This video is incredibly helpful. Other videos on this topic do not interpret the results as well, and none explain how to report the results. Please continue to include this in your videos! Thank you so much, James!
Dear Professor, I am grateful to you for your efforts here. I learned AMOS by watching and practicing your videos and used it in analysis section of my PhD thesis. God bless you and your family.
+Irfan Saleem, for some reason RUclips wouldn't let me reply directly, so I'll reply here. Moderated mediation is when you compare mediated effects across multiple groups. Mediated moderation is when you mediate an interaction effect. I've never used PROCESS, but I hear it is cool.
Dear Dr James. It was a great video. But I do have 1 question. How can we know exactly the indirect effect if there are 2 meditor tested at the same time to dependent variable from two different independent variable. For example. The mediator is trust. The dependent variable is intention. The independent variables are attitude and social pressure. So i want to taste the mediating effect of trust to intention, from attitude and social pressure. thanks in advance
Dear James, Your videos are very helpful! Thank You! I have questions for You: 1. Do I need do test correlation or regressions between mediators on one side and IV an DV on the other side before final model? 2. If I have more mediators should I calculate effect separately (delete paths with other mediators) or I can calculate mediation with many mediators in the same time?
A really helpful video..can you please help, is to correct to perform hypothesis testing by scale free method and check mediation effect through bootstraping for a data tht does not fit normal distribution
Dear james should we check the mediating passes standardized regression weights and their P value to control that if they are significant or not in the first step before mediation test. because I think if there is an unsignificant p value basicly there will be no mediation effect. best regards
Hey James, is it mandatory to test measurement model before structural model if we are interested to see mediation/moderation. I mean if the objective of the research is to assess mediating/moderating effect not the the reliability and validity issues of the scales.
If you have latent factors measured by multiple indicators, then you should do measurement validation so that you know that what you are measuring what you think you're measuring.
Great video and very understandable. I have a question in mind. In this video you don't apply Baron and Kenny approach. Is your way in the video supported by any scholars in the literature? When we use this method in our articles, which sources should we cite in method part? Thanks.
here is the most recent mediation video: ruclips.net/video/41XgTZc66ko/видео.html The approach here is based on Preacher and/or Hayes, or nearly any of the literature here: statwiki.kolobkreations.com/index.php?title=References#Mediation
Hye Mr. James. Your video is very good as it is precise and easy to understand. But, i want to ask. What can be explained and discussed with that evidence in your table. I mean that the number of estimate. What does it mean?
+Saudara Anda The standardized regression weight is the beta or slope (standardized slope). It tells us the amplitude of the effect between two variables. The p-value then tells us whether that estimate is significant.
Dr James, you are very helpful in the area of research. my question is this; can I use this bootstrapping method in the same steps when testing moderation effect? ( my model contain mediator and moderator). thank you
You can bootstrap for a moderator, although you don't have to. Here is a section of the statwiki that provides videos to interaction moderation: statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling#Interaction
One more question ,I have is that, if the variance of one of the constituent of my second order construct is low, then can I remove the second order construct and can I draw model directly to the constituent. In my model...Training-->ethics-->workplace behavior(second order construct with satisfaction and attrition). So if theory allows can I remove workplace behavior and draw a model Training-->ethics-->attrition.
I think I understand. I think you are asking whether you can model your 2nd order factor instead as just the first order dimensions, and then just have predictive arrows point at those dimensions rather than at the 2nd order factor. The answer is yes. Here is a paper where we do that: www.researchgate.net/profile/Paul_Lowry/publication/264623594_2013-Nov-JAIS_HMSAM_PUBLISHED_APPENDICES/data/53e9b4960cf2fb1b9b671a95/2013-Nov-JAIS-HMSAM-PUBLISHED-APPENDICES.pdf
Thank you very much for the video, very useful. However, I have two questions: how can I have access to the 95%CI of the indirect effects, please? And is there a textual way of reporting the results without the table? I really appreciate your help.
If you use an estimand and perform a bootstrap (as shown in this video: ruclips.net/video/ICnh3s2FG14/видео.html) then you can view the upper and lower CI. As for reporting results, you would just say that the indirect effect was observed, as indicated by the confidence interval not spanning zero (LB: [put your lower bound CI here], UB: [put your upper bound CI here]).
Hello James. I have a question. I have a model with one IV and one Mediator and four DVs. Do I do the Structural regression first then do the confirmatory or do the opposite. If you this I should start with structural, should I remove the effects that are insignificant and then do the confirmatory. Thank you
Dear James Gaskin, thanks for the video, it's very helpful! I have a question: when AMOS estimates the regression weights, including direct and indirect effects, of my model before bootstrapping (2000 resamples, CI at 95% level), they are significant, but after bootstrapping the indirect and direct effects are not significant anymore, so that I can't carry out the mediation analysis anymore. Why is that so and what can I do to make the mediator analysis possible? Thanks in advance!
+Uirassu Borges Junior Bootstrapping will often inflate the p-values a bit. You can try increasing the number of resamples. I just did a test with 200, 2000, and 20000. The p-values for generally decreased as I did this.
