I really appreciated your working and detailed explanation. It just very helpful for me especially testing assumption. Because I can easily access entire assumption analysis. Thank you so much from Turkey.
Just spent four hours on this whole thing and I cannot thank you enough for the detailed approach you took to explaining not just the process of PROCESS, but what everything means right from the outliers and data screening to the interpretation of the results! Thank you!
Hi Dr. Buchanan, do you know whether we can specify which variable the covariate is for in this version of process? In the previous version, we can specify whether the covariate is for M, for Y or for both. There must be a reason the newer version does not provide that function anymore.
I have a question, before we add covariates to our model should we run a preliminary test to see if we should put them into the model or not? If so what analysis should we run of these?
That's an entirely theoretical question. I would not recommend testing with and without unless you planned to do that before because it encourages p-hacking.
So....if direct and indirect effects are significant but the mediation effect isn't, what does that mean? Does that mean it's not even a partial mediation?
For the G*Power example, can you explain a) why you used "from correlation coefficient" and how you determined how .06 is a medium effect size for R2? (can you provide the citation?) and b) why did you not just use the conventions for f2 .02 (small), .15 (medium), and .35 (large) in the main window?
Those suggestions are from Cohen 1988, 1992. I easily remember these and not the f values but you can use either one. You should be able to justify the numbers you used based on previous research, but for these examples I just use the "standards".
@@StatisticsofDOOM Thank you for the quick reply! Can you please explain in detail why you went to "determine" and calculated from the correlation coefficient and used the squared multiple correlation of .06?
@@kellyberthiaume8094 Because, as I said on the last comment, I can remember the rules for R2, and it's the effect size I normally use. I have never seen an f2, so I would have trouble knowing what those numbers meant. It's really just about preference.
Hi Erin, thanks for the video! Quick question: when all paths show significance except the covariate (but insignificance when the covariate is not included), does this mean that the covariate actually influences the effects/ model or not? In short: does at least 1 covariate have to show significance to be able to interpret a Model 4 including covariates? Thanks a lot already!!
That sounds like the covariate is covering some of the error variance and so the other predictors are significant when included. Their effect sizes are likely small, so you should look at that for interpretation.
Hey, I have a question about the path from the mediator to the dependent variable. Assuming I want to test a model with multiple mediators using PROCESS and I have formulated a hypothesis that states that there is an association between one of the mediators and the DV. Is it then the right way to report the beta coefficient (for the relationship between mediator and DV) from the PROCESS output? I am asking myself this question because this beta coefficient is influenced by the IV and the other mediators... Alternatively, one could also calculate a separate simple linear regression (mediator as IV, DV as DV), so one could obtain a beta coefficient that isn't influenced by the IV of the overall model and the other mediators. Thanks in advance :)
I would report the coefficient in the model with other variables controlled for - so the final one (path b). I think there's a bit of black and forth on if it should be unstandardized b or beta, so doesn't hurt to report both.
@@pnaracet1562 modelinin ne kadar açıklama gücü olduğunu göstermek için (en basit haliyle daha yüksek açıklama gücü = daha iyi model, daha iyi model = daha genellenebilir model)
Hi Dr. Buchanan. I'm a graduate student and your videos are helping me a lot! I have one question. In this video you mentioned a "rule of thumb" for deleting/excluding multivariate outliers than violate the three tests (Mahalanobis, Cook, and Leverage). Do you have a reference I can read for that? I need to explain why I'm excluding some cases in my data using that rule. Thank you.
I use Tabachnick and Fidell 2012 for Mahalanobis and Cohen, Cohen, Aiken, and West (08? can't remember) for the others. These just explain the rules, so you just have to justify the number of strikes they get. I like using two strikes you are out because each is a bit too sensitive on their own, but taken together it's usually a good indicator they are outliers.
Hi Erin, you are a lifesaver! I just couldn't wrap my head around mediation before I stumbled on your video. I have a question though, how can I calculate effect size? You be=riefly mentioned effect size in the video but didn't explain what analysis to run. Also, it would be great if you could write out the APA results as a full sentence not just the numeric results. Do you also have an APA table or figure for summarizing the mediation results?
I would report R2 for the overall effect size, but then the indirect effect would be the "effect size" for the mediation. There used to be effect sizes embedded into Process but they were removed after several flaws were discovered. You can view some example write ups/pictures by looking through the files on our OSF page.
No real reference other than using multiple indicators is a good idea - you can use Tabachnick and Fidell and/or Cohen Cohen Aiken and West for the limitations/sensitivity of each of these, so it's better to combine several indicators (because one may be too liberal or conservative).
@@StatisticsofDOOM how do you personally cite this method when you use it in research articles? I followed along your videos and used this method for my thesis, but now can't find a good way to support it in my manuscript and my advisor is not happy haha
@@roxannefelig6101 If you want to cite process, use Hayes (afhayes.com/introduction-to-mediation-moderation-and-conditional-process-analysis.html), data screening is Tabachnick and Fidell (2012) and Cohen Cohen Aiken and West (2008? I think).
Thank you so much for your wonderful video! You are legend!! I've got one question. If the indirect effect is significant, which means the mediation exists, and meanwhile, one covariate also has significant effect on the dependent variable (based on results from the process). Thus, what conclusion could we have?
@@StatisticsofDOOM Thank you. I was also wondering if I would like to use G*power to calculate the sample size for mediation analysis, what are the acceptable power and effect size? Thank you very much!! I used 0.15 as effect size and 0.9 as power but my tutor told me that the effect size was too small. But I noticed that you were using 0.063 as effect size cutoff.
@@tingyuesun5017 I mean effect size is what it is - it should be best on some educated guess on what you think the size of the R2 is for example. If you were using .15 that would be considered a large effect in R2. Often, people make power only .80 as well, so it just depends on what's normal for your area of research.
@@magillroad Yes and no? Like mathematically they don't really change, but if you are just using them as a control, then you might just mention it and move on. However, if you think they are important, you might spend more time on it. The idea of b is the same either way though - if continuous, then for every one unit increase in X, you to get B changes in Y, etc. If b is categorical you would need to use dummy coding interpretation (one group versus another).
