Timeline for Identical coefficients estimated in Poisson vs Quasi-Poisson model
Current License: CC BY-SA 3.0
21 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 17, 2017 at 10:53 | history | edited | amoeba |
edited tags
|
|
| Nov 10, 2017 at 11:22 | history | edited | amoeba | CC BY-SA 3.0 |
edited tags; edited title
|
| Apr 13, 2017 at 12:44 | history | edited | CommunityBot |
replaced http://stats.stackexchange.com/ with https://stats.stackexchange.com/
|
|
| Oct 15, 2015 at 7:57 | history | tweeted | twitter.com/StackStats/status/654566697794465792 | ||
| Oct 14, 2015 at 19:08 | history | edited | Frank H. | CC BY-SA 3.0 |
deleted 63 characters in body
|
| Oct 14, 2015 at 19:07 | comment | added | Frank H. | In my original post, I added links to useful SE threads on this topic. The 2nd thread notes that coefficients should be the same if look at rate versus using exposure as an offset, yet in my case the coefficients are different. Any idea why this might be? | |
| Oct 14, 2015 at 18:51 | history | edited | Frank H. | CC BY-SA 3.0 |
added 545 characters in body
|
| Oct 14, 2015 at 16:21 | history | edited | Frank H. | CC BY-SA 3.0 |
brought additional information from comments into original post; re-ordered notes
|
| Oct 14, 2015 at 16:21 | comment | added | Ben Bolker | It's practical, although most important when doing Poisson (not quasi-Poisson) modeling. I don't know of a good reference offhand; if you can't find a relevant answer here on CrossValidated, it would make a fine follow-up question. | |
| Oct 14, 2015 at 16:20 | history | edited | Frank H. | CC BY-SA 3.0 |
brought additional information from comments into original post
|
| Oct 14, 2015 at 16:14 | comment | added | Frank H. | I did some googling but couldn't find out what the difference is between scaling the dependent variable and using an offset. (The best I found so far is casact.org/pubs/forum/09wforum/yan_et_al.pdf.) Is the difference practical or theoretical? Please point me to a useful link - Thank you! | |
| Oct 14, 2015 at 15:54 | comment | added | Ben Bolker |
you should not model frequencies/rates by computing ratios of counts/exposure. Rather, you should add an offset (offset(log(exposure))) term to your models.
|
|
| Oct 14, 2015 at 15:52 | comment | added | Frank H. | I am specifically modeling claim frequency, which is the actual claim counts divided by an underlying exposure e.g. revenue, employees, vehicle counts. | |
| Oct 14, 2015 at 15:49 | vote | accept | Frank H. | ||
| Oct 14, 2015 at 15:37 | comment | added | Ben Bolker | can you explain a bit more about where the non-integer values in your data come from ?? | |
| Oct 14, 2015 at 15:29 | answer | added | Ben Bolker | timeline score: 27 | |
| Oct 14, 2015 at 15:27 | history | edited | Frank H. |
edited tags
|
|
| Oct 14, 2015 at 15:26 | comment | added | Frank H. | Tried Tweedie from the get-go but our loss data is not ground-up, but rather on an excess basis. Also tried Negative Binomial, ZIP, and hurdle models to address the count dispersion. | |
| Oct 14, 2015 at 14:37 | history | migrated | from stackoverflow.com (revisions) | ||
| Oct 14, 2015 at 14:19 | comment | added | duffymo | Wouldn't a Tweedie distribution be a better idea? | |
| Oct 14, 2015 at 14:16 | history | asked | Frank H. | CC BY-SA 3.0 |