Hello James, the video is very useful. I have 2 queries and hope that you can help. First, can I use the same method for mediation with 2 mediators? Secondly, using this method, how do I get the 95% confidence interval for the indirect effect? Thank you! =)
+Happy Snappy When bootstrapping, AMOS will produce the 95% confidence intervals (see the upper and lower tables). If you have two mediators, if they are both connected to the IV, then you will need to do a sobel test instead of bootstrapping. If they are serial, and not both connected to the IV, then you can bootstrap the same as I've shown here.
Dr James, you are very helpful with videos. I am confused about some point. The regression weight is give to use a direct effects but when we use the boostrap methods this values are changing. I can not decide which p values is true. The problem; this regression weight and boostrap direct effects significant is not the same statistical result. Regression weight is significant but the boostrap direct effects is not significant. What is the problem with this case? Thanks in advance!
Hello James, Thank you for the video. I was wondering if there is mediated mediation? If yes how is it different from sequential mediation? And how can we do mediated mediation in Smartpls? Thanks
I think you mean serial mediation. In SmartPLS, I'm not sure if there is a way. Here is my latest video on SmartPLS specific mediation: ruclips.net/video/z1h6qMcMhWk/видео.html. And here is one for AMOS that does cover serial mediation: ruclips.net/video/KgtgQO9zQK0/видео.html
Dear Prof, I have a question regarding number of bootstrap samples. Is it the same one with our research samples? So, it means, if we have 600 samples, or 1500 samples, do we need to insert the same number?
No, it does not need to be the same. It is just the amount of resampling that is occurring during the bootstrap procedure. Typical bootstrap samples are 2000, 5000, and 10000. However, 10000 gets very arduous due to computational requirements if you also have a complex model or a large sample size. I usually just use 2000. If you do some tests, you'll notice that the variance in results tapers off quickly as you get past 2000.
Hi james, thank you for your video. I learned a lot by watching your video. But I have a question. What if my model shows that the direct effect is sig. but the indirect effect is not significant? Is there still a mediation effect if the output like that? (There is one IV, one M, and one DV. just for information path c, IV to DV is sig. path a, IV to M not sig. and path b, M to DV is sig) I hope you could help me james, thank you
First of all, thank you very much for your great video. Could you please help me with my problem: in my model i have 1 independent variable, 6 mediators and 1 dependent variable. However, I could only perform bootstrap in Amos without the standardized estimated values because there is a negative correlation between 2 mediators. So I only get the results for the total, direct and indirects effects without the lower or upper bound and p values. How can I interpret the results in this case? Is there any other resolution for me to test the mediators in Amos?
If you standardized the variables in SPSS before using them in AMOS, then the unstandardized estimates will be effectively similar to the standardized estimates.
Dear Professor, I really appreciate your efforts here but I have a question regarding the bootstrapping method. I would like to ask how to decide the optimal number of bootstrap samples as I have watched several youtube videos regarding the bootstrapping in AMOS. Almost all of them used 2000 times and therefore, I have used 2000 times in my paper as well. But I cannot find a good explanation on why I choose 2000 instead of 1000 or else. I would be grateful if you can answer this. Thank you!
Generally speaking, more is better. However, for decades, the number of samples was limited by computing power. Now that computers are so powerful, it is easy to do a few thousand. Hair et al and some other scholars, such as Ringle, suggest a minimum of 5000 when the sample size and model complexity allow it. I ran an analysis the other day on a sample size of 53000 though, and so I could only do a bootstrap of 500 (or else the computation would never finish...).
Dear professor It was Great, Thanks I ran with your method but an error message came up: D uring the analysis of a bootsrap sample ,an attempt was made to compute the correlation between two variable, one of whose stimated variance faild to be positive. Please help me solve the problem۱
Check the notes for the model. It should identify which error had a negative error variance. You can double-click on that error bubble and go to the Parameters tab and then set the variance to equal a small positive number, like 0.05.
Prof, can you help me? in my result, for direct effect I get .393 (sig) and for indirect effect I get -.036 (sig). Is it normal if our indirect effect show negative?
excuse me sir, I would like to ask what if the indirect effect is significant at p-value < 0,01 (p-value=0.004) and direct effect is also significant but only at p-value < 0,1 (p-value = 0.051). Do we call this effect partial or full mediation? Thank you in advance
In this case, it probably depends on how much the direct effect changed after adding the mediator. If the direct effect changed a lot, then I would call this full mediation. If not, then it is probably partial.
Hi Doctor. in the case where all paths in the structural model are the significant but direct effect is greater than the indirect effect. do we still have a potential mediation? thank you
Thanks. Can you please help to understand what about the the sign of the direct and the indirect effect. Please how can I interpret an unexpected negative direct effect ?
+SAID ETTIS Sometimes when the direction of the direct effect is unexpected in a mediation model, it is because the IV has a dual effect on the DV. The dual effect is split by the mediator so that the positive effect travels one way, and the negative effect travels the other way. If the mediator has a positive effect on the DV, then usually the positive effect from the IV will travel through the mediator, leaving the direct effect of IV on DV negative.
