Regression Discussion > Hierarchical regression and dummy coding
Hi Natalie, thanks for your question!
A few things:
When you say:
"time 1 outcomes entered in step 1, treatment entered in step 2, moderator in step three, and the interaction term in step 4"
I'd recommend entered both the treatment and moderator variables together at step 2 and the interaction terms at step 3, giving you only 3 steps.
Also, you said:
"Our advisor told us to dummy code the treatment such that both career interventions are compared to the control group and then to run another regression where the reference group is one of the treatment groups and compare the two treatments to one another. If we do that, we get results that don't make sense. For example, for one particular outcome, the control group = 3.0, treatment 1 = 3.1, and treatment 2 = 4.0. We found a very significant difference between control group and treatment 2 but not a significant difference between treatment 1 and treatment 2. This doesn't make sense. Can you tell us what the problem might be?"
How you described your dummy codes is not how I'd recommend doing them. When doing dummy codes to compare categorical/nominal variables, I'd recommend doing a "simple contrast", which entails making K-1 dummy variables (with K equalling the number of levels of your treatment variable).
In your case, if you have the following levels (hypothetical codes in parenthesis):
Control (1)
Treatment A (2)
Treatment B (3)
Then you'd need to create K-1 variables (i.e. 3-1=2; 2 dummy variables). They would be created using the following syntax in SPSS (assuming your current treatment variable is labeled "Treatment"):
Do IF (Treatment=2).
Recode Treatment (2=1)(else = 0) into TreatA.
Else IF (Treatment=3).
Recode Treatment (3=1)(else = 0) into TreatB.
END IF.
EXECUTE.
That syntax will create two new dummy variables (TreatA and TreatB), from which you'll also need to make new interaction terms with the dummy variables and the moderators. After that is complete (per your advisors request), you'll need to do the whole process again, but replace one of the Treatment Dummy variables with the "Control" dummy variable.
Jeremy Taylor
Thanks for the speedy reply! You seem to be suggesting the same approach as our advisor suggested - I.e., creating 2 sets of variables - 1 with the control group as the reference and one with a treatment group as a reference. Am I understanding you correctly? If so, this produces results that don't seem to make sense (see the example above). Do you know why that might be?
Natalie
Hey Natalie,
I'm afraid the info in your earlier post is not enough to make clear to me what might be going on, but if you like to send me your output file, I'd be happy to take a look and see if anything unusual drops-out at me. If you'd like to do that, you can send it to:
jtaylo20@me.com
Jeremy Taylor
Thanks, Jeremy! I will send the output right now.
Natalie
Hello,
We are trying to determine the effect of 2 career-related interventions on career and academic outcomes over time. We have three levels of the iv (control, online intervention, online + face-to-face intervention), 7 dvs, and several moderators and we are running a bunch of hierarchical regressions with time 1 outcomes entered in step 1, treatment entered in step 2, moderator in step three, and the interaction term in step 4. Our advisor told us to dummy code the treatment such that both career interventions are compared to the control group and then to run another regression where the reference group is one of the treatment groups and compare the two treatments to one another. If we do that, we get results that don't make sense. For example, for one particular outcome, the control group = 3.0, treatment 1 = 3.1, and treatment 2 = 4.0. We found a very significant difference between control group and treatment 2 but not a significant difference between treatment 1 and treatment 2. This doesn't make sense. Can you tell us what the problem might be?