Improvement of predicitve power in logistic regression - when is it significant?

[Lara87]

I've posted this in another forum before but didn't get answers, so I try it here again

I'm running a hierarchical logistic regression to find out if/how block 2 improves the predicted power in addition to block 1, so what I need is the value that is equal to ΔR2 in blockwise multiple regression.

Is it right that the difference from the 2-log-likelihood (2-ll) of my second block and the 2-ll from the first block tells me about the gain of predictive power? How can I find out if this gain is significant?

Thanks in advance for your help!

 

[Englund]

Is it of theoretical value to decide whether it is statistically significant? You could always cross-validate the models on out-of-sample data in order to determine if the second block improves the predictive power.

 

[Lara87]

Yes, it is very important. I don't have much clue about statistics so I don't really understand what you mean by crossvalidating on a out-of-sample data. Does it mean that I need to find another sample and take the same measurements? That is unfortunately not possible.
Isn't it also possible, to crossvalidate by splitting the original sample? I just don't know if it makes sense, because my sample is very small (n=42).

Another thing I've read, which is about the 2-log-likelihood again. The difference of the 2-ll from block 1 to block 2 is a chi-square distribution, so it would be possible to find out if it's significant or not through that value. I just don't know HOW. The difference between both 2-ll is 20.01 with 2 degrees of freedom. But then what?

 

[Karabiner]

Is it right that the difference from the 2-log-likelihood (2-ll) of my second block and the 2-ll from the first block tells me about the gain of predictive power? How can I find out if this gain is significant?

Do a search for the likelihood ratio test.

With kind regards

K.