Hello sir, I've been following your videos for more than a year to learn AMOS. I deeply appreciate them! I have recently encountered something that baffles me. When I do a path analysis in AMOS, the direct effect between A and B is nonsignificant (in "estimates", the p-value of "A --> B" is larger than .05). However, when I do linear regression in SPSS with A being the independent and B being the dependent, the result is significant (p < .05). And this happened to several other pairs of variables in my path. How could this happen? Which one is correct? Thank you in advance for your help!
I have a moderates Mediation like process number 10 But with 3 moderators. Problem is my Model is identified But Not over identified, so i dont get results on the Model fit. What can I do?
When is it advisable to include or remove the intercepts from the model specification. I am using JMP SEM fit. The most obvious effects of not including the intercepts is much higher R2, loss of some indirect effects (which would really make the results easier to interpret), and too many deep red squares (which do not show when intercepts are included) in the normalized heat unit map. Your advice would be greatly appreciated. There are obvious tradeoffs.
what is the easiest way to identify an exogenous variable in path analysis? Because sometimes,l look confuse in identifying the right variable for exogenous in path analysis.
Dear Dr. if you have videos about moderated mediation model with control using Amos, please send me the link. and how can we know CMB in AMOS except Herman's single factor method?
If modification indices suggest direct effects of both variable A predicted variable B, and variable B predicting variable A...how would you do revise the model? Do you allow for correlation (i.e. double headded arrow)? Or would this suggest adding a latent variable?
Hi Fabricio, I'm not sure what you are meaning by "spillover effect". Basically in mediation analysis, you are testing whether the effect of an exogenous variable on an endogenous variable is 'transmitted' by way of an intervening (mediating) variable. Direct effects propose that the effect of X on Y is not mediated. I don't know if this answers your question (I hope it does). Anyway, I do have a more recent video on mediation analysis (where we are testing direct and indirect effects using AMOS) here: ruclips.net/video/gYE4yIjfFIA/видео.html I hope you check it out! Cheers!
Thanks Mike Crowson, very informative video. I just want to ask that if I understand correctly, Model Fit is not necessary when doing a mediation analysis in SEM , right?
Thanks Mike for the amazing video. I have questions regarding the chi2 value that you got there. your chi1 value is 4.415 and p value is 0.036 which means that your model is differ significantly from the saturated/just identified model. Do you think that this can be a problem? If I am correct, it means that the structure that we make is not "correct". I see others also said that the model is acceptable if the p value is above 0.05. Could you give some insights please? thanks.
Do you have a way to do this with serial mediation (2 mediators)? I can't get the 3 specific indirect effects :( (And can't do it in PROCESS using Model 6 because one mediator is dichotomous and it doesn't like that).
Hi Jane, there is a way of obtaining specific indirect effects via AMOS by using the Define Estimands approach. Go here for examples (ruclips.net/video/9jGL45NuVAA/видео.html). The bigger problem, however, is that your mediator is dichotomous. Process can handle a dichotomous Y, but not a dichotomous mediator. About the only thing I can suggest is to use a structural equation modeling approach that would allow you to account for the categorical variable. For example, the R program 'lavaan' would allow you to do this. See (lavaan.ugent.be/tutorial/cat.html) and (ruclips.net/video/nx7PYvczXWg/видео.html). Mplus also has the for analyzing your data and obtaining the specific indirect effects you are seeking. Hope this helps.
@@mikecrowson2462 Hello, can we also do that using AMOS. I think you have a next video showing for mediation for categorical variable in AMOS using baysein distribution.
Dear Mike, I have problem with latent variable as the value squared multiple correlation is zero. So I do not know why that. and how to interpret it? Please give me suggestion for this matter.
