Odds ratio and confidence interval both 0?

#3
The Exp(B) = 0 in a logistic regression model and the lower limit of the confidence interval = .000 with no upper limit. The p-value is 0.99 and the Wald = .000. Any ideas what's going on?
 

hlsmith

Not a robit
#4
I can think of a few things that could have gone wrong, but likely the model had convergence issues related to small sample or sparsity problems. The only way for us to help diagnose the issue is if you post the output and code, and any errors (if generated). Feel free to provide context as well, which usually helps in understanding the inputs.

Thanks.
 
#5
I can think of a few things that could have gone wrong, but likely the model had convergence issues related to small sample or sparsity problems. The only way for us to help diagnose the issue is if you post the output and code, and any errors (if generated). Feel free to provide context as well, which usually helps in understanding the inputs.

Thanks.
I just checked and you were bang on. Thanks kindly! It was related to a very small number of people in one category of the independent variable with the binary outcome of interest. So one more question if I may. This independent variable was highly associated with other secondary outcomes so, for consistency, I really wanted to report it in this one instance too but not sure about reporting an Exp(B)= 0, with a lower limit CI of 0. Thoughts on reporting it?
 

hlsmith

Not a robit
#6
Yeah this scenario alone is a major contributor towards the reproducibility crisis. You see results and then change what you report. So it depends on what you are doing with the results. If you are going to share them with others, i suppose you could drop the non-significant term, but every time you present you have to disclose this wasnt your intended priori model.

also are you saying you want to change your primary outcome? If so this is an even bigger issue and you have to report that and maybe think about perhaps using a slightly lower alpha value for the confidence intervals to control for accidental findings.

Side comment, i am an epi/biostat person as well. Are you still in school or working?
 
#7
No my primary and secondary outcomes were decided a priori so I wouldn't change those. Just wondered how odd it would be to keep an Exp(B) of 0 in a table? Again, the same variable was associated with my other outcomes so I wanted to report it for consistency. I am working.
 

hlsmith

Not a robit
#8
You can keep it if there is theoretical reason for it to effect the exposure/outcome relationship. But i would likely drop it otherwise. If it theoretically has no relationship it is just creating an over complex model. Is your sample size large enough to support superfluous variables?
 
#9
All awesome suggestions. I have dropped it from the model as there were only 2 subjects with the outcome of interest which is why the OR, CI, and SE went crazy. Thanks very much!
 

ondansetron

TS Contributor
#10
I would look at the log again for errors. Just because the program gave you a number doesn't mean it's the answer. Many of the lower-level programs (point-and-click) will give you a number without making clear that it's not valid. Different programs will behave differently around numerical issues. Some programs might tell you there's and issue and refuse to report numbers, but other programs might be less obvious by providing a number from the last iteration that isn't the answer but is an intermediate step in the iterative process.

This is a case where I would be highly skeptical of the printed output in terms of whether the number(s) are even real. It's a huge red flag to have an infinite CI and like you're saying to see small numbers in groups. This needs more digging in my opinion before anything can be touted as a "result".
 

hlsmith

Not a robit
#11
Yeah, some programs force out an output and others say things like 'quasi-seperation', which I think perhaps Sander Greenland wrote a paper about recently. As noted, if this is not an important variable, its exclusion may be alright. If it is an important variable, it is very important to disclose its exclusion and you can even force it into the model by using exact or penalized approaches, but its CI's wont be infinity but equally obnoxious, and would also be a red flag for sparsity. With perhaps not the conditioning on the variable of interest that you would desire.