Hi James, When I do analysis with moderator, the result is like below: H1: A ---> D - Supported H2: B ---> D - Supported H3: C ---> D - Supported H4: X does not moderate any of the above However, when I analysed again without moderator (I did not create the composite variables because I did not use moderator this time), the result is different: for example: H1: A ---> D - Not supported H2: B ---> D - Support H3: C ---> D - Support My question is: Can that happen? Because I am expecting that without moderator, the result should be exactly the same as the first case. Thank you
This is perfectly logical. SEM is an interrelated network of effects. WHen you remove one variable (or add one), it can change all the other effects in the model.
Hi James, Thank you so much for your video! I noticed that you use path analysis (using the mean value for each variable) rather than the full SEM in most of your videos. I did the same with my analysis in order to balance the sample size and the complexity of my model (I also have moderators for my model, I don't think using full SEM model will give me stable results). I understand that it is not statistically justifiable to use path analysis due to the assumption of the error free measurement. I wondered given the issues I have whether it is justifiable to perform path analysis like this. Is there any papers to justify the use of path analysis for social science? Many thanks, Cherry
Here is a link to some relevant literature: scholar.google.com/scholar?q=impute+%22factor+score%22&btnG=&hl=en&as_sdt=0%2C45 The first result looks useful.
You are really awesome Dr. Anyway, i have one question to ask. In my dataset i have a categorical experience variable (1. 2 years or above, 2. 1 year and 11 months, 3. Between 6-11 months, 4. Less than 5 months) and i am wandering how to code this variable so that i perform a interaction of other variable with experience (Low experience and high experience). Thank you so much.
If you just want low and high experience, then you could recode this variable to be two groups (less than a year and more than a year). Then do a multigroup analysis. I would not try an interaction, since the intervals are not equal between the group values.
Dear Dr. Gaskin, thank you so much for your great work! I would be great full if you find the time to answer one more question: In my model there is one independent dichotomous variable (X) pointing on two dependent variables (Y). The effect is moderated by three independent Variables (M1/M2/M3). In my understanding z-standardisation is usually not appropriate for dichotomous variables (or am I wrong here?). Do you have any advice on how to create the "X*Mx"-Values for testing moderations with AMOS? Should I standardise only the Moderators to build "X*M1", "X*M2" and "X*M3"? Thank you so much in advance! BR
Dear Dr James Gaskin, happy new year! Hope this year would be even a better one for you and your loved ones. Quick question. You mentioned that alpha 0.1 can still be considered for interaction terms as it is usually weaker than regular bivariate relationship. May I know the reference for that? Or do you have any other source to validate that? Thank you, and looking forward to hearing from you. Alvin Hadiono University of Glasgow
Hi Alvin, The most current school of thought on this topic is that using p-values for interaction effects is not useful. Instead, experts suggest a floodlight analysis, achieved through a Johnson-Neyman plot, like in this video: ruclips.net/video/Rb_lAmjHd8s/видео.html
Hi James, Thank you for the videos. They are very useful. I have one question about the interaction effect. The analysis results show that the causal relationship between AtypicalUse and Usefulness is not significant. Can we conclude that the moderation effect of Experience on AtypicalUse - Usefulness causal relationship is not significant and so the hypethesis is not supported?
Dear Dr. Gaskin. Thank you always for your support. My question on this video is for the intercept value, 3. Is that because you used a 5-likert scale to measure the variables and the median value from 1-5 is 3?? If I used 7 likert scale, then do I have to change it to 4 instead?
It is just to place the lines in the center of the graph. The important part of the graph to interpret is the differing slopes and relative amplitudes (along the y-axis).
Hi James, Thank you so much for all of your videos - they are amazingly helpful. I have used your "myindirecteffects" estimand to look at a multiple mediation model within an APIM framework (with distinguishable dyads) and I have a couple of questions: 1- If I examine the direct effects in the model (e.g., same actor effect in each member of the dyad) and find that they are not significantly different, it is possible for me to incorporate this information when using your estimand? (i.e., if one of the 2 equivalent paths is path a of the mediation I am testing) 2- Do you know of a research paper that has given credit for using your estimand (and might be a good example of how to write up these analyses? Thank you so much again! Best, ~Roanne Millman
1. I’m not sure I understand, but it sounds like you are saying that you have two groups, and for one of the groups, one of the paths (A or B) is not significant. Traditionally (e.g., Baron and Kenny), we would say that both paths must be significant in order to consider mediation. However, I’m not sure if the more recent literature allows mediation simply if the indirect effect is significant (ignoring the A and B paths on their own). Some of the most recent literature is cited here: statwiki.kolobkreations.com/index.php?title=References#Mediation 2. As for a study using my estimand approach that is published, I’m not sure. It usually takes some time to get papers published, and I introduced the estimand only last year. However, the approach simply takes advantage of bootstrapping to generate standard errors and confidence intervals for indirect effects. This is essentially what the most recent literature recommends anyway. Even though that literature uses other tools, it is the same approach.
Hi doctor, is it acceptable to have a correlated structural error term between variables?? lets say i have a model consisting from A,B, and C variables A act as Exogenous variable B act as Endogenous variable for A but at the same time its act as Exogenous for C. C act as Endogenous variable Note : B is not acting as a mediator due some theoretical limitations. In the model you have discussed in this video the nearest scenario for my case as following: Is it acceptable to have a correlated structural error terms between Useful and InfoAcq ?? Your help is really appreciated Dr. James
Hi James, Thank you very much for the video, it is very clear and helpful! Just one quick question. I wonder if we need to mean center the IV, moderator and the interaction term in path analysis? I realize that when we need the simple slope analysis to further interpret the interaction relationship at each level of the IV, we may need to mean center the variables.