Thank you do much for this video. I am using model 4 for a mediation analysis with covariates. My Y variable is dichotomous. Are the methods you are using to test for assumption also applicable when the Y variable is dichotomous? (i.e., a. Mahalanobis, Cooks, Leverage, looking at zpred and zresid).
No, you should not use linear regression for that analysis. You should use logistic regression (although the same steps can apply, but I am unsure process does this at the moment).
@@StatisticsofDOOM Thank you for your reply. I am wondering about checking for outliers. Can I still look at those Mahalanobis, Cooks, Leverage values? Are there other indicators for logistic regression for outliers?
@@mys1990 I think generally I don't do a lot for outliers in logistic regression, but you could check their influence on the solution (cooks/leverage) and mahalanobis pattern of scores. This website gives some thoughts too: stats.stackexchange.com/questions/26930/residuals-for-logistic-regression-and-cooks-distance
Hi, thanks a lot for the explanation! It's great. If I compare two Mediations. One with covariates the other without leaving all IV, DV and the Mediator the same which values do I have to report for the effect of the Covariates on the DV? And how would I explain the difference in the indirect effect because of the covariates adjusting for error type II? Would be great if you can help me.
If there's a reason to use the CVs I would only report the analysis with the CVs...if you are trying to show how the CVs affected the analysis, then I might report both and stick everything in the table. If the CVs are related to the DV, then the changes between analyses would be because of error reduction in the DV by including those CVs.
Hi Dr. Erin. Thank you for the video. I am wondering with the degree of freedom for mahalanobis , do we need to consider covariance as the predictors? I am running moderated mediation (model 8) and have a case that has mahalanobis distance of over 20. If I just use my IVs as my predictors it is considered as outliers but if we consider also covariance as predictors then, the case is not outliers. What do you think is more robust or I should do?
First of all, thank you for those great videos, they are really really lifesavers. I would like to ask one question. I have conducted different mediation analyses with 111 cases [continuous Xs, continuous Ms, and continuous Ys (only one of them was measured with one item, 0-5 range). mediation was not sig, but what is most interesting is that Rsquare values were too low ( about .06 mostly). What could be the reason behind low Rsquares? I would appreciate your answer.
Hello Dr Buchanan. Your videos are very helpful, to interpret my output, thank you!!! Although I have a question... my SPSS book explains that the ZPRED goes in X and ZRESID in Y, however you do it the other way around. Is there a specific reason? Thank you!!!
With the way we interpret the pictures, it doesn't matter if you flip them. Mainly, you are concerned with the pattern/spread around zero and the shape of the dots. So, if it's bad, it's bad both ways.
@@StatisticsofDOOM Thank you for this great video. In the Linear Regression to check for outliers, you don't dummy code the categorical variables. The regression seems to work perfectly fine. Is the dummy coding solely necessary for interpretational purposes then and not for the feasability statistical procedures?
@@shannawielinga8207 haven’t watched this one in a while but generally you don’t have to check the categorical variables because they naturally cannot be normal - now they can be outliers but SPSS does not make this easy. I don’t believe there are any in this video? Either way you can dummy code them and then screen them just like in the video.
Hi! In a statistical diagram the relation between the covariate and the dependent variable is sometimes called the d-path. My supervisor has asked me to report the results in the diagram (instead of textual). You report about the d-path multiple times, which results am I supposed to report? It can be found both in the total effect model and the model where you also get the c' from (I think you called it the full model or direct effect model). Which one would be the "d-path" I'm looking for? Also if I would want to report the results in a table rather than a diagram, is this possible? It's hard to find articles online where they do this.....
Certainly you can put things into a table. If you are listing the covariates as a d path, you probably have to have a d path (just the x to y model) and a d' path (the x and m to y model). I might recommend a table of all the coefficients for each model to make things a bit easier to understand (as that might get confusing on a picture).
@@StatisticsofDOOM Thank you so much for your quick reply! I'll share a link to a document I found online. On page 24 and 25 the writer reported it with tables for each model, would you say that's the right way of reporting? : arno.uvt.nl/show.cgi?fid=141735
Hi Erin, thanks so much for your informative videos. I'm looking at longitudinal data, and want to see whether change in one variable mediates the change from X --> Y. Can you use change/delta scores as the mediator? When I run it in SPSS, the program seems to have trouble with a negative value as the mediator (i.e., it's negative because i'm looking at change/ improvement over the course of treatment). Any advice?
Dr. Erin M. Buchanan, thank you for this video! I would like to ask that can we use a centered variable as a covariate in Mediation analysis model 4. I calculated SES score as a composite in my data set and since they are in different range I used ZSES as a covariate. Thank you
Hi, These videos are useful and accessible, so thanks for creating them. Do you have any advice for mediation analysis for longitudinal repeated measures data? I have 3 data sets with the same mediator and a potential categorical covariate.
Unclear - depends on what you want to do with those predictors. I would decide how and what you expect to mediate and draw out your "triangle" diagrams to help you pick a model. There's a templates file (or it's in the book) to help visualize.
Hi! Thanks for your explanation! I have one quick question: what do I have to do if I expect that my continuous covariate has only an effect on Y (and not on M)?
The CV only goes into the model for Y I believe ... can't totally remember but if not, you might have to run each step separately, so it doesn't predict M.
@@aesthetics1110 Covariates are other variables you include the model that adjust the DV for their known variance so you can examine the impact the other variables have without them.
Hey Erin, will this work in SPSS if I want to use glm not lm for the regressions? If one of my variables is skewed (eg. reaction time data) - can it do glm, or do I have to transform (would rather not do that)? Thanks!
@@StatisticsofDOOM Lol I am not sure what I was asking. In R, I run my regressions as glm as often have skewed data. In SPSS, Process assumes linear regression, so wondered if it would work if my variables do not meet assumptions. Also had a binary mediator in one instance which Process freaks out about.
@@brittjane4486 OH! I get it now. If you only wanting to use process, I would suggest transformation, as I don't think it allows for anything other than least squares assumptions. OR just do the whole thing in R - I have been writing a package for it, but got sidetracked with other projects. I do think I have this type of thing written though! It doesn't do glm, but that might be something I can add?