If your latent variable is an endogenous variable (i.e., being predicted by one or more other variables) in your model, then the R-square value is the proportion of variation accounted for by its predictors. An R-square=0 indicates that its predictors account for 0% of the variance. I don't usually run across R-square values where they are exactly 0. One very basic question I have is whether you have your decimal (rounding) settings. If you are rounding to the nearest whole number then a very low non-zero value may be rounded to 0 in your output. Chances are this is not what you have (but I like to rule out this possibility since I don't have your output in front of me). Another question: Are you obtaining estimates for all estimated parameters in the model? Including the coefficients for the structural portion? If not, then that could be a source of your "problem". You may have some type of model identification problem. But if this is the case, I'd expect an R-square not to be produced at all. But I'll be honest. I've never paid close attention to R-square in Amos when I encounter identification problems. I would expect that non-significant predictors could translate into a very low (near zero) R-square value. If you are finding significant predictors and R-square at 0, then this would be a curious situation indeed. These are just a few thoughts I had in response to your question.
Thanks Mike for your amazing videos. Please, I need your help. My model is a latent model. i follow the steps you used here in this video to test mediation relationship in my model. what I should write in my paper about this method (what is its name)? Also, is there any reference I can cite for this method?
Why model is not fit in SEM after run on imputed data_c. I have to MI value in My case, Two IVs, one mediator and One DV. OR Let me know if I want to check mediating role after Run CFA, in the first step IVs To DV, Second step. IVs to MV, and in third step MV to DV (here i have to include IVs or not?)
Hi Michael, how would you report this in your methods and results? Do you just report that for example total, or indirect effects were significant at the 0.05 level, or should you report how you found that each effect was significant or not using bootstrapping etc. Is there a reference for this? Thank you.
If I was writing this up, I might say something along the lines (at some point) of: "...I utilized the bootstrapping module in AMOS to generate standard errors and 95% confidence intervals to test the indirect effects and total effects for statistical significance." Then you could report on effects, along with the significance test results. Hope this helps.
Thank you very much. Can the bootstrapping method be used for multiple mediation models where mediators are either in series, parallel or both? (Preacher and Hayes 2008).
Thanks for asking this great question! The short answer is that AMOS can handle these kinds of models, but it takes a bit of finesse. To provide you with a couple of strategies for breaking down the total indirect effect into separate indirect effects via separate mediators, I created a video on this topic. I discuss two options. One is using Andrew Hayes' Process add-in (www.processmacro.org/index.html) and the other covers an approach using AMOS that involves a series of steps involving fixing and freeing various parameters. You can go: ruclips.net/video/gYEUvpaqwH8/видео.html
Hi, thank you for your video. Very useful. I am facing a problem because I have a model where there is only one path between 2 variables (no way of indirect effect) and my direct effect is significant (p
Hi Renata, thanks for your post. I'm not really clear on what model you are running. Are you basically running a simple or multiple regression? (meaning there are no mediators between your X and Y variables? If not, then the direct effect between X and Y should equal the total effect and any significance tests should be the same as well.
Yeah, sorry I didn´t explained myself quite well. It is a cross-lagged with 3 waves and two variables. The analysis in AMOS is completely crazy: the regression coefficient is significant at .02 but the direct effect (it was supposed to be the same) with bootstrapping analysis at 95% is not significant therefore the total effect it is not significant as well. The analysis in R seems to me more reasonable but I am still doubting about the way AMOS calculate the direct effect and the total effect. I can´t paste any image of the model in AMOS, so here the model in R: #Regressions CE3 ~a* CE1 CE5 ~b* CE3 TM3 ~c* TM1 TM3 ~d* CE1 CE3 ~e* TM1 TM5 ~f* TM3 TM5 ~g* CE3 CE5 ~h* TM3 #indirect effects EICE1aCE5 := a*b EITM1aTMT5 := c*f EICE1aTM5 := a*g + d*f EITM1aCE5 := b*d + c*f #total effects ETCE1aCE5 := a*b + d*h + b+h ETTM1aTM5 := c*f+e*g+g+f ETTM1aCE5 := e*b + c*h + b+h ETCE1aTM5 := d*f + a*g+ f+g ETTM1aCE5 := e*b+ c*h + b+h #Covariances CEEMO2DIM1 ~~ TOTAL_TM_VAL1 CEEMO2DIM3 ~~ TMTOTALVAL3 CEEMO2DIM5 ~~ TMTOTALVAL5 Thank you so much for your support.
