think that a poisson model or a binomial model are inappropriate. However, it's also not quite a Gaussian model as most test outcomes were negative (ie proportion positive is often zero).

Thanks in advance!

[Carlo Lazzaro]

Saskia:
is something similar in line with your research goals?

. xtset physicianid year

. xtpoisson test i.positive_test##i.educated_physician i.year, fe vce(cluster physicianid)

[Saskia Bogers]

Dear Carlo,
Thank you for your reply. Indeed, I have already ran such a model. However, this uses the absolute number of positive tests in Poisson model. Epidemiologically speaking, the proportion tested, perhaps even weighted by the nuber tested (denominator), would be more interesting, but I am looking for the appropriate way to model this, using either the #positive and #ordered or the proportion positive (#pos/#ordered; range 0-1).