This is freaking amazing!!! You have a gift for breaking down analyses step by step. Thank you!!! I am using lots of exclamation points because I am in the middle of running analyses for my honours thesis and have had a lot of caffeine today!!!.
I have a question regarding the signs in the interaction term and the effects. If I'm correct, the negative sign in the interaction term means that the relationship between IV and DV is weakened by the moderator. How do we interpret the signs of the effects then? I have a negative interaction and negative effects. My IV is stress, DV is well-being and M is self-regulation. I've seen examples, where the interaction term is negative, yet the effects have positive values. In my case, both interaction and effects have negative values, and I'm wondering what it tells me. What is the general rule for interpretation here? Also, my effects decrease mathematically: when my moderator is below 1SD it is -1, then above 1 SD it's -2. So the effects become weaker or stronger with an increase in the moderator value? I'd really appreciate your detailed explanation because I got a bit lost here. Thanks!
I never usually interpret the sign/effect of the interaction because it's not that clear what it should mean based on a few factors. I would run the simple slope analysis and interpret those effects. Any estimate moving away from 0 in absolute value is a "stronger" coefficient - I would just say it's larger so as X increases Y decreases 1 point here, 2 points there, etc.
Thank you so much for this video. What if my moderator is not continuos (Gender) but the rest of the predictors are continuos. What would you do differently?
Hi Dr Buchanan, I am wondering how the rule for outliers came to be "2 strikes, you're out" at 14:05. Is there a source I could go to so I can understand this? Your video has been very helpful for me in completing my dissertation. Thank you.
Hi, very nice video :) Unfortunately, my model contains a nominal Covariate/ Control Variable (7 categories) Do I have to dummy code it manually and then Insert the dummy variables into the covariate-box in PROCESS? And if Yes: do I have to leave one of the dummies out as a reference category? (so: leave Dummy_1 and insert Dummy_2 to Dummy_7 into the box)
Hi! Great video. It really helps with my work. You are a saint. I have a brief question. Would it help if I used MANOVA prior to the PROCESS to check for correlations between the variables? Thank you.
Thank you for the great video! Is it possible to only plot the interaction terms of the linear regression models without including the main effects using PROCESS?
Very useful for acquiring skills in moderation analyses! Dear Dr. Buchanan,m I would like to ask a small question. I have one moderator (continuous) and 5 independent variables (all subscales of spirituality). Is there a way to test for this model in PROCESS with all the main effects and interactions with each of them? I think it is not correct to test for five separate models (one with each indep. variable and interaction with moderator) as the subscales are correlated to each other, am I correct? Thanks in advance for any help you would give me. Paolo
I don’t know how many 2 way moderators you can do in process but unlikely you can do all of those. If they are super highly correlated, then maybe consider just doing one of them because of multicollinearity.
Hi, is it possible to use Process with 2 independent variables and 1 moderator ? In this case I look into the effect of two types of slack resources on CSR performance. The moderator is corporate ownership structure
Yes! I think the newest version of process has this option, but I haven't looked at in a while so don't quote me on that. Otherwise, you'll have to do the steps manually using the logistic regression functions in SPSS.
When performing moderation analysis, testing the moderation effect (W) of a given moderator (M, e.g., sex) on the association of a independent variable (X, e.g., age) with a dependent variable (Y, e.g., height): ==> Must X have a statistically significant simple (main) effect on Y on step 1, so that a statistically significant effect of the interation term (W=X*M) is interpreted? Is this a condition for a mdoeration analysis? Note: Model 1, A. Hayes' PROCESS. e.g.: - Imagine that we are studying if age (X) predicts height (Y); - And that indeed, age predicts height just in men, but not in women (which stands for a moderation effect of sex on the association between the X and the Y); - The simple/main effect of age on height in step 1 (which considers the entire sample, men + women) might not be statistically significant because the non statistically significant observed on women may cancel/surpress/blur, the significant effect observed on men, and the interaction/moderation effect on step 3 is statistically significant.
Great video with easy explaination. I have two questions. Please give me response on both of them 1- I wanted to study 10 covariates which are nominal and having multicategories, can I include them all along in model 5, with one mediator,one moderator,one IV,one DV? I ran model 5 in spss and it didnot give any error, although i didnot transform multicategorical covariates into dummy variabels. I still need to transform these 10 covariates. Which will make my analysis so complex. What do you suggest? 2- I ran analysis, interaction result isnot significant but conditional direct effect are significant. How can i interpret? Also line graph of this output shows divergent at lower and higher values of moderator ,but convergent at mean. Regards
I think you can use process to make the dummy coding for you - but yes all the multicategoricals need to be dummy coded. I don't have a good comment on having a lot of variables - they should all be included for a reason, so consider if they are all theoretically important. If the interaction is not significant, then I would not interpret the simple slopes, as that result implies they are all statistically the same.
Hi Erin - great video, thank you so much for uploading! I just wanted to check my understanding regarding covariates / control variables within this model. Does the inclusion of a covariate control for its effect on X and Y or just on Y? Thank you in anticipation :)
Hi Erin! Thank you for such a good talk about moderation using PROCESS macro. Two questions: 1) Does it make sense to compute the standardized betas of simple slopes or conditional effects? Usually only standardized betas are reported in papers, but not unstandardized ones. 2), If so, how can we do it? How can we compute them from the unstandardized coefficients for simple slopes or conditional effects? PROCESS macro seems not to compute them at all, but reviewers are asking for them.
