Moderation Analysis in SPSS with a categorical IV and continuous moderator

[waimanchan2]

Hi everyone, I am helping out a friend with the data analysis portion of her master thesis. I have been doing research on the topic of moderation and mediation (and learned quite a bit), but so far I have not being able to figure out how to perform moderation analysis for her study.

In my friend's study, she is testing whether perception (the moderator) moderates the relationship between eating disorder (the Dependent Variable) and family cohesion (the Independent Variable). All three variables are measured on a continuous scale, which is the ideal situation. However, her thesis adviser insists that family cohesion be categorized into 3 groups: low, medium, and high. I told my friend this results in a significant loss in information, but she said the categorization would allow her to test for differences between the 3 groups. In summary, her model involves a continuous DV, a categorical IV, and a continuous moderator.

I understand in the case where all variables are continuous, the analysis would entail a multiple regression that regresses the DV on the IV, the moderator, and the product term between the IV and the moderator. And significance of the product term is an indication of a moderating effect. But in the model specified by my friend, how do we go about testing for moderating effects? (I realize that the categorical IV must be represented by 2 dummy variables in the regression model. Do I multiply each of them by the moderator?)

Any suggestions would be much appreciated!

---

Members don't see this ad.

 

[PerhapsMaybeOk]

do we get a cut of your fee for doing her stats?

 

[aequitasveritas]

Hi everyone, I am helping out a friend with the data analysis portion of her master thesis. I have been doing research on the topic of moderation and mediation (and learned quite a bit), but so far I have not being able to figure out how to perform moderation analysis for her study.

In my friend's study, she is testing whether perception (the moderator) moderates the relationship between eating disorder (the Dependent Variable) and family cohesion (the Independent Variable). All three variables are measured on a continuous scale, which is the ideal situation. However, her thesis adviser insists that family cohesion be categorized into 3 groups: low, medium, and high. I told my friend this results in a significant loss in information, but she said the categorization would allow her to test for differences between the 3 groups. In summary, her model involves a continuous DV, a categorical IV, and a continuous moderator.

I understand in the case where all variables are continuous, the analysis would entail a multiple regression that regresses the DV on the IV, the moderator, and the product term between the IV and the moderator. And significance of the product term is an indication of a moderating effect. But in the model specified by my friend, how do we go about testing for moderating effects? (I realize that the categorical IV must be represented by 2 dummy variables in the regression model. Do I multiply each of them by the moderator?)

Any suggestions would be much appreciated!

if low, med, and high are accurately reflected in 1,2,3 then you can multiple the IV by the Mod to create an interaction term that you would enter into a second block for a linear regression (alternate way of doing it). Just enter each IV and the Mods independently in the first block, and the second block would be the interaction term.

 

[Ollie123]

With only 3 conditions it is tough to justify treating it as continuous rather than categorical. I'd recommend dummy-coding them (probably with 1 as the reference) and creating the interaction terms, though with interactions in there this does inflate the number of variables in the model which isn't ideal. Hopefully you have a good sample size.

While its taboo, I've actually become increasingly convinced that median splits are okay depending on what exactly you are looking at, largely by our current statistician. Its true you lose information but I often think we are stronger believer in questionnaires (what we typically use them on) than is justifiable. Once response bias, etc. is taken into account its unlikely that the small differences being tested when you treat it continuously are meaningfully different from one another, which adds substantial noise to the model. Of course, where to define the cutpoints (arbitrary), the balance of scores in each group and treatment of scores close to the mean are still valid points. If the latent structure is non-normal and gives some indication of categories (e.g. bimodal, trimodal) that might help justify it.