Message/Author 

Dieter Urban posted on Wednesday, December 01, 1999  10:47 am



Does Mplus offer an easy way to obtain the values of total effects (i.e. the sum of all direct and indirect effects)? 


No, Mplus does not give the values of total effects. The Sobel (1986) formulas would have to be applied. 

Anonymous posted on Thursday, May 17, 2001  5:09 am



Does MPLUS plan, in the near future, to support the automatic calculation of total and indirect effects? 


It is on our list of things to add to Mplus, but this will not happen before the end of the year. 

Anonymous posted on Friday, October 19, 2001  6:37 am



I am conducting a combination of continuous and categorical variables and am trying to find how to get additional fit indices other than chisquare and RMSEA with the original version of Mplus. Is this possible? If so, how? 


Version 1 did not provide fit statistics other than chisquare and RMSEA. Version 2 does. You would have to compute them yourself if you want other fit measures and are using Version 1. There are several postings on Mplus Discussion that give formulas for alternative tests of fit. 

marycampa posted on Thursday, January 24, 2002  12:48 pm



What is the status of having indirect / total effects added to MPLUS. In the mean time could you give the full reference for hand calculating them? Thank you. 


They are on the list to do. They just haven't risen to the top yet. I think you can find the information that you need in Ken Bollen's book. The reference is on the Mplus website. 

duckhye posted on Tuesday, May 07, 2002  8:24 pm



I am considering path analysis with a couple of binary endogenous variables. Can we calculate direct/indirect/total effects in the path model involving binary dependent variables? Thanks 


Mplus does not compute the indirect effects automatically. You can do this for categorical outcomes in the same way as it is done for continuous outcomes. 


A student with whom I am working is doing a path analysis with all of the variables in her model being binary. She has estimated logistic regression coefficients with SPSSPC, which does not provide standardized coefficients. Is it possible to calculate indirect and total effects with the unstandarized coefficients, given the variables' metrics are all 1/0? If not, could you direct me to a reference that would show how standardized coefficients could be calculated? Thank you. Leonard Bloomquist Kansas State University 

bmuthen posted on Sunday, August 04, 2002  5:20 pm



The analogy with regular continuous dependent variable calculations of direct and indirect effects can be used when applying Mplus. Mplus considers latent continuous response variables y* (one behind each categorical dependent variable). The unstandardized estimates would be used here. 

Tauhid posted on Thursday, April 24, 2003  6:39 am



Hello: I am Ph.D. candidate at Washington State University, Pullman. Currently I am working a paper where I am using MIMIC model (Multiple Indicators and Multiple Causes (Goldbergers' MIMIC Model). Can I estimate this MIMIC model using Mplus? I am new to Mplus, and therefore I am unable to use it. Please advice me how? PS: My MIMIC model is as follows: Suppose I have a single latent variable which we call health status (h). Latent health status has many observable indicators say, (Y, which is a vector). Moreover, the latent health status has many cause variables say (Xvector). So we have a problem where the latent variable (h) is a linearly determined by vectors of cause variables (X), and the latent variable (h) in turn determines the vector of indicators variables (Y). (1) h= alpha*x+ error term (2) y= beta*h+ error term Then the reduced model is: y=alpha*beta*x+ error term Now, I want to estimate the latent variable h, coefficients alpha and beta. With Best Regards! Tauhid 


Chapter 18 of the Mplus User's Guide has examples showing how to set up a MIMIC model. You can also find examples at www.statmodel.com under Examples, Continuous Outcomes, Factor analysis with covariates. A MIMIC model is a confirmatory factor analysis model with covariates. 

Anonymous posted on Monday, March 22, 2004  6:48 pm



I am considering a mediating effect model across multigroups.How can I test the difference of mediating effect across groups. Many thanks. 

bmuthen posted on Tuesday, March 23, 2004  7:11 am



I think you refer to an indirect effect, that is a product of two slopes. If that is correct, Mplus Version 3 gives the standard error for this indirect effect. Testing difference (or rather equality) across groups would seem straightforward if there are no acrossgroup parameter equality constraints because then the difference test is based on uncorrelated parameter estimates. So you only use the SEs for each indirect effect. If there acrossgroup constraints, you would need to use covariances between parameters in the different groups which isn't readily available for indirect effects, but would have to be worked out via the Delta method. 

