Im attempting to do a multivariable logistic regression looking at the independent association of variables for an adverse event, however the result for my variable of interest is reading as unstable. I believe this is because the variable actually has a 100% sensitvity for my outcome (adverse event) on univariate analysis. If this is this case, does anyone have suggestions on what to do?
You have a binary response (adverse event present or not). You have multiple predictors in your model. Your estimates are unstable when you use nominal logistic regression. You might try instead to use GLM with the binomial distribution and logit link function. Be sure to select the Firth adjustment option as this will help if you have a separate problem.
Thank you for this answer!
I have a few follow up questions. When using GLM with binomial, is it alright that some of the my 2x2 tables have cells with values < 5?
Also, in this model, would you just report the estimate and SE and not an OR (like you would with nominal LR)?
One more question- does jmp have a way to report a RR if there are zeros in the 2x2? I have seen that some programs will add 0.5 or 1 to each cell.
Thank you again for your time
I am just wondering if anyone has any insight into this question. I have used GLM with the binomial distribution and logit link function and no longer have an "unstable" model. The primary predictor I am interested in is independently associated with my reponse variable (Adverse event vs No event) however now there are no CI reported. I have a total of 8 predictor variables for only 69 patients... When I remove some of the predictor variables from the model then the confidence intervals appear. My questions are...
1. Are the CI initially not being reported because there are too many predictor variables in the model
2. Is this issue still stemming from having 100% sensitivity for the outcome (the report states "evidence of perfect fit for some data points detected, and the Hessian matrix suggests quasi-complete separation of the data. Fit and results are of questionable value- proceed with caution".
3. Is there another more appropriate model I should be using?
Thank you again for all of you help
Just checking a few points:
Some thoughts are:
It could also be a problem with the data or the assumptions of the model and estimation method that you are using. Can't say for sure at this point.
I forgot to say that you might have a problem because of rare events, that is, response levels that are not observed for some of your predictor levels. It is related to the separation problem. The Firth adjustment to the estimation helps, but it can't always eliminate all of the problems if the rare event problem is severe.