Assessing multiple mediation in AMOS (testing total and specific indirect effects)

[@smaximok1]

Thank you very much, Mike. I never assessed direct & indirect effects before. You opened my eyes.

[@jintukurian621]

Well explained and it really resulted in easing my work. Many Thanks.

[@navjiwanhira3888]

very very informative, useful and interesting video. Thank you

[@chakibbrahimi4722]

Thank you very much for this excellent video

[@mikecrowson2462]

You are very welcome! Thanks for visiting!

[@godfreytumwesigye]

Thanks Prof Mike for this wonderful video. I am using version 20 of AMOS software. I followed your steps very well and tried to put them in practice. But when I try to create the estimand, it does not give me a blank window; instead it gives me a window with the following: Then I go ahead to create the indirect paths as guided by you Prof. and everything seems Ok until I click in the bottom right corner. The feed back I get is that: Just asking, anybody who can help me identify my error, I will be profoundly grateful.

[@mikecrowson2462]

Hi there. It looks like the message you were trying to copy and paste did not come through in your posting, so I can't really comment on that. I will say that you might try a more recent video (made a week or two ago) I put out on the topic here: kzbin.info/www/bejne/nYqoZax_n8t5f6M . If you are still having problems try posting there too (but make sure to include the error message you are receiving). Cheers!

[@josaschamusic]

@@mikecrowson2462 I have a similar problem, professor. After I name the paths, I try to create estimands but instead of the syntax window I get an error window and the message: "The program editor may take several seconds to start the first time. It will start much faster the next time". It's been like that for a while, what could I do?

[@mikecrowson2462]

@@josaschamusic Oh gosh, I don't know what that could be. I've never seen that before. You might see if there is something different between what I mention in the video and what you have done. Sometimes even very small differences can produce different results. By the way, I have a much newer video here on the topic: kzbin.info/www/bejne/nYqoZax_n8t5f6M If you are doing everything exactly the same, then I'm not sure what it could be other than a glitch with your program. Perhaps you might need to reinstall it (but be careful with this option; I'd check with the folks who sold it to you first!) I'm sorry I couldn't be more help on this. best wishes

[@sperrotta91]

I have another question. Is there a way to get user-defined standardised indirect effects? Thank you for your help!

[@smaximok1]

I do it manually in MS Excel

[@lesleysmith9377]

Amazing video thank you very very much!

[@mikecrowson2462]

You are very welcome!

[@josaschamusic]

Thanks for your video. One question, my reviewers asked me to do Monte Carlo. Can I just click the Monte Carlo option in AMOS and do everything else like you did here?

[@mikecrowson2462]

I believe that is right. Funny though. The main use I have had for using Monte Carlo in AMOS is when I am testing indirect effects but with summary data that has been input into AMOS: kzbin.info/www/bejne/iKGUnZitd7Caga8 But it looks to work with when raw data is used in SPSS. My preference is still bootstrapping. Cheers!

[@atnguyen-sg8xk]

Thank you, very informative video, I just have one question. Usually when analyzing a model in SEM, we should report the model fit. However, I notice that when conducting mediation analysis using SEM, we don't need to report the model fit since sometimes the model is too simple (e.g: 2 IVs, 1 M and 1DV), is that correct?

[@MonjiGhanem]

Very helpful. Thank you 🌷

[@김나경-b4y8d]

thank U so much,,,, -from south korea

[@mikecrowson2462]

You are very welcome! Best wishes!

[@talzabidi1569]

Thanks Dr What about if we have latent constructs? Can we use same process? And why do you use observation variables in this video?

[@navjiwanhira3888]

If the indirect effect is negative, what does it mean??

[@spindarella4353]

Why are the isolated indirect effects only available in unstandardized form? I reported the direct effects using standardized estimates so that I would be able to make inferences of weighted effect sizes. Is it reasonable to report both standardized (for direct effects and total indirect effects) and unstandardized values (for the isolated indirect effects)? Thank you!

[@spindarella4353]

Also, would you be able to provide an example of how you would interpret and report one of your isolated indirect effects?

[@mariataimoor1591]

Please tell that the estimands or specific indirect effects generated are unstandardised? Can we generate standardised specific indirect estimates also

[@iqrasherazi2569]

hey, your video is very informative. but i have a question. how to fix syntax error while using amos-21

[@xuannhat115]

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, 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?

[@prayaansood-akidwhosnotkid4334]

Dear Mike thanks for this video..I have a query here..if you may help me with it When I am trying to create the estimand with the multiplication of the named path reg weights, I am getting an syntax error DECLARATION EXPECTED. What might be the reason behind this error..how do I resolve it?

[@범고래씨]

Hi Mike! I always found your videos helpful, thank you for sharing! One question. I want to compare indirect effects in a model and test whether two indirect effects are significantly different (one bigger than the other). I heard that it is possible to do it in Mplus but is it also possible to test this in AMOS?

