Correlation of different types of variables

[Kim Boehm]

Dear statalist,

I want to investigate possible relationships between different types of variables. If have got some continuous, some ordinal and one dichotomous (nominal with two options) variables. To be more precise, variables describe persons by for example their age, personality items (with 5 or 7 point Likert scale) and their gender (dichotomous). Moreover, I have got a mean of decisions to act over periods. For example a value of 0.50 for the case, that this person decided to do something in half of all periods.

I thought about using the spearman correlation as a non-parametric test. However, I'm unsure if I can do it with a dichotomous or in fact any nominal variable. What is your opinion on that?

Thank you very much!
Kim

[Clyde Schechter]

Yes, you can use Spearman with dichotomous and ordinal variables, but you cannot use it with nominal variables. In fact, you cannot do any kind of "correlation" with nominal variables: it's completely meaningless.

You might also want to look at tetrachoric and polychoric correlations. -tetrachoric- is an official Stata command. -findit polychoric- will show you a link where you can get Stas Kolenikov's polychoric.ado.

[Kim Boehm]

Hello Clyde,

thanks for your fast reply!

If I have got a correlation of gender (0 male / 1 female) and any ordinal variable of let's say -.38, it would be fair to say, that there is a relationship between being female and that certain score, which is negative. Therefore, women tend to have lower scores than men (or persons with lower scores tend to be female). Is that right? (So basically, I interpret the value like a pair of ordinal and/or continuous variables.)

If I got the ideas of tetrachoric and polychoric correlations right and due to the fact that I have got only one dichotomous and none nominal variables, I would stick with the spearmen correlation. Nonetheless, thanks for the idea!

[Clyde Schechter]

If I have got a correlation of gender (0 male / 1 female) and any ordinal variable of let's say -.38, it would be fair to say, that there is a relationship between being female and that certain score, which is negative. Therefore, women tend to have lower scores than men (or persons with lower scores tend to be female). Is that right? (So basically, I interpret the value like a pair of ordinal and/or continuous variables.)

Well, that is correct. But that is probably the most opaque way to state this relationship I can think of. Why would you not simply contrast the distributions of the score variable in the two genders using, say the Mann-Whitney U-test, or the median test, or something like that?

[Kim Boehm]

Simply because I have about 20+ continuous and ordinal variables for which I test a relationship with gender and between each other. So testing for this relationship is only a very small part of what I test for. Alternatively, I also thought about a U-test, but I would have to report 20 values for gender and still would need the spearman correlation matrix for the other variables. So I wanted to combine that.

Thanks again for your fast reply and help!