Hello Adele Dahmani. First, the 4 groups you describe result from the 2x2 combination of two variables, sex (M v F) and age group (Senior v Adult). So I suggest including those two indicator variables and their interaction in your model. You would also need to include your focal explanatory variable, and its interactions with sex, age group, and sex*age group.

Second, what is your focal explanatory variable (i.e., the variable you called the independent variable)? Is it categorical or quantitative?

Hello Bruce,

Thank you for taking the time to answer my question.

When you write: include the variables and their interaction in your model. What model do you mean? Can I do that using PROCESS?

My focal explanatory variable is a quantitative continuous variable.

Thank you again,

Adele

Hello Adele Dahmani. I do not use the PROCESS macro, so I do not know if there is a PROCESS "model" that fits your situation. But given that "moderation" as many people use that term these days just means "interaction", why do you feel a need to use PROCESS? Can you not estimate the desired model directly via REGRESSION or UNIANOVA?

Adele Dahmani , I agree with Bruce Weaver that you do not need PROCESS at all, it just can make things easier for you, if you know what you are doing.

To me, it is not clear what your model is. What is your DV and more importantly, what is your IV? Is it a continuous or a nominal variable? If the latter one, I suggest to follow Bruce's advice and simply use ANOVA (if your DV is suitable). Otherwise, you can use a regression approach to handle continuous and nominal variables in one model.

But as Bruce also already mentioned, you seem to have not 1 moderator with 4 levels, but 2 moderators with 2 levels, each. Therefore, you should incorporate this into your model, but you have to think about it. Do you only want the two-way interactions

IV + sex + age_group + IV*sex + IV*age_group

or are you also interested in the three-way interaction (as I would think, if you were thinking about using a single variable with 4 groups), which would look like this

IV + sex + age_group + IV*sex + IV*age_group + sex*age_group + IV*sex*age_group

Both models are possible in PROCESS (or manually of course!!!), namely Model2 and Model3, respectively.

With 2 dichotomous moderators, you do not have to tell PROCESS anything additionally. If you have more than two levels, you need to tell PROCESS that this variable is multinominal and it will create appropriate dummy variables itself. Therefore, the answer to your original question (which should be obsolete by now, since you have two moderators) is: "yes it is possible to do this in PROCESS, but you do not have to create the dummy variables manually!".

Dear Bruce Weaver and Rainer Duesing ,

Thank you for your answers.

I thought PROCESS would be an easier way to test my interactions, but I understand thanks to your answers that using ANOVA might actually be better.

My IV is a dummy variable (Fear = 1 and Love = 0) and my DV is continuous (click through rates).

I want to see if gender and age combine have a moderating effect on the impact fear has on CTR. That is why I had made 4 groups (ie : Female Seniors, Female Adults, Male Seniors, Male Adult)

I already had made 2 dummy variables (for another hypothesis of my study) that are :

Gender : Female = 1, Male = 0

Age group : Senior = 1, Adult = 0

If i do 4 ANOVA tests in SPSS with this equation:

IV + sex + age_group + IV*sex + IV*age_group + IV*sex*age_group.

(so for example for Female Senior :

Fear + female + senior + Fear * Female + Fear*Senior + Fear*Female* Senior

and for Male Adult :

Fear + male + adult + Fear * male + Fear * adult + Fear * male* adult

and the same logic for Male Senior and Female Adult)

Can i read the unstandardized beta of Fear * sex * age that the ANOVA table gives me, as the interaction coefficient? And can i compare the 4 of these coefficient together when discussing my results?

Thank you again for taking the time to answer,

Adele Dahmani

Adele Dahmani

I think you did not understand the ANOVA approach (or similarly the regression approach) correctly. Why do you think that you need 4 ANOVAs??? All information you need is in the one equation

IV + sex + age_group + IV*sex + IV*age_group + IV*sex*age_group

There is simply no variable "female" vs "male" or "Senior" vs. "adult", there are only the variables "sex" and "age_group", respectively. These variables already contrast "female" vs "male" or "Senior" vs. "adult" against each other. In case of the ANOVA, the output will not give you the corresponding regression coefficients, but the mean values for each group. You will get 7 effects:

3 main effects: Fear, sex, and age_group

3 two-way interactions: Fear*sex, Fear*age_group, and sex*age_group

1 tree-way interaction: Fear*sex*age_group

In addition, you may want to use post hoc tests to see, if the differences of interest show the expected (significant) results.

Since there may be a lot of options how to conduct this analysis (e.g. how to handle variance heterogeneity etc) I would suggest to contact a local statistician to help you, since there seem to be some misconceptions how to analyze and interpret the results in general.

Oluwaseyi Ayorinde Mohammed I would contradict some of your statements:

1) There is no need that "all" variables have to be significant to conduct a moderation analysis. What would that even mean? You can only tell if a variable is significant AFTER you have conducted the analysis in the first place. Maybe you mean that you should only probe (as Hayes terms it) an interaction, if that interaction is significant. But this is clearly not "all" variables.

2) Of course, an ANOVA is just a linear model and hence can also be estimated with a multiple regression. But if you are interested in the group means and the differences between specific groups, the ANOVA approach, as implemented in typical software packages, is MUCH more convenient. With the regression approach, you need to do this by hand, to get estimates for each group (or in R with some packages like "emmeans" or "marginaleffects").

3) It is simply not true that ANOVA does not generate the interactions. Quite the contrary! In the default settings of most software implementations is to include ALL interactions, whereas in regression you are able to specify each term on your own, like my example in the first post, where I left out the three-way interaction.

Dear Rainer Duesing ,

I'm not sure I quite understand. If I only do for the three way interaction: fear*sex*age, I will have the interaction effect of female senior only, as my sex is female coded as 1 and my age is senior coded as 1.

Does that make sense or am I completely misunderstanding the whole thing?

Thank you

Adele

Adele Dahmani yes I think that you misunderstood the whole thing. Regression parameters for nominal/dichotomous variables are not absolute values, but reflect the difference, i.e. the slope between the groups. For example (without interaction, this makes it only more complicated to do manually and then it has to be interpreted as conditional effects, not as main effects anymore), if you want to predict your DV (for example height in cm) with gender alone as predictor:

height ~ gender

then you get an estimate for the intercept b0 and and a regression coefficient b1 for "gender". If If you have used dummy coding, i.e. Female = 1, Male = 0 as you did, then the intercept b0 is the group mean value for the reference group, here male, and b1 gives you the difference between male and female participants, for example:

height ~ b0 + b1*gender

height ~ 180 + (-15)*gender

This means, the mean height of male participants is 180cm and the mean height of female is 180cm + (-15cm) = 165cm

This happens if you insert the corresponding values for each group into the ONE equation

Male: 180 + (-15)*0 [since males are coded 0] --> 180

Female: 180 + (-15)*1 [since female are coded 1] --> 165

As you can see, you do not need two equations for two groups, but only one. Likewise, you can follow the same logic with more than one IV (age_group with children vs adults, where you may expect that female children are larger than male and vice versa for adults):

height ~ gender + age_group + gender*age_group

To get the mean height of each group, you need to insert only the corresponding values into the variable. But with presence of interactions, this is not so straight forward to do and to interpret. That's why I mentioned that in such cases the ANOVA approach as implemented in most software packages is much nicer to handle.

Oluwaseyi Ayorinde Mohammed please elaborate in detail, where I contradicted myself and please provide adequate references. Thank you.

Rainer Duesing

Thank you for being so elaborate and understandable.

I now understand better, and I especially understand that I need to go back to my statistics book, haha.

This really helped me, thank you again

Adele

You are welcome, feel free to drop a message, if needed.