[@Nbl-369]

What if perfgoal and anxiety would have one moderator suppose that's low self-esteem. Mastery and interest also has a moderator that perhaps be skills. Now how to run the overall model.??? This would be something like dual moderated parallel meditation model? Can you tell me how to run this specific model. 2.moderator 2.mediator .independent variable and one Dv. Anxiously awaiting plz

[@vlogrezalborz8551]

Thank you so much, it was sooooooo useful

[@donkisott]

Hi Mike, I'm using Amos 21 and I could not write the syntax for indirect effect estimand. It gave an error (sier1 is not declared. It may be inaccessible due to its protection level. par1 isnot declared. It may be inaccessible due to its protection level). So, I'm stuck with the estimating user defined estimand level. May you explain how you did it.

[@dr.denilajinny1454]

Hi, can the variables involved in SIE be categorical too? Thank you very much for the time and effort you have put into creating such videos. Truly helpful.

[@mikecrowson2462]

Hi Dj. A categorical variable can be included in your mediation model so long as you it is an exogenous variable (i.e., a variable that is not predicted by others in the model). Basically, you can use dummy or effect coding of that variable as you would in the case of multiple regression. On the other hand Endogenous variables (such as your mediating variables and outcomes) in your model must be continuous. So, yes, you can test specific indirect effects involving categorical predictors so long as they are exogenous. [You may sometimes hear folks refer to multivariate normality in SEM. Multivariate normality is an important assumption when using ML estimation. However, the assumption pertains mainly to the endogenous variables in your model; see Kline 2015 for discussion in his book Principles and Practice of Structural Equation Modeling; he refers to multivariate normality of 'continuous outcome variables']. You will also find that Andrew Hayes Process macro (which can also be used to test for mediation models, moderated regression, and even moderated mediation) also allows the inclusion of categorical predictors, where the categorical predictor can be recoded automatically based on the coding scheme you request; download at processmacro.org/download.html). The downside is that the models are estimated using OLS regression (although indirect effects are tested with bootstrapping), thus you can't obtain global measures of fit like in the case of using SEM with maximum likelihood estimation. Hope this helps. Best wishes!

[@dr.denilajinny1454]

​@@mikecrowson2462 It does help a lot. A comprehensive explanation. Thank you very much. I now understand that not just the specific path, but the model itself should not have an endogenous categorical variable. But one of the endogenous variables in my model is categorical. Is there any way I can estimate SIE? I can try and learn another software too. I really need to understand SIE to explain my results. Thank you again.

[@vincenzocalvo5618]

Thank you, very useful

[@hishonline]

I dont have a status bar at all in amos 18. Cant find it your way!

[@i.g.8206]

Hey, I have a question, I am struggelling with for days now. My task is to turn the following theoretical background into a graphical causal model following these rules and I just dont understand which variables are needed and which I can cut: Maybe s.o. has an idea. Following the rules for building DAGs for a simplyfied but sufficient model: 1. Variables for which causal relations are of interest. 2. Common causes of these variables 3. Variables that are conditioned on in the statistical analysis or during data 4. Causal relations between all variables Theoretical backround: Why expect an effect of having a university degree on anti-refugee-sentiment? There are several reasons to expect an effect of having a university degree on anti-refugee-sentiment. Additional education may increase tolerance and reduce fears of different cultures. An elevated average living standard (as indicated by higher income and life satisfaction) resulting from completing tertiary education may lower the general susceptability to xenophobia. Finally, individuals with a university degree may not consider refugess as competition in the labour market, thus harbouring less resentment against their arrival. What are alternative explanations for an association between having a university degree and anti-refugee-sentiment? There are several factors that may lead to a noncausal association. For example, we know that fewer people in East Germany have attained a university degree than in West Germany and that antirefugee- sentiment is more pronounced in the East. Furthermore, there is/was a gender difference in attaining tertiary education and men are generally more likely to espouse xenophobic views than women. Likewise, younger people are more likely to have obtained a university degree than older people. Finally, both general intellectual competence and social origin (i.e., parental class/education) may affect both educational attainment and anti-refugee-sentiment.

[@sperrotta91]

Could I ask how you establish the significance of indirect effects when there is missing data? I cannot perform the bootstrap...

[@mikecrowson2462]

Unfortunately, Amos does not have the capability to do bootstrap with missing data. One option is to create a new dataset in spss with no missing data - listwise deleting cases with missing data on the variables you are analyzing - and then perform the analysis with bootstrapping on that reduced dataset. This is actually the default with Hayes process macro. A second option is to create a complete dataset using an imputation approach such as regression imputation and analyze the data with that new dataset. Amos has an imputation option to do this. There are downsides to both approaches that are too involved to get into here. But these are the main approaches available to address your problem. Cheers

[@sperrotta91]

@@mikecrowson2462 Appreciate your response. I deleted missing cases and reran the model accordingly... didn't lose any significance so seemingly success!

