My question is similar to Formula for 95% confidence interval for $R^2$ (I indeed want to corelate predicted with true values to obtain $R^2$).

The problem is their distribution is not normal. I wanted to use Cohen et al. (2003) approach. However I dig through Correlations redux Olkin and Finn (1995) and The mean and second moment coefficient, in samples from a normal population Wishart (1931), from which the formula originates. Wishart states explicitely:

[...] althought the utility of this quantity, for a distribution which is far from normal, is not so great as would at first sight appear.

Is there known a formula for such CI, or do I have to resort to bootstrap-like approach?

  • $\begingroup$ The formula, as you already note, would have to depend on the distribution: there is no universal formula, just as there isn't even a universal formula for the CI of a mean. Would you like to provide some specific distributional assumptions? $\endgroup$
    – whuber
    Feb 7 '18 at 21:56
  • $\begingroup$ Unfortunately the only thing I can assume is that number of samples is large (>> 100). In case of mean I think it may be enough due to the central limit theorem. But I have no idea how to use it in case of corelation. $\endgroup$
    – abukaj
    Feb 7 '18 at 22:06
  • $\begingroup$ @whuber I may also assume that both variance and mean of distributions are finite $\endgroup$
    – abukaj
    Feb 7 '18 at 22:25
  • $\begingroup$ That's not going to get you very far. In fact, you already made that assumption by supposing $R^2$ exists. $\endgroup$
    – whuber
    Feb 7 '18 at 22:44
  • $\begingroup$ @whuber then it seems bootstrap is the only option $\endgroup$
    – abukaj
    Feb 8 '18 at 13:26

Your Answer

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

Browse other questions tagged or ask your own question.