Hi there. Ok, it sounds like you are comparing the normal theory significance tests from your "estimates" table against the significance test of the direct effect using bootstrapped standard errors. If you click on estimates and then look at the boostrap standard errors (and 95% CI for the direct effect that are impacted by the standard error) for your regression coefficients, then I suspect you will see that the direct effect and total effect will be the same (including the bootstrapped standard error and confidence interval). It's totally possible to arrive at different conclusions about the significance of an effect based on whether you are using a normal theory method (which standard ML relies on) or using an empirical re-sampling strategy (e.g., bootstrap). I could see how this might occur if your data are non-normal. One effect of non-normality on ML estimates is that it results is that it results in smaller standard errors (resulting in an increased Type 1 error probability; here's a good discussion of this: web.pdx.edu/~newsomj/semclass/ho_estimate2.pdf). Assuming that bootstrapping gets around that problem, then I could see that you might have a difference in conclusions about statistical significance for your path, where one approach might yield a conclusion of significance and the other does not. I don't know the nuances of your data, but this seems like a reasonable possibility for what you are seeing.
Dear MIke, thank you very much!! I really appreciate your comments. And yes, I didn´t realise that...and now what gives us the best interpretations (the one that can fit better the "reality") the normal theory or bootstrapping? I guess this has not an easy answer!!! :) :) Once more, thank you for your attention.
Hello. When you have a just-identified model (this will occur when the df=0) then you cannot rely on traditional fit indices such as chi-square gof, chi-square/df, TLI, CFI, RMSEA, etc to evaluate overall model fit. In cases such as this, the best you can do is examine model fit in terms of its individual components (i.e., path coefficients and indirect effects). Hope this helps! best wishes!
Hi, Mike, thank you very much for your video. I have a little question about evaluating indirect effect using Amos, that how to compare the significant differences between two indirect effect in different condition using Amos, my reviewer told me to use nest-model comparison approach, but I'm not sure how to conduct it in Amos. I would be much appreciated if you can answer my question. thanks again.
Hi, great video. I have a question. If my mediation variables are continuous but my independent variable is categorical, can I use the method that you talked about in the video, or do I need to do any adjustments? Thank you!
Hi there. If you have a categorical independent variable, you can still run the mediation analysis. However, you'll need to recode it (e.g., dummy code) into binary predictors. If your IV only has 2 groups, then you can run the mediation analysis with that variable as IV. If you have an IV with 3 groups, you'll need 2 dummy variables, which is effectively a re-packaging of the original IV into those new variables. In regression analysis, the relationship between each dummy variable in a model (reflecting a single categorical IV) represents the same thing as an orthogonal contrast (ie., a test of the difference in a comparison group and reference group on the mediator). Extending that logic, if you have more than one dummy exogenous variable and you are predicting the mediating variable, then the indirect effect being computed per dummy variable will represent part of the total mediated effect between the original IV and the DV. FYI, some folks might complain that this violates the assumption of 'multivariate normality'. However, this assumption really only pertains to the endogenous variables in your model (Rex Kline, 2015, provides a good discussion of this assumption; see www.guilford.com/books/Principles-and-Practice-of-Structural-Equation-Modeling/Rex-Kline/9781462523344). I hope you find this helpful. Best wishes!
Hi Samra, I'm not sure what exactly your question pertains to. In this illustration, I'm demonstrating evaluation of direct, indirect, and total effects in a path model with measured/observed variables. Typically, the term "indicator variables" refer to measured variables that serve as "indicators" of latent variables in SEM models, such as CFA models and path models using latent variables. Technically, the measured variables in path models - such as the one shown in the video where there are no latent variables being modeled - are indicators of latent variables too; but are only single indicators of underlying concepts. Moreover, in models such as these we do not account for measurement error when estimating path coefficients. I hope this is helpful.
The video only focused on using measured variables. You can easily extend the logic to running tests of indirect effects using path analysis involving latent variables. Basically, start with CFA model to test the fit of your measurement model. When that model appears to fit the data well, then you can move to testing the proposed structural relationships among your latent variables. You evaluate the overall fit of the model and can also use the procedures shown in this video to test for the direct and indirect effects in your latent variable model.