The standardized coefficients will be the same either way. For the interpretation of the simple slopes analysis, you can use either, as long as you account for the scale size when making your interpretation.
Dear Dr. James, A question: How do we interpret the interaction effect with second order factor? For ex: SYQ and IFQ (1st order) create second order factor with the name SIQ. Then, I tested the interaction and found that SIQ (2nd order) is significant. If I test the 1st order factor interaction, SYQ is significant, but IFQ is not significant. Which one should I write for my report? SYQ (1st order factor) or SIQ (2nd order factor) ? Also, what should I put for the 2-way-interaction from excel (Stats Tools Package)? the 1st order or 2nd order? *Note: the theory said that 2nd order or 1st order is acceptable. I can use SYQ (1st order) or SIQ (2nd order). However, my hypothesis is using SYQ (1st order), not SIQ (2nd order).
Dr. Gaskin thanks so much for your very helpfull videos! In my path analysis I find CMIN/DF=1.697 GFI=.991 RMR=.071 CFI= .995 PCLOSE= .391 and RMSEA=.053 Can I assume that I have a good model to support my hypothesis? My sample is 250 cases. Thank's in advance.
Hi James, many thanks for your useful lesson. Can I ask, when you mention that a 90% confidence level is fine for interactions, do you happen to know some sources that explain it? It would help me a lot, been looking around but couldn't find any. Many thanks!
Here are some references on moderation: statwiki.gaskination.com/index.php?title=References#Moderation_and_Multigroup I'm not sure if any of them talk about a more liberal confidence level. You can just argue that interactions are generally more exploratory than confirmatory, so a more liberal confidence level is merited.
Dear Dr. Gaskin, for moderated mediations, I tried out your ModMedEstimands plug-in - and I love it. Is it flawed (the code looks straightforward to me) or why is there no reference to it in your videos or elsewhere? Thanks a lot for the great help and inspiration you provide. Best, Leo
It works just fine I think. I haven't made a video about it yet though. All it does is calculate the difference between two indirect effects and creates confidence intervals and p-values for that difference.
Thank you for this, Dr Gaskin. If my hypothesis says that a high level of the moderator affects the relationship between the independent and the dependent, can I only test the moderation at a high level of the moderator (-1SD)?
It simply refers to the direction of the moderation. The plotter should interpret it for you, something like: "M strengthens the positive effect of X on Y"
@@Gaskination Thank you for your reply. The problem is that the moderation is insignificant. However, I am only interested in the moderation at high levels of the moderator! (the theory assumes that high levels of the moderator affect the relationship... not all levels).
@@mohmoh7078 Ah, ignore the significance. To assess the moderation at high levels of the moderator, you can us the interaction plotter, as it plots three lines, one for mean, and one for -1 SD and for +1 SD. So, this should tell you if moderation occurs at high levels of the moderator. Another option would be to split the moderator into low/high groups and then conduct a multigroup analysis.
@@Gaskination Many thanks! So, If the moderation at high levels were significant, then my hypos are supported regardless of the insignificance of the moderation at mean levels? (since I am concerned only about high moderation)
@@mohmoh7078 It's all about the slopes. If the slopes are parallel, then there is no moderation. If the slopes diverge or converge, then their is moderation.
Hi Dr. Gaskin, I found this technique is pretty useful for testing the moderation effects among the observed exogenous variables. However, what if I want to test the moderation effects between latent variables, should I have to use the factor score from my CFA instead to compute the X*Y(they are both latent variables)? Thanks in advance!
You can keep them latent if you multiply every pair of indicators across the IV and moderator. Here is an example (but with smartPLS): ruclips.net/video/upEf1brVvXQ/видео.html I also show how to do it in AMOS during my most recent SEM Boot Camp videos. It is on day three: ruclips.net/video/ViQITM1htjU/видео.html
Hi James, thank you for the whole SEM Series! It is awesome. Yet I struggle a lot with writing down my results in my thesis. With the interaction, I can't use the Stats tool, as I have to do all calculations and plots by myself. How can I plot the two way interaction using the regression coefficients? What if, only the moderator and the interaction term are statistically significant but the relationship of IV on Y is not? This is the case for my analysis (I have stimuli situations as the IV and the M - the IV would be that respondents in the survey saw an affective ad claim and others a cognitive ad claim and the moderator is, whether respondents see a speech of Greta Thunberg or not). So for me it doesn't make sense to standardize them. Can I use the regression coefficients of the normal AMOS latent regression with the two dichotomous variables as observed variables? Thanks in advance for any help!
Also, I wonder why you can include the coefficients of the non-significant relationships (AtypUse --> Useful). Do you have a suggestion for any book where this procedure is explained? So far unfortunately I have not found any sufficient source...
@@CarinanaLaa 1. Why would you have to do the calculations yourself? That seems like a waste of effort. That's like continuing to walk to campus rather than riding the bus. For interactions, you can look at the formulas in the stats tools package. I don't think they are restricted. You just have to find them below the line. I think I made the text invisible... but, you can still see the formula in the formula bar at the top. 2. If the IV to DV direct effect is not significant, but the interaction and moderator are, then this is very interesting. This means that the effect from IV to DV is only meaningful when considering the moderating effect. That's great. 3. You are correct that standardizing a binary variable isn't very useful. But, if it is binary, then just use multigroup analysis instead of interaction. Here is a video on how to plot MGA using google: ruclips.net/video/ZnptsnQ2VT4/видео.html 4. Retaining non-significant relationships is wise. If an effect is theorized, it should be included in the final model, even if not significant. This way you can account for the nominal effect it has when you are testing other hypotheses. No citation needed for including a theorized nonsignificant effect.
