3
$\begingroup$

More precisely I mean "All other things being equal, what impact does including more predictors ('IVs') have on the CI for R squared?"

Say we'll get an R squared of .30 no matter whether we include 4 predictor variables or 7 predictor variables. Would one definitely produce a more precise CI than the other, or tend to produce a more precise CI than the other? Or is the number of predictor variables unrelated to the precision of the CI?

$\endgroup$
3
$\begingroup$

There is an approximate formula for the standard error of $R^2$ in the CI.Rsq function in the psychometric package in R (and probably other places, too).

It uses the s

$(\frac{4R^2(1-R^2)^2(n-k-1)^2)}{((n^2-1)*(n+3))})^.5$

where k is the number of predictors. So, with the same sample size and $R^2$, more predictors reduces the numerator, making the SE smaller.

Intuitively, if we had $n-1$ predictors we would have a 'perfect' model and the SE would be 0; but this would be grossly overfit to the sample.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.