I'm ambivalent on the betas for moderation. They make interpretation of the coefficients difficult. I feel like I see just centered more, but you can do either based on what you see more in your literature. If you would like those, then z-score your X and M columns, but be sure to turn off the mean center option in PROCESS. (so you don't do it twice basically).
Hi Erin, I was wondering why you use the Normal P-P Plot of Regression Standardized Residuals to test the linearity assumption. I learnt that the plot is used for testing the assumption that the residuals of the dependent variable are normally distributed. I am confused as to why the plot can be used for both purpose. Maybe you can help me understand it.
I've seen both - you can either look at a regression line plotted on the residuals or a PP/QQ plot. Deviations from the line can indicate non-linearity. Residual scatterplots are also pretty telling that they make rainbows. If my memory serves me right, I took this from Tabachnick and Fidell.
I think process can handle categorical variables just fine for moderation. I would put that category variable as M to get the different slopes for each group.
Hey, thanks for your great step by step tutorials, you're definitely saving my master thesis! When I add a dichotomous covariate such as gender (m/f), I can do this without dummy coding, right? And how do I interpret a significant effect of this covariate then?
Yes you can do it by just coding each gender as 0 and 1. A significant dummy coded variable implies that there is a difference in the means for each group on Y after covarying out the other X variables.
Hi Erin, Thanks a lot for this video! I have a question. My model is significant, but the only significant "thing" after that is one CV or my 2 CV (all others interactions are not significant). For my master degree , I did many analyses with 1 IV, 1 DV, 1 moderator and 2 covariable and really often the model is significant, but only the CV is also significant. In my results section, i report all that but i end up talking A LOT about covariables and not so much about IV and DV (wich are the real thing). Do i have to report all significant model that have just CV ? Do you have a solution so i'm not talking just about CV? Thanks to you. G
Hi Erin! My output shows everything except the "conditional effects of the focal predictor at the values of the moderator." I still get the graph syntax code, so I'm able to graph low/average/high rates. I would like to know the p-value for each of my simple slopes, is there any way to do that without getting that same output?
In the newer process, you have to force it to show you those effects if your interaction was significant. Don't remember exactly which window it is but change the p-value to something higher so they print out.
Dear Erin, thank you so much for another great video. I have a question about your interpretation. In the interpretation of the main effect, you mention first that Illiteracy isn't a great predictor of income (p = .508). But then you say, if I suppose to interpretate these results, I would have to say that as illiteracy increases, income decreases. Isn't that a bit of contradictory? The p value greater than .05 states that not enough evidence is found that the b that is found differs significantly from null.
Right but it helps to know how to interpret b as well - many people only focus on p and don’t actually know what the coefficient is. I usually try to teach both so you know what to say either way.
HI there, I came to this page for learning moderation first. But I really like your approach for hadingling outliers. Is there any particular references explains your approach (i.e. using Mahalanobis distance, Cook's D and Leverage values simultaneously? Thank you!
@@morgandossey9877 Yes - dummy code is a way to represent categorical variables in regression analyses. You can learn how to do it here: blogs.perficient.com/2012/04/25/dummy-coding-with-ibm-spss/ or I have a few videos that cover the concept I can link. Mainly, you are creating unique values for each category, which creates pairwise comparisons when you run a regression.
Great step-by-step video. I've also appreciated your talks on SEM. Thank you for these resources! I have two issues I've run into as someone new to PROCESS: (1) I've noticed that the previous version of PROCESS allows for four moderator variables but the most recent version 3.3 only allows two (variables W and Z). I am trying to analyze a model that includes three moderators. Is there any way around this limitation in PROCESS? A similar limitation exists for use of multiple independent variables in, and I know running PROCESS iteratively (rotating predictors of interest through the IV and covariate slots) is the workaround for such models. Would it yield mathematically identical results if I ran all moderators simultaneously vs. running PROCESS twice, first with moderators 1 and 2 and then with moderator 3? Or is there a way to include a third moderator via syntax? I have one IV, one DV, 1 mediator, 3 moderators, and a couple of covariates in the model. If I ran the model iteratively, I assume I should include omitted moderators as covariates? (2) Is there an option to select in PROCESS a mediation model with two moderators directed at the direct effect of X on Y? I couldn't find that option from among the model templates, but it is a fairly basic model that I would expect to be preset so I wonder whether I am missing something. I had originally had in mind a concept like Model 10 but with no arrows from the mediators to the X->M path. Do you recommend just using Model 10 or customizing it?
I don't know about what's going on with the updates, as I only use SPSS for videos :). Definitely run a model with all the moderators at once, as it's not the same mathematically if you are rotating some in and out (remember that coefficients are "controlling for all other variables, X predicts Y with b slope). Also if you limit to two moderators and actually want the three way moderation, you will not get that. You can do them manually in SPSS without process! For the second one, I think customization might work - the bootstrapped CI is the harder part in SPSS, which is why I teach a lot of materials in R for more flexibility.
How would you interpret a non-significant interaction but a significant covariant of ethnicity? I previously recoded this variable in to values 1 and 2. Thank you :)
Does a dichotomous covariate have to be dummy coded? E.g. I have gender -with values of 1 for female and 2 for males already? Can this variable be placed in to the covariate box as it is?
Dear Erin, thank you so much for these PROCESS videos! I have a question regarding the significance of simple slopes. So the interaction is significant but none of the slopes are, the case that you already mention in the video. However, if I further check the Johnson-Neyman output, I see that when the moderator value is either very high or very low, X actually does predict Y. My question is: is it okay then to report that there is a moderation effect, but only when M is x-times above/below the mean?