Hey Mike, really helpful tutorial so far! Can you tell me if there is a way to calculate effect sizes for the direct effects in the model? Thanks a lot!
Hi Mike, thank you for a very informative explanation. I have a question for you. Just say in the model there are two exogenous variables (X1 and X2), with both X1 and X2 directly effecting Z and indirectly effecting Z through Y. How would you calculate the total effect of both X1 AND X2 on Y? (je can you not only calculate the total effect of X1 and the total effect of X2 on Y, but the Total effects of both X1 AND X2 on Y?
Hi there. Thanks for your question. It is an interesting one. As far as I know, you are computing indirect, direct, and total effects using a single X variable at a time. It seems you are asking about some type of joint indirect effect of multiple X's and Y, which is not something I've ever heard of. best wishes
the best video series of SEM and AMOS ever!
Thank you so much for your video! I would like to know how can I know the significance of the total effects?
Hello sir, I've been following your videos for more than a year to learn AMOS. I deeply appreciate them! I have recently encountered something that baffles me. When I do a path analysis in AMOS, the direct effect between A and B is nonsignificant (in "estimates", the p-value of "A --> B" is larger than .05). However, when I do linear regression in SPSS with A being the independent and B being the dependent, the result is significant (p < .05). And this happened to several other pairs of variables in my path. How could this happen? Which one is correct? Thank you in advance for your help!
I have a moderates Mediation like process number 10 But with 3 moderators. Problem is my Model is identified But Not over identified, so i dont get results on the Model fit. What can I do?
When is it advisable to include or remove the intercepts from the model specification. I am using JMP SEM fit. The most obvious effects of not including the intercepts is much higher R2, loss of some indirect effects (which would really make the results easier to interpret), and too many deep red squares (which do not show when intercepts are included) in the normalized heat unit map. Your advice would be greatly appreciated. There are obvious tradeoffs.
Great tutorial. Can anyone help me with how to write the model M1 and M2 at 1:52 in a form of an equation?
Hello, thanks for the video. Please, can I know what is the name of this method?
what is the easiest way to identify an exogenous variable in path analysis?
Because sometimes,l look confuse in identifying the right variable for exogenous in path analysis.
Thank you Mike for your detailed explanation with example.
I appreciate you watching!
Dear Dr. if you have videos about moderated mediation model with control using Amos, please send me the link. and how can we know CMB in AMOS except Herman's single factor method?
Thanks teacher. It is very important and useful.😊😊😊
Hey, why does the number of distinct sample moment differ?
If modification indices suggest direct effects of both variable A predicted variable B, and variable B predicting variable A...how would you do revise the model? Do you allow for correlation (i.e. double headded arrow)? Or would this suggest adding a latent variable?
How can I get AMOa Application Deara
Hello, is it possible to have insignificant direct and indirect effects but significant total effects? what does it mean?
Hi Mike Crowson, great video. Question: Direct & Indirect effects are like the "spillover effect"? Thanks
Hi Fabricio, I'm not sure what you are meaning by "spillover effect". Basically in mediation analysis, you are testing whether the effect of an exogenous variable on an endogenous variable is 'transmitted' by way of an intervening (mediating) variable. Direct effects propose that the effect of X on Y is not mediated. I don't know if this answers your question (I hope it does).
Anyway, I do have a more recent video on mediation analysis (where we are testing direct and indirect effects using AMOS) here: ruclips.net/video/gYE4yIjfFIA/видео.html
I hope you check it out! Cheers!
Hello sir, for the mediation value in the bootstrap P-value = 0.021.
Lower bound=0.351
Upper bound=1.251.
Is this significant or not?
Thanks Mike Crowson, very informative video. I just want to ask that if I understand correctly, Model Fit is not necessary when doing a mediation analysis in SEM , right?
Thanks Mike for the amazing video. I have questions regarding the chi2 value that you got there. your chi1 value is 4.415 and p value is 0.036 which means that your model is differ significantly from the saturated/just identified model. Do you think that this can be a problem? If I am correct, it means that the structure that we make is not "correct". I see others also said that the model is acceptable if the p value is above 0.05. Could you give some insights please? thanks.
