Thanks for the advice Jay.
Regarding the listserv: I remembered not being able to post in the past because I was trying to send rich text and I thought that maybe sending a link was unacceptable as well. I usually stay pretty calm when I don't instantly see my question posted.
Regarding variance inflation factors: I just tried out the -collin- function for fun. Is there a generally accepted VIF value that should raise a red flag?... Or is this something that people argue about too? I seem to recall that VIF's above 10 are supposed to be bad, right?
----- Original Message ----
From: "Verkuilen, Jay" <[email protected]>
To: [email protected]
Sent: Thursday, November 20, 2008 2:31:36 PM
Subject: RE: st: multicollinearity
Chris Witte wrote:
>>Thanks Michael. I knew about the multiple posts, but after having waited for over 5 hours without my question being posted I thought that maybe the server did not like my inclusion of the link, which is why my third attempt did not include the question about the UCLA example.<<
Sometimes the listserv is a little slow but it gets there eventually. Also, answers can come slowly too depending on when people read the listserv.
>>Thanks for the multicollinearity information! It makes sense that Stata would take this approach, as the acceptable amount of multicollinearity seems rather subjective. I've been taught that correlations > 0.70 is something to be concerned with, but I'm sure that many other people would suggest different values. I'm in the field of fisheries biology, and usually deal with relatively small sample sizes.<<
I work in psych where sample sizes are also relatively small. I wouldn't recommend looking at correlations of .70 or more though pro forma because the situation is rather more complex than that. For a simple treatment look at the variance inflation factors (VIFs), which are calculated by various collinearity diagnostics such as are found in -collin-. Cook's distance measures are helpful for identifying problem cases.
The big problem you have with collinearity is separating out the effects of different variables on the DV. In a standard factorial experiment this comes about by the way you did the design because orthogonality imposes 0 collinearity---that's the whole point! With observational variables, though, it's not going to happen.
JV
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