Yes that's fine! I would just write up that you used the Johnson-Neyman approach to find the regions of significance, rather than the simple slopes approach.
Hi! I have a question about moderation. I have a moderated moderation model (model 3) with a continuous DV, a dichotomous IV and 2 dichotomous moderators. Under Options, should I click mean centering or not? I find contrasting answers on the internet and I don't really know what to do. The main effect between the IV and DV is very insignificant with mean centering but significant without it. Thank you so much in advance!
Hi Erin, Please help!! I am doing an MR moderation analysis using PROCESS for my masters thesis. I am unable to obtain the Hayes Book as there's only 1 hard copy in reference at the Uni library and due to restrictions I am not allowed on site to view it. The people I have approached to discus this are unsure about what to do. My model has 1 IV, 1DV, 1M - which are all continuous and 3 covariates - 2 are continuous (age and number of training sessions) and 1 is categorical "position". 1) Can I use 3 covariates in PROCESS model 1? 2) Can I use a multi-categorical variable (i.e. 3 groups) instead of a dichotomous variable (dichotomous does not work conceptually and would not make sense) in PROCESS model 1 and get meaningful results from the output? 2) If so, do I assume I would create dummy variables for group 2 and 3 (against reference variable 1), and enter these into PROCESS as the 2 covariates representing "position"? Thus, I'd have 4 covariates in total in the model. 3) How would I interpret the results, particularly from the covariates?... especially If all relationships came out as not signifiant except for one or both of the 2 "position" covariates 4) If this is not valid for use in PROCESS , what other analysis would be MOST valid here instead please to determine a moderation effect of M on the IV/DV relationship, taking account of the other covariate variables? 5) In a second situation it may be that 2 of my covariates are actually control variables and 1 is a covariate. How would I then treat and enter these 2 control variables into a moderation analysis (with the same IV/M/DV as above) using PROCESS? Would this second model still come under model 1? I'd really appreciate any help with thsi asap please. Thanks in anticipation!
1) You can use the covariates - the categorical one may need to be dummy coded first depending on the number of categories. 2) Yes but dummy code it first stats.idre.ucla.edu/other/mult-pkg/faq/general/faqwhat-is-dummy-coding/ 3) I have a few videos on dummy coding in regression, I would recommend those for understanding how to interpret dummy coded variables (basically they are measures of group differences) 4) Process is just a software plug in, categorical variables can be used in regression just fine because things like t test and ANOVA are just special versions of regression with categorical predictors 5) Do not mix terms here moderation and mediation are different things. I would treat these as separate hypotheses unless you are doing moderated mediation. Practically, there isn't a mathematical difference in regression if one is a control and one is a covariate. They are all entered into the equation as predictors.
Hi Erin, very good explanation! It is completely clear what to do in Process version 3.1 and how to interpret the outcome. Most video's use old Process versions with different options. However, my model summary is significant, however the interaction in the model under Model Summary wasn't (p = .0706). So, the rest of the output is not interesting anymore, unfortunately, because I did fixed it with a nice graph :-). Now I must report not only in writing what I have been doing also in a table. Have you got an APA style table for this outcome to report? Or where can I find one. I didn't yet find in on Google. Thanks in advance.
So, I have an experiment and want to see if participants' score on a test moderates the impact of the independent variable on the dependent variable. I am able to follow these instructions for Model 1 in Process. Conceptually I want to understand how it is different from using the univariate GLM in SPSS...when I use SPSS GLM and enter my IV as a fixed factor and test score as covariate (and then go to "model" and make sure the IV, the test score, and the IV test score interaction are all in the model) -- the F statistic and p value are the same for the interaction term and the IV in GLM and in Process. But the F statistic for the continuous moderator is different between Process and SPSS GLM. Do you have any idea why this might be?
Do you mean the t-score is different? The F score is for the overall model, so I'm not sure if I'm following. If the t-score is different (i.e. the coefficient test statistic), I believe that's because process uses a correction for heteroscedasticity on the standard errors (and the formula for t is b / SE, so that will change the values between that and the normal GLM).
@@StatisticsofDOOM Thank you so much, and sorry for the lack of clarity! When I do SPSS, Analyze, General Linear Model, and then Univariate and put in my IV (with two conditions) and my moderator (as a covariate) and their interaction into the model, there is a table called "test of between subject effects" that gives me a F value for the entire corrected model that matches what process gives me. Then there is an F score for my IV, my continuous variable, and their interaction. For the IV and the interaction, the sig. values EXACTLY match the t score sig values for the IV and interaction. (And the F score is the square of the t-score in process). But the sig value for the continuous predictor does not match what is in process and the F score is not the square of the t score
Ah, oops. I realized that I confused the GLM test of between subject effects with the GLM parameter estimates. These DO match what Process spits out -- except I have to reverse the coding of the independent variable and put the reversed one into Process to get it to match.
Hello, first of all i would like to thank you for comprehensive guidance. I want to ask if i have one moderator, one mediator, 1 dv, 1 IV, along with 3 covariates in the same model then how to get the results using hayes process and how to interpret the results.
That really depends on where the moderation is. There's a file (templates.pdf or the back of the book) that includes the different pictures and how the moderation and mediation interact. You'd have to pick which one theoretically matches your hypotheses. The CVs can go in the covariates box.
