Decision Between Multiple Regression and Canonical Correlation
Locked
 
|
I'm assisting a colleague with a study of the relationship of ethnic identity and traditionality-modernity to acculturation. Acculturation is the dependent variable. The scores
used in prior research include two subscales - traditionality and modernity; and a full scale acculturation score, which is determined by modernity minus traditionality. Not at all surprising is the fact that while the two subscales have a very low correlation,
they both correlate highly to the overall acculturation score. So, because of the colinearity aspect, I'm concerned about including them all as possible DV's in a canonical regression analysis. If parsimony is an issue, then I could simply use the overall
score and do a multiple regression. I'm also aware that selection of scales to use based on colinearity is discussed only for IV's. So...my moderately informed opinion is to go with the two subcales in the canonical correlation analysis. Thanks for any input.
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD
Brian
|
Re: Decision Between Multiple Regression and Canonical Correlation
|
|
Canonical gives you a test for "any combination of the IVs" as criterion.
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD
Is that useful for you? If you have oodles of excess power, then the canonical correlation is fine for an overall test. Coefficients will show you if the first root is similar to the Acculturation Score -- which you are then justified in analyzing as a follow-up -- and you also have the possibility of defining a second root. Obviously, a regression with one DV is more powerful than having two. Do you have enough power so that your regression would be powerful, even if it had twice the IVs as what you have? - That is a crude guess at how much power is lost by using 2 d.f. for an outcome rather than one. If power is a problem and you don't mind missing the wider test, then you want the single regression Besides power, the other thing that you gain from a regression is a clear statement of "Acculturation" as a criterion for your hypothesis. Is that something you want? The main thing that you get from the two DVs being /highly/ correlated (when they are) is that it might be clearer to the Investigator on a-priori grounds whether a composite outcome is expected: whether that is the sum or the difference score. -- Rich Ulrich Date: Wed, 14 Oct 2015 22:04:02 +0000 From: [hidden email] Subject: Decision Between Multiple Regression and Canonical Correlation To: [hidden email] I'm assisting a colleague with a study of the relationship of ethnic identity and traditionality-modernity to acculturation. Acculturation is the dependent variable. The scores
used in prior research include two subscales - traditionality and modernity; and a full scale acculturation score, which is determined by modernity minus traditionality. Not at all surprising is the fact that while the two subscales have a very low correlation,
they both correlate highly to the overall acculturation score. So, because of the colinearity aspect, I'm concerned about including them all as possible DV's in a canonical regression analysis. If parsimony is an issue, then I could simply use the overall
score and do a multiple regression. I'm also aware that selection of scales to use based on colinearity is discussed only for IV's. So...my moderately informed opinion is to go with the two subcales in the canonical correlation analysis. Thanks for any input.
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARDBrian
|
|
|
Thanks, Rich. I’m going with the multiple regression because I don’t have enough power with the canonical. One of the tests has been revised, which I found
out after I sent my message yesterday. The revision has correlations between the two subscales and among the two subscales an total score greater than .95, and the VIF’s are all above 5. It’s the same for the intercorrelations among the three possible dependent
variables, so I went with one DV. Thanks again, Rich. Brian Brian Dates, M.A. From: Rich Ulrich [mailto:[hidden email]]
Canonical gives you a test for "any combination of the IVs" as criterion. Date: Wed, 14 Oct 2015 22:04:02 +0000 I'm assisting a colleague with a study of the relationship of ethnic identity and traditionality-modernity to acculturation. Acculturation is the dependent variable.
The scores used in prior research include two subscales - traditionality and modernity; and a full scale acculturation score, which is determined by modernity minus traditionality. Not at all surprising is the fact that while the two subscales have a very
low correlation, they both correlate highly to the overall acculturation score. So, because of the colinearity aspect, I'm concerned about including them all as possible DV's in a canonical regression analysis. If parsimony is an issue, then I could simply
use the overall score and do a multiple regression. I'm also aware that selection of scales to use based on colinearity is discussed only for IV's. So...my moderately informed opinion is to go with the two subcales in the canonical correlation analysis. Thanks
for any input. Brian ===================== To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
| Free forum by Nabble | Edit this page |