Timeline for Logistic regression: Strange standard errors from glm() in R
Current License: CC BY-SA 3.0
10 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| May 23, 2014 at 5:02 | history | edited | Sergio | CC BY-SA 3.0 |
deleted 11 characters in body
|
| May 22, 2014 at 22:29 | vote | accept | Hans Ekbrand | ||
| May 22, 2014 at 22:29 | comment | added | Hans Ekbrand | Thanks for the explanation and the links. Even if I don't understand the details, I do understand where I went wrong. To summarize: I used a faulty method when trying to calculate confidence intervals from the intercept model, which in turn lead me to belive that there was something wrong the standard errors of the non-intercept model. | |
| May 22, 2014 at 22:04 | comment | added | Sergio | Joint confidence "intervals" are actually confidence regions (see, e.g., here and here), this is why sampling is so convenient. | |
| May 22, 2014 at 21:43 | comment | added | Hans Ekbrand | Because I want to learn. What brought me into this was the desire to understand when I could do without an intercept term. Do you imply that there is no formula for me to understand in this case? Is sampling the only way? | |
| May 22, 2014 at 21:22 | comment | added | Sergio | Sorry, but... why? The models are numerically equivalent (this is what I wanted to highlight), but statistically different, they address different scientific questions. If you are interested in "relative" effects you have to a) choose a baseline (Buddhism or something else, but you shoud explain why one or another) and b) run the first model and not bother to run the second one. If you are interested in "absolute" effects (no baseline) you should run the second model and not bother to run the first one. IMHO you can and should compare several models only if they address the same question. | |
| May 22, 2014 at 20:54 | comment | added | Hans Ekbrand | Your answer is useful, and I am thankful for it, but to close the question I would like to see a formula that arrives at the lower bound for "Christianity" from the estimates in the intercept model. | |
| May 22, 2014 at 20:49 | comment | added | Hans Ekbrand | No, I don't. Thanks for the code, it gives a practical way to get to my goal, and it proves that both sets of standard errors are correct. However, I would like to have the formula too. For the second, non-intercept model, getting eg. the lower bound for "Christianity" is simple enough: $ -2.390448 = -2.378317 - 1.96 \times 0.006189045 $. My naive idea was to create the "combined" interval for the first model by $ -2.8718056 + 0.4934891 - 1.96 * 0.03234887 $, but that gave a much larger confidence interval. | |
| May 22, 2014 at 10:29 | history | edited | Sergio | CC BY-SA 3.0 |
added 11 characters in body
|
| May 22, 2014 at 8:49 | history | answered | Sergio | CC BY-SA 3.0 |