Timeline for Problem calculating, interpreting regsubsets and general questions about model selection procedure
Current License: CC BY-SA 3.0
18 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Dec 26, 2012 at 22:00 | answer | added | Greg Snow | timeline score: 8 | |
| Sep 27, 2012 at 0:09 | comment | added | gung - Reinstate Monica | @user1690846, if you want to understand better why this is not a strategy that is likely to work well in the long run, you might want to read my answer here: algorithms-for-automatic-model-selection. | |
| Sep 27, 2012 at 0:07 | history | edited | gung - Reinstate Monica | CC BY-SA 3.0 |
added tags; loaded figure; fixed code to show in window; formatted; light editing
|
| Sep 26, 2012 at 13:43 | answer | added | Michael R. Chernick | timeline score: 2 | |
| Sep 26, 2012 at 11:41 | comment | added | user1690846 | well actually I want to reconstruct the regsubsets stuff because I do not understand what it actually does and I could not find a good description (yeah I now, there is a manual but this does not help me that much) | |
| Sep 26, 2012 at 11:38 | comment | added | mark999 |
summary(lm(Gesamt ~ ExpHealth, data = subset(olympiadaten, !is.na(CommunistSocialist) & !is.na(CountrySize) & !is.na(GNI) & !is.na(Lifeexp) & !is.na(Schoolyears) & !is.na(ExpMilitary) & !is.na(Mortality) & !is.na(PopPoverty) & !is.na(PopTotal) & !is.na(ExpEdu) & !is.na(ExpHealth))))
|
|
| Sep 26, 2012 at 11:34 | comment | added | mark999 |
I'm guessing that the difference is because your data frame contains more variables than just the ones you used in regsubsets, so you've removed too many rows. And anyway, I don't see why you would want to replicate the adjusted $R^2$ that regsubsets gives, unless it's just to understand how it was obtained.
|
|
| Sep 26, 2012 at 11:26 | comment | added | user1690846 |
ok @mark999, I did the estimation with the missing values deleted in the following way: > olympiadaten2<-na.omit(olympiadaten2) > attach(olympiadaten2) > Gesamt<-olympiadaten2$Gesamt > ExpHealth<-olympiadaten2$ExpHealth > summary(lm(Gesamt~ExpHealth)) but the problem is, now I get an adj. R squared of 0.009202, which is still not correct (and even more worse)?
|
|
| Sep 26, 2012 at 11:21 | comment | added | mark999 | @user1690846 See Peter Flom's answer as suggested above, and/or look at Frank Harrell's "Regression Modeling Strategies" book, and/or google "harrell stepwise". | |
| Sep 26, 2012 at 11:18 | comment | added | user1690846 | @mark999 first of all thanks for an answer, but why is this a poor method? And is selecting with AIC better? So should I fitt the model by using na.omit(olympiadaten) ? If anyone has an answer to the other questions any futher answers would be very appreciated, thanks | |
| Sep 26, 2012 at 11:14 | comment | added | mark999 | @user1690846 I recommend looking at Peter Flom's answer to stats.stackexchange.com/questions/8303/… | |
| Sep 26, 2012 at 11:11 | comment | added | mark999 | Thanks @MichaelChernick but I prefer just to leave it as a comment. | |
| Sep 26, 2012 at 11:11 | history | edited | Michael R. Chernick | CC BY-SA 3.0 |
added 29 characters in body
|
| Sep 26, 2012 at 11:06 | comment | added | Michael R. Chernick | @mark999 Your comments are good and it looks like it gives the right answer. You should convert it to an answer. | |
| Sep 26, 2012 at 10:29 | comment | added | mark999 | The difference in adjusted $R^2$ is because some of the variables have missing values. I believe you would get the same adjusted $R^2$ if you fitted the model "manually" just using the subset of the data for which all the variables (in the formula in regsubsets) are non-missing. Note: choosing your model using regsubsets is considered to be a poor method. | |
| Sep 26, 2012 at 10:00 | history | edited | Peter Flom | CC BY-SA 3.0 |
fixed spelling
|
| Sep 26, 2012 at 8:56 | review | First posts | |||
| Sep 28, 2012 at 16:36 | |||||
| Sep 26, 2012 at 8:53 | history | asked | user1690846 | CC BY-SA 3.0 |