Fw: Re: Comparing regression coefficients in related samples

[Mike]

Here is another perspective on the situation:

 

(1)  In the old BMDP series of stat program, the program "1R" (for

Multiple Linear Regression) allowed one to calculate a regression

equation for the total sample as well as regressions for groups

within the total sample.  An "omnibus" test was conducted to

test whether a common regression equation held for both groups.

If the omnibus test was significant then the groups differed in

either intercepts and/or regression slopes.  One could follow the

procedures suggested by Pedhazur (1997, 3rd Ed; Mult Regr in

Beh Res) in his chapter 14 to determine which slopes differed

and so on.  Pedhazur shows how to do a comparable analysis

in SPSS (see p562-572).  Although SPSS allows one to use the

"Select" subcommand in the Regression procedure (see the

syntax manual, page 1695, for SPSS v18), it only calculates the

regression equation for the selected group (one would have

to do this for both groups).  This might be somewhat simpler

than adding the interaction terms suggested by Gene below.

 

(2) Structural Equation Modeling (SEM) can also be used if

you access to AMOS, Lisrel, Mplus, or similar software.

In old AMOS manuals (I got started with AMOS v3.6; I haven't

seen recent manuals) you would look at "Simultaneous analyses

of several groups".  The logic is similar to that used in BMDP

but much more flexible.  You would specify a model where all of

the coefficients in the two groups are set equal to each other.

This is comparable to an omnibus test and if the chi-square

for model fit is nonsignificant, then the coefficient for the two

groups are all equal.  If the chi-square is significant, then one

can change the relationship between coefficients in the two

groups from equal to not equal.  Do this one at a time for

coefficient until the model chi-square is nonsignificant. The

final model will identify which coefficients are equal in the

two groups as well as which coefficient are different (if

memory serves).

 

The SEM approach is probably the way to go because it is

likely that your errors may be correlated with and across

time periods as well as having different variances.  Talking

to an SEM'er with experience with these types of models should

help you develop a strategy for testing the different models.

 

-Mike Palij

New York University

[hidden email]

 

 

 

----- Original Message -----

From: [hidden email]

To: [hidden email]

Sent: Friday, June 10, 2011 11:06 AM

Subject: Re: Comparing regression coefficients in related samples

Tony,

 

One way to do this is through interaction terms. Add the two files together along with a variable indicating which file contributed the case (the IN subcommand is useful here). You have the same model for both files. Systematically add interaction terms between model variables and file indicator variable to the baseline model.

 

Gene Maguin

---

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Baglioni, Tony
Sent: Friday, June 10, 2011 9:42 AM
To: [hidden email]
Subject: Comparing regression coefficients in related samples

All,

 

I know this is a bit non-SPSS but I have a simple question.  I have data from 1990 and 2000 census.  I have developed the same model on both sets of data and I’d like to compare the regression coefficients.  I have searched around archives and cannot find an example that matches this exactly.  Any suggestions?

 

TIA,

Tony

 

A J Baglioni jr

McIntire School of Commerce

University of Virginia

Rouss & Robertson Halls, East Lawn

P.O. Box 400173

Charlottesville, Va  22904-4173

434.924.4961