4
$\begingroup$

I have a basic question, which hopefully you all can resolve for me. What is the best way to determine correlations between 3 or more variables? I have a dataset in which 5 continuous variables each correlate (Pearson, R ~= 0.7) with another continuous variable (which I'll call Z).

I know I can use partial correlation between groups of 2 variables and Z, but I don't think this is exactly what I'm looking for, if I understand it correctly.

$\endgroup$
6
  • 1
    $\begingroup$ You may be looking for the coefficient of determination of a regression of $z$ on $x_1, \dots, x_5$: en.wikipedia.org/wiki/Coefficient_of_determination But beware, interpreting it can be misleading. $\endgroup$ Jan 17, 2013 at 19:30
  • $\begingroup$ If I read that correctly, the coefficient of determination gives an idea about how several measurements correlate together with a single variable (Z in this case). What if I want to get a feel for the relationship between these other variables, without reference to Z? $\endgroup$
    – learner
    Jan 17, 2013 at 21:01
  • 2
    $\begingroup$ @learner, if you want summarize the strengths of intercorrelations among p variables in a single number you could compute, for example, (1) average abs. correlation or (2) geometric mean abs. correlation, or (3) determinant of abs(corr. matrix). These three are very different ways to conceptualize p-variate association by a single value. $\endgroup$
    – ttnphns
    Jan 18, 2013 at 9:45
  • $\begingroup$ @ttnphns, I have heard that the difference between the arithmetic mean and the geometric mean is that the geometric mean reduces the impact of unusually high and low values in the quantities examined (correlations, here). What insight is the determinant of the correlation matrix providing? $\endgroup$
    – learner
    Jan 18, 2013 at 16:32
  • 1
    $\begingroup$ Geometric mean has also another obvious sense: the product is really big when its factors are big jointly. Determinant is the "volume" of the correlation matrix, - the product of its eigenvalues. When correlations are small, det is high; when either they are big or there is collinearities, det is close to 0. $\endgroup$
    – ttnphns
    Jan 18, 2013 at 18:06

1 Answer 1

4
$\begingroup$

Since you seem to be saying that a matrix of all 10 correlations (those that don't involve z) is insufficient, it sounds as if you want to distill all these correlations into some underlying themes. The way to do that might be with factor analysis or principal component analysis, depending largely on whether your variables are measured with or without much error. Multidimensional scaling could be used as well. All of these techniques typically require a fair amount of expertise, so you may want to ask the help of a mentor or consultant.

$\endgroup$
2
  • $\begingroup$ Am I right in thinking that I would want to use factor analysis in instances where I expect a high degree of error? $\endgroup$
    – learner
    Jan 18, 2013 at 16:26
  • 1
    $\begingroup$ Right, relatively speaking. See stats.stackexchange.com/questions/1576/… $\endgroup$
    – rolando2
    Jan 18, 2013 at 17:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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