Anonymous posted on Monday, November 08, 2004  5:26 am



I've heard about the possibilty to compute 'indiect effects' now. If this is correct, can you tell me what is to do in the Syntax to examine this effect? May be you have some 'zipped' examples here in a chapter on the homepage? 

bmuthen posted on Sunday, November 14, 2004  11:21 am



Take a look at example 3.16 from the Version 3 User's Guide  it is also on the web. This uses "MODEL INDIRECT". 

Anonymous posted on Tuesday, March 15, 2005  11:29 am



I am getting 999.000 for estimates of Std StdYX. Can you tell me why this might be happening? See output below: TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT, AND DIRECT EFFECTS Estimates S.E. Est./S.E. Std StdYX Effects from A1 to A3 Sum of indirect 0.795 0.060 13.155 999.000 999.000 Specific indirect A3 A2 A1 0.795 0.060 13.155 999.000 999.000 


I would need to see the full output at support@statmodel.com. 

Anonymous posted on Monday, April 18, 2005  10:49 pm



Hi, I would like to know how the indirect effects are computed for non recursive systems. What formula or equation is being used?. Thanks 


See Bollen's SEM book for a description of how indirect effects are estimated for nonrecursive systems. 

Anonymous posted on Monday, April 25, 2005  4:10 pm



Does MPLUS computes marginal effects for tobit equations? Thanks 

bmuthen posted on Monday, April 25, 2005  4:21 pm



Mplus does censorednormal modeling (= Tobit), and also allows for "inflation" at the censoring point (so a 2class model in line with ZIP). I am not sure what you mean by marginal effects  the modelestimated means and variances are produced. 

Anonymous posted on Tuesday, May 03, 2005  7:49 am



Hi, Linda, I am getting 999.00 for the std and stdyx when requesting indirect effects, the same situation as the authoer of Mar 15, 2005. Did you find out what the problem was for that post? 


You probably have a negative residual variance in your results. If not, send your output and license number to support@statmodel.com. 


Also, be sure you are using Version 3.12 and not an earlier version. The update is available under Product Support and is free. 

Anonymous posted on Monday, June 13, 2005  8:22 am



Hi, I have two ordinal (ordered categorical) endogenous variables Y1 and Y2: ANALYSIS: PARAMETERIZATION=THETA; TYPE=MEANSTRUCTURE; Y1 ON Y2 X1 X2 Y2 ON X2 X3 MODEL INDIRECT: Y1 IND Y2 X2 The indirect effect is significant. I know how MPLUS gets the coefficient of the indirect effect, but I don't know how it gets the standard error. Could you please answer this question for me? Thanks a lot! 


The default is to use the Delta method. You can also request bootstrap standard errors. See the Mplus User's Guide for further information. 

Anonymous posted on Wednesday, June 15, 2005  12:45 pm



Hello. I am running a path model in MPlus v. 3.12 and am examining indirect effects. I am confused by some of the results. I am finding many significant indirect effects (as gauged by the Est./S.E. values), but all of these effects have extremely low coefficients. For example, many of the significant indirect effects have STD and StdYX values of .0002 or .0004. I do not understand how such small coefficients can be significant. Am I missing something? Thank you. 

Anonymous posted on Wednesday, June 15, 2005  2:06 pm



A quick followon to my question above. First, I made a mistake above  the effects are reported to three not four decimal places. So, I meant to say that there are several effects around .002 or .004. Second, I am noticing that there are several situations where the indirect effect is larger (and significant) than the total effect (which may be nonsignificant). I thought the total effect included the indirect effect, so I don't understand how this can be. Can you clarify this for me. Thanks again. 


You will need to send your output, data, and license number to support@statmodel.com. Be sure you are using Verson 3.12 before you do this. 

Anonymous posted on Friday, June 17, 2005  9:27 am



Hi, thanks a lot for your reply to my previous inquiry on June 14. Back to my model with two ordinal (ordered categorical) endogenous variables Y1 and Y2: ANALYSIS: PARAMETERIZATION=THETA; TYPE=MEANSTRUCTURE; Y1 ON Y2 X1 X2 Y2 ON X2 X3 MODEL INDIRECT: Y1 IND Y2 X2 In the output, Mplus gives Estimates, S.E., and Est./S.E. for both direct and indirect effects. Is Est./S.E. the zscore, from which we can calculate probabilities? For both direct and indirect effects? And, for both DELTA and THETA PARAMETERIZATION? I want to confirm this because Mplus doesn't report probibilities for one or two tailed tests. Thanks again! 


