I am using a multilevel (3 level) logistic regression model to compare the proportions of statin use in an earlier quarter with the current quarter. level 3 - practice level; level 2 pt level.
I built a mulitlevel model, including a level of practices and my ORs were of moderate size. But realising the data was not independent, I then included patient level as an additional level.
The output has generated quite large ORs. For instance, you are 30 times more likely to be prescribed statins in the current quarter, if you are diabetic. When i ran a logistic regression and a 2 level model, being diabetic yielded high ORs, but not as high as 30.
I have checked the data/code and everything seems to be in order.
How can I interpret these inflated ORs, in relation to including patient level to my model?
Look forward to hearing from you
Thanks
This is possible, but I'd want to check the model (e.g., for convergence etc.) and also look closely at the interval estimates.
I have displayed the model here:
https://www.statalist.org/forums/forum/general-stata-discussion/general/1443467-inflated-ors
Hi Luca,
Sorry my email did not prompt me that anyone had replied until now.
My dependent variable is binary 0/1
if i understand correctly you have, on average, 1.4 occasions per patient, meaning that the majority of your level-2 clusters (patient level) are only observed once. i think this is the reason why your ORs are so extreme.
I have included level 2 pt level to account for my data being non independent
Can you provide a simple cross tab of binary statin against binary diabetes as that is the relation in question , is it not?
ta statins diabetes
| diabetes
statins | 0 1 | Total
-----------+----------------------+----------
0 | 1,480 249 | 1,729
1 | 3,341 3,802 | 7,143
-----------+----------------------+----------
Total | 4,821 4,051 | 8,872
- Badges
- Science topic
- Similar topics
- Correction