Dear Dr. Erin, thank you so much for this video, I got one questions to confirm: I got significant indirect effect and nonsignificant direct effect, it indicates full mediation, but the total effect is not significant, can I still make this conclusion, i.e., whether a significant total effect is a pre-requirement to test mediation effect, i cannot find a conclusive arguement on this issue, thanks a lot!
You should have a theoretical reason for the variable you decide is X, as it's part of the conceptual mediation portion of the analysis. Covariates should be variables that wish to control the variance for.
Hi Dr. Buchanan - Thank you for the video! If my total effects model summary and full model with mediator and covariates account for significant variance but my paths (total effect and indirect effect) contain zero, what can I conclude? That the variables in my model account for significant variance but do not mediate the relationship between X and Y?
Hi Erin! Thank you so much for this video, I still have a doubt though. If my IV has two conditions should I insert the computed variable in the process analysis or should I perform model 4 analysis twice, one per condition? Thank you again
I believe those numbers are based on Cohen's effect size papers, but I would use what is normal in your field (as he argues, that average sizes should be based on the field)
Thank you so much for the video, really helps for my work! May I ask if I find the covariant is significant, is there anything else I need to do? Or just need to write it is significant and do nothing.
Thank you so much Erin! Just another super useful video on testing mediation. I just have a question: so I find X predicting M (a path significant), M predicting Y (b path significant), but indirect effect not significant (direct effect also not significant). So apparently there is no mediation effect. But can I still report the significant a and b path? And what might be possible explanations for the non significant mediation?
Sure, I mean you would just report what you found. Can't help with the mediation non-sig question, could be that it's truly not there (doesn't exist), you didn't have enough power to find the effect (sample too small to see the meditation), or some other theoretical reason.
HI, i want to know if control variables are related to only DV in model 4, then how to do that in process 3 and above version as both do not give us the option to select "if covariate is linked to only dv or to both mediator and dv" just like the option that "process v 2" used to give us in model 4.
@@StatisticsofDOOM e.g. i want to test the relationship of role conflict with turnover intention through mediating role of anger. Previous research has shown that work overload is related to turnover intention. Now i want to control work overload to see the pure relationship between role conflict and turnover intention through anger. Now in "process v 2" model 4 gives the option "covariates in models of (a) both M and Y, (b) M only (c) Y only". whereas in process 3 and 3.2 model 4 doesn't give us this option and output contain control variables in first output (a path) as well, which otherwise do not appear if we run model 4 in process 2 by selecting option "covariates in model(s) of (c) Y only". I like to know that how to have this "covariate in model of Y only"? as this command is not available in process 3 and latest version. Regards
Ok, I think I get what you are saying - I think in process 3 you just have to pick the model that includes the variables in the right places...so for it to covary with Y AND M, you would need to find one that includes it as a variable in the template pictures. Otherwise, it will only predict Y.
Thank you!! This is extremely helpful. Looking at the very end of the output - the section titled "Total, direct and indirect effects of x on y" : How does one interpret the raw magnitude of the two numbers under "total effect" and "direct effect" respectively? In my data, I am also seeing a change, from .0120 (total effect) to -.4246 (direct effect). Unlike in the data you used here, I do not have the value of 0 between my two confidence, indicating that mediation is happening. BUT I do not know what to make of the numbers .0120 and -.4246? Struggling to understand what they mean.
The total effect is just X on Y (c path). The direct effect is X + M on Y (c' prime). These are coefficients, so you interpret them as for everyone one unit increase in X, you get c or c' units increase/decrease in Y.
No - Hayes has suggested these are not appropriate for analyses in his books. If you still want to calculate them for comparison purposes, you could z-score your variables first, then run PROCESS on it again.
Thank you so so much for your videos, they are extremely helpful! I appreciate you get a lot of comments and are super busy but I have a challenging question that is driving me a bit crazy. I have a multicategorical IV, and want to run a mediation analysis with 3 mediators, a moderator, and about 4 covariates. I'm wondering what PROCESS model to use and how to do it? Thank you so so much and I really hope you can give me a bit of advice ♥️
Hi Erin, please if you can just give me a direction I would be really grateful. I'm on my final PhD year and am really really stuck on this, and can't find anyone familiar with mediation and process
I’d recommend checking out the possibilities in the diagrams that are available with the book - hard to know since what you are describe is pretty complex and where the moderator goes. You’ll have to see if one of the model numbers matches the conceptual diagram.
This is going to save my master's thesis, thank you so much!! One quick question: I work with a model just like you, but I use two different DVs to capture the same construct in the same sample. So there are actually two mediation models that are conceptually related. Do I have to worry about alpha inflation and correct for it? Thanks in advance!
Dear Arin, thank you for the videos. I want to conduct this commend process vars=y1 iv1 iv2 iv3 med1/y=y1/x=iv1/m=med1/boot=10000/model=4 /seed=5235. But I cannot, it does give me neither a result nor an error. I want to do that to have bootstrap confidence intervals based on the same set of the simulated sample from the data at each time I run my analyses.
Hi! I found a significant indirect effect of my mediator. However, when I perform the sobel test with the t-value of the A and B path, the aroian p-value is 0.089. Does that mean that my mediator isn't a mediator after all? Or can I choose not to report the Sobel test if it isn't always necessary?
I'm not sure what you mean by "significant indirect effect" - do you mean that the indirect confidence interval does not include zero? How did you run the sobel test?
@@StatisticsofDOOM yes the confidence interval did not include zero, I ran the test through the website you shared, and used the t-value of the a and b path, all p-values were > 0.05 (but < 0.09)
@@dunja2610 Got it! Thanks for the clarification. I would say more people like the bootstrapped CI to determine if the mediation "occurred" and like the Sobel test less. It does sound like the CI is probably close zero though with those results.
Hi! Your videos are very helpful, I love them. But what I need the most is a tutorial video about Model 5. It is impossible to find one on RUclips like nobody uses Model 5? It would be great if you could share a video about that or a helpful link that you know. Thanks a lot in advance!