The estimate/standard error is always a zscore is all contexts. The tests are always twotailed. 


When standardized coefficients for indirect effects are calculated, which variables provide the standard deviations for standardization? For example, if A > B > C, the indirect effect of A on C is beta(AB)* beta(BC). Are the standardized coefficients (StdXY) to be interpreted as the change in C, expressed in terms of a fraction of one standard deviaiton, that occurs as a result of one standard deviation change in A? Or is B involved somehow also? Thank you in advance for your response. Inna 


It is for a one standard deviaton change in A. 

Roger Brown posted on Monday, April 14, 2008  11:38 am



Linda, Sorry to bother you again. Just a quick question. Have you incorporated total effect calculations into Mplus with SE's. yet? The new manual seems to indicate it but I can only see total indirects. Thanks. Roger 


We have always given the total effect with a standard error. It depends on how you specify MODEL INDIRECT. In some specifications, it is referred to as sum of indirect effects. If you want more information, please send your input, data, output, and license number to support@statmodel.com. 

Roger Brown posted on Tuesday, April 15, 2008  8:39 am



Linda, Sorry, I must not have made myself clear. Take for example this simple model: TITLE: this is an example of a path analysis with total effects DATA: FILE IS my.cov; type is covariance; nobservations are 100; VARIABLE: NAMES ARE y1 y2 x1; MODEL: y1 on y2 x1; y2 on x1; MODEL INDIRECT: y1 ind y2 x1;  The total effect of X1 on Y1 would be the direct effect of X1 on Y1 = 0.060 plus the indirect effect of X1 on Y1 = 0.190, with the total effect of X1 on Y1 as 0.250 (se= 0.117). How would I get these total effects and standard errors? Sorry for the hassle. Thanks. Roger 

Roger Brown posted on Tuesday, April 15, 2008  8:50 am



Linda, Opps, here is the covariance matrix if you need it. 1.400 .600 1.500 .250 .500 1.00 Roger 


You will get the total effect if you say: MODEL INDIRECT: y1 ind x1; The way you are specifying it, you are asking only for a specific indirect effect. 

Roger Brown posted on Tuesday, April 15, 2008  9:17 am



Linda, Aha yes, thank you, didn't think of that. Roger 

C. Sullivan posted on Friday, August 01, 2008  6:57 am



I'm running a path model with a set of endogenous variables both continous and binary. In such cases, how are the total and indirect effect estimates scaled? The outcome is dichotomous so I believe they are logits, but wanted to double check. 


With weighted least squares the coefficients are probit. With maximum likelihood, the default is logits. 

jtw posted on Sunday, March 29, 2009  9:06 am



Typically, I have found that when there is a statistically significant total effect the direct, indirect, or both are also statsitically significant. However, I have a situation in which I have a statistically significant total effect (p<0.05) but both a marginally significant direct and indirect effect (p<0.10). Hence, at the conventional level of significance, neither the direct nor indirect effects are statistically significant. How would you suggest interpreting this situation? 


I believe the sum of two small effects can be significant. I would simply report the results as found. 

Eulalia Puig posted on Thursday, November 12, 2009  11:00 am



Hello, What happens when the other way around is the case, that is, we have two significant coefficients (one is .015 and the other .001; both positive) in a simple path model A>B>C, but the indirect (total) effect is only marginally significant (.06). Sample size is small (n=129) but then the that should play in the coefficients as well  and it doesn't. Please, advice. Thanks. 


Once little piece of info. When running the models with each item individually for variable C (as opposed to C being a construct of 3 items) none of the paths (B>C) are significant... That is certainly a strange result.... Any ideas? Thanks again. 


I don't think it is surprising that indirect effects are insignificant when the 2 component effects are significant. Substantively you can think of the A effect as not reaching as far as to C. Statistically, the SE for the product can be larger than for the parts. Also, if this is seen in your construct model version but not in the itemspecific versions, then model misspecification may play in as well. The model chisquare may not indicate good fit. 