Hello Prof James Gaskin, My model is following Path std. dev t-value p-value X->Y= 0.338 0.077 4.400 0.000 X->M 0.581 0.053 10.981 0.000 M->Y 0.347 0.061 5.713 0.000 i have calculated indirect path coefficient manually path = 0.581*0.347= 0.201 (Correct according to SmartPLS) std.dev = 0.053*0.061= 0.0032 (Incorrect) t-value= 10.981*5.713= ---- (Incorrect) Please help me in this regard how to calculate std.dev, t-values and p-values for indirect effect
This plugin will do some of this for you: ruclips.net/video/-xzkjqKhWHE/видео.html. The t-value is just the unstandardized estimate divided by the standard error.
dhruv dutta you have to have a good theory to support your modeling choices. If you do it just based on the statistics, it's hard to tell which model is best. It really should be based on your theory.
Hi professor, How if my mediator did not shown direct effect into the dependent variables and Independent variable has shown direct effect into the dependent variable. Does it still considered as Partial Mediation?
Dear Prof, If I have 2 IV with 7 dimensions and 2 mediators and one moderator and 1 DV. So, How to check the mediation and moderation effect on AMOS? Looking for your kind help. Thanks. Suliman
Hi Suliman, I have several videos on mediation and moderation in AMOS. If you click on my avatar/profile, you'll be taken to my channel where you can search for these videos using the search icon (magnifying glass next to the About tab). Here is the channel url: ruclips.net/user/Gaskination
Hi James, my Amos does not bring out the Estimate/Bootstrap results window for me to check the two tail significance. Is there anything I am doing wrong? My measurement model is a second-order model. This happens when I add control variables to the model.
If you are running a bootstrap (which must be selected on the bootstrap tab of the analysis properties), and if you have checked the box for indirect effects (output tab), and it still isn't working, then I'm not sure. It should definitely show up if you've checked those boxes.
@@Gaskination strangely it disappears when I add in controls to the model and there are arrows from the controls to both the mediator and the DV. When the arrows only go to the DV suddenly bootstrap results boom! show up. I also face this problem: when I bootstrap 2000 samples in the mediation analysis I receive an error msg that "During the analysis of a bootstrap sample, an attempt was made to compute a standardized regression weight between two variables, one of whose estimate failed to be positive. The attempt was made because ' Standardized Estimates' in the 'analysis properties' window was checked or because the Standardized method was used." when I bootstrap 30 samples the analysis runs just fine. Have you encountered this before? How may I overcome this if I want over 500 bootstrap samples?
@@DK-em6oz That's odd about the controls... It may have something to do with missing data in the control variables. I have never seen the second error either... I'm impressed you've stumped me twice. I've been doing this for a decade and have never seen those... The error is weird though, because it is fine if a standardized estimate is not positive. I hope you figure it out! Best of luck to you!
Hi James, why did you not test the model without the mediator first? What would it mean for a model when 3 out of 4 direct relationships from IV to DV in a model without the mediator are not significant. I think it means no mediation, but does this mean your research should go in the bin? When the moderator is added, only 1 out of 4 is not sig.
+Sarah Preis More recent literature suggests that we do not need to test the model without the mediator. It suggests that if the indirect effect is significant, then we have mediation, regardless of whether the direct effect was significant before adding the mediator. So, mediation is determined purely by the bootstrapped indirect effect.
@@Gaskination Sir, in your stats wiki website, you have given the lesson pdf under mediation in which you have specified that we have to report direct effect without mediator and then direct effect with mediator along with the indirect effect. So, sir, my doubt is which approach should I use?
@@dr.anupreetkaurmokha5792 The best approach is full disclosure and transparency with more information. So, I would recommend reporting all effects (direct and indirect). The indirect effect is what we use to determine whether there is mediation present. The other effects provide additional context. It is no longer required to test with and without the mediator (Barron and Kenny approach). Here are some helpful references: statwiki.gaskination.com/index.php?title=References#Mediation
+Cherry Zhang I might in the future, but for now, just know that you can simply do a bootstrap and look at the indirect effects. If your model looks like this: A->B->C->D, you can look at the indirect effect matrix which will show you the indirect effects from A->C, A->D, and B->D.
Hello, Dear James, and all interested in mediation analysis. :) I have a question again. Its: when I look at regression model (direct relationship) from A, B, C to D it was significant. But when I add into the model two mediators X and Y the relationship between A,B, C to D was unsignificant (it looks like full mediation from first sight BUT, A, B, C to X and X to D is also UNSIGNIFICANT (Y was significant full mediator). What does it mean and how should I interpret results?
+Tija Ragelienė This can happen if the mediator is a permeating variable that floods out other effects in the model. Some known variables like this are self-efficacy, confidence, optimism, etc.
+James Gaskin Thank You James. My mediator x is empathy, which was supposed to mediate the relationship between family relationships characteristics and adolescent identity. How I should interpret results in such a case? does it mean that effect of permeating variable is the bigger than other mediators in some way? Maybe do You any article or etc. where I could read more about it? Cause mediator Y which was full mediator was much more strongly supported by theoretical background...
+Tija Ragelienė Sounds like the other mediator is just very strong, or that empathy is not a good mediator. You can try it without the other mediator to see if empathy works.
