Timeline for Is a super high odds ratio unusable? [duplicate]
Current License: CC BY-SA 4.0
13 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 16, 2022 at 18:31 | history | closed | kjetil b halvorsen♦ | Duplicate of Enormous coefficients in logistic regression - what does it mean and what to do? | |
| Nov 16, 2022 at 18:31 | comment | added | kjetil b halvorsen♦ | This is probably complete or quasi-complete separation, as in the duplicate. If you think that is not the case, report back! | |
| Nov 11, 2019 at 13:04 | comment | added | James Connor | OK, so I'm not using R for my model, but I managed to figure out how to use glmnet. The outputted coefficients show only V4 with a value (all the others are dots) Correct me if I'm interpreting this wrong, but does that mean that the other three exposures don't cause a statistically significant difference? | |
| Nov 11, 2019 at 8:17 | comment | added | German Demidov | @JamesConnor I think it is correct. Try glmnet - if it will not work, then well, there are no meaningful predictors in your sample... (follow the tutorial and perform a cross-validation too) | |
| Nov 10, 2019 at 16:39 | comment | added | James Connor | So I check my VIFs, and they're all around 1.5. While this suggests some correlation, I don't think that's enough to be concerned about? Am I checked for mutlicollinearity the wrong way? | |
| Nov 10, 2019 at 16:15 | comment | added | German Demidov | Agree with @PeterFlom-ReinstateMonica . As far as I know, this approach is more stable against multicollinearity: sthda.com/english/articles/36-classification-methods-essentials/… - try glmnet (or just ridge approach). | |
| Nov 10, 2019 at 16:09 | comment | added | Peter Flom | It's quite likely that the independent variables are collinear. Have you checked? | |
| Nov 10, 2019 at 16:04 | comment | added | James Connor | Huh. So basically, those CIs mean that none of the four inputs have a statistically significant effect on the output? | |
| Nov 10, 2019 at 15:23 | comment | added | German Demidov | I think you already see that all of your CIs are including 1, no matter how huge the actual point estimate is. "Is this unusable?" - kind of, you may say that "there is no statistically significant difference in ORs" (you should think about the correct formulation of the conclusion, I provided a rough idea) | |
| Nov 10, 2019 at 15:21 | comment | added | James Connor | Thanks, I fixed my question to include the CIs. | |
| Nov 10, 2019 at 15:21 | history | edited | James Connor | CC BY-SA 4.0 |
added 207 characters in body
|
| Nov 10, 2019 at 15:00 | comment | added | German Demidov | The point estimates of ORs are not good. Always provide at least 95% Confidence Intervals - they also have drawbacks, but they are obviously better than point estimates. | |
| Nov 10, 2019 at 14:56 | history | asked | James Connor | CC BY-SA 4.0 |