I have a few that cover model 7, which is similar but the moderation is in a different place. Same concept though: ruclips.net/video/DOQkBcT-Zzk/видео.html
Many thanks for your videos. They are really very helpful especially the interpretation of outputs. I have a quick question.What happens if total and direct effect are non-significant but direct effect is? How do we interpret this?
Thank you so much for this helpful video! I had a question about my results. I first ran results without covariates, and the total effect was significant, as was the indirect effect (direct was non-significant). Reviewers suggested I add in covariates, so I ran it again with covariates, and now the total effect is not significant, but the indirect effect continues to be significant (again, direct effect was non-significant). How do I interpret this? Are the covariates explaining more than my X variable? Thank you!
I also want to add that the indirect effect size has decreased with the addition of covariates - I am not sure if this means that although the mediation still exists, it is likely that the mediator is also interacting with the covariates, thereby reducing the effect size of the relationship between IV and mediator?
@@anushaa You'll need to check the sobel test or the bootstrapping results to know what your "mediation" effect is, not just the total/direct effect. The addition of the CVs just controls for their variance - so the changing values indicates that they predicted some of the same variance in the DV that your IV did.
Thank you so so much for these videos! It helps me so much with my thesis! I have a little question regarding the p-values: do I need to split the p-value if I have an one-sided hypothesis? and if so, is there a shorter way in the calculation process? e.g. by putting the alpha level on 10% rather than 5%?
If you wanted to make a directional prediction, you could divide p value by 2. I'm not sure that makes a lot of sense in a mediation model, as mediation includes a two sided bootstrap confidence interval.
@@StatisticsofDOOM Thank you so so much. I have read about this analysis in a research paper, however, the authors performed it through SAS. I would be really grateful to you if you can teach me the same analysis through SPSS.
@@StatisticsofDOOM Hi. I am extremely sorry for bothering you again and again. If you know MLM for logistic regression only in R, could you upload tutorial for that? Please? Also, along with that tutorial, could you also guide me how to run "Commonality Analysis"? If you need any research studies I am replicating, please let me know, I will share with you. Again sorry for the inconvenience caused. And waiting for a tutorial video soon. Bundle of thanks.
@@khadeejamunawar2529 I'm not sure if I know what a communality analysis is? I will work on a log MLM sometime soon, as I have a bunch of deadlines over the coming weeks.
@@StatisticsofDOOM This is so kind of you. I am anxiously waiting for your SPSS tutorial for MLM when outcomes are categorical (3 or more categories). And good luck for your deadlines. Thank you
Hello Dr. Erin, thank you for the insightful tutorial. I had a quick question on how to report the tests' results. I have ran a dozen of tests using Model 4 and all models were insignificant. In other words, the models' summary shows that the model was insignificant. Can I directly conclude that there is no mediation effect, without having to go through all details in my thesis? Also do you think that this insignificance could be the result of the data sample I have? I have only 101 respondents for my thesis. Thank you for your time!
Unfortunately, that's the problem with null results - it could be one of many reasons. Power (too small of a sample to detect this effect) could be the issue, could also be that there's no mediation, could be that there's no mediation in this particular sample or with these particular variables, etc. I usually just say that the hypothesis of mediation was not supported (which is not the same thing as no mediation at all, just that no mediation was seen in this data with this analysis).
Hi Dr. Erin M. Buchanan, Does paralel mediation also use model 4? For the present example does it mean after controlling covariates mediation did not occur?
if covariates are added to the mediation model, are they excluded from the mediation? or are we able to say that they predict the mediator / outcome variable and do influence the mediation (if the indirect effect of the mediation is also significant)
I really appreciated your working and detailed explanation. It just very helpful for me especially testing assumption. Because I can easily access entire assumption analysis. Thank you so much from Turkey.
Appreciate the kind words!
Just spent four hours on this whole thing and I cannot thank you enough for the detailed approach you took to explaining not just the process of PROCESS, but what everything means right from the outliers and data screening to the interpretation of the results! Thank you!
Glad to be help!
Same here. Thanks a lot from EGYPT Prof. I'm a Genetic Anthropologist
great job on this. You've been piggy backing me through my thesis analyses!
Thanks for sharing this, save my thesis!! Really helps a lot!😀
Thank you so much! The most helpful video I've ever watched for analyzing process model4
Thank you!
Hi Dr. Buchanan, do you know whether we can specify which variable the covariate is for in this version of process? In the previous version, we can specify whether the covariate is for M, for Y or for both. There must be a reason the newer version does not provide that function anymore.
I'm not sure - I would think that all covariates would be for Y in this type of analysis.
Very clear, thank you very much
Glad to be of help!
Thank you so much for this video! You are a queen, Dr. Buchanan.
Glad to be of help!
I have a question, before we add covariates to our model should we run a preliminary test to see if we should put them into the model or not? If so what analysis should we run of these?
That's an entirely theoretical question. I would not recommend testing with and without unless you planned to do that before because it encourages p-hacking.
Thank you so much. I have one question if the CV is the multi-categorical variable. Do we need to switch to dummy variables?
Yes, you'll need it to be effects coded in some way: stats.idre.ucla.edu/other/mult-pkg/faq/general/faqwhat-is-effect-coding/
This video is really helpful! Thank you so much and I really appreciate it.
Thanks for the kind words!
So....if direct and indirect effects are significant but the mediation effect isn't, what does that mean? Does that mean it's not even a partial mediation?
Correct.
For the G*Power example, can you explain a) why you used "from correlation coefficient" and how you determined how .06 is a medium effect size for R2? (can you provide the citation?) and b) why did you not just use the conventions for f2 .02 (small), .15 (medium), and .35 (large) in the main window?
Those suggestions are from Cohen 1988, 1992. I easily remember these and not the f values but you can use either one. You should be able to justify the numbers you used based on previous research, but for these examples I just use the "standards".
@@StatisticsofDOOM Thank you for the quick reply! Can you please explain in detail why you went to "determine" and calculated from the correlation coefficient and used the squared multiple correlation of .06?
@@kellyberthiaume8094 Because, as I said on the last comment, I can remember the rules for R2, and it's the effect size I normally use. I have never seen an f2, so I would have trouble knowing what those numbers meant. It's really just about preference.