All right. I've just tried and without the other mediator empathy works. So what should I do with my results? Should I interpret as two separate models with different mediators? But then I am confused with theory because empathy and other mediator could be both predicted by family charakteristics and from my own theoretical assumptions based on theory predict identity diffusion. Or shoud I just write in discussion section that empathy effect could be not such as expected because the other mediator was much stronger?
+Tija Ragelienė I would include both but then say that when tested separately, empathy works. This implies that the other mediator is the stronger one and the one that should receive the attention in the literature and in practice.
I would recommend a minimum of 2000. The trouble with such high sample size is that if you do very high bootstrap as well, it will take a very long time to run...
Thanks, professor. Is the high sample size not good? Another issue is the scatter plot in regression has some values above 3 on the x-axis and y-axis. Is that mean the data is unsuitable for SEM?
@@nargisali7298 high sample size is generally good. Its only downsides are a deflation of p-values (i.e., everything, even small effects, are "significant") and the time it takes to run a bootstrap. Otherwise, it is very good to have a high sample size. As for the scatterplot, I'm not sure what the problem is. The axis size on a scatterplot differs from dataset to dataset, based on the size of the scale for the measures. So, I'm not sure if values above 3 is a bad thing.
Thanks, I am using a five point likert scale. Scales with mostly disagree and agree are positively related with the DV, significant p-value and scales with disagree or strongly disagree are negatively related with DV, sig. p-value. However, two of the scales are insignificant in the model which I think I will have to exclude from the model. Is it okay to keep a few negatively and few positively related variables in a model? Sorry, I am asking too many questions and taking your time :(
@@nargisali7298 As long as the outcomes make sense, then it doesn't matter which direction the effect is. For example, if the DV is job satisfaction, then predictors like management fairness, participation, and autonomy will have positive effects, but predictors like burnout, ambiguity, and bullying will have negative effects. I would recommend also keeping non-significant predictors in the model if you theorized they should have an effect.
Nice video. Sir, I have 2 Question, Please 1. While testing Mediator, we should interpret in research only the final model in which direct and indirect impact were tested i.e. partial or full mediation or to interpret also the initial model tested separate for the purpose that there is mediation or not such as IVs to DV (separate checked), IVs to Mediator and Mediator to DV.? 2. In final model (partial or full mediation model), we should interpret only direct and indirect impact OF standarized indirect impact or unstandarized Impact value, and we should also interpret Total impact?. will be thanked.
Muhammad Anwar there are different schools of thought on this. The most recent literature recommends you just assess the final model. No need to assess a separate IV to DV direct effects model. As for the total effects, some journals do you like you to report the total effects. They are most concerned about whether that value is significant.
@@Gaskination Hi there, tnx for all your help. I have some related question on this, is it possible when our total effects not sig we interpret based on direct and indirect impacts in our model ? In our model we have 7IV, 1mediation and 1DV. After bootstrapping(5000sample) in both standardized total effects and standardized direct effects only one IV to DV had acceptable p value(different path), while six standardized indirect effects were acceptable. If possible, please cite the source for reference in the article. Tnx again.
@@psycho2758 I think most papers argue that total effects are important. However, I think Zhao et al 2010 allows for indirect effects only. (please make sure to double-check it) Zhao, X., Lynch, J. G., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and truths about mediation analysis. Journal of consumer research, 37(2), 197-206.
+sugashwar prashanth Same approach. Just do a bootstrap and then check the standardized indirect effect and two-tailed p-value. You can check it for the IV to M2 and for the IV to DV, and for the M1 to DV.
This video is incredibly helpful. Other videos on this topic do not interpret the results as well, and none explain how to report the results. Please continue to include this in your videos! Thank you so much, James!
Dear Professor,
I am grateful to you for your efforts here. I learned AMOS by watching and practicing your videos and used it in analysis section of my PhD thesis. God bless you and your family.
Murat unanoğlu My sentiments exactly! James's videos have helped me tremendously as I am writing my thesis. 😊
Dear James Gaskin
I can't thank you enough for this video. It's very helpful for my PhD thesis.
This is the best video on youtube so far that elaborate the mediating effect and bootstrapping. Thank you so much
This video just saved my dissertation. Thank you!
Keep up the good work! Thank you!
Many thanks again for this informative explanation which is easy to understand...
Sir thaaank u so kuch for wonderful life saving videos
+Irfan Saleem, for some reason RUclips wouldn't let me reply directly, so I'll reply here. Moderated mediation is when you compare mediated effects across multiple groups. Mediated moderation is when you mediate an interaction effect. I've never used PROCESS, but I hear it is cool.
+James Gaskin hI just need to ask if u have any SEM Model and data on Lean Engineeing please reply me
+Zee Asim I don't. I only have psychometric data.
Thanks For reply where i can find such data
+Zee Asim I really have no idea. I'm not in that field. Best of luck to you.
thanks for your support and advice do u also have some work Related to DEMATEL and ANP interconnection
Thank you so much. I really like your videos. It is so helpful
Thanks. Can you help to understand & differentiate moderated mediation vs mediated moderation along practical usage of AMOS and or Process Macro
Dear Dr James. It was a great video. But I do have 1 question. How can we know exactly the indirect effect if there are 2 meditor tested at the same time to dependent variable from two different independent variable. For example. The mediator is trust. The dependent variable is intention. The independent variables are attitude and social pressure. So i want to taste the mediating effect of trust to intention, from attitude and social pressure. thanks in advance
You will want to look at the specific indirect effect: ruclips.net/video/41XgTZc66ko/видео.html
Dear James, Your videos are very helpful! Thank You! I have questions for You:
1. Do I need do test correlation or regressions between mediators on one side and IV an DV on the other side before final model?
