Glad to be of help!

Glad it was helpful!

Here's an example of a paper we wrote with some crazy interactions: osf.io/k7dx5

Thank you Erin. Is there model 7 video (moderated mediation model) as you noted above. Also in this example, what would be the sample size. thanks

Thank you.

I mean, you wouldn't know if M is high or low, you would get the b value for M at that level of W. So you would say, as M goes up/down at less than 3 of W, then Y does blah blah blah. Or am I misunderstanding your question?

I don't think so? Maybe? I think only categorical IVs can be used in Model 1 and 4? I would just try it and see what happens.

yes! even though it took e a while to notice you had responded.

I don't at the moment, but it is a popular request, so I will add it to my to do list.

I don't! I will add it to the request list. I do think they finally worked out my license, so I can make videos again!

Not in process that I'm aware of ... but it's been a while since I looked at it. You'd have to break it down into the smaller interactions and rerun it would be my guess.

Would it be considered okay to analyse each moderator to the 1 DV & 1 IV link; and run PROCESS for each moderator and the two DVs?

Cohen, Cohen, Aiken, and West: www.amazon.com/Multiple-Regression-Correlation-Analysis-Behavioral/dp/0805822232

Yes in the covariates box (re-adding my answer back from when changed over youtube accounts, and they were deleted).

A number typed wrong maybe?

You will not see the moderator simple slopes if the interaction is not significant. Try changing the options to always show the simple slopes.

+Erin Buchanan yes that is what I was thinking as well (that I should not see a difference). And yes I did center the variables (by subtracting the mean) before making my product terms. Do I have to use the centered variables for the single variables as well or only to form the product terms? Because for the main variables I did not use the centered variables. Could this explain the difference?

+Erin Buchanan great I will try that. And Thank you for your very fast reply! This helped me a lot.

Sure - you could create the two way interaction instead - it wouldn't be covaried, so much as averaged across Math Efficacy. You would say at average levels of ME (so 0 in your scale), here's what context by gender looks like for GPA.

You can check out the templates.pdf that comes with process - it appears to me that only three way interactions are supported at this time.

To me, that implies that there's a lot of error that the values are not significantly different from zero. So, the output you are describing is possible yes.

Yes, thanks again ! :D

I would just denote that the covariate was a significant adjustor of the DV - and then interpret the direction of the coefficient (i.e., positive or negative). If you want to interpret the model, you could say the adjusted Y is mediated by M etc.

I'd mostly just check out guides for moderation on reporting - same idea, but with 3 ways interactions.

I would probably try to power the highest interaction or the overall R2 using gpower in that case if you aren't familiar with simulation.

Thank you. But PROCESS did not seem to differentiate categorical vs. continuous variables here when I use Model 3. How would you be able to let PROCESS work on categorical variables and simple slopes for each category？ Many many thanks.

And what if one of the IV is a categorical variable with three levels?

Yes it can, but I'm pretty sure PROCESS cannot handle that yet. You would have to do it manually.

+Erin Buchanan Thank you for your prompt reply and advice.

If they are basically the same variable (i.e. super correlated), then consider dropping one of them.

That will happen since there are different variables when you only do simple versus 3 way. If the three way is significant, then you would say there’s moderation.

@@StatisticsofDOOM Thank you so much for your help. I really wanna appreciate you. Hope you have a wonderful day!

Generally most people consider above .9 as multicollinear. (Cohen Cohen Aiken and West).

Thank you very much, it had been quite useful!!! Happy new year :)

+Efraín García Hi Efraín, may I ask which way you finally chose for your data? I am curious because me too I am trying to run a 3way interaction with a categorical variable with three levels as moderator. Are u on linkedin? Maybe we can connect there. I would highly appreciate an exchange on that matter with you. Thank you so much in advance!

Svenja Hi Svenja, sorry for the delayed answer. I run the model suggested by Erin, and also did the moderated moderation with the SPSS Macro Process (Hayes). I found both ways quite useful to conduct the analysis. I'm on LinkedIn with this same name, so we could exchange ideas there.

+Efraín García Hi Efraín, sorry, I guess I wasn't able to find you on linkedin. My last name is ohlemann. Maybe this way around it might be easier. :-)

The outlier ideas are from Cohen Cohen Aiken and West regression book, along with Tabachnick and Fidell multivariate book.

Thank you you are an amazing professor

For that version of process, you'll need to check out the templates - they may not be the same numbers as the one I am using in this video.

Hi, i have the same problem. Did you figure out how to do it?

Yes - they are listed under mediation with process on the channel, there are a couple of them.

:( not that I'm aware of. You might be able to use the SEs from the B values? I'm not 100% sure how you'd do it honestly.