[@Rxze-]

Hi Mike, Is it possible to assess a moderator on a multiple mediation pathway? Say you have the indirect effect of mastery to interest to anxiety to achieve, and you wanted to know whether anxiety moderated between interest and anxiety, would this be possible to work out via SEM? Thanks

[@Rxze-]

Not to worry - found your video. Thanks

[@hasank83]

Hi. Thanks a lot Mike. I've really learned much about SEM, AMOS, mediation, and moderation by watching your videos. Many people asked for the estimation of the standardized specific indirect effects. Do you have any new information about this? How can we get the standardized estimates for the specific indirect effects? I heard it is possible in Mplus, but not in AMOS. I have a model just like the one you used n this videos (2 independents, 2 mediators, 1 dependent). I just reported the unstandardized estimates.

[@samlee5247]

I have the same question too. Hope that anyone can see our question and answer it

[@freepolca]

Firstly I am so thankful for your video, it was really helpful in running the serial meditations which I have in my Multi variate Model for my PhD. one little thing can we use this for getting the confirmation of either an indirect effect is significant or not and in place of Un standardized coefficients we can multiply the Standardized effect from the main model and get standarzied effects ?

[@mikecrowson2462]

Hi Muhammad, if you go under Options and select Standardized Estimates, then it will calculate the standardized direct and indirect effects for the model. Thanks for watching!

[@mahakhawajah]

Hi Mike, this is so useful. Is there a reference for this technique?

[@mikecrowson2462]

Hi Maha, the use of User-Defined Estimands is covered in the AMOS user's guide, chapter 38 I believe: www.csun.edu/itr/downloads/docs/IBM_SPSS_Amos_User_GuideV24.pdf From there it's just simply creating the correct estimands for the indirect and total effects. by the way, here is a much newer video on this topic: kzbin.info/www/bejne/nYqoZax_n8t5f6M best wishes!

[@Ibrahimalnawas4773]

Hi Mike... Many thanks for your helpful video... I have a question please... the indirect effect of a specific relationship is a*b, correct? the thing is that when I multiplied a*b it gave me a number that was less than the number that I obtained from Estimand as you kindly explained in the video. Aren't both number supposed to be equal?

[@mikecrowson2462]

Hi there. I don't really know why that would be the case. I just went back and computed everything and things looks to be in order. If you multiply the unstandardized path coefficients for mastery->interest (par 1) and interest->achieve (par 2), those product of those paths will be the same as the outcome user the user-defined estimands reviewed in the video. Make sure that you haven't inadvertently misnamed a path (as you are following along with the video). If you mislabeled something then you may think you are computing the same indirect effect I'm talking about, but then you are actually computing something else. Also, make sure to check that you aren't multiplying paths associated with a standardized solution and then comparing against the indirect effect computed based on unstandardized coefficients. Best wishes.

[@nilayacat5516]

Thank you very much for video. I have a question. 0.03350 for perfgoal ->achieve relation in regression weights table. That’s mean is significant and you said before this table is same with direct effects. But if I look at the bootstrap value for perfgoal and achieve (standardized direct effect - two tailed significance) is 0.164. So, this effect is not a significant. These value is not match each other. Which one is true value for this model ? Significant or not? I am really confused. Thank you.

[@mikecrowson2462]

Hi there. The difference in results you are raising actually has less to do with differences between tests of unstandardized and standardized coefficients being tested. The p-value you are noting in the regression estimates table is computed using a normal theory method (and assumes multivariate normality of your endogenous variables), whereas the p-value you see reflected when using bootstrap (btw, it would be something around .151 if comparing with the bootstrap test of the unstandardized indirect effect) is based on a re-sampling approach. When I just-reran this, the standard error which was originally was .004 was larger (i.e., .006) when I requested it in the bootstrap results. So, yes it is possible for the results between the normal theory approach and the bootstrap approach to be different. Looking at the multivariate kurtosis (I requested normality tests when I re-ran this), it looks to be significant, suggesting a significant departure from multivariate normality (again, this is a fundamental assumption of ML estimation). Given that a violation of multivariate normality could contribute to downwardly biased standard errors (if using the normal theory method), it may be that the bootstrap approach corrected for this. Obviously, I didn't spend any time on this issue because this would take folks off track from the main thrust of the video. However, I do have some discussion on this topic here (kzbin.info/www/bejne/kJyydJqfqNOFnLM), as bootstrapping is one possible approach to dealing with non-normality. Thanks for your question! Best wishes.

[@nilayacat5516]

@@mikecrowson2462 Thanks for your answer, that was very helpful :)

[@eslamafifi9699]

Hi Mike, Many thanks for your video. It helps A LOT. I have a silly question (hope you do not mind), I need to report the t value for the indirect effect for par 1 * par 2. where can i find it?