2. If I have more mediators should I calculate effect separately (delete paths with other mediators) or I can calculate mediation with many mediators in the same time?
1. Not that I'm aware of.
2. You can do multiple mediators simultaneously. Here is a video: ruclips.net/video/ICnh3s2FG14/видео.html
Thank you prof !
A really helpful video..can you please help, is to correct to perform hypothesis testing by scale free method and check mediation effect through bootstraping for a data tht does not fit normal distribution
I'm not familiar with "scale free" methods. But yes, bootstrapping does mitigate some of the effects of non-normal distributions.
Dear james
should we check the mediating passes standardized regression weights and their P value to control that if they are significant or not in the first step before mediation test. because I think if there is an unsignificant p value basicly there will be no mediation effect.
best regards
The direct effect can be non-significant and still there may be mediation. The two indirect paths (X-->M and M-->Y) should be significant though.
Hey James, is it mandatory to test measurement model before structural model if we are interested to see mediation/moderation. I mean if the objective of the research is to assess mediating/moderating effect not the the reliability and validity issues of the scales.
If you have latent factors measured by multiple indicators, then you should do measurement validation so that you know that what you are measuring what you think you're measuring.
Great video and very understandable. I have a question in mind. In this video you don't apply Baron and Kenny approach. Is your way in the video supported by any scholars in the literature? When we use this method in our articles, which sources should we cite in method part? Thanks.
here is the most recent mediation video: ruclips.net/video/41XgTZc66ko/видео.html The approach here is based on Preacher and/or Hayes, or nearly any of the literature here: statwiki.kolobkreations.com/index.php?title=References#Mediation
That was very useful. Thanks. Greetings from Turkey
Hye Mr. James.
Your video is very good as it is precise and easy to understand. But, i want to ask.
What can be explained and discussed with that evidence in your table. I mean that the number of estimate. What does it mean?
+Saudara Anda The standardized regression weight is the beta or slope (standardized slope). It tells us the amplitude of the effect between two variables. The p-value then tells us whether that estimate is significant.
Thank you very much.
Dr James, you are very helpful in the area of research. my question is this; can I use this bootstrapping method in the same steps when testing moderation effect? ( my model contain mediator and moderator). thank you
You can bootstrap for a moderator, although you don't have to. Here is a section of the statwiki that provides videos to interaction moderation: statwiki.kolobkreations.com/index.php?title=Structural_Equation_Modeling#Interaction
thank you Sir
One more question ,I have is that, if the variance of one of the constituent of my second order construct is low, then can I remove the second order construct and can I draw model directly to the constituent. In my model...Training-->ethics-->workplace behavior(second order construct with satisfaction and attrition). So if theory allows can I remove workplace behavior and draw a model Training-->ethics-->attrition.
I think I understand. I think you are asking whether you can model your 2nd order factor instead as just the first order dimensions, and then just have predictive arrows point at those dimensions rather than at the 2nd order factor. The answer is yes. Here is a paper where we do that: www.researchgate.net/profile/Paul_Lowry/publication/264623594_2013-Nov-JAIS_HMSAM_PUBLISHED_APPENDICES/data/53e9b4960cf2fb1b9b671a95/2013-Nov-JAIS-HMSAM-PUBLISHED-APPENDICES.pdf
@@Gaskination Excellent Professor, I have m defense tomorrow and this is extremly important for me to cite.
Thank you very much for the video, very useful. However, I have two questions: how can I have access to the 95%CI of the indirect effects, please? And is there a textual way of reporting the results without the table? I really appreciate your help.
If you use an estimand and perform a bootstrap (as shown in this video: ruclips.net/video/ICnh3s2FG14/видео.html) then you can view the upper and lower CI. As for reporting results, you would just say that the indirect effect was observed, as indicated by the confidence interval not spanning zero (LB: [put your lower bound CI here], UB: [put your upper bound CI here]).
@@Gaskination Perfect, thank you very much!!!
Hello James. I have a question. I have a model with one IV and one Mediator and four DVs. Do I do the Structural regression first then do the confirmatory or do the opposite. If you this I should start with structural, should I remove the effects that are insignificant and then do the confirmatory. Thank you
Dear James Gaskin, thanks for the video, it's very helpful!
I have a question: when AMOS estimates the regression weights, including direct and indirect effects, of my model before bootstrapping (2000 resamples, CI at 95% level), they are significant, but after bootstrapping the indirect and direct effects are not significant anymore, so that I can't carry out the mediation analysis anymore. Why is that so and what can I do to make the mediator analysis possible?
Thanks in advance!
+Uirassu Borges Junior Bootstrapping will often inflate the p-values a bit. You can try increasing the number of resamples. I just did a test with 200, 2000, and 20000. The p-values for generally decreased as I did this.