Thank you for this video. Got two questions according to issues with the linear dependency error. One of my moderators is only measured when X='Yes', this leads to a linear dependency error, so I can not run the PROCESS. Some say you have to add zero's for the structural missings (X='No'). This doesn't work, same error. There has to be a way to measure this, right? Do you know how? The same problem occurs when you ad multiple dummy variables as covariates. For example, when age is a categorical variable. Maybe you can help me, I hope so. Thank you!
If the variable is only measured for that group, then you could do a simple linear regression of M to Y for that group. You wouldn't be able to add it as an interaction because it doesn't exist for the whole sample.
A new subscriber here! Great video! I hope that you could make a video illustrating Model 14 with some covariates (e.g., 1 dichotomous and 1 continuous). I have only found 5 videos describing Model 14.. I am not sure why. Anyways, thank you for sharing your knowledge! ^^×
Hello. Thank you so much for this video. It helps me a lot to finish my assignment. I have a question. What do you think if IV is correlated with DV in Pearson correlation, but not in Process Hayes? Can you help me to explain/interpret this case for me? For the context, I have one IV, one DV, one moderator, and one covariate.
Hi Erin, Thank you for the video! I was wondering how you chose your covariates in the first place (i.e. is it simply a matter of choosing them based on common sense or are there statistical tests you need to run to determine them as covariates before including them in your main analyses?). Thank you :)
Hi Dr Erin, I'm computer engineer and now doing my masters in engineering management. I'm investigating the impact of Supply chain integration(customer Integration, Supplier Integration & Internal Integration) on supply chain performance with a moderating role of enterprise systems. In this way, I have 3 IVs, 1 DV and one moderator. Since I'm new to the social sciences field, Could you please suggest and guide me how can I use SPSS to analyze data and interpret results ? What tests do I need to carry out for my research to be legit and authentic?
I'm not sure I follow the question - this video shows you have to do how to run the moderation .... kind of depends on where you want the moderation to be (i.e. only IV1 or all IVs).
Hi Erin, if i have 4 IVs and 1 CV, do i need to run this method 4 times in PROCESS, with the same CV each time? If yes, what in your view is the best way to report the results of 4 IVs + 1 CV? I've not been able to find videos of how to report this, so your input would be extremely helpful!
That depends on the hypothesis really - if you are proposing four separate interactions, then you can run them all separately or run each two way interaction within the same model (not with process though). With that many variables, I would just put all the B SE t p values into a table. The reporting is the same if significant or not, but you will not report simple slopes if the interaction is not significant.
Thank you for your reply, Erin! I followed your advice and put the B, SE, etc. values into one table. I have a 2-part follow-up question with regard to covariates: 1) what does it mean if a significant interaction becomes insignificant after adding in a CV? (The CV was included in the regressions because it is theoretically related to the DV, even though it's not part of the main model.) 2) conversely, what does it mean if an insignificant interaction becomes significant after including the CV? For this, I came across an explanation for suppression - in your opinion, would that be a plausible explanation?
1) that usually indicates that the CV and the interaction were predicting the same variance (i.e. those are correlated as well). Or that the variance the interaction predicts is not more than zero after accounting for the CV. 2) Same idea but in reverse. After account for the CV, the interaction is significant because controlling for the error that was actually the CV increased the usefulness of the interaction. Suppression is what folks normally call this effect yes.
Statistics of DOOM thank you so much again, Erin! Your explanations in plain language have really helped a non-statistician like me to understand what I’m seeing and writing. I think my dissertation has some glimmer of hope now because of your help. Thank you!
This is freaking amazing!!! You have a gift for breaking down analyses step by step. Thank you!!! I am using lots of exclamation points because I am in the middle of running analyses for my honours thesis and have had a lot of caffeine today!!!.
I am always heavily caffeinated so I feel that!!! :)
30:44 for simple slopes analysis.
Great explanations. A huge thank you!
Thanks for the kind words!
THANK YOU THANK YOU THNAK YOU
needed this for my PhD!!
Glad it helped!
20:00 for the moderation analysis
I looooooove you Erin!!!!! This is so easy to understand!
Thanks for the kind words!
I have a question regarding the signs in the interaction term and the effects. If I'm correct, the negative sign in the interaction term means that the relationship between IV and DV is weakened by the moderator. How do we interpret the signs of the effects then? I have a negative interaction and negative effects. My IV is stress, DV is well-being and M is self-regulation. I've seen examples, where the interaction term is negative, yet the effects have positive values. In my case, both interaction and effects have negative values, and I'm wondering what it tells me. What is the general rule for interpretation here? Also, my effects decrease mathematically: when my moderator is below 1SD it is -1, then above 1 SD it's -2. So the effects become weaker or stronger with an increase in the moderator value? I'd really appreciate your detailed explanation because I got a bit lost here. Thanks!
I never usually interpret the sign/effect of the interaction because it's not that clear what it should mean based on a few factors. I would run the simple slope analysis and interpret those effects. Any estimate moving away from 0 in absolute value is a "stronger" coefficient - I would just say it's larger so as X increases Y decreases 1 point here, 2 points there, etc.
thank you so much! 100% recommend this video!
Thank you!
Thank you for the video!
Glad to be of help!
Thank you so much for this video. What if my moderator is not continuos (Gender) but the rest of the predictors are continuos. What would you do differently?
Not a whole lot - the simple slopes are often separated by the categorical moderator, rather than by standard deviation.
