I have a correlation between two variables, and 40 groups in which I would like to compare it. Then extract several larger groups (reduce the number of subgroups) in which the correlations are "similar" as a result. I would like to use some technique that would define similar correlations in groups, I don't really have an idea now, and will check any suggestions.
I prefer SPSS / AMOS, but R also comes second. I have some experience in CFA / PCA, and some techniques based on the k-means method, and simple classifications.
Of course, the problem is more complex because I have many such variable systems (finally, 2x4 correlation matrix in 80 subgroups), but if it can be done simply, then it can be done on a more complex set of variables.
Thank you in advance for your suggestions.
Hi Thomas, great question!
I was faced with a similar problem a few years ago as I was interested in the association between symptoms of mania and depression in the population. My hunch was that the correlation between mania and depression would be determined by their level, that is, the correlation would be higher when the level of mania and depression was low, and the correlation would be low when the level was high - a kind of level by correlation interaction.
My solution was to use mixture modelling. I had a model with the two variables and allowed them to correlate and tested for different classes. I allowed the means of the variables and the correlation between them to vary across classes. What I found was 5 ordered classes where the strength of the correlation did indeed get lower for those classes that had higher mean scores on the two variables.
There's probably other, neater, ways to approach this problem, but this worked for me.
best wishes
mark
Dear Mark Shevlin
Thank You for that answer. I will try this method, although it seems inaccessible from the level of visual tools like SPSS, I will probably have to rummage in R. Do I understand right now that I would have to average the results in groups so that the analysis creates classes from among these groups, or can I try on raw data?
Hello Tomasz,
You could as well explore the use of agglomerative clustering, using the five variables of Mean of variable X, Mean of variable Y, SD of X, SD of Y, and XY correlation to identify subsets that are more or less homogeneous with respect to these characteristics.
Good luck with your work.
Hi Thomas , I'd use the raw data, and David's post alerted me to an error in my post. It should have read-
"I allowed the means and variances of the variables and the correlation between them to vary across classes".
OK, thank You for information. All these methods seems to be kinda too complicated. I think - i will just use obtained coefficients as a new dataset and then make clustering. :)
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS