I’m still learning regression, and I’m sure I’m not alone in this struggle. When working with logistic regression models, have you ever encountered extremely large odds ratios (ORs) and standard errors (SEs), accompanied by uninformative confidence intervals (CIs)? I did, and I’m here to share my experience and the lessons I learned.
In my case, I was fitting a generalized logistic regression model in R using the glmer function. The model looked like this: glmer(ftm ~ ft * condition + (1|Subject)). The data came from a mixed-methods design with repeated measures, so I included a random effects variable to account for the variation within subjects.
The model summary results were alarming: ORs in the millions, SEs in the hundreds of thousands, and infinite CIs. Something was clearly wrong. After digging deeper, I suspected that the issue might be due to quasi-separation in the data, which is a common problem in logistic regression.
Quasi-separation occurs when there’s a strong association between the predictor variables and the outcome, leading to unstable estimates and inflated ORs. In my case, the data showed near-separation in some cells, which was expected based on the experiment design.
So, what can you do when faced with such issues? Here are a few suggestions:
* Check for quasi-separation in your data. If you find it, consider using alternative models or techniques that can handle this issue, such as Bayesian logistic regression or penalized likelihood methods.
* Verify that your model is correctly specified. Make sure you’re not missing any important predictor variables or interactions.
* Consider using a different link function or a different type of regression model altogether.
Remember, regression models are only as good as the data and assumptions that go into them. By being aware of common pitfalls like quasi-separation, you can take steps to ensure that your results are more reliable and interpretable.
I hope this helps, and I’d love to hear about your experiences with logistic regression in the comments!