Hello, I am running a path model with two mediators, and have used model indirect to get estimates of the specific and total indirect effects. I am puzzled because I have one specific indirect path that is significant (and one not significant), but the total indirect effect is not significant. Is it legitimate to interpret a specific indirect path that is significant, if the total indirect effect is not significant? 


This seems perfectly acceptable. 


What is the estimate listed in the STDXY for an indirect effect? When reporting the EST/SE from the indirect effect, what is the appropriate statistical symbol? Thank you. 


The indirect STDXY does the usual  divide by the estimated SD for the DV and multiply by the SD for the IV. With the usual x, m , and y notation the DV is y and the IV is x. It is an approximately normal (z) test. 

Emily Blood posted on Wednesday, July 07, 2010  11:41 am



Is there a way to test (i.e. get SE's for) a combination of parameters? I am interested in testing the effect of z on y through m (direct + indirect) plus the effect of z*t on y (g1*lam1 + k1 + k2). Relevant part of model is: y2 on z2*0.0244 (k1); y2 on zt2*0 (k2); y2 on m2*0.16 (lam1); m2 on z2*0.16 (g1); [If this matters, I am doing this within a LGC and also MONTECARLO type of data]. 


You can use MODEL CONSTRAINT to create new parameters and obtain standard errors for these new parameters. 

Emily Blood posted on Wednesday, July 07, 2010  1:56 pm



Thank you. I created a new parameter (g1*lam1 + k1 + k2) and the value is estimated along with SEs, but for the MONTECARLO data, the true value for the combined effect (against which coverage prob and power are calculated) is not correct in the output. The true value of the combined effect is 0.16*0.16+0.0244+0=0.05, but in the output is listed as 0.5. Is there a way to set the true value for the combined effect other than what I've done above? 


I think you can give the value using the NEW option of the MODEL CONSTRAINT command. NEW (f*.5); where .5 is the true value. 

Emily Blood posted on Wednesday, July 07, 2010  3:00 pm



Yes, NEW (TOT1*.05) worked. Thanks!! 


Hello, I have an interpretation question for a simple mediated model with a single mediator. When I request: y IND m, it shows the 'total' effect from x to y is nonsignficant, but the specific effects show a significant positive indirect effect and a nonsignificant negative direct effect. My question is, can I still conclude that x is relevant to y (even though the 'total' effect is nonsignificant), in that there was an indirect effect through m? Thank you. 


You can report that the indirect effect is significant even when the direct effect is not. 


Thank you for your prompt response. 

Eddie Ng posted on Tuesday, April 19, 2011  10:11 pm



Hi, I have a model in which I want to test both the total and specific indirect effect from X>Y through 3 mediators via 4 indirect paths: 1) X>M1>Y 2) X>M2>Y 3) X>M1>M3>Y 4) X>M2>M3>Y The syntax of the last part is as follow: MODEL INDIRECT: Y IND M1 X; Y IND M2 X; Y IND M3 M1 X; Y IND M3 M2 X; ANALYSIS: BOOTSTRAP = 1000; OUTPUT: CINTERVAL(BCBOOTSTRAP); There are a few questions I would love to have your help: 1) Do I need to include "standardized"in the output row? 2) When reporting the confidence intervals of total/ specific indirect effect, is the normal confidence intervals data or STDYX standardization data to be reported? 3) If STDYX standardization data is to be reported, then I have a significant total indirect effect. In addition, I have 2 indirect path (Path 1 & Path 3) significant at 95% CI and 2 two paths (Path 2 & 4) marginally insignificant at 95% CI (but significant at 90% CI). Regarding the two marginally insignificant indirect path, what is interesting is that all the component effect (i.e. X>M2, M2>Y, X>M2, M2>M3, M3>Y) are all significant (<.05). In this case, can I still claim that M2 mediate the effect of X on Y? Or how will you interpret the result? Thanks 


1) If you want standardized results as well, yes. 2) That depends largely on the journal you submit to. Personally, I try to avoid standardization when it is not essential. Also, you don't want to use STDYX if X is categorical (see UG). Regarding your last paragraph, I would report all that. It is the significance of the indirect effect that is the key so I would report that it doesn't reach significance at the 5% level but at the 10% level. 