@@Gaskination Thank you very much for the fast reply! I know, its stupid, but I have to show in the appendix in my thesis how I calculated it or at least my source of calculations which needs to be a book or a paper. I can't use multigroup analysis as the sample size is definitely too small for that. For the non-significant effect I meant, that I thought in die Interaction-Plot it might need to be put to 0 as it has no significant effect. Do you have a source for the calculations of the plot?
Thanks for posing all these videos, very helpful. I have a question: What if the path between the mediator variable (M) and the dependent variable (Y) is moderated by a factor (W)? There is no factor moderating the path from independent variable (X) to M. When doing moderation test (after testing mediation effect), shall I just add a path from W*M to Y and W to Y, co-variate X, W*M and M and then run the analysis? Thanks a lot.
Dear Sir, Like others I also struggle to cite the methods suggested by you and your videos. The approach is unclear. For example, Preacher, Rucker, and Hayes (2007) suggests testing two regression equations, the “mediator model” and the “dependent variable model” to test for moderated mediation. The ppt provided on statwiki suggests Preacher and Hayes paper (link not working though). Assuming Preacher and Hayes (2008) paper, it talks about multiple mediators.. Not asking you to spoon feed us but requesting to clarify the approach you are using.. Rest readings are on us.... but as of now it is unclear what to read to refer the methods used by you. Regards
For interactions, I'm just using a standard approach from any textbook, like Hair et al 2010. In this approach, you standardize the IV and moderator and then multiply them. For mediation, my most current videos use Zhao et al 2010 (reconsidering baron and kenny) - although there is a newer approach that I need to make a video for. For moderated mediation, whether it is indirect effects across multiple groups or whether it is the mediated interaction effect, I think Preacher and Hayes cover all forms.
Hi Prof Gaskin, First of all, thank you very much for the video. I am having difficulties following the variables which previously have a lot of indicators now only having one indicator per LV. How do you process those indicators into one LV? Did I miss an important step? How to moderate 2nd order variable instead? Thank you for response
Here is the missing link: ruclips.net/video/dsOS9tQjxW8/видео.html If you use this linked video to create factor scores for the 2nd order factor, the process is the same.
@@Gaskination Thank you very much. The series is very helpful. Maybe could run the same data set and model in other CB-SEM such as Lisrel for comparison sake? Just my 2 cents.
Hi Gaskin, what does this interaction method called? is it two-stage approach to interaction as suggested by Kenny and Judd (1984)? Would you please refer to any study that used this method? Thanks
Hello James..here I am again to thank you for such saviour videos and impose another query as follows In my A-B-C type of model where I have two Moderators M1 and M2 at A-B and B-C level resp ..how do I test interaction M2 at B-C level..? In A-B level interaction M1 we covary all exogenous variables including variables at A level, control variables and interaction variables.. But what do we do to test M2 interaction in the second part (B-C) of the model.. Can I covary M1 with A and control variables, and M2 with only control variables and proceed with the interaction analysis I tried to look for a solution at many places before asking you but couldn't find anything..pls help
I don't know of an article to support the approach, but, practically, you can just covary the error of the mediator with the interaction and moderator.
Hi James, Thank you for your videos. they're really helpful in understanding SEM. However I do have one question about this moderation technique in AMOS. Currently my moderator is a latent variable that is loaded by 9 different subscales and is not represented by a total composite score. How should I go about running moderation analysis in this scenario? Much thanks, Melvin
If you want to do interactions, it is very tricky with latent variables. You would have to cross multiply all items from the IV and the moderator. Then each of these new product items would become the indicators for a latent interaction.
Just a basic question, should we must have a variable that is categorical in nature to act as a moderator or we can use interval data like that in social support variable or any other such variable can be used as moderator in interaction based moderation?
Ahmad Usman for interactions, just best to have non-categorical variables as shown in this video. When you have categorical variables, it is better to do a multiple groups analysis for your moderation.
Hi James, Thank you for this video. ... Can the relationship (M-DV) between the Moderator variable (M) and the Dependent Variable (DV) be significant, given that the Interaction effect (MxID) is significant on the DV?
Unfortunately there is not. In AMOS, I always use a path model with composites when testing interactions. It is just way simpler. However, when I do interactions in SmartPLS, I use a latent model. Here is a video of that: ruclips.net/video/PnPfOGtl-lc/видео.html
Hello Dr. Gaskin, a kind request can you please suggest an empirical paper of interaction (moderation) conducted in SEM (AMOS) for reference, as I am not able get any paper on interaction (moderation) that was conducted in AMOS.
Hi James. Thank you once again for this amazing video. For my paper I want to test the impact of Institutional factors on Innovation. I have done EFA in SPSS and found institutional variables that fit in two factors. According to the literature I have named them Formal Institutions and Informal Institutions. As per your advice I run Cronbarch Alpha and it is higher than 0.7 Now I want to save the factors so I can use them in AMOS SEM model and check interaction effect as well but I am not sure if I should save factors like this in SPSS or use AMOS Data Imputation option to save the factors. I tried to do both and compare the result of factor saving and data imputation but I did not get similiar results. In this case I do not know which option I should use and I would be extremely thankful if you can help me! Lena
@@AshutoshPandeymentor Yes, if they're not already standardized or mean-centered. Check their distribution just to be sure. You don't want to standardize twice.
If you mean you want to use an ordinal (likert scale) measure as a moderator for grouping, then you can simply split the data into two groups where 1 and 2 are part of the low group, and 4 and 5 are part of the high group (and you remove 3). Or, if you want to do an interaction, you can create a factor score as shown in this video above, and then standardize and multiply with an IV to create an interaction variable. If you mean that your variable is 2nd order, then you can still create a factor score for it in AMOS during the CFA.