Hello James, the video is very useful. I have 2 queries and hope that you can help. First, can I use the same method for mediation with 2 mediators? Secondly, using this method, how do I get the 95% confidence interval for the indirect effect? Thank you! =)
+Happy Snappy When bootstrapping, AMOS will produce the 95% confidence intervals (see the upper and lower tables). If you have two mediators, if they are both connected to the IV, then you will need to do a sobel test instead of bootstrapping. If they are serial, and not both connected to the IV, then you can bootstrap the same as I've shown here.
+Amit Agrawal sorry, that was me. I was signed in to one of my other youtube accounts...
Dr James, you are very helpful with videos. I am confused about some point.
The regression weight is give to use a direct effects but when we use the boostrap methods this values are changing. I can not decide which p values is true. The problem; this regression weight and boostrap direct effects significant is not the same statistical result.
Regression weight is significant but the boostrap direct effects is not significant. What is the problem with this case?
Thanks in advance!
bootstrap is more accurate.
@@Gaskination thank you :)
Hello James,
Thank you for the video.
I was wondering if there is mediated mediation? If yes how is it different from sequential mediation? And how can we do mediated mediation in Smartpls?
Thanks
I think you mean serial mediation. In SmartPLS, I'm not sure if there is a way. Here is my latest video on SmartPLS specific mediation: ruclips.net/video/z1h6qMcMhWk/видео.html. And here is one for AMOS that does cover serial mediation: ruclips.net/video/KgtgQO9zQK0/видео.html
Dear Prof,
I have a question regarding number of bootstrap samples. Is it the same one with our research samples?
So, it means, if we have 600 samples, or 1500 samples, do we need to insert the same number?
No, it does not need to be the same. It is just the amount of resampling that is occurring during the bootstrap procedure. Typical bootstrap samples are 2000, 5000, and 10000. However, 10000 gets very arduous due to computational requirements if you also have a complex model or a large sample size. I usually just use 2000. If you do some tests, you'll notice that the variance in results tapers off quickly as you get past 2000.
@@Gaskination oh, I see. thank you for the explanation, prof. 🙏
Hi james, thank you for your video. I learned a lot by watching your video. But I have a question. What if my model shows that the direct effect is sig. but the indirect effect is not significant? Is there still a mediation effect if the output like that? (There is one IV, one M, and one DV. just for information path c, IV to DV is sig. path a, IV to M not sig. and path b, M to DV is sig) I hope you could help me james, thank you
If the indirect effect is not significant, then there is no indirect (mediation) effect.
First of all, thank you very much for your great video. Could you please help me with my problem: in my model i have 1 independent variable, 6 mediators and 1 dependent variable. However, I could only perform bootstrap in Amos without the standardized estimated values because there is a negative correlation between 2 mediators. So I only get the results for the total, direct and indirects effects without the lower or upper bound and p values. How can I interpret the results in this case? Is there any other resolution for me to test the mediators in Amos?
If you standardized the variables in SPSS before using them in AMOS, then the unstandardized estimates will be effectively similar to the standardized estimates.
Dear Professor,
I really appreciate your efforts here but I have a question regarding the bootstrapping method. I would like to ask how to decide the optimal number of bootstrap samples as I have watched several youtube videos regarding the bootstrapping in AMOS. Almost all of them used 2000 times and therefore, I have used 2000 times in my paper as well. But I cannot find a good explanation on why I choose 2000 instead of 1000 or else. I would be grateful if you can answer this. Thank you!
Generally speaking, more is better. However, for decades, the number of samples was limited by computing power. Now that computers are so powerful, it is easy to do a few thousand. Hair et al and some other scholars, such as Ringle, suggest a minimum of 5000 when the sample size and model complexity allow it. I ran an analysis the other day on a sample size of 53000 though, and so I could only do a bootstrap of 500 (or else the computation would never finish...).
@@Gaskination Thank you for your reply and it helps me a lot!
Dear professor
It was Great, Thanks
I ran with your method but an error message came up:
D uring the analysis of a bootsrap sample ,an attempt was made to compute the correlation between two variable, one of whose stimated variance faild to be positive.
Please help me solve the problem۱
Check the notes for the model. It should identify which error had a negative error variance. You can double-click on that error bubble and go to the Parameters tab and then set the variance to equal a small positive number, like 0.05.
Thanks alot
I do it.
Prof, can you help me? in my result, for direct effect I get .393 (sig) and for indirect effect I get -.036 (sig). Is it normal if our indirect effect show negative?
Dear Dr.Gaskin: Is this way to test mediation different from the way you've explained in SEM Series (2016) 8. Mediation??
The only difference is that in the sem series I show how to use the estimand. Otherwise it is the same.
excuse me sir, I would like to ask what if the indirect effect is significant at p-value < 0,01 (p-value=0.004) and direct effect is also significant but only at p-value < 0,1 (p-value = 0.051). Do we call this effect partial or full mediation?
Thank you in advance
In this case, it probably depends on how much the direct effect changed after adding the mediator. If the direct effect changed a lot, then I would call this full mediation. If not, then it is probably partial.
Hi Doctor. in the case where all paths in the structural model are the significant but direct effect is greater than the indirect effect. do we still have a potential mediation? thank you
yes. It is just partial mediation
Thanks. Can you please help to understand what about the the sign of the direct and the indirect effect. Please how can I interpret an unexpected negative direct effect ?