Can nominal data/categorical independent variable be used as covariates?
Yes, and there's a place somewhere to denote that they are categorical in the process windows.
Hi Erin, thanks for the video!
Quick question: when all paths show significance except the covariate (but insignificance when the covariate is not included), does this mean that the covariate actually influences the effects/ model or not?
In short: does at least 1 covariate have to show significance to be able to interpret a Model 4 including covariates?
Thanks a lot already!!
That sounds like the covariate is covering some of the error variance and so the other predictors are significant when included. Their effect sizes are likely small, so you should look at that for interpretation.
Hey, I have a question about the path from the mediator to the dependent variable. Assuming I want to test a model with multiple mediators using PROCESS and I have formulated a hypothesis that states that there is an association between one of the mediators and the DV. Is it then the right way to report the beta coefficient (for the relationship between mediator and DV) from the PROCESS output? I am asking myself this question because this beta coefficient is influenced by the IV and the other mediators... Alternatively, one could also calculate a separate simple linear regression (mediator as IV, DV as DV), so one could obtain a beta coefficient that isn't influenced by the IV of the overall model and the other mediators. Thanks in advance :)
I would report the coefficient in the model with other variables controlled for - so the final one (path b). I think there's a bit of black and forth on if it should be unstandardized b or beta, so doesn't hurt to report both.
thank you so much
Your videos are great. Thank you very much for your effort. Could you please give an example of CHAID analysis?
I had to look that one up - it does look neat - I will add it to the list of things to add examples for.
@@pnaracet1562 modelinin ne kadar açıklama gücü olduğunu göstermek için (en basit haliyle daha yüksek açıklama gücü = daha iyi model, daha iyi model = daha genellenebilir model)
Hi Dr. Buchanan. I'm a graduate student and your videos are helping me a lot! I have one question. In this video you mentioned a "rule of thumb" for deleting/excluding multivariate outliers than violate the three tests (Mahalanobis, Cook, and Leverage). Do you have a reference I can read for that? I need to explain why I'm excluding some cases in my data using that rule. Thank you.
I use Tabachnick and Fidell 2012 for Mahalanobis and Cohen, Cohen, Aiken, and West (08? can't remember) for the others. These just explain the rules, so you just have to justify the number of strikes they get. I like using two strikes you are out because each is a bit too sensitive on their own, but taken together it's usually a good indicator they are outliers.
Hi Erin, you are a lifesaver! I just couldn't wrap my head around mediation before I stumbled on your video. I have a question though, how can I calculate effect size? You be=riefly mentioned effect size in the video but didn't explain what analysis to run. Also, it would be great if you could write out the APA results as a full sentence not just the numeric results. Do you also have an APA table or figure for summarizing the mediation results?
I would report R2 for the overall effect size, but then the indirect effect would be the "effect size" for the mediation. There used to be effect sizes embedded into Process but they were removed after several flaws were discovered. You can view some example write ups/pictures by looking through the files on our OSF page.
Do you have a reference/cite for the 2 out of 3 outlier rule? I can't find mention of it
No real reference other than using multiple indicators is a good idea - you can use Tabachnick and Fidell and/or Cohen Cohen Aiken and West for the limitations/sensitivity of each of these, so it's better to combine several indicators (because one may be too liberal or conservative).
@@StatisticsofDOOM how do you personally cite this method when you use it in research articles? I followed along your videos and used this method for my thesis, but now can't find a good way to support it in my manuscript and my advisor is not happy haha
@@roxannefelig6101 If you want to cite process, use Hayes (afhayes.com/introduction-to-mediation-moderation-and-conditional-process-analysis.html), data screening is Tabachnick and Fidell (2012) and Cohen Cohen Aiken and West (2008? I think).
Thanks
Thank you!
Thank you so much for your wonderful video! You are legend!! I've got one question. If the indirect effect is significant, which means the mediation exists, and meanwhile, one covariate also has significant effect on the dependent variable (based on results from the process). Thus, what conclusion could we have?
Just exactly how you've described it. There's mediation, and a significant covariate that is related to the DV.
@@StatisticsofDOOM Thank you so much Erin!
@@StatisticsofDOOM Thank you. I was also wondering if I would like to use G*power to calculate the sample size for mediation analysis, what are the acceptable power and effect size? Thank you very much!! I used 0.15 as effect size and 0.9 as power but my tutor told me that the effect size was too small. But I noticed that you were using 0.063 as effect size cutoff.
@@tingyuesun5017 I mean effect size is what it is - it should be best on some educated guess on what you think the size of the R2 is for example. If you were using .15 that would be considered a large effect in R2. Often, people make power only .80 as well, so it just depends on what's normal for your area of research.
Great video!
So does the interpretation of the CV depend on the type of question you pose?
@@magillroad Yes and no? Like mathematically they don't really change, but if you are just using them as a control, then you might just mention it and move on. However, if you think they are important, you might spend more time on it. The idea of b is the same either way though - if continuous, then for every one unit increase in X, you to get B changes in Y, etc. If b is categorical you would need to use dummy coding interpretation (one group versus another).
Thank you do much for this video. I am using model 4 for a mediation analysis with covariates. My Y variable is dichotomous. Are the methods you are using to test for assumption also applicable when the Y variable is dichotomous? (i.e., a. Mahalanobis, Cooks, Leverage, looking at zpred and zresid).
No, you should not use linear regression for that analysis. You should use logistic regression (although the same steps can apply, but I am unsure process does this at the moment).
@@StatisticsofDOOM Thank you for your reply. I am wondering about checking for outliers. Can I still look at those Mahalanobis, Cooks, Leverage values? Are there other indicators for logistic regression for outliers?
@@mys1990 I think generally I don't do a lot for outliers in logistic regression, but you could check their influence on the solution (cooks/leverage) and mahalanobis pattern of scores. This website gives some thoughts too: stats.stackexchange.com/questions/26930/residuals-for-logistic-regression-and-cooks-distance
Is there a video that explains how to report the results according to APA (model 4)? helppp
If you check out the osf page, you can find example write ups in APA style.
If you check out the osf page, you can find example write ups in APA style.