Hi Gaskin, I am wondering what if I have a mediated moderation model as follows X: independent var Z: Moderator var and XZ: the interaction M: mediator and Y: dependent variable. I found the interaction (XZ) significantly related to the mediator (M) but the independent (X) and the moderator (Z) were not significantly related to the mediator (M). The moderator (Z) and Mediator (M) positively related to dependent variable (Y). Is the interaction effect/the moderation still exist and how to interpret it? many thanks
This is very interesting. This means the IV and moderator are not sufficient alone, and are only good predictors when combined. This is the best kind of moderation result.
+miracle ozzoude I think you mean greater than 0.05. In this case, I am accepting effects at the 90% confidence level, so a p-value less than 0.100 is fine enough.
Hi James, is the 90% confidence level the mostly accepted (compared to 95% level) standard? Is there any requirement to sample size if I use the 90% level not the 95% level? Thanks.
It depends on your desired alpha threshold. If you're okay with 90% confidence in the estimates, then a p-value less than 0.100 is fine. If you require 95% confidence, then p-values must be less than 0.05. For interactions, I'm usually a little more lenient and choose 90% confidence.
@@Gaskination Hi James. Thanks for the Video, your explanations are amazing! Do you have a paper to cite for the fact that 90% confidence level is enough for moderators? I don't find any...
@@m__h5506 No need to cite your choice to accept a different confidence interval. This is just an analytical choice that depends on your research goals.
I need an urgent suggestion: Why you are not including items of study variables? Can we do simply adding the average of each variable like this? Please help.
@@Gaskination Sir can you please share the link where you talked about factor scores?? Also, can you please help me. In my research work I have one IV, two DVs, and with two moderators. Suppose one IV has 6items, can I take the average of all the six items and name it as AVGIV and then run the full model?? Similarly with other variables also?
@@Gaskination Hi James, Sorry for this, I sent you a message mas for some reason the text didn't go. Thanks a lot for your videos and explanations! They help a lot. Would you be so kind to help me a little further? Doing my analysis, the following situation came up: Results after testing the causal latent model: H1:A-->C (expected negative) - significant (pC (expected positive) - not significant (p=0,07) and negative (beta=-0,065), not supported Results after testing moderation (with composite variables) of B on the relation A-->C (H3: moderation should be negative, e.g. B expected to dampen the negative effect from A-->C): A-->C - not significant (p=0,75) and positive (beta=0,019) B-->C - significant (pC - significant (p=0,013) and positive (beta=0,139) I understand that being SEM an interrelated network of effects, adding or removing variables can change the overall effects, but in this case the relations A-->C and B-->C generally went opposite! So my doubts are: 1. Following moderation test, I would say H3 was not supported (when plotting, it says "B strengthens the positive relation between A and B"), am I correct? 2. Do my conclusions about H1 and H2 after testing the causal latent model have to change due to the corresponding estimates obtained in the moderation test? Thanks a lot! Ligia
@@ligianascimento1923 Ha! That's rough. Generally speaking, latent models are more reliable and produce better estimates. With interactions, latent models are unwieldy, so we often use composites as you have. The trouble is that these often alter the estimates too much. So, in your case, I would recommend to either use a different approach, such as the J-N floodlight approach: ruclips.net/video/Rb_lAmjHd8s/видео.html or a different software, such as Mplus: ruclips.net/video/lmMbEI_Kw2I/видео.html or SmartPLS: ruclips.net/video/PnPfOGtl-lc/видео.html. Mplus and SmartPLS allow latent interaction analysis.
Hi James,
When I do analysis with moderator, the result is like below:
H1: A ---> D - Supported
H2: B ---> D - Supported
H3: C ---> D - Supported
H4: X does not moderate any of the above
However, when I analysed again without moderator (I did not create the composite variables because I did not use moderator this time), the result is different: for example:
H1: A ---> D - Not supported
H2: B ---> D - Support
H3: C ---> D - Support
My question is: Can that happen? Because I am expecting that without moderator, the result should be exactly the same as the first case.
Thank you
This is perfectly logical. SEM is an interrelated network of effects. WHen you remove one variable (or add one), it can change all the other effects in the model.
@@Gaskination I see, thank you :)
I am really grateful for all of these videos and materials you are offering Dr James. You are really awesome.
Hi James,
Thank you so much for your video! I noticed that you use path analysis (using the mean value for each variable) rather than the full SEM in most of your videos. I did the same with my analysis in order to balance the sample size and the complexity of my model (I also have moderators for my model, I don't think using full SEM model will give me stable results).
I understand that it is not statistically justifiable to use path analysis due to the assumption of the error free measurement. I wondered given the issues I have whether it is justifiable to perform path analysis like this. Is there any papers to justify the use of path analysis for social science?
Many thanks,
Cherry
Here is a link to some relevant literature: scholar.google.com/scholar?q=impute+%22factor+score%22&btnG=&hl=en&as_sdt=0%2C45
The first result looks useful.
You are really awesome Dr. Anyway, i have one question to ask. In my dataset i have a categorical experience variable (1. 2 years or above, 2. 1 year and 11 months, 3. Between 6-11 months, 4. Less than 5 months) and i am wandering how to code this variable so that i perform a interaction of other variable with experience (Low experience and high experience).
Thank you so much.
If you just want low and high experience, then you could recode this variable to be two groups (less than a year and more than a year). Then do a multigroup analysis. I would not try an interaction, since the intervals are not equal between the group values.
Dear Dr. Gaskin,
thank you so much for your great work!
I would be great full if you find the time to answer one more question:
In my model there is one independent dichotomous variable (X) pointing on two dependent variables (Y). The effect is moderated by three independent Variables (M1/M2/M3). In my understanding z-standardisation is usually not appropriate for dichotomous variables (or am I wrong here?).
