Timeline for Logistic Regression: Does my model selection process make sense?
Current License: CC BY-SA 3.0
7 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jun 30, 2016 at 20:14 | comment | added | ybeybe | My understanding is, if interpretability is important, do not drop some levels from the categorical variable - it'll make the coefficients for the rest of the levels not interpretable. If prediction accuracy is key, and cross-validation suggests dropping some levels, you can probably do that. | |
| Jun 30, 2016 at 19:32 | comment | added | RobertF | VIF is variance inflation factor, and indicates how much the stnd error of variable coefficients is "inflated" due to multicollinearity between variables. Sometimes researchers will drop one or more variables with high VIF values. VIF is not as useful in predictive models, where multicollinear variables are often retained in the model (usually with smaller coefficients) to increase predictive accuracy. | |
| Jun 30, 2016 at 19:25 | comment | added | vdiddy | Okay thanks a lot. A very short and high-level explanation is that VIF values over 10 are indicators of multicollinearity. For a further deeper question on the categorical variables, I decided to ask another question. Maybe you can help with further clarity on that (minus the VIF parts) here: stats.stackexchange.com/questions/221540/… | |
| Jun 30, 2016 at 19:22 | comment | added | Kodiologist | (2) I'm not familiar with VIF. But, it's probably better to use a method that can handle collinearity (e.g., ridge regression) than to try to remove it. (6) Don't use area under the ROC curve. It's not good for anything. Proper scoring rules are a good replacement for model comparison. But yeah, if you don't have enough 1s for your purposes, there's not much you can do short of collecting more data. | |
| Jun 30, 2016 at 19:16 | comment | added | vdiddy | Thanks again for the note. Made some edits to my original post as well. | |
| Jun 30, 2016 at 19:10 | comment | added | vdiddy | Hey thank you so much. That was very helpful! Here are a few clarifying points: (2.) If I calculate VIFs for my variables (say ten categorical variables for one feature) and find high VIFs, that indicates multicollinearity, right? So I should take those out or make interaction variables? But I don't know how this applies to categorical variables. (6.) Let's say that my productivity (number of 1's) is so low that my results are bad. So basically I'm SOL, right? No way to "enhance" the data? | |
| Jun 30, 2016 at 18:57 | history | answered | Kodiologist | CC BY-SA 3.0 |