Re: Comparing Spearmann coefficients from 2 independent samples

On May 16, 4:03 am, silent...@xxxxxxxxx wrote:
Dear all,
i have 2 independent samples (US and German). Both groups have been
asked the same set of questions to measure trust (t1, t2, t3) and
concern (c1, c2, c3). All variables (1, t2, t3, c1, c2, c3) were asked
on a 7-p Likert scales. I now need to make inferences on whether there
are significant differences in correlations (trust-concern) between US
and German samples.

To do it:
1) i form an index: Index_Trust=MEAN (t1, t2, t3) and
Index_Concern=MEAN (c1, c2, c3)
2) I checked whether my indecies are normally distributed: FOr US
sample they are, for German no. Hence i use SPEARMANN correlation.
IS IT CORRECT?

3) I get two correlation coefficients : one for US and one for
Germany. How can i compare them? If that would be Pearson correlations
then i could run this test http://faculty.vassar.edu/lowry/rdiff.html

But I have Spearmann correlation. How should i proceed?

I also read abstract of this publication:
http://mrw.interscience.wiley.com/emrw/9780471667193/ess/article/ess5050/current/abstract
and seems that they allow to "pretend" Spearmann is Pearson in
calculating significance of differences.....

Thank You!!!!!

I would recalculate the composites scores as sums rather than means,
so that they would have integer values (19-point scales), and then
treat them as having interval properties. The fact that the
distributions may be non-normal is irrelevant.

Look at the two 19 x 19 contingency tables and their marginals. Use
all 19 scale points. Don't let the computer pool them on you. Look at
both the actual numbers and graphic representations (histograms for
the marginals, jittered scatter plots for the joint distributions).

If the marginals differ much then any difference in the correlations
is going to be very difficult to interpret. (In both cases, the
differences I'm talking about are effect-size-type differences,
that do not depend on the sample sizes.) You may have to abandon
correlation as a measure of the abstract "strength of relation"
notion and try a regression approach.

Unless there is obvious nonlinearity, I would use the usual least-
squares linear regression/correlation statistics, with bootstrapped
standard errors if the non-normality is severe.

If you must resort to ordinal-level analyses then use Kendall's tau,
not Spearman's rho. For standard errors see

Cliff, N., & Charlin, V. (1991).
Variances and covariances of Kendall's tau and their estimation.
Multivariate Behavioral Research, 26, 693-707.
.

Relevant Pages