Do you have any advice on how to create the "X*Mx"-Values for testing moderations with AMOS? Should I standardise only the Moderators to build "X*M1", "X*M2" and "X*M3"?
Thank you so much in advance!
BR
Correct. No need to standardize dichotomous variables. Just do as you've indicated with the multiplication.
@@Gaskination thank you so much!
Dear Dr James Gaskin, happy new year! Hope this year would be even a better one for you and your loved ones.
Quick question. You mentioned that alpha 0.1 can still be considered for interaction terms as it is usually weaker than regular bivariate relationship. May I know the reference for that? Or do you have any other source to validate that? Thank you, and looking forward to hearing from you.
Alvin Hadiono
University of Glasgow
Hi Alvin, The most current school of thought on this topic is that using p-values for interaction effects is not useful. Instead, experts suggest a floodlight analysis, achieved through a Johnson-Neyman plot, like in this video: ruclips.net/video/Rb_lAmjHd8s/видео.html
@@Gaskination Thanks Dr James!!!!!!!
Hi James,
Thank you for the videos. They are very useful. I have one question about the interaction effect. The analysis results show that the causal relationship between AtypicalUse and Usefulness is not significant. Can we conclude that the moderation effect of Experience on AtypicalUse - Usefulness causal relationship is not significant and so the hypethesis is not supported?
Dear Dr. Gaskin. Thank you always for your support. My question on this video is for the intercept value, 3. Is that because you used a 5-likert scale to measure the variables and the median value from 1-5 is 3?? If I used 7 likert scale, then do I have to change it to 4 instead?
It is just to place the lines in the center of the graph. The important part of the graph to interpret is the differing slopes and relative amplitudes (along the y-axis).
Hi James,
Thank you so much for all of your videos - they are amazingly helpful.
I have used your "myindirecteffects" estimand to look at a multiple mediation model within an APIM framework (with distinguishable dyads) and I have a couple of questions:
1- If I examine the direct effects in the model (e.g., same actor effect in each member of the dyad) and find that they are not significantly different, it is possible for me to incorporate this information when using your estimand? (i.e., if one of the 2 equivalent paths is path a of the mediation I am testing)
2- Do you know of a research paper that has given credit for using your estimand (and might be a good example of how to write up these analyses?
Thank you so much again!
Best,
~Roanne Millman
1. I’m not sure I understand, but it sounds like you are saying that you have two groups, and for one of the groups, one of the paths (A or B) is not significant. Traditionally (e.g., Baron and Kenny), we would say that both paths must be significant in order to consider mediation. However, I’m not sure if the more recent literature allows mediation simply if the indirect effect is significant (ignoring the A and B paths on their own). Some of the most recent literature is cited here: statwiki.kolobkreations.com/index.php?title=References#Mediation
2. As for a study using my estimand approach that is published, I’m not sure. It usually takes some time to get papers published, and I introduced the estimand only last year. However, the approach simply takes advantage of bootstrapping to generate standard errors and confidence intervals for indirect effects. This is essentially what the most recent literature recommends anyway. Even though that literature uses other tools, it is the same approach.
Hi doctor,
is it acceptable to have a correlated structural error term between variables??
lets say i have a model consisting from A,B, and C variables
A act as Exogenous variable
B act as Endogenous variable for A but at the same time its act as Exogenous for C.
C act as Endogenous variable
Note : B is not acting as a mediator due some theoretical limitations.
In the model you have discussed in this video the nearest scenario for my case as following:
Is it acceptable to have a correlated structural error terms between Useful and InfoAcq ??
Your help is really appreciated Dr. James
Hi James,
Thank you very much for the video, it is very clear and helpful! Just one quick question. I wonder if we need to mean center the IV, moderator and the interaction term in path analysis?
I realize that when we need the simple slope analysis to further interpret the interaction relationship at each level of the IV, we may need to mean center the variables.
The standardized coefficients will be the same either way. For the interpretation of the simple slopes analysis, you can use either, as long as you account for the scale size when making your interpretation.
Dear Dr. James,
A question:
How do we interpret the interaction effect with second order factor?
For ex: SYQ and IFQ (1st order) create second order factor with the name SIQ.
Then, I tested the interaction and found that SIQ (2nd order) is significant. If I test the 1st order factor interaction, SYQ is significant, but IFQ is not significant.
Which one should I write for my report? SYQ (1st order factor) or SIQ (2nd order factor) ?
Also, what should I put for the 2-way-interaction from excel (Stats Tools Package)? the 1st order or 2nd order?
*Note: the theory said that 2nd order or 1st order is acceptable. I can use SYQ (1st order) or SIQ (2nd order).
However, my hypothesis is using SYQ (1st order), not SIQ (2nd order).
Your test should follow your hypothesis. So, I would recommend using SYQ, not SIQ.
@@Gaskination understand. I will try to write it based on hyphothesis.
Dr. Gaskin thanks so much for your very helpfull videos! In my path analysis I find CMIN/DF=1.697 GFI=.991 RMR=.071 CFI= .995 PCLOSE= .391 and RMSEA=.053
Can I assume that I have a good model to support my hypothesis? My sample is 250 cases. Thank's in advance.
Dio Trikilis That looks great.
Hi James, many thanks for your useful lesson. Can I ask, when you mention that a 90% confidence level is fine for interactions, do you happen to know some sources that explain it? It would help me a lot, been looking around but couldn't find any. Many thanks!
Here are some references on moderation: statwiki.gaskination.com/index.php?title=References#Moderation_and_Multigroup
I'm not sure if any of them talk about a more liberal confidence level. You can just argue that interactions are generally more exploratory than confirmatory, so a more liberal confidence level is merited.
