Moderation - help!

#1
** Oops I meant to post this is the 'Psychology Statistics' forum. I've posted there, but cannot figure out how to delete this thread. Apologies for double posting! **

Hi everyone,

I am new and hoping that you will be able to help me figure out which analysis is most appropriate.

My experiment has:
- 2 IVS - 1 continuous, 1 categorical (3 levels)
- 1 DV (continuous)

I am interested in
- the main effects of the IVs on the DV
- finding out if the continuous IV moderates the relationship between the categorical IV and the DV

Should I use multiple regression of ANCOVA to do this? I am very confused because I have read different opinions in different resources and cannot get a clear answer.

Thanks in advance for your help.
 
Last edited:

Jake

Cookie Scientist
#2
Probably one of the major reasons you haven't been able to find a clear answer is because, statistically speaking, multiple regression and ANCOVA are quite literally the same thing. You're going to get the same answer no matter what you choose to call the statistical model when you're writing up the results.

Basically you just want a model that looks like:

Y = b0 + b1Cat1 + b2Cat2 + b3Cont + b4(Cat1*Cont) + b5(Cat2*Cont)

Where Y is the continuous dependent variable, Cat1 and Cat2 are the two predictors that code your categorical variable, Cont is your continuous variable, the products thereof are the interaction terms, and the bs are the parameter estimates. You probably want to code Cat1 and Cat2 as orthogonal contrasts to ease interpretation of b3.

Now, is this model an ANCOVA or a multiple regression? Well, it's both. If your primary interest is in the categorical variable, it is customary (but ultimately arbitrary) to think of it as an ANCOVA because in that context the continuous variable is "merely" a covariate. But if you are not especially more interested in the categorical variable over the continuous variable, you might instead think of it as a multiple regression because the continuous variable has more substantive interest. But clearly your opinions about which variables are interesting should not (and indeed, will not) affect the estimates from the model. The model is just, well, the model, and your answer is not going to change just because you called it one thing versus another thing. (If the results do change, then something has gone wrong!)

If you have more specific questions about implementing the model I'm sure many of us will be happy to help.