Eddie Ng posted on Saturday, April 23, 2011  1:21 am



Dear Bengt, Thank you very much for your reply, but it brings me to a more basic question: Re 2, I like your point of avoiding the STDYX. I also have read some articles (introducing me to doing specific mediation using Mplus) which do not report STDYX, but just the normal confidence interval. Why it confuses me is the different result when normal CI or STDYX CI is used. In my case, both CI and STDYX CI report the consistent pattern of result of 90% CI (both do not cover 0, i.e. significant) and 99% CI (both cover 0, i.e. nonsignificant) for my two ¡°borderline¡± mediation path. But it differs in the 95% CI. If the normal CI is used, then the range of 95% CI will not cover 0 (i.e. significant mediation effect), but the 95% CI will cover 0 if I report the STDYX result (i.e. nonsignificant). So which one is correct? If STDYX is not compulsory and essential, do you think it will cause problem if I just report normal CI and claim the mediation effect is significant? Or is it better to report the sig. level at all level of CI for the reader's reference? (The journal that I plan to submit is psychology journal rather than statistic journal) Many thanks 


There could be small differences between the results for standardized and unstandardized effects. This is due to them having different sampling distributions, with one possibly being more normal than the other. I would report as much as possible of what you have found to give the reader an appreciation for the results. I would also try Bayesian analysis here (just say Estimator = Bayes). See my paper Muthén, B. (2010). Bayesian analysis in Mplus: A brief introduction. Technical Report. Version 3. which is on our website under Papers, Bayesian Analysis. See especially the first path analysis examples for the ATLAS and Firefighter data. With Bayes you can see the whole parameter estimate distribution (the "posterior"), for both unstandardized and standardized estimates, and you can judge which approximates the assumed normal distribution the best. This is a good alternative to bootstrapping when estimate distribution are possibly not normal. 

Eddie Ng posted on Sunday, April 24, 2011  8:51 am



Thank you very much Bengt for the reference and kind suggestion. Also appreciate your persistence to perfection. 


Dear Dr. Muthén, I want to estimate some total indirect effects in a two wave latent different score model. I have three latent variables, each measured two times (af2, af3, depr2, depr3, sk2, sk3) Given the following model specification: DeltaAF by af3@1; af3 on af2@1; af3@0; [deltaAF af2]; ... .. I estimate the following paths: DeltaSK ON sk2 af2 deltaAF; sk2 ON af2; DeltaAF ON af2 depr2 deltaDEPR; af2 ON depr2; deltaDEPR ON depr2; When I want to estimate the indirect path from depr2 to DeltaSK, the total effect is confounded with (for example) depr2 > af2 > deltaAF > DeltaSK. This effect has the opposite direction than depr2 > af2 > DeltaSK, because af2 > DeltaAF is negativ (cealing and floor effects). So the total indirect effects does not make sense. Is it possible to exclude such indirect effects? Or do I have to calculate the total indirect effects? And if so, how can I calculate the SE for a sum of specific indirect effects? Thanks and best regards Christoph Weber 


If you want to sum certain indirect effects that are not automatically summed using MODEL INDIRECT, you can use MODEL CONSTRAINT to specify these indirect effects and then create a sum using the NEW option. You will obtain a standard error in that way. 


Thanks, it works perfect for the unstand. effect. But how can I get the stdyxsolution? I used additional constraints for the variance of the indep. variable and the dependent variable. But the contraint for the dep. var refers to the residual variance, so this does not work. Are there other possibilities to obtain the stdyx solution? Thanks Christoph Weber 


You would need to create the variance as a new parameter in MODEL CONSTRAINT. 

chidi O posted on Monday, May 14, 2012  1:42 pm



Hi, In addition to my SEM model am also trying to get direct or indirect effects. I am hoping to get direct effect from x2 to x4. My observed variables are categorical variables. But mplus keeps rejecting with this warning "Unknown group name INDRIECT specified in groupspecific MODEL command" MODEL: f1 BY x1 x2 ; f2 BY x3 x4; model indirect: x4 IND x2... (gets rejected) f2 IND x4 x2;...(also gets rejected) Not sure what am doing wrong. Thanks, Chidi 


Are you using an older version of Mplus that did not have MODEL INDIRECT? 

chidi O posted on Monday, May 14, 2012  2:49 pm



Am using version 6.12. The below specifications ran but gave 0 estimates and SE. MODEL: f1 BY x1 x2 ; f2 BY x3 x4; model indirect: x4 ind x2; I am trying to get a direct path for the 2 observed variables (x2, x4) to avoid letting them covary. Thanks 


Please send the output and your license number to support@statmodel.com. 