@@Gaskination Amazing, many thanks James!
Dear Dr. Gaskin,
for moderated mediations, I tried out your ModMedEstimands plug-in - and I love it. Is it flawed (the code looks straightforward to me) or why is there no reference to it in your videos or elsewhere?
Thanks a lot for the great help and inspiration you provide.
Best,
Leo
It works just fine I think. I haven't made a video about it yet though. All it does is calculate the difference between two indirect effects and creates confidence intervals and p-values for that difference.
Thank you! It is a very nice time-saver!
Thank you for this, Dr Gaskin. If my hypothesis says that a high level of the moderator affects the relationship between the independent and the dependent, can I only test the moderation at a high level of the moderator (-1SD)?
It simply refers to the direction of the moderation. The plotter should interpret it for you, something like: "M strengthens the positive effect of X on Y"
@@Gaskination Thank you for your reply. The problem is that the moderation is insignificant. However, I am only interested in the moderation at high levels of the moderator! (the theory assumes that high levels of the moderator affect the relationship... not all levels).
@@mohmoh7078 Ah, ignore the significance. To assess the moderation at high levels of the moderator, you can us the interaction plotter, as it plots three lines, one for mean, and one for -1 SD and for +1 SD. So, this should tell you if moderation occurs at high levels of the moderator. Another option would be to split the moderator into low/high groups and then conduct a multigroup analysis.
@@Gaskination Many thanks! So, If the moderation at high levels were significant, then my hypos are supported regardless of the insignificance of the moderation at mean levels? (since I am concerned only about high moderation)
@@mohmoh7078 It's all about the slopes. If the slopes are parallel, then there is no moderation. If the slopes diverge or converge, then their is moderation.
Hi Dr. Gaskin, I found this technique is pretty useful for testing the moderation effects among the observed exogenous variables. However, what if I want to test the moderation effects between latent variables, should I have to use the factor score from my CFA instead to compute the X*Y(they are both latent variables)? Thanks in advance!
You can keep them latent if you multiply every pair of indicators across the IV and moderator. Here is an example (but with smartPLS): ruclips.net/video/upEf1brVvXQ/видео.html I also show how to do it in AMOS during my most recent SEM Boot Camp videos. It is on day three: ruclips.net/video/ViQITM1htjU/видео.html
Hi James, thank you for the whole SEM Series! It is awesome. Yet I struggle a lot with writing down my results in my thesis. With the interaction, I can't use the Stats tool, as I have to do all calculations and plots by myself. How can I plot the two way interaction using the regression coefficients? What if, only the moderator and the interaction term are statistically significant but the relationship of IV on Y is not? This is the case for my analysis (I have stimuli situations as the IV and the M - the IV would be that respondents in the survey saw an affective ad claim and others a cognitive ad claim and the moderator is, whether respondents see a speech of Greta Thunberg or not). So for me it doesn't make sense to standardize them. Can I use the regression coefficients of the normal AMOS latent regression with the two dichotomous variables as observed variables? Thanks in advance for any help!
Also, I wonder why you can include the coefficients of the non-significant relationships (AtypUse --> Useful). Do you have a suggestion for any book where this procedure is explained? So far unfortunately I have not found any sufficient source...
@@CarinanaLaa
1. Why would you have to do the calculations yourself? That seems like a waste of effort. That's like continuing to walk to campus rather than riding the bus. For interactions, you can look at the formulas in the stats tools package. I don't think they are restricted. You just have to find them below the line. I think I made the text invisible... but, you can still see the formula in the formula bar at the top.
2. If the IV to DV direct effect is not significant, but the interaction and moderator are, then this is very interesting. This means that the effect from IV to DV is only meaningful when considering the moderating effect. That's great.
3. You are correct that standardizing a binary variable isn't very useful. But, if it is binary, then just use multigroup analysis instead of interaction. Here is a video on how to plot MGA using google: ruclips.net/video/ZnptsnQ2VT4/видео.html
4. Retaining non-significant relationships is wise. If an effect is theorized, it should be included in the final model, even if not significant. This way you can account for the nominal effect it has when you are testing other hypotheses. No citation needed for including a theorized nonsignificant effect.
@@Gaskination Thank you very much for the fast reply! I know, its stupid, but I have to show in the appendix in my thesis how I calculated it or at least my source of calculations which needs to be a book or a paper. I can't use multigroup analysis as the sample size is definitely too small for that. For the non-significant effect I meant, that I thought in die Interaction-Plot it might need to be put to 0 as it has no significant effect. Do you have a source for the calculations of the plot?
@@CarinanaLaa The calculation of the plot is just a regression equation. y=mx+b
Thanks for posing all these videos, very helpful. I have a question: What if the path between the mediator variable (M) and the dependent variable (Y) is moderated by a factor (W)? There is no factor moderating the path from independent variable (X) to M. When doing moderation test (after testing mediation effect), shall I just add a path from W*M to Y and W to Y, co-variate X, W*M and M and then run the analysis?
Thanks a lot.
yes, except AMOS will not let you covary an endogenous variable (M). So, instead, you'll have to covary the error of M with the others.
Thanks a lot James.
Sorry James I have another question. How to decide the value of the intercept/constant when plotting moderation effect? Thank you.
Just use zero. The intercept in the plotter is just the vertical position.
OK thanks James.
Dear Sir,
Like others I also struggle to cite the methods suggested by you and your videos. The approach is unclear. For example, Preacher, Rucker, and Hayes (2007) suggests testing two regression equations, the “mediator model” and the “dependent variable model” to test for moderated mediation. The ppt provided on statwiki suggests Preacher and Hayes paper (link not working though). Assuming Preacher and Hayes (2008) paper, it talks about multiple mediators..