+SAID ETTIS Sometimes when the direction of the direct effect is unexpected in a mediation model, it is because the IV has a dual effect on the DV. The dual effect is split by the mediator so that the positive effect travels one way, and the negative effect travels the other way. If the mediator has a positive effect on the DV, then usually the positive effect from the IV will travel through the mediator, leaving the direct effect of IV on DV negative.
Hello Prof James Gaskin,
My model is following
Path std. dev t-value p-value
X->Y= 0.338 0.077 4.400 0.000
X->M 0.581 0.053 10.981 0.000
M->Y 0.347 0.061 5.713 0.000
i have calculated indirect path coefficient manually
path = 0.581*0.347= 0.201 (Correct according to SmartPLS)
std.dev = 0.053*0.061= 0.0032 (Incorrect)
t-value= 10.981*5.713= ---- (Incorrect)
Please help me in this regard how to calculate std.dev, t-values and p-values for indirect effect
This plugin will do some of this for you: ruclips.net/video/-xzkjqKhWHE/видео.html. The t-value is just the unstandardized estimate divided by the standard error.
Hi, Could you please explain the answer of statement from this video (run time @ 3.32-3.54 min.).. "Which model to chose" ?
dhruv dutta you have to have a good theory to support your modeling choices. If you do it just based on the statistics, it's hard to tell which model is best. It really should be based on your theory.
Hi professor, How if my mediator did not shown direct effect into the dependent variables and Independent variable has shown direct effect into the dependent variable. Does it still considered as Partial Mediation?
Just look at the indirect effect. Generally speaking though, the mediator must have an effect on the DV to consider mediation.
Dear Prof,
If I have 2 IV with 7 dimensions and 2 mediators and one moderator and 1 DV. So, How to check the mediation and moderation effect on AMOS?
Looking for your kind help.
Thanks.
Suliman
Hi Suliman,
I have several videos on mediation and moderation in AMOS. If you click on my avatar/profile, you'll be taken to my channel where you can search for these videos using the search icon (magnifying glass next to the About tab). Here is the channel url: ruclips.net/user/Gaskination
Hi James, my Amos does not bring out the Estimate/Bootstrap results window for me to check the two tail significance. Is there anything I am doing wrong? My measurement model is a second-order model. This happens when I add control variables to the model.
If you are running a bootstrap (which must be selected on the bootstrap tab of the analysis properties), and if you have checked the box for indirect effects (output tab), and it still isn't working, then I'm not sure. It should definitely show up if you've checked those boxes.
@@Gaskination strangely it disappears when I add in controls to the model and there are arrows from the controls to both the mediator and the DV.
When the arrows only go to the DV suddenly bootstrap results boom! show up.
I also face this problem: when I bootstrap 2000 samples in the mediation analysis I receive an error msg that "During the analysis of a bootstrap sample, an attempt was made to compute a standardized regression weight between two variables, one of whose estimate failed to be positive. The attempt was made because ' Standardized Estimates' in the 'analysis properties' window was checked or because the Standardized method was used."
when I bootstrap 30 samples the analysis runs just fine. Have you encountered this before? How may I overcome this if I want over 500 bootstrap samples?
@@DK-em6oz That's odd about the controls... It may have something to do with missing data in the control variables. I have never seen the second error either... I'm impressed you've stumped me twice. I've been doing this for a decade and have never seen those... The error is weird though, because it is fine if a standardized estimate is not positive. I hope you figure it out! Best of luck to you!
Hi James, why did you not test the model without the mediator first? What would it mean for a model when 3 out of 4 direct relationships from IV to DV in a model without the mediator are not significant. I think it means no mediation, but does this mean your research should go in the bin? When the moderator is added, only 1 out of 4 is not sig.
+Sarah Preis More recent literature suggests that we do not need to test the model without the mediator. It suggests that if the indirect effect is significant, then we have mediation, regardless of whether the direct effect was significant before adding the mediator. So, mediation is determined purely by the bootstrapped indirect effect.
+James Gaskin \\ Thank you very much for the reply. Very much appreciated.
+Amit Agrawal This should help: scholar.google.com/scholar?q=mediation+bootstrap+%22indirect+effect%22&btnG=&hl=en&as_sdt=0%2C45
@@Gaskination Sir, in your stats wiki website, you have given the lesson pdf under mediation in which you have specified that we have to report direct effect without mediator and then direct effect with mediator along with the indirect effect. So, sir, my doubt is which approach should I use?
@@dr.anupreetkaurmokha5792 The best approach is full disclosure and transparency with more information. So, I would recommend reporting all effects (direct and indirect). The indirect effect is what we use to determine whether there is mediation present. The other effects provide additional context. It is no longer required to test with and without the mediator (Barron and Kenny approach). Here are some helpful references: statwiki.gaskination.com/index.php?title=References#Mediation
Hi James, could you post a video on serial mediation with multiple independent variables in AMOS, please?
+Cherry Zhang I might in the future, but for now, just know that you can simply do a bootstrap and look at the indirect effects. If your model looks like this: A->B->C->D, you can look at the indirect effect matrix which will show you the indirect effects from A->C, A->D, and B->D.
