[EMAIL PROTECTED] wrote:
> A more technical set of questions
>
> (1) Is it proper to talk about independent and dependent variables in a 
> correlational study? And to what extent? Isn't it *more* correct to call the 
> variables predictor and criterion variables?What is the current status of 
> this 
> language?
>
>   
Experimental psychologists tend to be rather "fastidious" about this 
issue, but their usage is idiosyncratic. One doesn't see it ias much in 
other natural sciences, and it is completely absent in mathematics 
(where anything on the x axis is "independent" and anything on the y 
axis is "dependent).
> (2) I have learned that a rule of thumb for evaluating the effect size of a 
> significant correlation is to square r and this is a crude indicator of how 
> much of 
> the variability in the criterion variable comes from the predictor variable. 
> I'd like 
> to hear if this is too crude to be useable. Is there another, readily 
> calculable 
> effect size? I am very bothered by studies that make a big deal of a 
> significant 
> correlation of .2 or .3.
>   
This is the standard interpretation of r. It is difficult for most 
people not trained in stats to understand. As for low-but-significant 
correlations, you should be bothered by this (though there are 
situations where low- r0squared is misleading). On the other hand, these 
are the kinds of effect psychologists typically discover. Should they go 
unreported? 

In some circumstances (restricted range, nonlinearity, etc.), however, 
it is better to use a different measure of correlation altogether. 
Kendall's tau is best for ordinal data (much better than the "r-ish" 
Spearman coefficient, but its interpretation is entirely different (not 
based on variance, which makes no sense in the ordinal context). When 
one of the two variables is dichotomous but can be viewed as being a 
crude measurement of an underlying continuous variable, it is better to 
use the biserial (rather than the "r-ish" point-biserial). Better still 
is Lord's modification of the biserial, but it is rarely used and so 
needs to be explained. In 2x2 situations, especially where the cells are 
greatly unbalanced, it is often better to use the odds ratio rather than 
the "r-ish" phi. Its interpretation (increased probability of X also 
being a case of Y) is much more understandable for people not trained in 
stats. One finds it used in prospective medical research quite commonly 
(in which cells are often badly unbalanced because developing any 
particular disease -- e.g., a heart attack -- in a given period of time 
is a highly unlikely event).

Check out David Howell's chapter on "Alternative Correlational 
Techinques" for a good overview.

Regards,
Chris
-- 

Christopher D. Green
Department of Psychology
York University
Toronto, ON M3J 1P3
Canada

 

416-736-2100 ex. 66164
[EMAIL PROTECTED]
http://www.yorku.ca/christo/

==========================


---
To make changes to your subscription contact:

Bill Southerly ([EMAIL PROTECTED])

Reply via email to