Not asking you to spoon feed us but requesting to clarify the approach you are using.. Rest readings are on us.... but as of now it is unclear what to read to refer the methods used by you.
Regards
For interactions, I'm just using a standard approach from any textbook, like Hair et al 2010. In this approach, you standardize the IV and moderator and then multiply them. For mediation, my most current videos use Zhao et al 2010 (reconsidering baron and kenny) - although there is a newer approach that I need to make a video for. For moderated mediation, whether it is indirect effects across multiple groups or whether it is the mediated interaction effect, I think Preacher and Hayes cover all forms.
Hi Prof Gaskin,
First of all, thank you very much for the video.
I am having difficulties following the variables which previously have a lot of indicators now only having one indicator per LV. How do you process those indicators into one LV? Did I miss an important step?
How to moderate 2nd order variable instead?
Thank you for response
Here is the missing link: ruclips.net/video/dsOS9tQjxW8/видео.html
If you use this linked video to create factor scores for the 2nd order factor, the process is the same.
@@Gaskination Thank you very much. The series is very helpful.
Maybe could run the same data set and model in other CB-SEM such as Lisrel for comparison sake? Just my 2 cents.
@@arifnugroho73 I've never used LISREL. I've thought of learning some others though. Thanks!
@@Gaskination Thank you very much Prof. The videos are a life saver for me.
Hi Gaskin,
what does this interaction method called? is it two-stage approach to interaction as suggested by Kenny and Judd (1984)? Would you please refer to any study that used this method?
Thanks
Correct. It's just the two-stage approach. Here are some papers: statwiki.gaskination.com/index.php?title=References#Moderation_and_Multigroup
what if we don't take the zscore, just run the model with mean values of construct?
Just make sure to check multicollinearity if you do not use zscore. Sometimes using just the mean value results in very high multicollinearity.
Hello James..here I am again to thank you for such saviour videos and impose another query as follows
In my A-B-C type of model where I have two Moderators M1 and M2 at A-B and B-C level resp
..how do I test interaction M2 at B-C level..?
In A-B level interaction M1 we covary all exogenous variables including variables at A level, control variables and interaction variables..
But what do we do to test M2 interaction in the second part (B-C) of the model..
Can I covary M1 with A and control variables, and M2 with only control variables and proceed with the interaction analysis
I tried to look for a solution at many places before asking you but couldn't find anything..pls help
I don't know of an article to support the approach, but, practically, you can just covary the error of the mediator with the interaction and moderator.
Ok thanks a lot..I'll do that..n look for supportive work myself..All I wanted to know was the process which would make sense
so only moderator variables and the interaction effects uses the standardized?
Correct. Only the variables used directly in the multiplication need to be standardized.
Hi James,
Thank you for your videos. they're really helpful in understanding SEM. However I do have one question about this moderation technique in AMOS. Currently my moderator is a latent variable that is loaded by 9 different subscales and is not represented by a total composite score. How should I go about running moderation analysis in this scenario?
Much thanks,
Melvin
If you want to do interactions, it is very tricky with latent variables. You would have to cross multiply all items from the IV and the moderator. Then each of these new product items would become the indicators for a latent interaction.
James Gaskin hi dr. Can you upload a Video showing interaction among latent variables? Than you very much
Just a basic question, should we must have a variable that is categorical in nature to act as a moderator or we can use interval data like that in social support variable or any other such variable can be used as moderator in interaction based moderation?
Ahmad Usman for interactions, just best to have non-categorical variables as shown in this video. When you have categorical variables, it is better to do a multiple groups analysis for your moderation.
Hi James, Thank you for this video. ... Can the relationship (M-DV) between the Moderator variable (M) and the Dependent Variable (DV) be significant, given that the Interaction effect (MxID) is significant on the DV?
It is possible, yes.
Dear Professor, is there any video of yours for testing moderation in a full latent model? Thank you
Unfortunately there is not. In AMOS, I always use a path model with composites when testing interactions. It is just way simpler. However, when I do interactions in SmartPLS, I use a latent model. Here is a video of that: ruclips.net/video/PnPfOGtl-lc/видео.html
thanks a lot for the stats tool package on the website. I have a question could I make moderator test for a variable like "gender" using amos
Yes. See the next video in this playlist. It's about multigroup analysis.
Hello Doctor James Gaskin,
If p-values of dependent v
If you are using interactions, and the effect of the interaction is not significant, then it means there is no moderating effect.
Thank you, doctor.
Hello Dr. Gaskin, a kind request can you please suggest an empirical paper of interaction (moderation) conducted in SEM (AMOS) for reference, as I am not able get any paper on interaction (moderation) that was conducted in AMOS.
When all else fails, you can cite Barbara Byrne (she wrote the AMOS guide book).
Thank you
www.researchgate.net/post/How_to_test_moderation_effect_in_AMOS
Hi James. Thank you once again for this amazing video. For my paper I want to test the impact of Institutional factors on Innovation.
I have done EFA in SPSS and found institutional variables that fit in two factors. According to the literature I have named them Formal Institutions and Informal Institutions.
As per your advice I run Cronbarch Alpha and it is higher than 0.7
Now I want to save the factors so I can use them in AMOS SEM model and check interaction effect as well but I am not sure if I should save factors like this in SPSS or use AMOS Data Imputation option to save the factors.
I tried to do both and compare the result of factor saving and data imputation but I did not get similiar results.
In this case I do not know which option I should use and I would be extremely thankful if you can help me!
Lena
+lena kadriu The results should be similar if the EFA and CFA were the same models. Regardless, the CFA should provide more accurate factor scores.
