Re: Best Fit for Classification
- From: Graham Ashe <knight_armour@xxxxxxxxx>
- Date: Mon, 30 Jun 2008 18:41:15 EDT
Yes, with a caution. The sample correlation
coefficient is itself
random (a point estimate of the true correlation).
One can test whether
sample correlation A is really bigger than sample
correlation B. Just
eyeballing it and saying 0.15 is larger than 0.1
ignores the possibility
that random error inflated the 0.15 and deflated the
0.1. So you might
weasel by saying "appears to correlate more highly";
or you might do the
test.
What test is this? I thought determining the Spearman rho for each sample (i.e. set) against human assessment *was* the test. Is there another test to determine if the coefficient for one sample is indeed greater than that of another set?
Be careful not to confuse "statistical significance"
with "practical
significance". If a correlation of 0.1 is
statistically significant,
that means it is likely to point to a genuine
positive correlation
(which may or may not be around 0.1) and is unlikely
to be an artifact
of sampling error. It does not say that the two
measures covary enough
that anybody besides you would care. :-)
This is why I thought, based on what I've read, that a (statistically significant) correlation of r=0.6 or better is "generally considered" to be a "good" correlation. I suppose this is necessary when you're presenting information to non-statisticians. I was just curious if it's the same with the Spearman rho. I think the coefficient still needs to be "interpreted" somehow at the end of the day.
.
- Follow-Ups:
- Re: Best Fit for Classification
- From: Paul Rubin
- Re: Best Fit for Classification
- Prev by Date: Re: Relationship between Pearson rho and Kendall tau
- Next by Date: Re: Nonparametric covariance equality test
- Previous by thread: Re: only significant with interaction
- Next by thread: Re: Best Fit for Classification
- Index(es):
Relevant Pages
|
