I am dealing with a logistic regression with quasi-complete separation. I used Firth's logistic regression to deal with this, and everything seems to be right, except that I have two predictors with the upper confidence limit of Odds Ratio being quite large (i.e. around 120 in both cases). The punctual value of Odds Ratio seems to be "normal" (i.e. around 5 in both cases).
Is this correct or is there something wrong with my regression?
Thanks
If you ran the model correctly then yes the upper 95% confidence interval should be the same as if you stratified the data and calculated the odds ratio. Watch the stansard errors. If they blow up, your model is not specified correctly.
DId you check the relationship between those two predictors and the outcome? This could be the result of a near perfect relationship (where everyone in the particular cell has positive values for both the predictor and outcome).
If ALL your predictors have wide confidence intervals then you just have too little data. If JUSTthese two explanvars have extremely wide confidence intervals then I am thinking that these two explanvars are too strongly correlated. Even categorical variables can have colinearity problems.
I am not familiar with R, but in Stata you will get a notification if there is collinearity between covariates in the model and one of them will be automatically excluded.
It's important to keep in mind that the ORs and their CIs are just exp version of the LR coefficients and their standard errors. In the spirit of what's already been suggested I would do a stepwise model to see what happens to each of the LR coefficients and their standard errors. It should become immediately apparent! Good luck!
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R Statistical Package