Is it possible to use process macro if my IV is measured on a likert scale (ordinal), my DV is continuous and my moderator is categorical (yes/no)?
I'm not sure process handles ordinal IVs at the moment.
Hi Dr Buchanan, I am wondering how the rule for outliers came to be "2 strikes, you're out" at 14:05. Is there a source I could go to so I can understand this?
Your video has been very helpful for me in completing my dissertation. Thank you.
2 out of 3 is just something that I use, but it's based on Cohen et al.'s linear regression book.
Thanks for your great explanation, how could I estimate the model fit?
You would use the overall F test results for that and R2.
Hi, very nice video :) Unfortunately, my model contains a nominal Covariate/ Control Variable (7 categories)
Do I have to dummy code it manually and then Insert the dummy variables into the covariate-box in PROCESS?
And if Yes: do I have to leave one of the dummies out as a reference category? (so: leave Dummy_1 and insert Dummy_2 to Dummy_7 into the box)
Yes, and yes - check this out: statistics.laerd.com/spss-tutorials/creating-dummy-variables-in-spss-statistics.php
Hi! Great video. It really helps with my work. You are a saint. I have a brief question. Would it help if I used MANOVA prior to the PROCESS to check for correlations between the variables? Thank you.
Or you could just run a correlation table, which seems much simpler?
OMG THANK YOU SOOOOO MUCH!!!
You are welcome!
Thank you for the great video! Is it possible to only plot the interaction terms of the linear regression models without including the main effects using PROCESS?
I think the graphs are basically on the simple slopes, so I'm not sure I understand exactly what you want to do?
Very useful for acquiring skills in moderation analyses! Dear Dr. Buchanan,m I would like to ask a small question. I have one moderator (continuous) and 5 independent variables (all subscales of spirituality). Is there a way to test for this model in PROCESS with all the main effects and interactions with each of them? I think it is not correct to test for five separate models (one with each indep. variable and interaction with moderator) as the subscales are correlated to each other, am I correct? Thanks in advance for any help you would give me. Paolo
I don’t know how many 2 way moderators you can do in process but unlikely you can do all of those. If they are super highly correlated, then maybe consider just doing one of them because of multicollinearity.
Hi, is it possible to use Process with 2 independent variables and 1 moderator ? In this case I look into the effect of two types of slack resources on CSR performance. The moderator is corporate ownership structure
Like one of them is a control variable or two two-way interactions?
thanks for your video. I have a question. Can plug-ins also test the moderating effect in logistic regression？
Yes! I think the newest version of process has this option, but I haven't looked at in a while so don't quote me on that. Otherwise, you'll have to do the steps manually using the logistic regression functions in SPSS.
When performing moderation analysis, testing the moderation effect (W) of a given moderator (M, e.g., sex) on the association of a independent variable (X, e.g., age) with a dependent variable (Y, e.g., height):
==> Must X have a statistically significant simple (main) effect on Y on step 1, so that a statistically significant effect of the interation term (W=X*M) is interpreted?
Is this a condition for a mdoeration analysis?
Note: Model 1, A. Hayes' PROCESS.
e.g.:
- Imagine that we are studying if age (X) predicts height (Y);
- And that indeed, age predicts height just in men, but not in women (which stands for a moderation effect of sex on the association between the X and the Y);
- The simple/main effect of age on height in step 1 (which considers the entire sample, men + women) might not be statistically significant because the non statistically significant observed on women may cancel/surpress/blur, the significant effect observed on men, and the interaction/moderation effect on step 3 is statistically significant.
No the interaction is the important part, you can have interactions with no main effects
Great video with easy explaination. I have two questions. Please give me response on both of them
1- I wanted to study 10 covariates which are nominal and having multicategories, can I include them all along in model 5, with one mediator,one moderator,one IV,one DV? I ran model 5 in spss and it didnot give any error, although i didnot transform multicategorical covariates into dummy variabels. I still need to transform these 10 covariates. Which will make my analysis so complex. What do you suggest?
2- I ran analysis, interaction result isnot significant but conditional direct effect are significant. How can i interpret? Also line graph of this output shows divergent at lower and higher values of moderator ,but convergent at mean.
Regards
I think you can use process to make the dummy coding for you - but yes all the multicategoricals need to be dummy coded. I don't have a good comment on having a lot of variables - they should all be included for a reason, so consider if they are all theoretically important.
If the interaction is not significant, then I would not interpret the simple slopes, as that result implies they are all statistically the same.
Hi Erin - great video, thank you so much for uploading! I just wanted to check my understanding regarding covariates / control variables within this model. Does the inclusion of a covariate control for its effect on X and Y or just on Y? Thank you in anticipation :)
They control for their effect on Y - which reduces the error and can change the effect of X on Y.
Hi Erin! Thank you for such a good talk about moderation using PROCESS macro.
Two questions:
1) Does it make sense to compute the standardized betas of simple slopes or conditional effects? Usually only standardized betas are reported in papers, but not unstandardized ones.
2), If so, how can we do it? How can we compute them from the unstandardized coefficients for simple slopes or conditional effects? PROCESS macro seems not to compute them at all, but reviewers are asking for them.
I'm ambivalent on the betas for moderation. They make interpretation of the coefficients difficult. I feel like I see just centered more, but you can do either based on what you see more in your literature. If you would like those, then z-score your X and M columns, but be sure to turn off the mean center option in PROCESS. (so you don't do it twice basically).
