Re: Is there a statistical test for comparing two sets of values (which are statistics)?

James wrote:
Hi, Ray,

If what you want to talk about are the means of the two populations
that you have sampled then the t-test will probably be OK. The fact
that the population elements are correlations is relevant only to the
extent that it might suggest heteroscedasticity. If you describe the
data in more detail, you might get a more definite answer.

Thanks for concuring that "it it might suggest heteroscedasticity".
This is the reason why I suggest using Mann-Whitney test to compare
the two.

The Mann-Whitney test responds to heteroscedasticity the same way that
the t-test does; that is, it makes the same kind of errors.


Anyway, I don;t understand you mean by "If you describe the
data in more detail, you might get a more definite answer."

Can you enlighten me on this? Thx

I think that what was uppermost in my mind was the relative sizes of
two sources of variability: error in assessing each person's(?) true
correlation, and within-group person-to-person variability in the true
correlations. If the person-to-person variability ought to be close to
zero then it might be advisable to Fisher-transform (i.e., arctanh) the
correlations before testing. But if the person-to-person variability is
large then transforming would be a bad idea.

.

Relevant Pages

  • Re: Is there a statistical test for comparing two sets of values (which are statistics)?
    ... that you have sampled then the t-test will probably be OK. ... that the population elements are correlations is relevant only to the ... Thanks for concuring that "it it might suggest heteroscedasticity". ...
    (sci.math)
  • Re: PCA: uncorrelated items with high loadings on the same component
    ... is that they explain variability AFTER the variability explained by ... dimensions 1 and 2 have their variability removed from the data, ... the residuals after 2 dimensions and see if x1 and x10 are now highly ...
    (sci.stat.consult)
  • Re: efficient computation of many univariate least squares fits
    ... The suggestion in a different subthread to use the master correlation matrix may be a good alternative, particularly if you're using something that has a built-in routine for calculating correlations and it has a built-in t-test for the significance of the correlations. ...
    (sci.stat.math)
  • Re: tails and variance
    ... >> I am doing correlations as well, but I still want to do a t-test to see ... > particular age-related cutoff point for clinical purposes? ...
    (sci.stat.consult)
  • Re: tails and variance
    ... > I am doing correlations as well, but I still want to do a t-test to see ... particular age-related cutoff point for clinical purposes? ...
    (sci.stat.consult)