Do you have a way to do this with serial mediation (2 mediators)? I can't get the 3 specific indirect effects :( (And can't do it in PROCESS using Model 6 because one mediator is dichotomous and it doesn't like that).
Hi Jane, there is a way of obtaining specific indirect effects via AMOS by using the Define Estimands approach. Go here for examples (ruclips.net/video/9jGL45NuVAA/видео.html). The bigger problem, however, is that your mediator is dichotomous. Process can handle a dichotomous Y, but not a dichotomous mediator. About the only thing I can suggest is to use a structural equation modeling approach that would allow you to account for the categorical variable. For example, the R program 'lavaan' would allow you to do this. See (lavaan.ugent.be/tutorial/cat.html) and (ruclips.net/video/nx7PYvczXWg/видео.html). Mplus also has the for analyzing your data and obtaining the specific indirect effects you are seeking. Hope this helps.
@@mikecrowson2462 Hello, can we also do that using AMOS. I think you have a next video showing for mediation for categorical variable in AMOS using baysein distribution.
Just amazing and specific. I like your tutorials. Mike, can you teach us Mplus path analysis? I need your suggestion
Than you very much dr Michael Crowson
Thanks for watching!
Dear Mike, I have problem with latent variable as the value squared multiple correlation is zero. So I do not know why that. and how to interpret it? Please give me suggestion for this matter.
If your latent variable is an endogenous variable (i.e., being predicted by one or more other variables) in your model, then the R-square value is the proportion of variation accounted for by its predictors. An R-square=0 indicates that its predictors account for 0% of the variance. I don't usually run across R-square values where they are exactly 0. One very basic question I have is whether you have your decimal (rounding) settings. If you are rounding to the nearest whole number then a very low non-zero value may be rounded to 0 in your output. Chances are this is not what you have (but I like to rule out this possibility since I don't have your output in front of me). Another question: Are you obtaining estimates for all estimated parameters in the model? Including the coefficients for the structural portion? If not, then that could be a source of your "problem". You may have some type of model identification problem. But if this is the case, I'd expect an R-square not to be produced at all. But I'll be honest. I've never paid close attention to R-square in Amos when I encounter identification problems. I would expect that non-significant predictors could translate into a very low (near zero) R-square value. If you are finding significant predictors and R-square at 0, then this would be a curious situation indeed. These are just a few thoughts I had in response to your question.
Thanks Mike for your amazing videos.
Please, I need your help. My model is a latent model. i follow the steps you used here in this video to test mediation relationship in my model. what I should write in my paper about this method (what is its name)? Also, is there any reference I can cite for this method?
Why model is not fit in SEM after run on imputed data_c. I have to MI value in My case, Two IVs, one mediator and One DV.
OR
Let me know if I want to check mediating role after Run CFA, in the first step IVs To DV, Second step. IVs to MV, and in third step MV to DV (here i have to include IVs or not?)
try to run the data without imputing the data.
what do you mean by pergoal or performance goal
It is one of the variables being analyzed in the model.
Hi Michael, how would you report this in your methods and results? Do you just report that for example total, or indirect effects were significant at the 0.05 level, or should you report how you found that each effect was significant or not using bootstrapping etc. Is there a reference for this? Thank you.
If I was writing this up, I might say something along the lines (at some point) of: "...I utilized the bootstrapping module in AMOS to generate standard errors and 95% confidence intervals to test the indirect effects and total effects for statistical significance." Then you could report on effects, along with the significance test results. Hope this helps.
Thank you very much. Can the bootstrapping method be used for multiple mediation models where mediators are either in series, parallel or both? (Preacher and Hayes 2008).
