3
$\begingroup$

For the descriptive purposes, I am looking at the (Pearson) correlation coefficients between 6 variables over 4 periods separately. Thus, I computed the correlation between 15 pairs. I am doing a cross-sectional study before I start the longitudinal analysis (i.e. over 4 periods). However, as I am currently reporting the correlation coefficients for each time point (i.e. "cross-section"), it feels weird not to say anything about how they seem to evolve over time.

Now, it is easy to miss something or give a biased / subjective view when comparing 15 coefficients with 15 coefficients for example. Hence, I thought it would probably make sense to take the sum of the correlations for each variable and plot it as a line graph over time.

Does it make sense to do this; purely for descriptive purposes? So I get a picture of each variable how its overall correlation (relative to other items at the same time point) changes over time?

Improve this question
$\endgroup$
2
  • 3
    $\begingroup$ Instead of just downvoting, it'd be helpful to actually give a reason for it. Imo, the question is clear and I searched on stackexchange for similar questions but could not find an answer there. If it's unclear, I am happy to reformulate.. $\endgroup$
    – Amonet
    Commented May 20, 2018 at 11:45
  • $\begingroup$ It's not clear to me why your question attracted a downvote without a comment; while there are ways the question might be better, it's not clear what the downvoter's particular objection might have been. $\endgroup$
    – Glen_b
    Commented May 21, 2018 at 12:12

1 Answer 1

Reset to default
2
$\begingroup$

No, I recommend against that. Correlation coefficients are not on an interval scale, so summing, say, .5+.5 will reflect something different from summing .2 + .8, much less 0+1.0.

If you want to do this, computing Fisher's z first would help it make more sense. If you want to keep it in correlation metric, you could average the zs (instead of summing) and back-translate the mean to Pearson's r.

Improve this answer
$\endgroup$
2
  • 3
    $\begingroup$ The Fisher's z is used in meta-analysis for this very reason. However, in this application I think summing the correlations or Fisher's z is problematic as it assumes independence of the 15 correlations (which is clearly not the case). As such, the sum has no clear interpretation. I recommend just plotting a separate line for each correlation over the 15-time points. $\endgroup$
    – dbwilson
    Commented May 20, 2018 at 13:04
  • $\begingroup$ Thanks for your suggestions. @dbwilson there are four time points, but indeed there is clear dependence over time since I measure the same variables over time. I thought of plotting a separate line for each pair over time as well, but that'd be 15 lines (as I think you meant to suggest). However, if that's the only thing that makes sense, I will do that :) $\endgroup$
    – Amonet
    Commented May 20, 2018 at 13:10

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.