Hi Erin, I was wondering why you use the Normal P-P Plot of Regression Standardized Residuals to test the linearity assumption. I learnt that the plot is used for testing the assumption that the residuals of the dependent variable are normally distributed. I am confused as to why the plot can be used for both purpose. Maybe you can help me understand it.
I've seen both - you can either look at a regression line plotted on the residuals or a PP/QQ plot. Deviations from the line can indicate non-linearity. Residual scatterplots are also pretty telling that they make rainbows. If my memory serves me right, I took this from Tabachnick and Fidell.
Within the Process options, how do you account for a moderation that is a continuous variable and an iv that is dichotomous (yes/no)?
I think process can handle categorical variables just fine for moderation. I would put that category variable as M to get the different slopes for each group.
Hello, when I conduct it using HC4 the moderation becomes insignificant, when I chode "none" it is significant. it means my data heteroscedastic?
I would say that indicated some adjustment for heteroscedasticity yes - or that the coefficient is only slightly different from zero.
Hey, thanks for your great step by step tutorials, you're definitely saving my master thesis! When I add a dichotomous covariate such as gender (m/f), I can do this without dummy coding, right? And how do I interpret a significant effect of this covariate then?
Yes you can do it by just coding each gender as 0 and 1. A significant dummy coded variable implies that there is a difference in the means for each group on Y after covarying out the other X variables.
OK, that makes sense! Thank you very much :)
Hi Erin, Thanks a lot for this video! I have a question. My model is significant, but the only significant "thing" after that is one CV or my 2 CV (all others interactions are not significant). For my master degree , I did many analyses with 1 IV, 1 DV, 1 moderator and 2 covariable and really often the model is significant, but only the CV is also significant. In my results section, i report all that but i end up talking A LOT about covariables and not so much about IV and DV (wich are the real thing). Do i have to report all significant model that have just CV ? Do you have a solution so i'm not talking just about CV? Thanks to you. G
My vote it always just explain the model. If the CV is the only thing that’s predictive, then you say that.
Hi Erin! My output shows everything except the "conditional effects of the focal predictor at the values of the moderator." I still get the graph syntax code, so I'm able to graph low/average/high rates.
I would like to know the p-value for each of my simple slopes, is there any way to do that without getting that same output?
In the newer process, you have to force it to show you those effects if your interaction was significant. Don't remember exactly which window it is but change the p-value to something higher so they print out.
Dear Erin, thank you so much for another great video. I have a question about your interpretation. In the interpretation of the main effect, you mention first that Illiteracy isn't a great predictor of income (p = .508). But then you say, if I suppose to interpretate these results, I would have to say that as illiteracy increases, income decreases. Isn't that a bit of contradictory? The p value greater than .05 states that not enough evidence is found that the b that is found differs significantly from null.
Right but it helps to know how to interpret b as well - many people only focus on p and don’t actually know what the coefficient is. I usually try to teach both so you know what to say either way.
HI there, I came to this page for learning moderation first. But I really like your approach for hadingling outliers. Is there any particular references explains your approach (i.e. using Mahalanobis distance, Cook's D and Leverage values simultaneously?
Thank you!
It's a combination of the Tabachnick and Fidell and the Cohen, Cohen, Aiken, and West book.
This is so great! Does it matter if the covariates are continuous or categorical?
I believe I answered this on your other comment, but they can be either. If they are categorical, you will need to create dummy coded columns first.
@@morgandossey9877 Yes - dummy code is a way to represent categorical variables in regression analyses. You can learn how to do it here: blogs.perficient.com/2012/04/25/dummy-coding-with-ibm-spss/ or I have a few videos that cover the concept I can link. Mainly, you are creating unique values for each category, which creates pairwise comparisons when you run a regression.
Statistics of DOOM do you have to add all the dummies as covariates? I made dummies for educational level am now confused on how to add them
Great step-by-step video. I've also appreciated your talks on SEM. Thank you for these resources! I have two issues I've run into as someone new to PROCESS:
(1) I've noticed that the previous version of PROCESS allows for four moderator variables but the most recent version 3.3 only allows two (variables W and Z). I am trying to analyze a model that includes three moderators. Is there any way around this limitation in PROCESS? A similar limitation exists for use of multiple independent variables in, and I know running PROCESS iteratively (rotating predictors of interest through the IV and covariate slots) is the workaround for such models. Would it yield mathematically identical results if I ran all moderators simultaneously vs. running PROCESS twice, first with moderators 1 and 2 and then with moderator 3? Or is there a way to include a third moderator via syntax? I have one IV, one DV, 1 mediator, 3 moderators, and a couple of covariates in the model. If I ran the model iteratively, I assume I should include omitted moderators as covariates?
(2) Is there an option to select in PROCESS a mediation model with two moderators directed at the direct effect of X on Y? I couldn't find that option from among the model templates, but it is a fairly basic model that I would expect to be preset so I wonder whether I am missing something. I had originally had in mind a concept like Model 10 but with no arrows from the mediators to the X->M path. Do you recommend just using Model 10 or customizing it?
I don't know about what's going on with the updates, as I only use SPSS for videos :).
Definitely run a model with all the moderators at once, as it's not the same mathematically if you are rotating some in and out (remember that coefficients are "controlling for all other variables, X predicts Y with b slope). Also if you limit to two moderators and actually want the three way moderation, you will not get that. You can do them manually in SPSS without process!
For the second one, I think customization might work - the bootstrapped CI is the harder part in SPSS, which is why I teach a lot of materials in R for more flexibility.