Thanks for asking this great question! The short answer is that AMOS can handle these kinds of models, but it takes a bit of finesse. To provide you with a couple of strategies for breaking down the total indirect effect into separate indirect effects via separate mediators, I created a video on this topic. I discuss two options. One is using Andrew Hayes' Process add-in (www.processmacro.org/index.html) and the other covers an approach using AMOS that involves a series of steps involving fixing and freeing various parameters. You can go: ruclips.net/video/gYEUvpaqwH8/видео.html
@@mikecrowson2462 bootstrapping does not generate standard error, how can I calculate SE here?
Hi, thank you for your video. Very useful.
I am facing a problem because I have a model where there is only one path between 2 variables (no way of indirect effect) and my direct effect is significant (p
Hi Renata, thanks for your post. I'm not really clear on what model you are running. Are you basically running a simple or multiple regression? (meaning there are no mediators between your X and Y variables? If not, then the direct effect between X and Y should equal the total effect and any significance tests should be the same as well.
Yeah, sorry I didn´t explained myself quite well. It is a cross-lagged with 3 waves and two variables. The analysis in AMOS is completely crazy: the regression coefficient is significant at .02 but the direct effect (it was supposed to be the same) with bootstrapping analysis at 95% is not significant therefore the total effect it is not significant as well.
The analysis in R seems to me more reasonable but I am still doubting about the way AMOS calculate the direct effect and the total effect. I can´t paste any image of the model in AMOS, so here the model in R:
#Regressions
CE3 ~a* CE1
CE5 ~b* CE3
TM3 ~c* TM1
TM3 ~d* CE1
CE3 ~e* TM1
TM5 ~f* TM3
TM5 ~g* CE3
CE5 ~h* TM3
#indirect effects
EICE1aCE5 := a*b
EITM1aTMT5 := c*f
EICE1aTM5 := a*g + d*f
EITM1aCE5 := b*d + c*f
#total effects
ETCE1aCE5 := a*b + d*h + b+h
ETTM1aTM5 := c*f+e*g+g+f
ETTM1aCE5 := e*b + c*h + b+h
ETCE1aTM5 := d*f + a*g+ f+g
ETTM1aCE5 := e*b+ c*h + b+h
#Covariances
CEEMO2DIM1 ~~ TOTAL_TM_VAL1
CEEMO2DIM3 ~~ TMTOTALVAL3
CEEMO2DIM5 ~~ TMTOTALVAL5
Thank you so much for your support.
Sorry, the problematic value is the path g (CE3 -->TM5). in AMOS the standardized regression coefficient is .22 (p
Hi there. Ok, it sounds like you are comparing the normal theory significance tests from your "estimates" table against the significance test of the direct effect using bootstrapped standard errors. If you click on estimates and then look at the boostrap standard errors (and 95% CI for the direct effect that are impacted by the standard error) for your regression coefficients, then I suspect you will see that the direct effect and total effect will be the same (including the bootstrapped standard error and confidence interval). It's totally possible to arrive at different conclusions about the significance of an effect based on whether you are using a normal theory method (which standard ML relies on) or using an empirical re-sampling strategy (e.g., bootstrap). I could see how this might occur if your data are non-normal. One effect of non-normality on ML estimates is that it results is that it results in smaller standard errors (resulting in an increased Type 1 error probability; here's a good discussion of this: web.pdx.edu/~newsomj/semclass/ho_estimate2.pdf). Assuming that bootstrapping gets around that problem, then I could see that you might have a difference in conclusions about statistical significance for your path, where one approach might yield a conclusion of significance and the other does not. I don't know the nuances of your data, but this seems like a reasonable possibility for what you are seeing.
Dear MIke, thank you very much!! I really appreciate your comments. And yes, I didn´t realise that...and now what gives us the best interpretations (the one that can fit better the "reality") the normal theory or bootstrapping? I guess this has not an easy answer!!! :) :) Once more, thank you for your attention.
Hey Mİke. Thank you. What about model fit indices? Mediation in AMOS does not give any coz of df is 0? thanks again
Hello. When you have a just-identified model (this will occur when the df=0) then you cannot rely on traditional fit indices such as chi-square gof, chi-square/df, TLI, CFI, RMSEA, etc to evaluate overall model fit. In cases such as this, the best you can do is examine model fit in terms of its individual components (i.e., path coefficients and indirect effects). Hope this helps! best wishes!
