Regression with a categorical moderator (with 3 catoegories)

#1
Hi All.

I am wanting to run a moderating regression. But my moderator has 3 categories(Black, Asian, White).

Usually, with a dichotomous moderator, I would just make an interaction term and see if it was significant (when controlling for IV and moderator).

But I have no idea how to do it with 3 categories.

I've seen somewhere that said run 3 different regressions, one for each group and compare the unstandardised coefficients, but it also said don't do this if the error of measurement is different for each group (which it is).

So I'm confused, any help would be appreciated.

I'm using SPSS 18 (hopefully).
 

Lazar

Phineas Packard
#3
Another approach that is used is to test whether the relationship between continuous variable x predicting outcome y is invariant across the three groups (Black, Asian, White). In this case you run a model in which the relationship of y on x is constrained to be equal across all three groups and compare it to a model in which that y on x is allowed to vary across groups. If they are significantly different using a Chi-squared difference test then you know that the relationship between y on x is not the same across your groups of interest.

This approach is easiest to achieve using a program like LISREL or MPLUS.
 
#4
I *think* I understand your last point. But how would you go about finding out where the difference lies?

I am expecting that the Asian and White groups are similar, whereas for the Black group the predictors operate differently (actually I thin they don't predict the outcome for the black group at all). Would I just have to do plot it?
 

Lazar

Phineas Packard
#5
I *think* I understand your last point. But how would you go about finding out where the difference lies?

I am expecting that the Asian and White groups are similar, whereas for the Black group the predictors operate differently (actually I thin they don't predict the outcome for the black group at all). Would I just have to do plot it?

Easy enough. You can hold various groups invariant and leave the other free to vary. For your example you can hold the Asian and white group invariant and then leave the remaining group free to vary and compare the fit indices.

For the most part i think this approach is easier to understand as you simply report the regression co-efficients for the different groups and indicate which ones are significantly different. If you are looking for further help google 'multi-group analysis'.
 
#6
cheers, will do. I think I've found a book which explains clearly. Problem is I'm an SPSS girl, and I don't "do" equations, so will have to ask my maths PhD husband to take a look at what the heck its going on about later tonight!
 

Lazar

Phineas Packard
#7
cheers, will do. I think I've found a book which explains clearly. Problem is I'm an SPSS girl, and I don't "do" equations, so will have to ask my maths PhD husband to take a look at what the heck its going on about later tonight!
:D this is why it is easier to use a package like LISREL or MPLUS. No equations just a few lines of syntax. Everything else is easy enough. If you got AMOS with your copy of SPSS this should also do it (not sure never used it). MPLUS is easiest in my experience.

The only formula you need is a Chi squared difference test which is just
the difference of the chi-square values and then look up a chi-square table where the df is just the different in the df between the models.