How would you interpret a non-significant interaction but a significant covariant of ethnicity? I previously recoded this variable in to values 1 and 2. Thank you :)
Does a dichotomous covariate have to be dummy coded? E.g. I have gender -with values of 1 for female and 2 for males already? Can this variable be placed in to the covariate box as it is?
Dear Erin, thank you so much for these PROCESS videos! I have a question regarding the significance of simple slopes. So the interaction is significant but none of the slopes are, the case that you already mention in the video. However, if I further check the Johnson-Neyman output, I see that when the moderator value is either very high or very low, X actually does predict Y. My question is: is it okay then to report that there is a moderation effect, but only when M is x-times above/below the mean?
Yes that's fine! I would just write up that you used the Johnson-Neyman approach to find the regions of significance, rather than the simple slopes approach.
@@StatisticsofDOOM Perfect, thank you so much for your reply!
Hi! I have a question about moderation. I have a moderated moderation model (model 3) with a continuous DV, a dichotomous IV and 2 dichotomous moderators. Under Options, should I click mean centering or not? I find contrasting answers on the internet and I don't really know what to do. The main effect between the IV and DV is very insignificant with mean centering but significant without it. Thank you so much in advance!
The main effect is insignificant when doing an ANOVA or t-test btw haha!
No, categorical moderators should not be centered. Centering is only for continuous variables.
@@StatisticsofDOOM okay thank you!!!
Hi Erin,
Please help!! I am doing an MR moderation analysis using PROCESS for my masters thesis. I am unable to obtain the Hayes Book as there's only 1 hard copy in reference at the Uni library and due to restrictions I am not allowed on site to view it. The people I have approached to discus this are unsure about what to do. My model has 1 IV, 1DV, 1M - which are all continuous and 3 covariates - 2 are continuous (age and number of training sessions) and 1 is categorical "position".
1) Can I use 3 covariates in PROCESS model 1?
2) Can I use a multi-categorical variable (i.e. 3 groups) instead of a dichotomous variable (dichotomous does not work conceptually and would not make sense) in PROCESS model 1 and get meaningful results from the output?
2) If so, do I assume I would create dummy variables for group 2 and 3 (against reference variable 1), and enter these into PROCESS as the 2 covariates representing "position"? Thus, I'd have 4 covariates in total in the model.
3) How would I interpret the results, particularly from the covariates?... especially If all relationships came out as not signifiant except for one or both of the 2 "position" covariates
4) If this is not valid for use in PROCESS , what other analysis would be MOST valid here instead please to determine a moderation effect of M on the IV/DV relationship, taking account of the other covariate variables?
5) In a second situation it may be that 2 of my covariates are actually control variables and 1 is a covariate. How would I then treat and enter these 2 control variables into a moderation analysis (with the same IV/M/DV as above) using PROCESS? Would this second model still come under model 1?
I'd really appreciate any help with thsi asap please. Thanks in anticipation!
1) You can use the covariates - the categorical one may need to be dummy coded first depending on the number of categories.
2) Yes but dummy code it first stats.idre.ucla.edu/other/mult-pkg/faq/general/faqwhat-is-dummy-coding/
3) I have a few videos on dummy coding in regression, I would recommend those for understanding how to interpret dummy coded variables (basically they are measures of group differences)
4) Process is just a software plug in, categorical variables can be used in regression just fine because things like t test and ANOVA are just special versions of regression with categorical predictors
5) Do not mix terms here moderation and mediation are different things. I would treat these as separate hypotheses unless you are doing moderated mediation. Practically, there isn't a mathematical difference in regression if one is a control and one is a covariate. They are all entered into the equation as predictors.
@@StatisticsofDOOM Thank you so much for your prompt response. I will take a look at the videos and link you have given here. Much apprececiated!
Hi Erin, very good explanation! It is completely clear what to do in Process version 3.1 and how to interpret the outcome. Most video's use old Process versions with different options. However, my model summary is significant, however the interaction in the model under Model Summary wasn't (p = .0706). So, the rest of the output is not interesting anymore, unfortunately, because I did fixed it with a nice graph :-). Now I must report not only in writing what I have been doing also in a table. Have you got an APA style table for this outcome to report? Or where can I find one. I didn't yet find in on Google. Thanks in advance.
We've made some regression tables in this paper: osf.io/fcesn/ - hope that helps!
Sure thing - here's an example of a published paper using moderation we did: www.aggieerin.com/pubs/schnetzer%2012.pdf
So, I have an experiment and want to see if participants' score on a test moderates the impact of the independent variable on the dependent variable. I am able to follow these instructions for Model 1 in Process. Conceptually I want to understand how it is different from using the univariate GLM in SPSS...when I use SPSS GLM and enter my IV as a fixed factor and test score as covariate (and then go to "model" and make sure the IV, the test score, and the IV test score interaction are all in the model) -- the F statistic and p value are the same for the interaction term and the IV in GLM and in Process. But the F statistic for the continuous moderator is different between Process and SPSS GLM. Do you have any idea why this might be?
Do you mean the t-score is different? The F score is for the overall model, so I'm not sure if I'm following. If the t-score is different (i.e. the coefficient test statistic), I believe that's because process uses a correction for heteroscedasticity on the standard errors (and the formula for t is b / SE, so that will change the values between that and the normal GLM).