Hi, thanks a lot for the explanation! It's great. If I compare two Mediations. One with covariates the other without leaving all IV, DV and the Mediator the same which values do I have to report for the effect of the Covariates on the DV? And how would I explain the difference in the indirect effect because of the covariates adjusting for error type II? Would be great if you can help me.
If there's a reason to use the CVs I would only report the analysis with the CVs...if you are trying to show how the CVs affected the analysis, then I might report both and stick everything in the table. If the CVs are related to the DV, then the changes between analyses would be because of error reduction in the DV by including those CVs.
Hi Dr. Erin. Thank you for the video. I am wondering with the degree of freedom for mahalanobis , do we need to consider covariance as the predictors? I am running moderated mediation (model 8) and have a case that has mahalanobis distance of over 20. If I just use my IVs as my predictors it is considered as outliers but if we consider also covariance as predictors then, the case is not outliers. What do you think is more robust or I should do?
You could easily go either way - I would personally run the model with and without the outlier and report both results.
First of all, thank you for those great videos, they are really really lifesavers. I would like to ask one question. I have conducted different mediation analyses with 111 cases [continuous Xs, continuous Ms, and continuous Ys (only one of them was measured with one item, 0-5 range). mediation was not sig, but what is most interesting is that Rsquare values were too low ( about .06 mostly). What could be the reason behind low Rsquares? I would appreciate your answer.
The variables aren’t related? You are not predicting Y very well if r2 is low.
@@StatisticsofDOOM Mostly they have good correlations. İ.e. for one model X and M .35**, X and Y .23*, only M and Y .01
Hello Dr Buchanan. Your videos are very helpful, to interpret my output, thank you!!! Although I have a question... my SPSS book explains that the ZPRED goes in X and ZRESID in Y, however you do it the other way around. Is there a specific reason? Thank you!!!
With the way we interpret the pictures, it doesn't matter if you flip them. Mainly, you are concerned with the pattern/spread around zero and the shape of the dots. So, if it's bad, it's bad both ways.
@@StatisticsofDOOM Thanks a lot for taking your time to answer!!
What to do with multicategorical covariates (some with more than 8 levels) and some with only two levels?
Check out my videos on dummy coding!
If my covariate is a categorical variable, would I need to dummy code it and then enter each dummy-coded one as a covariate?
Yes probably - I think the X variable can be categorical and it will automatically dummy code for you, but not the CVs.
@@StatisticsofDOOM Thank you for this great video. In the Linear Regression to check for outliers, you don't dummy code the categorical variables. The regression seems to work perfectly fine. Is the dummy coding solely necessary for interpretational purposes then and not for the feasability statistical procedures?
@@shannawielinga8207 haven’t watched this one in a while but generally you don’t have to check the categorical variables because they naturally cannot be normal - now they can be outliers but SPSS does not make this easy. I don’t believe there are any in this video? Either way you can dummy code them and then screen them just like in the video.
Hi! In a statistical diagram the relation between the covariate and the dependent variable is sometimes called the d-path. My supervisor has asked me to report the results in the diagram (instead of textual). You report about the d-path multiple times, which results am I supposed to report? It can be found both in the total effect model and the model where you also get the c' from (I think you called it the full model or direct effect model). Which one would be the "d-path" I'm looking for? Also if I would want to report the results in a table rather than a diagram, is this possible? It's hard to find articles online where they do this.....
Certainly you can put things into a table. If you are listing the covariates as a d path, you probably have to have a d path (just the x to y model) and a d' path (the x and m to y model). I might recommend a table of all the coefficients for each model to make things a bit easier to understand (as that might get confusing on a picture).
@@StatisticsofDOOM Thank you so much for your quick reply! I'll share a link to a document I found online. On page 24 and 25 the writer reported it with tables for each model, would you say that's the right way of reporting? : arno.uvt.nl/show.cgi?fid=141735
@@dunja2610 Yep! Those folks at Tilberg know what they are doing!
Hi Erin, thanks so much for your informative videos. I'm looking at longitudinal data, and want to see whether change in one variable mediates the change from X --> Y. Can you use change/delta scores as the mediator? When I run it in SPSS, the program seems to have trouble with a negative value as the mediator (i.e., it's negative because i'm looking at change/ improvement over the course of treatment). Any advice?
I don't see why not ... negatives shouldn't be a big deal, as other variables can obviously be negative. What's the problem that you are seeing?
Dr. Erin M. Buchanan, thank you for this video! I would like to ask that can we use a centered variable as a covariate in Mediation analysis model 4. I calculated SES score as a composite in my data set and since they are in different range I used ZSES as a covariate. Thank you
Sure - you would just want to interpret the variable as a standardized score instead of the traditional b interpretation.
Hi, These videos are useful and accessible, so thanks for creating them. Do you have any advice for mediation analysis for longitudinal repeated measures data? I have 3 data sets with the same mediator and a potential categorical covariate.
That sounds like a multilevel model if you want to do them all together - which is not super easy in SPSS.
Thank you so much!
I have one quick question. If I have multiple independent predictors with covariates, which PROCESS Macro model I should use??
Unclear - depends on what you want to do with those predictors. I would decide how and what you expect to mediate and draw out your "triangle" diagrams to help you pick a model. There's a templates file (or it's in the book) to help visualize.
Hi! Thanks for your explanation! I have one quick question: what do I have to do if I expect that my continuous covariate has only an effect on Y (and not on M)?
The CV only goes into the model for Y I believe ... can't totally remember but if not, you might have to run each step separately, so it doesn't predict M.
What does covariance indicate here?
Do you meant covariate variable?
@@StatisticsofDOOM Yes
@@aesthetics1110 Covariates are other variables you include the model that adjust the DV for their known variance so you can examine the impact the other variables have without them.
Hey Erin, will this work in SPSS if I want to use glm not lm for the regressions? If one of my variables is skewed (eg. reaction time data) - can it do glm, or do I have to transform (would rather not do that)? Thanks!
Are you meaning that you want to use different linking functions? I'm not 100% following the question.