[@mikecrowson2462]

Glad you found it useful. You can compute the z-value for the indirect effect by dividing the indirect effect (under user-defined estimands) by its bootstrap standard error. Click on bootstrap standard error to get the SE. Then simply divide IE/SE.

[@eslamafifi9699]

Many thanks mate...Much appreciated

[@mikecrowson2462]

You're welcome! Good look with your research.

[@soehartosoeharto8471]

i think you have made a plugin to solve this case, or i am wrong? please let me know

[@mikecrowson2462]

I don't think of it as a plugin per se. However, you might find these page helpful: statwiki.kolobkreations.com/index.php?title=Plugins and kzbin.info/www/bejne/f3TRmWapZ6t6Z5Y. I hope this helps! Thanks for visiting!

[@siobhanmcdonnell4059]

THANK YOU for this video. I have been searching for days for a way to estimate a sequential indirect effect in AMOS so you might be saving my thesis right now. Would this also work when there is a latent variable in the model?

[@mikecrowson2462]

Hi there. Yes, it should work whether the path model only has manifest variables or incorporates the use of latent variables. Good luck with your thesis!

[@siobhanmcdonnell4059]

Thank you for your help! I actually also have a follow-up question. When I run the analysis with the user-defined estimands, it worked great for my first model, which only includes observed variables. On my second model with a latent variable, for each of the estimated specific indirect effects, one of upper/lower bound intervals was not displayed - just "..." in the place the estimate had appeared in the video and in the previous model. Do you know of something I can try to fix this?

[@mikecrowson2462]

Hi, are you getting stars, like "***"? When you get the "***" in AMOS, it's because the number is really small (close to 0) and you don't have the decimals rounded off at enough spaces to actually get the specific numerical value. For instance, this often occurs with p-values in AMOS when testing individual parameters. I don't know if this is what is going on with your confidence intervals. In my version of AMOS (25) it looks like you'll get a value such as 000. for the upper or lower bound when generating a confidence interval if you don't have rounding to enough decimal places. I find it odd that you could get a value for one or the other bound of the confidence interval but not both. Are you getting a p-value? If not, are you getting an error message of some sort?

[@siobhanmcdonnell4059]

Thank you for your reply. It is, unfortunately, not the asterisks, but an ellipsis, which I had never seen before (unlike the asterisks and such). Particularly because I didn't get an error message and the p-value was still computed, I was puzzled by it. Could it have something to do with the inclusion of the latent variable?

[@mikecrowson2462]

I don't really know what to tell you. I've never seen this problem before. If you are getting a p-value and only an upper (or lower) bound for the indirect effect, then it sounds like it has to do with the way the program is set up to present confidence intervals. If your fit statistics make sense and the path coefficients and significance tests all make sense, then it may just be a presentation issue. You could always check by setting number of decimals to a higher number to see if you actually get a number in your output. If there is no apparent error message anywhere - and one that doesn't throw your general results off, then you could always go with the p-value in your presentation of the test results for your indirect effect. Sorry I can't be more help on this issue.

[@sosyalmedyaeposta6071]

Thank you very much for this excellent video, this really helped me a lot. I have one problem about how to report the results. We get unstandardized values for specific indirect effects, so when reporting in a paper, should we report unstandardized effects for the whole model and direct effects as well? I am really confused about this. Thank you again.

[@mikecrowson2462]

Hi there. A lot of times folks report on the standardized effects (say in figures) because they can be viewed from an effect size perspective. I tend to do this in the figures I include in my research. On the other hand, others report on the unstandardized coefficients in their figures. Either approach is reasonable (but my preference is the former. I also include asterisks reflecting significance levels associated with the tests of the individual path coefficients). Another possibility is to report on standardized effects in the figure, and to include presentation of the unstandardized coefficients, s.e.'s, and p-values in the text (I often do this too). If you want to report on indirect effects that are standardized, you could simply request the standardized coefficients (through analysis properties) and then after the model is estimated simply multiply the standardized path coefficients associated with each mediation effect you were testing using the user-defined estimands. To compute an indirect effect of two paths (for instance), you multiply the path coefficients together - and this would be the same whether you are working with unstandarized or standardized path coefficients. Hope this helps!

[@sosyalmedyaeposta6071]

Hi Mike. Thank you very much for your kind reply. You really helped me a lot with this. I would really appreciate if you could provide me some references of your (or other) work related to this kind of model? Thank you again.

[@oluwaseunadeoyeoyebamiji3592]

I think for the indirect effect for interest and anxiety, the indirect equation should include path c and b.

[@tinakiki5507]

think you sooooooooooo much.

[@paulsauer1475]

P

[@mujganinozu2639]

Thanks so much. The video helped too much