@@mikecrowson2462 Thank you Mike
Hi, Mike, thank you very much for your video. I have a little question about evaluating indirect effect using Amos, that how to compare the significant differences between two indirect effect in different condition using Amos, my reviewer told me to use nest-model comparison approach, but I'm not sure how to conduct it in Amos. I would be much appreciated if you can answer my question. thanks again.
Yeah sure, but how do you report them.
What if the indirect effect is lower than the direct effect, does it mean no mediation, or it is partially mediating..?
Hi, great video. I have a question. If my mediation variables are continuous but my independent variable is categorical, can I use the method that you talked about in the video, or do I need to do any adjustments? Thank you!
Hi there. If you have a categorical independent variable, you can still run the mediation analysis. However, you'll need to recode it (e.g., dummy code) into binary predictors. If your IV only has 2 groups, then you can run the mediation analysis with that variable as IV. If you have an IV with 3 groups, you'll need 2 dummy variables, which is effectively a re-packaging of the original IV into those new variables. In regression analysis, the relationship between each dummy variable in a model (reflecting a single categorical IV) represents the same thing as an orthogonal contrast (ie., a test of the difference in a comparison group and reference group on the mediator). Extending that logic, if you have more than one dummy exogenous variable and you are predicting the mediating variable, then the indirect effect being computed per dummy variable will represent part of the total mediated effect between the original IV and the DV. FYI, some folks might complain that this violates the assumption of 'multivariate normality'. However, this assumption really only pertains to the endogenous variables in your model (Rex Kline, 2015, provides a good discussion of this assumption; see www.guilford.com/books/Principles-and-Practice-of-Structural-Equation-Modeling/Rex-Kline/9781462523344).
I hope you find this helpful. Best wishes!
@@mikecrowson2462 Super helpful! What if I have a dependent variable that's a categorical or dummy variable?
Can I still use the SEM command in STATA? Can SEM in STATA deal with logistic regression/multinomial regression?
thanks for the video. Could you please tell me should not we include indicator variables in the model?
Hi Samra, I'm not sure what exactly your question pertains to. In this illustration, I'm demonstrating evaluation of direct, indirect, and total effects in a path model with measured/observed variables. Typically, the term "indicator variables" refer to measured variables that serve as "indicators" of latent variables in SEM models, such as CFA models and path models using latent variables. Technically, the measured variables in path models - such as the one shown in the video where there are no latent variables being modeled - are indicators of latent variables too; but are only single indicators of underlying concepts. Moreover, in models such as these we do not account for measurement error when estimating path coefficients. I hope this is helpful.
The video only focused on using measured variables. You can easily extend the logic to running tests of indirect effects using path analysis involving latent variables. Basically, start with CFA model to test the fit of your measurement model. When that model appears to fit the data well, then you can move to testing the proposed structural relationships among your latent variables. You evaluate the overall fit of the model and can also use the procedures shown in this video to test for the direct and indirect effects in your latent variable model.
Hey Mike, really helpful tutorial so far! Can you tell me if there is a way to calculate effect sizes for the direct effects in the model? Thanks a lot!
Hi Mike, thank you for a very informative explanation. I have a question for you. Just say in the model there are two exogenous variables (X1 and X2), with both X1 and X2 directly effecting Z and indirectly effecting Z through Y. How would you calculate the total effect of both X1 AND X2 on Y? (je can you not only calculate the total effect of X1 and the total effect of X2 on Y, but the Total effects of both X1 AND X2 on Y?
Hi there. Thanks for your question. It is an interesting one. As far as I know, you are computing indirect, direct, and total effects using a single X variable at a time. It seems you are asking about some type of joint indirect effect of multiple X's and Y, which is not something I've ever heard of. best wishes
Absolutely helpful. A great video for answering many of my questions from reading Kline 2016.
Many Thanks Mike for share your knowledge, I have a question: Is an indirect effect stadistically significat if the lower and apper not includ 0?
Yes. That is correct Gonzalo.
Hi, can we have mediating variables in a non-recursive model?
Thank you sir
Tq sir