@@StatisticsofDOOM Lol I am not sure what I was asking. In R, I run my regressions as glm as often have skewed data. In SPSS, Process assumes linear regression, so wondered if it would work if my variables do not meet assumptions. Also had a binary mediator in one instance which Process freaks out about.
@@brittjane4486 OH! I get it now. If you only wanting to use process, I would suggest transformation, as I don't think it allows for anything other than least squares assumptions. OR just do the whole thing in R - I have been writing a package for it, but got sidetracked with other projects. I do think I have this type of thing written though! It doesn't do glm, but that might be something I can add?
Dear Dr. Erin, thank you so much for this video, I got one questions to confirm: I got significant indirect effect and nonsignificant direct effect, it indicates full mediation, but the total effect is not significant, can I still make this conclusion, i.e., whether a significant total effect is a pre-requirement to test mediation effect, i cannot find a conclusive arguement on this issue, thanks a lot!
I would mostly only look at if the indirect effect is "significant" or the bootstrapped confidence interval does not include zero.
@@StatisticsofDOOM Dear Dr. Erin, thanks a lot for your reply, I got it!
Which item should we put as X and covariate? Will it yield different result?
You should have a theoretical reason for the variable you decide is X, as it's part of the conceptual mediation portion of the analysis. Covariates should be variables that wish to control the variance for.
Hi Dr. Buchanan - Thank you for the video! If my total effects model summary and full model with mediator and covariates account for significant variance but my paths (total effect and indirect effect) contain zero, what can I conclude? That the variables in my model account for significant variance but do not mediate the relationship between X and Y?
That is what it sounds like to me - that the R2 is greater than zero but mediation (and the X predictor) likely did not occur.
Hi Erin! Thank you so much for this video, I still have a doubt though. If my IV has two conditions should I insert the computed variable in the process analysis or should I perform model 4 analysis twice, one per condition? Thank you again
The answer depends on what that variable is for - if it's X or M, it should be included in the analysis, not separated by group.
Any suggestions for good journal articles that used model 4 for mediation analysis with covariates? I'm not sure how to report the results APA style.
A simple example write up is here: osf.io/vf5gh/
Statistics of DOOM thank you so much, really appreciate it!
Great video; do you have a citation for your how you define a medium effect size in g*power?
I believe those numbers are based on Cohen's effect size papers, but I would use what is normal in your field (as he argues, that average sizes should be based on the field)
Thank you so much for the video, really helps for my work! May I ask if I find the covariant is significant, is there anything else I need to do? Or just need to write it is significant and do nothing.
Just write it up!
@@StatisticsofDOOM thank you !
Thank you so much Erin! Just another super useful video on testing mediation. I just have a question: so I find X predicting M (a path significant), M predicting Y (b path significant), but indirect effect not significant (direct effect also not significant). So apparently there is no mediation effect. But can I still report the significant a and b path? And what might be possible explanations for the non significant mediation?
Sure, I mean you would just report what you found. Can't help with the mediation non-sig question, could be that it's truly not there (doesn't exist), you didn't have enough power to find the effect (sample too small to see the meditation), or some other theoretical reason.
HI, i want to know if control variables are related to only DV in model 4, then how to do that in process 3 and above version as both do not give us the option to select "if covariate is linked to only dv or to both mediator and dv" just like the option that "process v 2" used to give us in model 4.
I'm not totally following your question - if you want to use a CV, it only shows the relationship to the DV.
@@StatisticsofDOOM e.g. i want to test the relationship of role conflict with turnover intention through mediating role of anger. Previous research has shown that work overload is related to turnover intention. Now i want to control work overload to see the pure relationship between role conflict and turnover intention through anger. Now in "process v 2" model 4 gives the option "covariates in models of (a) both M and Y, (b) M only (c) Y only". whereas in process 3 and 3.2 model 4 doesn't give us this option and output contain control variables in first output (a path) as well, which otherwise do not appear if we run model 4 in process 2 by selecting option "covariates in model(s) of (c) Y only". I like to know that how to have this "covariate in model of Y only"? as this command is not available in process 3 and latest version. Regards
Ok, I think I get what you are saying - I think in process 3 you just have to pick the model that includes the variables in the right places...so for it to covary with Y AND M, you would need to find one that includes it as a variable in the template pictures. Otherwise, it will only predict Y.
Thank you!! This is extremely helpful. Looking at the very end of the output - the section titled "Total, direct and indirect effects of x on y" : How does one interpret the raw magnitude of the two numbers under "total effect" and "direct effect" respectively? In my data, I am also seeing a change, from .0120 (total effect) to -.4246 (direct effect). Unlike in the data you used here, I do not have the value of 0 between my two confidence, indicating that mediation is happening. BUT I do not know what to make of the numbers .0120 and -.4246? Struggling to understand what they mean.
The total effect is just X on Y (c path). The direct effect is X + M on Y (c' prime). These are coefficients, so you interpret them as for everyone one unit increase in X, you get c or c' units increase/decrease in Y.
hey, Can I get BETA values instead of b-values in Process, so the results are comparable?
No - Hayes has suggested these are not appropriate for analyses in his books. If you still want to calculate them for comparison purposes, you could z-score your variables first, then run PROCESS on it again.
Thank you so so much for your videos, they are extremely helpful! I appreciate you get a lot of comments and are super busy but I have a challenging question that is driving me a bit crazy. I have a multicategorical IV, and want to run a mediation analysis with 3 mediators, a moderator, and about 4 covariates. I'm wondering what PROCESS model to use and how to do it? Thank you so so much and I really hope you can give me a bit of advice ♥️
Hi Erin, please if you can just give me a direction I would be really grateful. I'm on my final PhD year and am really really stuck on this, and can't find anyone familiar with mediation and process
I’d recommend checking out the possibilities in the diagrams that are available with the book - hard to know since what you are describe is pretty complex and where the moderator goes. You’ll have to see if one of the model numbers matches the conceptual diagram.
This is going to save my master's thesis, thank you so much!!
One quick question: I work with a model just like you, but I use two different DVs to capture the same construct in the same sample. So there are actually two mediation models that are conceptually related. Do I have to worry about alpha inflation and correct for it? Thanks in advance!