Hi James,
How can we standardise the value of multi item Likert Scale data to run the moderation? Can we impute it's value from Amos analyze option.
Yes, you can impute first. That is the simplest method for interactions.
@@Gaskination So after imputing, then we can go for standardising the variables in SPSS (Analyse - descriptive statistics - Descriptive). Am I right?
@@AshutoshPandeymentor Yes, if they're not already standardized or mean-centered. Check their distribution just to be sure. You don't want to standardize twice.
Hello Dr Gaskin. How do I measure moderator variable that is a multi dimensional construct measure on a 5 Likert scale? Thank you
If you mean you want to use an ordinal (likert scale) measure as a moderator for grouping, then you can simply split the data into two groups where 1 and 2 are part of the low group, and 4 and 5 are part of the high group (and you remove 3). Or, if you want to do an interaction, you can create a factor score as shown in this video above, and then standardize and multiply with an IV to create an interaction variable. If you mean that your variable is 2nd order, then you can still create a factor score for it in AMOS during the CFA.
Hi Gaskin, I am wondering what if I have a mediated moderation model as follows X: independent var Z: Moderator var and XZ: the interaction M: mediator and Y: dependent variable. I found the interaction (XZ) significantly related to the mediator (M) but the independent (X) and the moderator (Z) were not significantly related to the mediator (M). The moderator (Z) and Mediator (M) positively related to dependent variable (Y). Is the interaction effect/the moderation still exist and how to interpret it? many thanks
This is very interesting. This means the IV and moderator are not sufficient alone, and are only good predictors when combined. This is the best kind of moderation result.
Wow..Thanks a lot, James..I am very excited right now. Happy new year
hey james. how is the interaction between atyp and experience significant when the p value is less that 0.05?
+miracle ozzoude I think you mean greater than 0.05. In this case, I am accepting effects at the 90% confidence level, so a p-value less than 0.100 is fine enough.
Hi James, is the 90% confidence level the mostly accepted (compared to 95% level) standard? Is there any requirement to sample size if I use the 90% level not the 95% level? Thanks.
90% is often used in exploratory research, but 95% is definitely the more widely used threshold.
wow... you made it so easy. Thank you James.
Hi,
Can you please guide how can i download the plugin for the moderation graph you used in this video?
it is on the homepage of statwiki
hello James. .07 is a significant interaction? I kinda got confused
It depends on your desired alpha threshold. If you're okay with 90% confidence in the estimates, then a p-value less than 0.100 is fine. If you require 95% confidence, then p-values must be less than 0.05. For interactions, I'm usually a little more lenient and choose 90% confidence.
@@Gaskination Hi James. Thanks for the Video, your explanations are amazing! Do you have a paper to cite for the fact that 90% confidence level is enough for moderators? I don't find any...
@@m__h5506 No need to cite your choice to accept a different confidence interval. This is just an analytical choice that depends on your research goals.
@@Gaskination Alright, thank you! :)
I need an urgent suggestion: Why you are not including items of study variables? Can we do simply adding the average of each variable like this? Please help.
In a previous video in this series, I imputed the factor scores for the latent factors. This simplifies interaction testing.
@@Gaskination Sir can you please share the link where you talked about factor scores?? Also, can you please help me. In my research work I have one IV, two DVs, and with two moderators. Suppose one IV has 6items, can I take the average of all the six items and name it as AVGIV and then run the full model??
Similarly with other variables also?
@@mantashafiroz4431 ruclips.net/video/dsOS9tQjxW8/видео.html Yes, you can use averages, but factor scores are more valid.
@@Gaskination Thank you so much. It really means a lot.
@@Gaskination Sir, can you please provide me the reference where it's mentioned that factor scores are valid. It will be a great help.
hi, i am struck in feedback loop, my model is causing lot many issues, can you please help me out ?
Don't include feedback loops. AMOS does not work well with feedback loops.
Thanks James ...
Hi James,
hi
@@Gaskination Hi James, Sorry for this, I sent you a message mas for some reason the text didn't go.
Thanks a lot for your videos and explanations! They help a lot. Would you be so kind to help me a little further? Doing my analysis, the following situation came up:
Results after testing the causal latent model:
H1:A-->C (expected negative) - significant (pC (expected positive) - not significant (p=0,07) and negative (beta=-0,065), not supported
Results after testing moderation (with composite variables) of B on the relation A-->C (H3: moderation should be negative, e.g. B expected to dampen the negative effect from A-->C):
A-->C - not significant (p=0,75) and positive (beta=0,019)
B-->C - significant (pC - significant (p=0,013) and positive (beta=0,139)
I understand that being SEM an interrelated network of effects, adding or removing variables can change the overall effects, but in this case the relations A-->C and B-->C generally went opposite! So my doubts are:
1. Following moderation test, I would say H3 was not supported (when plotting, it says "B strengthens the positive relation between A and B"), am I correct?
2. Do my conclusions about H1 and H2 after testing the causal latent model have to change due to the corresponding estimates obtained in the moderation test?
Thanks a lot!
Ligia
@@ligianascimento1923 Ha! That's rough. Generally speaking, latent models are more reliable and produce better estimates. With interactions, latent models are unwieldy, so we often use composites as you have. The trouble is that these often alter the estimates too much. So, in your case, I would recommend to either use a different approach, such as the J-N floodlight approach: ruclips.net/video/Rb_lAmjHd8s/видео.html or a different software, such as Mplus: ruclips.net/video/lmMbEI_Kw2I/видео.html or SmartPLS: ruclips.net/video/PnPfOGtl-lc/видео.html. Mplus and SmartPLS allow latent interaction analysis.
@@Gaskination Hi James, Thanks a lot. I will follow you recommendations. Kind regards