@@StatisticsofDOOM Thank you so much, and sorry for the lack of clarity! When I do SPSS, Analyze, General Linear Model, and then Univariate and put in my IV (with two conditions) and my moderator (as a covariate) and their interaction into the model, there is a table called "test of between subject effects" that gives me a F value for the entire corrected model that matches what process gives me. Then there is an F score for my IV, my continuous variable, and their interaction. For the IV and the interaction, the sig. values EXACTLY match the t score sig values for the IV and interaction. (And the F score is the square of the t-score in process). But the sig value for the continuous predictor does not match what is in process and the F score is not the square of the t score
Ah, oops. I realized that I confused the GLM test of between subject effects with the GLM parameter estimates. These DO match what Process spits out -- except I have to reverse the coding of the independent variable and put the reversed one into Process to get it to match.
Hello, first of all i would like to thank you for comprehensive guidance. I want to ask if i have one moderator, one mediator, 1 dv, 1 IV, along with 3 covariates in the same model then how to get the results using hayes process and how to interpret the results.
That really depends on where the moderation is. There's a file (templates.pdf or the back of the book) that includes the different pictures and how the moderation and mediation interact. You'd have to pick which one theoretically matches your hypotheses. The CVs can go in the covariates box.
@@StatisticsofDOOM I checked that pdf, process model 5 fits my research model
Thank you for this video.
Got two questions according to issues with the linear dependency error.
One of my moderators is only measured when X='Yes', this leads to a linear dependency error, so I can not run the PROCESS. Some say you have to add zero's for the structural missings (X='No'). This doesn't work, same error.
There has to be a way to measure this, right? Do you know how?
The same problem occurs when you ad multiple dummy variables as covariates.
For example, when age is a categorical variable.
Maybe you can help me, I hope so.
Thank you!
If the variable is only measured for that group, then you could do a simple linear regression of M to Y for that group. You wouldn't be able to add it as an interaction because it doesn't exist for the whole sample.
A new subscriber here! Great video! I hope that you could make a video illustrating Model 14 with some covariates (e.g., 1 dichotomous and 1 continuous). I have only found 5 videos describing Model 14.. I am not sure why. Anyways, thank you for sharing your knowledge! ^^×
Thanks! I'll see if I can add that to the list!
@@StatisticsofDOOM Thank you! I have found your videos very clear en I hope to learn more from your upcoming videos!
감사합니다~^^ 도움이 되었어요!
아니에요 (my best google translate!)
Hello. Thank you so much for this video. It helps me a lot to finish my assignment. I have a question. What do you think if IV is correlated with DV in Pearson correlation, but not in Process Hayes? Can you help me to explain/interpret this case for me? For the context, I have one IV, one DV, one moderator, and one covariate.
That implies that the IVs are correlated with each other, so the individual IV isn't predictive.
@@StatisticsofDOOM Thank you for your explanation. But I'm really sorry that I still confuse and don't understand. Can you explain more about it?
Hi Erin, Thank you for the video! I was wondering how you chose your covariates in the first place (i.e. is it simply a matter of choosing them based on common sense or are there statistical tests you need to run to determine them as covariates before including them in your main analyses?). Thank you :)
I usually recommend picking these based on theory - known variables that you should control for (which might be common sense!).
Hi Dr Erin,
I'm computer engineer and now doing my masters in engineering management. I'm investigating the impact of Supply chain integration(customer Integration, Supplier Integration & Internal Integration) on supply chain performance with a moderating role of enterprise systems.
In this way, I have 3 IVs, 1 DV and one moderator.
Since I'm new to the social sciences field, Could you please suggest and guide me how can I use SPSS to analyze data and interpret results ?
What tests do I need to carry out for my research to be legit and authentic?
I'm not sure I follow the question - this video shows you have to do how to run the moderation .... kind of depends on where you want the moderation to be (i.e. only IV1 or all IVs).
@@StatisticsofDOOM I want the moderation on all IVs.
Pls help me with this
Regards
Hi Erin, if i have 4 IVs and 1 CV, do i need to run this method 4 times in PROCESS, with the same CV each time? If yes, what in your view is the best way to report the results of 4 IVs + 1 CV? I've not been able to find videos of how to report this, so your input would be extremely helpful!
Also, how to report if the IVs' main effects are significant, but the interactions are not significant?
That depends on the hypothesis really - if you are proposing four separate interactions, then you can run them all separately or run each two way interaction within the same model (not with process though). With that many variables, I would just put all the B SE t p values into a table. The reporting is the same if significant or not, but you will not report simple slopes if the interaction is not significant.
Thank you for your reply, Erin! I followed your advice and put the B, SE, etc. values into one table. I have a 2-part follow-up question with regard to covariates:
1) what does it mean if a significant interaction becomes insignificant after adding in a CV? (The CV was included in the regressions because it is theoretically related to the DV, even though it's not part of the main model.)
2) conversely, what does it mean if an insignificant interaction becomes significant after including the CV? For this, I came across an explanation for suppression - in your opinion, would that be a plausible explanation?
1) that usually indicates that the CV and the interaction were predicting the same variance (i.e. those are correlated as well). Or that the variance the interaction predicts is not more than zero after accounting for the CV.
2) Same idea but in reverse. After account for the CV, the interaction is significant because controlling for the error that was actually the CV increased the usefulness of the interaction. Suppression is what folks normally call this effect yes.
Statistics of DOOM thank you so much again, Erin! Your explanations in plain language have really helped a non-statistician like me to understand what I’m seeing and writing. I think my dissertation has some glimmer of hope now because of your help. Thank you!