I think you'd be ok with only 2 analyses, but if you had like 5-10 mediations, then I would consider controlling for Type 1 error yes.
@@StatisticsofDOOM Percent! Thank you so much for your answer
Dear Arin, thank you for the videos. I want to conduct this commend
process vars=y1 iv1 iv2 iv3 med1/y=y1/x=iv1/m=med1/boot=10000/model=4
/seed=5235.
But I cannot, it does give me neither a result nor an error. I want to do that to have bootstrap confidence intervals based on the same set of the simulated sample from the data at each time I run my analyses.
Not sure honestly, I haven’t used process code functions before.
Hi! I found a significant indirect effect of my mediator. However, when I perform the sobel test with the t-value of the A and B path, the aroian p-value is 0.089. Does that mean that my mediator isn't a mediator after all? Or can I choose not to report the Sobel test if it isn't always necessary?
I'm not sure what you mean by "significant indirect effect" - do you mean that the indirect confidence interval does not include zero? How did you run the sobel test?
@@StatisticsofDOOM yes the confidence interval did not include zero, I ran the test through the website you shared, and used the t-value of the a and b path, all p-values were > 0.05 (but < 0.09)
@@dunja2610 Got it! Thanks for the clarification. I would say more people like the bootstrapped CI to determine if the mediation "occurred" and like the Sobel test less. It does sound like the CI is probably close zero though with those results.
@@StatisticsofDOOM very true, the CI is indeed close to zero. Can I still conclude there's a mediating effect?
@@dunja2610 I would cautiously go that route - you might mention that it's support for a small mediation effect since the CI is .
Hi! Your videos are very helpful, I love them. But what I need the most is a tutorial video about Model 5. It is impossible to find one on RUclips like nobody uses Model 5? It would be great if you could share a video about that or a helpful link that you know. Thanks a lot in advance!
I have a few that cover model 7, which is similar but the moderation is in a different place. Same concept though: ruclips.net/video/DOQkBcT-Zzk/видео.html
Many thanks for your videos. They are really very helpful especially the interpretation of outputs. I have a quick question.What happens if total and direct effect are non-significant but direct effect is? How do we interpret this?
You simply interpret it as a slope of X predicting Y with M in the equation.
Thank you so much for this helpful video! I had a question about my results. I first ran results without covariates, and the total effect was significant, as was the indirect effect (direct was non-significant). Reviewers suggested I add in covariates, so I ran it again with covariates, and now the total effect is not significant, but the indirect effect continues to be significant (again, direct effect was non-significant). How do I interpret this? Are the covariates explaining more than my X variable? Thank you!
I also want to add that the indirect effect size has decreased with the addition of covariates - I am not sure if this means that although the mediation still exists, it is likely that the mediator is also interacting with the covariates, thereby reducing the effect size of the relationship between IV and mediator?
@@anushaa You'll need to check the sobel test or the bootstrapping results to know what your "mediation" effect is, not just the total/direct effect. The addition of the CVs just controls for their variance - so the changing values indicates that they predicted some of the same variance in the DV that your IV did.
@@StatisticsofDOOM thank you!!
Thank you so so much for these videos! It helps me so much with my thesis!
I have a little question regarding the p-values: do I need to split the p-value if I have an one-sided hypothesis? and if so, is there a shorter way in the calculation process? e.g. by putting the alpha level on 10% rather than 5%?
If you wanted to make a directional prediction, you could divide p value by 2. I'm not sure that makes a lot of sense in a mediation model, as mediation includes a two sided bootstrap confidence interval.
@@StatisticsofDOOM thank you very much for your answer! does it makes sense to double the alpha value e. g. 10 instead of 5%?
@@seymatuerk Not normally - you should keep the idea of alpha (type 1 error control rate) and p value found in the study as separate things.
Hi, I am trying to learn "Multilevel multinomial mixed-effect model". Could you upload a tutorial for this? Thansk in advance
Somewhere I have a MLM for logistic regression - only in R though! I'll see if I can get it on the list.
@@StatisticsofDOOM Thank you so so much. I have read about this analysis in a research paper, however, the authors performed it through SAS. I would be really grateful to you if you can teach me the same analysis through SPSS.
@@StatisticsofDOOM Hi. I am extremely sorry for bothering you again and again. If you know MLM for logistic regression only in R, could you upload tutorial for that? Please? Also, along with that tutorial, could you also guide me how to run "Commonality Analysis"? If you need any research studies I am replicating, please let me know, I will share with you. Again sorry for the inconvenience caused. And waiting for a tutorial video soon. Bundle of thanks.
@@khadeejamunawar2529 I'm not sure if I know what a communality analysis is? I will work on a log MLM sometime soon, as I have a bunch of deadlines over the coming weeks.
@@StatisticsofDOOM This is so kind of you. I am anxiously waiting for your SPSS tutorial for MLM when outcomes are categorical (3 or more categories). And good luck for your deadlines. Thank you
Hello Dr. Erin, thank you for the insightful tutorial. I had a quick question on how to report the tests' results. I have ran a dozen of tests using Model 4 and all models were insignificant. In other words, the models' summary shows that the model was insignificant. Can I directly conclude that there is no mediation effect, without having to go through all details in my thesis?
Also do you think that this insignificance could be the result of the data sample I have? I have only 101 respondents for my thesis.
Thank you for your time!
Unfortunately, that's the problem with null results - it could be one of many reasons. Power (too small of a sample to detect this effect) could be the issue, could also be that there's no mediation, could be that there's no mediation in this particular sample or with these particular variables, etc. I usually just say that the hypothesis of mediation was not supported (which is not the same thing as no mediation at all, just that no mediation was seen in this data with this analysis).
Hi Dr. Erin M. Buchanan, Does paralel mediation also use model 4?
For the present example does it mean after controlling covariates mediation did not occur?
Parallel/double/serial mediation is model 6 I believe. And that's correct - because the confidence interval includes zero.
@@StatisticsofDOOM Thank you, I got it
if covariates are added to the mediation model, are they excluded from the mediation?
or are we able to say that they predict the mediator / outcome variable and do influence the mediation (if the indirect effect of the mediation is also significant)
argh, I need multiple IV's