3
$\begingroup$

I’m modelling tuberculosis (TB) case rates at the neighborhood level and trying to identify risk factors that are associated with higher rates. I would like to identify communities with below average case rates given their risk factor profile with the idea that these would be the best places for new case-finding. It’s a bit circular I realize to use the same units I’m identifying risk factors with to then identify areas with expected higher rates, but my thinking is that I could identify the communities which are at the lower end of the distribution for each risk factor.

My plan then would be to construct a Poisson model with significant risk factors, plug in each neighborhood and calculate the ‘expected rate’. Then calculate the difference by subtracting ‘observed’ from ‘expected’ rates, and then identify the communities with the highest difference.

Is this approach legitimate? Or is this using the model in a way it isn’t meant to be used?

$\endgroup$
6
  • 1
    $\begingroup$ @Andre: Tuberculosis? $\endgroup$ – Wayne Aug 26 '13 at 15:19
  • 1
    $\begingroup$ @Tom: I've not been able to figure out if your overall proposal is valid -- it seems like it is, but I can't prove or disprove it -- but when you say "rate", you do mean something like "3 per thousand per month" or something like that? Obviously different neighborhoods could have different populations, which would result in different rates per time period if you didn't also account for population. $\endgroup$ – Wayne Aug 26 '13 at 15:31
  • 3
    $\begingroup$ If I understood you correctly, you want to look at the “residuals” of the model. Until someone provides an answer, this term might help you find some relevant information. $\endgroup$ – Gala Aug 26 '13 at 15:38
  • 1
    $\begingroup$ @AndreSilva, sorry yes tuberculosis (shows my health-centric view!). I'll change that in my question. $\endgroup$ – Tom Aug 28 '13 at 0:03
  • $\begingroup$ @Wayne: Yes, I'm comparing proportional rates (e.g. tuberculosis cases per 100 000 population per year). $\endgroup$ – Tom Aug 28 '13 at 0:04
0
$\begingroup$

Gaël Laurans you are correct! It is all about residuals and apparently this is a completely legitimate exercise of identifying the units with the most extreme residuals. It still seems slightly odd that you can go out and more or less "manually" fix the variation in the data so the model fits better (i.e. find tuberculosis cases in communities where the model doesn't fit as well). But apparently this is okay!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.