No worries! I appreciate you taking the time to weigh in, and thanks again for the helpful video!

I kind of wanted to know, so I found this calculator: www.danielsoper.com/statcalc/calculator.aspx?id=90 which will let you calculate the CI values around the estimated scores (standard error of the estimate = mean square error / residual). However, it's still not totally clear to me how you'd add them into graph builder in SPSS - it seems to only want to calculate them ... you might be able to make the chart in Excel, which will let you add error bars manually.

Well that's quite helpful as I ended up making the original graph in excel anyways! One follow-up, if you don't mind: I'm assuming the MSE comes from the model summary portion of the process output. For the "residual" would I simply use the standard error of the three-way interaction term in the regression output?

So, you don't get this output in the version I'm using in the video (2.15) but if you have the newer process (2.16) you get the mean square error of the model (like the overall model not the predictor) under MSE. Different example below - but you'd use that number. Model Summary R R-sq MSE F df1 df2 p .1849 .0342 .2046 1364.2116 7.0000 269720.000 .0000

+Erin Buchanan Thank you, Erin!

I will like to opened this discussion. I was told to never interpret the main effect when interaction terms are put in the equation, since they are tainted by that matter which render them non-interpretable. Should we interpret main effect at step 1 only? What about simple interaction in step 2 before the triple interaction? Can we interpret the two-way interaction in the final model when the triple interaction is added to the model? Any thought?

I will like to opened this discussion. I was told to never interpret the main effect when interaction terms are put in the equation, since they are tainted by that matter which render them non-interpretable. Should we interpret main effect at step 1 only? What about simple interaction in step 2 before the triple interaction? Can we interpret the two-way interaction in the final model when the triple interaction is added to the model? Any thought?

These are for outlier cutoffs for leverage and cooks. It's how you know what scores are beyond what normal variation in the sample might produce.

No. You would need to use logistic regression for categorical outcomes, which requires less assumptions. You could still use cooks, leverage, and standardized residuals though.

Could you make a video on checking threeway interaction with logistic regression? Including checking the assumptions and all? I could not find this anywhere. Thanks a lot!

@@iliagugenishvili7976 I have not - could add it to the list (things are very busy right now though). Three way interactions are always tough!

If you look at the template documents provided with process, the variables used much match the variables in the template. Therefore, you must have an M and W if you use model 2, otherwise the macro will not know what to run to calculate the analysis.

Thank you!

One more question: If I use Model 2 and only one of the interactions was significant, am I supposed to re-run the moderation analysis but with only the significant moderator?

In some papers I have read they have "dropped" the non-significant moderator, but they seemed to be using regression analyses rather than the PROCESS plug-in

I wouldn't drop variables if they weren't significant, as you had them in there for a reason. People might consider that "cheating".

The new version of process does interactions with categorical variables (and dummy codes them for you), so you could use that (then just stick the variable in X or M).

Is there a link for a how-to? I’ve tried looking. I have a continuous variable that I am splitting up into 3 groups to check differences between the 33, 66, and 100 percentiles.

@@eaterrr Splitting a continuous variable always makes me nervous :| I believe you can just denote that it's categorical with the categorical button options that are available in process, but you would want to create fake category column if you are splitting it from a continuous column. I can't quite remember if the categorical options work for a three-way interaction model, but I think so.

I’m nervous about it, too! I’m looking at whether two continuous variables (one is a moderator) will predict performance on a task - and I want to see if there are differences with the moderator variable between high, middle, and low groups split by those percentiles, but I’m trying to figure out the best way to do so. Thank you!

@@eaterrr Ok here's what I might tell you to consider instead. Figure out the score that a person would have on that variable at those points - like what is 33%, 66%, etc. Then shift the data around that point - normal simple slopes are 1 SD above and below the mean, and these are created by adding and subtracting a SD from the data, rather than creating subgroups of people who are only 1 above or below. Therefore, you could apply the same idea to your data by manually creating columns that mimic a 33% group, 66%, etc. So, if the 33% group scores a 4 and the mean is 8, you need to shift them up 4 points to be centered around the mean (assuming mean centered regression for easier interpretation). If the 66% group scores a 12 and the mean is 8, you should shift them down 4 points to see their effect. Then you would run the analysis again with those columns (one at a time, instead of the original column for that variable) and check out what happens to the other variables in the analysis for interpreting simple slopes.

No not yet, but you could treat it much like two model ones! I can add model 2 to the list.

Yes? I believe all these are treated continuously, but either way yes.

@@StatisticsofDOOM thank you for your super fast response Statistics of DOOM, I really appreciate it

Thank you for your kind help. i will review and hope to learn how to use it

They are a little different yes - I have several videos on the new version, search Process 3 on my channel.