multivariate correlation

[#1]

Hi.

I have about 40 variable for analyzing in the multivariate correlation. some of them need parametric estimation and some of them need nonparametric estimation. which analysis should i preform (spearman, REML, pairwise etc...)  

thanks!

[#2]

Without a lot more details regarding the actual data and conclusions or information you are seeking in the context of your practical problem it's difficult to say 'which' method you should perform. And that in and of itself is a question...why limit yourself to 'one' method when you have lots of options for examining correlation among variables in a multivariate setting in JMP? So naively, I'd start with just a good old fashioned pairwise graphical approach and then add on methods from there as you explore your data. That's half the fun and great features of JMP...you can explore as you go with no 'one way' cookie cutter only approach. And don't discount higher order correlation analysis such as principal components...there might be informational gold there as well. And whatever you do, 'avoid' mononumerosis and visualize the relationships you seek graphically first, second, and third. Again this is where JMP shines.

[#3]

thank you very much for your replay. actually i am asking for publishing this results in a paper. if, for example one of the methods is more "conserved" or stringent,  i think it's may be better to use it.

basically the question which of the models is more "safe" to use.

thanks!  

[#4 · Accepted solution]

I encourage you to not think of statistical methods in general with words like '...conserved or stringent', or '...safe...' since these words have no meaning without context. I think your time can be better served by investigating the different methods you are considering wrt to their assumptions, methodology, in the context of your specific data (continuous/categorical/discrete numeric/text/multiple response and on and on,,,,JMP supports all these in the 'correlation' space), the method used to collect it, what are the representation risks (for example, relevance of the sample frame wrt to populations of interest) associated with the data, then pick one or more methods that are most aligned with your specific context.

The other completely separate issue you raise is you are looking to 'publish'. OK,  publish for what purpose? Reason I ask is if you say publishing for something like regulatory compliance there may be guidelines for analytical approaches that are either generally used or specified for compliance evaluation.

[#5]

thank you ! actually this point of view sounds reasonable, I think I will encourage it.  

[#6]

Just to add to Pete's comments.  There is no one "right way".  Different statisticians will have their bias as to preferred method.  Some publishing companies have their bias (usually a function of the statisticians that advise them).  That does not make it right or invalidate other methods.  Quantitative analysis is particularly impacted by assumptions.  Graphical less so.  Analysis should always follow this order:

1. Practical (does it makes sense from the SME point of view)

2. Graphical (looking for patterns and outliers in the data)

3. Quantitative