l***a 发帖数: 12410 | 1 with 24 variables I only got a R-sqr of 0.23. is it too low?
I know for linear regressions, I will feel confident if R-sqr is over 0.9 or
sometimes 0.8 or even 0.7.
normally how big R-sqr will make you feel comfortable? |
D******n 发帖数: 2836 | 2 logistic has r2 too?
or
【在 l***a 的大作中提到】 : with 24 variables I only got a R-sqr of 0.23. is it too low? : I know for linear regressions, I will feel confident if R-sqr is over 0.9 or : sometimes 0.8 or even 0.7. : normally how big R-sqr will make you feel comfortable?
|
d******r 发帖数: 1389 | 3 real data我觉得0.2不算太差了……
or
【在 l***a 的大作中提到】 : with 24 variables I only got a R-sqr of 0.23. is it too low? : I know for linear regressions, I will feel confident if R-sqr is over 0.9 or : sometimes 0.8 or even 0.7. : normally how big R-sqr will make you feel comfortable?
|
l***a 发帖数: 12410 | 4 yes. use rsquare option in model statement
【在 D******n 的大作中提到】 : logistic has r2 too? : : or
|
l***a 发帖数: 12410 | 5 thanks. the data is not too big (about 45k records) so I am afraid when I
regress on another data file with 4mil records, the r-sqr might get even
worse. plus I want to control the number independent var to under 20....
besides r-sqr, what other statistics you normally rely on measuring the fit
of the logistic model? some people like AIC but the AIC is just an absolute
value to me. i think relative number like xx% makes more sense....
【在 d******r 的大作中提到】 : real data我觉得0.2不算太差了…… : : or
|
B******5 发帖数: 4676 | 6 generalized吧。。。
【在 D******n 的大作中提到】 : logistic has r2 too? : : or
|
g*******y 发帖数: 380 | 7 I think I saw some posts recently, if it's not on this board.
The R2 in logistic model output is a pseudo-one.
【在 D******n 的大作中提到】 : logistic has r2 too? : : or
|
c****o 发帖数: 69 | 8 u can look at the area under ROC curve, and if it is close to 1, then the
model is a good fit
fit
absolute
【在 l***a 的大作中提到】 : thanks. the data is not too big (about 45k records) so I am afraid when I : regress on another data file with 4mil records, the r-sqr might get even : worse. plus I want to control the number independent var to under 20.... : besides r-sqr, what other statistics you normally rely on measuring the fit : of the logistic model? some people like AIC but the AIC is just an absolute : value to me. i think relative number like xx% makes more sense....
|
z**k 发帖数: 378 | 9 I believe ROC is better ... never touched the R-sq for logistic regression
before ...
btw, is that a film camera? looks like a fisheye?
0.9 or
【在 l***a 的大作中提到】 : with 24 variables I only got a R-sqr of 0.23. is it too low? : I know for linear regressions, I will feel confident if R-sqr is over 0.9 or : sometimes 0.8 or even 0.7. : normally how big R-sqr will make you feel comfortable?
|
l***a 发帖数: 12410 | 10 thanks. to be specific, just to observe the curve _senit_*_1mspec_, the
closer the curve approaches 1, the better the model, right?
【在 c****o 的大作中提到】 : u can look at the area under ROC curve, and if it is close to 1, then the : model is a good fit : : fit : absolute
|
|
|
l***a 发帖数: 12410 | 11 thank you as well.
yes the camera is a film one. contax G2 with Hologon 16/8 lens. It's an
ultra wide angle lens, not a fisheye tho :)
【在 z**k 的大作中提到】 : I believe ROC is better ... never touched the R-sq for logistic regression : before ... : btw, is that a film camera? looks like a fisheye? : : 0.9 or
|
D******n 发帖数: 2836 | 12 a curve doesnt approach any scalar value. a curve is a curve.
it is the AUC(area under the roc curve) should be as close as possible to
one.
【在 l***a 的大作中提到】 : thanks. to be specific, just to observe the curve _senit_*_1mspec_, the : closer the curve approaches 1, the better the model, right?
|
j*****e 发帖数: 182 | 13 The best curve is the line segment over the unit interval, who interceps
with the vertical axis at 1. It is only at this situation that the AUC is
one. |
j*****e 发帖数: 182 | 14 I forget to mention that the lowest value of AUC is 0.5. You can get AUC
directly from SAS proc logistic.
If your predictors are all discrete, you can test model goodness of fit
using the deviance. Otherwise, there is no test for it.
Also, deviance over the degree of freedom in SAS output offers sort of
measurement on how the model is. If the model fits, this ratio should be
close to one. |
l***a 发帖数: 12410 | 15 thanks a lot.
my sas is 9.1.3 so the logistic procedure doesn't have the ROC statment
integrated yet. and since we didn't buy all the softwares, the proc iml is
also not available, which is called in the ROC macro provided by SAS.
but I do see there is "c" shown in the output of proc logistic and it's
related to condordance rate. is this the AUC provided by proc logistic your
mentioned?
I also tried to do the simple calculation of AUC based on the outroc file by
interporation, the result looks c
【在 j*****e 的大作中提到】 : I forget to mention that the lowest value of AUC is 0.5. You can get AUC : directly from SAS proc logistic. : If your predictors are all discrete, you can test model goodness of fit : using the deviance. Otherwise, there is no test for it. : Also, deviance over the degree of freedom in SAS output offers sort of : measurement on how the model is. If the model fits, this ratio should be : close to one.
|
d******r 发帖数: 1389 | 16 偶的SAS也是9.1,俺有ROC curve来的
your
by
also
【在 l***a 的大作中提到】 : thanks a lot. : my sas is 9.1.3 so the logistic procedure doesn't have the ROC statment : integrated yet. and since we didn't buy all the softwares, the proc iml is : also not available, which is called in the ROC macro provided by SAS. : but I do see there is "c" shown in the output of proc logistic and it's : related to condordance rate. is this the AUC provided by proc logistic your : mentioned? : I also tried to do the simple calculation of AUC based on the outroc file by : interporation, the result looks c
|
j*****e 发帖数: 182 | 17 To get the ROC, you should use the outroc option in proc logistic and plot
the output data using proc gplot.
The c value in SAS output corresponds to AUC.
AUC is used to compares two models. It is hard to make a conclusion based a
single value. And it wouldn't give you a parcimonious model.
As I said, deviance over degree-of-freedom is the best way to measure fit.
The 90% you mentioned seems to comes out of a cross-classification table.
What cut-off value did you use to make a decision? Remember |