+James Gaskin thanks!!!
Hello, Dear James, and all interested in mediation analysis. :) I have a question again. Its: when I look at regression model (direct relationship) from A, B, C to D it was significant. But when I add into the model two mediators X and Y the relationship between A,B, C to D was unsignificant (it looks like full mediation from first sight BUT, A, B, C to X and X to D is also UNSIGNIFICANT (Y was significant full mediator). What does it mean and how should I interpret results?
+Tija Ragelienė This can happen if the mediator is a permeating variable that floods out other effects in the model. Some known variables like this are self-efficacy, confidence, optimism, etc.
+James Gaskin Thank You James. My mediator x is empathy, which was supposed to mediate the relationship between family relationships characteristics and adolescent identity. How I should interpret results in such a case? does it mean that effect of permeating variable is the bigger than other mediators in some way? Maybe do You any article or etc. where I could read more about it? Cause mediator Y which was full mediator was much more strongly supported by theoretical background...
+Tija Ragelienė Sounds like the other mediator is just very strong, or that empathy is not a good mediator. You can try it without the other mediator to see if empathy works.
All right. I've just tried and without the other mediator empathy works. So what should I do with my results? Should I interpret as two separate models with different mediators? But then I am confused with theory because empathy and other mediator could be both predicted by family charakteristics and from my own theoretical assumptions based on theory predict identity diffusion. Or shoud I just write in discussion section that empathy effect could be not such as expected because the other mediator was much stronger?
+Tija Ragelienė I would include both but then say that when tested separately, empathy works. This implies that the other mediator is the stronger one and the one that should receive the attention in the literature and in practice.
Dear Professor, How many bootstrap samples are required for 1100 observations?
I would recommend a minimum of 2000. The trouble with such high sample size is that if you do very high bootstrap as well, it will take a very long time to run...
Thanks, professor. Is the high sample size not good? Another issue is the scatter plot in regression has some values above 3 on the x-axis and y-axis. Is that mean the data is unsuitable for SEM?
@@nargisali7298 high sample size is generally good. Its only downsides are a deflation of p-values (i.e., everything, even small effects, are "significant") and the time it takes to run a bootstrap. Otherwise, it is very good to have a high sample size. As for the scatterplot, I'm not sure what the problem is. The axis size on a scatterplot differs from dataset to dataset, based on the size of the scale for the measures. So, I'm not sure if values above 3 is a bad thing.
Thanks, I am using a five point likert scale. Scales with mostly disagree and agree are positively related with the DV, significant p-value and scales with disagree or strongly disagree are negatively related with DV, sig. p-value. However, two of the scales are insignificant in the model which I think I will have to exclude from the model. Is it okay to keep a few negatively and few positively related variables in a model?
Sorry, I am asking too many questions and taking your time :(
@@nargisali7298 As long as the outcomes make sense, then it doesn't matter which direction the effect is. For example, if the DV is job satisfaction, then predictors like management fairness, participation, and autonomy will have positive effects, but predictors like burnout, ambiguity, and bullying will have negative effects. I would recommend also keeping non-significant predictors in the model if you theorized they should have an effect.
Professor, how I can add reference .Any paper which suggests bootstrapping
Many of these: statwiki.kolobkreations.com/index.php?title=References#Mediation
@@Gaskination Thankyou Professor.
Nice video. Sir, I have 2 Question, Please
1. While testing Mediator, we should interpret in research only the final model in which direct and indirect impact were tested i.e. partial or full mediation or to interpret also the initial model tested separate for the purpose that there is mediation or not such as IVs to DV (separate checked), IVs to Mediator and Mediator to DV.?
2. In final model (partial or full mediation model), we should interpret only direct and indirect impact OF standarized indirect impact or unstandarized Impact value, and we should also interpret Total impact?.
will be thanked.
Muhammad Anwar there are different schools of thought on this. The most recent literature recommends you just assess the final model. No need to assess a separate IV to DV direct effects model. As for the total effects, some journals do you like you to report the total effects. They are most concerned about whether that value is significant.
Thanks Sir, Noted.
@@Gaskination Hi there, tnx for all your help. I have some related question on this, is it possible when our total effects not sig we interpret based on direct and indirect impacts in our model ?
In our model we have 7IV, 1mediation and 1DV. After bootstrapping(5000sample) in both standardized total effects and standardized direct effects only one IV to DV had acceptable p value(different path), while six standardized indirect effects were acceptable.
If possible, please cite the source for reference in the article.
Tnx again.
@@psycho2758 I think most papers argue that total effects are important. However, I think Zhao et al 2010 allows for indirect effects only. (please make sure to double-check it) Zhao, X., Lynch, J. G., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and truths about mediation analysis. Journal of consumer research, 37(2), 197-206.
@@Gaskination That's something to me, thank you very much.
I am looking forward regarding the SEM Model on Lean or six sigma please if u have any data regarding to that please reply me
+Zee Asim I don't. Sorry.
How about serial mediation?? IV-M-M-DV. Help me.
+sugashwar prashanth Same approach. Just do a bootstrap and then check the standardized indirect effect and two-tailed p-value. You can check it for the IV to M2 and for the IV to DV, and for the M1 to DV.