Timeline for Multicollinearity between categorical and continuous variable
Current License: CC BY-SA 3.0
17 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Feb 23, 2013 at 15:44 | comment | added | Wayne | @tomas: I've thought and thought about your response to my comment and am still totally puzzled. The lmer I proposed will yield a fixed intercept, a by-month intercept, and a by-month effort slope. Add the fixed and by-month intercept and you have your monthly effect which I'd imagine is very close to what you get in your fixed effects model. There is no simple number for effort, it's broken out by month, so it's not as simple as a monthly fixed effect, but it gives you insight into whether the effect of effort varies by month which would seem to be important in my mind. | |
| Feb 2, 2013 at 14:16 | vote | accept | Tomas | ||
| Jan 29, 2013 at 1:30 | comment | added | Tomas | Thanks @Scortchi for useful links and for mentioning the problem with large coefficients! I have met the issue of large coefficients in another question, just when I thought that I filtered the multicollinearity out! Gosh!! | |
| Jan 28, 2013 at 17:43 | answer | added | Macro | timeline score: 6 | |
| Jan 28, 2013 at 17:05 | comment | added | Scortchi♦ | @Tomas: Well worry if you like, but you'll have some degree of collinearity in almost any observational data-set, so you'll be doing a lot of worrying. To my eyes this data-set doesn't look so bad, but if you want a quantitative assessment of the degree of collinearity calculate variance inflation factors & condition indices: stats.stackexchange.com/questions/16692/… stats.stackexchange.com/questions/4099/… | |
| Jan 28, 2013 at 15:29 | comment | added | Tomas | @Scortchi, I'm afraid you are wrong in this. The collinear variables will compete for the fraction of variability they explain. And I don't care just about how the whole model fits, but about the importance of each predictor variable. | |
| Jan 28, 2013 at 15:27 | comment | added | Tomas | @PeterFlom: actually, I do this, I haven't said that just to be brief. I have divided the year in two seasons, according to the biological sense. | |
| Jan 28, 2013 at 15:25 | comment | added | Peter Flom | Not related to your question, but I would think there would be a better way to model "month" than just as a factor. How exactly to do this would depend on why month has such an effect. | |
| Jan 28, 2013 at 15:21 | history | edited | Peter Flom | CC BY-SA 3.0 |
fixed typo
|
| Jan 28, 2013 at 14:33 | comment | added | Scortchi♦ | Both might be important predictors. There's no requirement for predictors to be perfectly orthogonal in generalized linear models, any more than there is in linear regression or ANOVA. If the predictors are too collinear the model fit will still be valid, but useless because the variances for individual model coefficients become enormous. When that doesn't happen you have no cause to worry. | |
| Jan 28, 2013 at 14:20 | history | edited | Tomas | CC BY-SA 3.0 |
added 11 characters in body
|
| Jan 28, 2013 at 14:12 | comment | added | Tomas | @Wayne, no, what you propose is something completely different. You have confused fixed effects for random. All these should be fixed, affecting the outcome, not its error structure. | |
| Jan 28, 2013 at 14:10 | comment | added | Tomas |
@Scortchi, not actually, pleasee see the updated answer. Why I do think there is a problem - this is obvious, isn't it? Effort and month will "compete" for variability that belongs to the other! So we don't know what is actually the important predictor.
|
|
| Jan 28, 2013 at 14:08 | comment | added | Wayne |
I would naively suggest trying package lme4's lmer, where something like lmer (cbind (young, adults) ~ (effort | month), family="binomial") might be what you want. Scortchi's comment sounds much more informed than I can be, though.
|
|
| Jan 28, 2013 at 14:07 | history | edited | Tomas | CC BY-SA 3.0 |
added 4614 characters in body
|
| Jan 28, 2013 at 13:59 | comment | added | Scortchi♦ |
Are you getting very high variances for coefficient estimates when both month & effort are in the model - compared to what you see when you put each in separately? If not, why do you think there's a problem?
|
|
| Jan 28, 2013 at 12:50 | history | asked | Tomas | CC BY-SA 3.0 |