Dear Professors Muthén, I would like to enquire about the indirect effects for a nominal dependent variable. I know, that MPlus v6 does not calculate it, but I am hoping you can advise me on literature regarding this problem. (The article I know from Sobel (1987) only deals with linear structural equations, unfortunately.) Here is my syntax: NOMINAL is jp; MODEL: EGAL by e1 e2 e3 e4 e5; EGAL on gen male educ inc empl; jp on EGAL male educ inc empl; jp on gen male educ inc empl; I am interested in the indirect effect of jp on gen via EGAL. Thank you and kind regards, Zsofia Ignacz 


See the following paper which is available on the website: Muthén, B. (2011). Applications of causally defined direct and indirect effects in mediation analysis using SEM in Mplus. 


Dear Professor Muthén, Thank you for your answer. Based on the recommendation of the paper, I tried to run a model. I think I prompted Mplus to calculate the indirect effects, but in the output I got some disturbing results. In the STDYX and STDX Output, my mediating variable seems to be fixed at 1: JP#1 ON EGAL 1.000 0.000 ********* 0.000 This suggests for me that I do not have a correct syntax. I ran the following model: [...] NOMINAL is JP; MODEL: EGAL by e1 e2 e3 e4 e5; EGAL on gen (p0) ; JP on EGAL (p1, p2, p3); MODEL CONSTRAINT: NEW (ind1, ind2, ind3); ind1=p0*p1; ind2=p0*p2; ind3=p0*p3; I would be very greatful, if you would advise me on what I need to fix. Also how I can calculate the standardized coefficients for the two indirect effects? (The example in UG 5.20 has me a bit confused) Thank you, Kind regards, Zsofia Ignacz 


With a nominal final outcome, indirect effects cannot be computed as a product. See again the paper I referred you to for the correct way to get indirect effects in this case. 


Hello I am trying to run a multiple mediation model (with model indirect) with bootstrapping. I am looking at three mediators (m1, m1, m3) and whether they mediate the relation between x (pretreatment level of problems) to y (posttreatment level of problems). I ran the following model indirect paths: Y IND M1 X ; Y IND M2 X ; Y IND M3 X ; In the output I requested : CINTERVAL (BCBOOTSTRAP) and SAMPSTAT I have a few questions about this: 1. Under the title of CONFIDENCE INTERVALS FOR TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT AND DIRECT EFFECTS ¡V I seem to only get the Sum of Indirect and Specific Indirect effects „³ How can I get the estimates and CIs for the Total and the Direct effects as well? 2. I get a significant CI for the Sum of Indirect, which I am interpreting as the Total Indirect suggesting that taken together the three tested mediators have an indirect impact on change from X to Y. Is this correct? 3. How can I get the variance accounted for the Sum of Indirect and/or each Specific Indirect? (I assume the R quared in the output pertains to the direct paths to Y?) 3. Two of the specific Indirect effects are also significant  is it okay to interpret these as significant mediators? 


Further to my question above, I am wondering whether I also need to say something about the direct effects of X on Y? a) where to I find information about whether X on Y is significant in this model? b) if X on Y is significant in the meditational model, yet there are also the indirect effects described above does this suggest partial mediation? Or? Thank you so much for your help! Monica 


Dear Professor Muthén, Thank you for your answer. I have reviewed the suggested paper several times, as well as corresponding papers and it is not clear to me: should I be building up my analysis for the continuous meditator with nominal outcome from the example, where there is a continous mediator and a binary outcome OR from the example, where a nominal variable is the mediator variable on a continuous outcome? (I have mainly experimented with the first version). Thank you again for your help, Zsofia Ignacz 


Monica: In the future, please limit posts to one window. 1. Use the specification y IND x where only the final outcome and covariate are mentioned. 2. Yes. 3. This is not available. 4. We can't provide this detailed interpretation of the results. 


Zsofia: Create a set of dummy variables for the nominal variable and use the continuous mediator and binary outcome. 


Dear Prof. Muthen, thank you so much for your response and my appologies for the multiple messages! I just have a followup question  would you recommend calculating the total effect (sum of the indirect and direct effect) given that without a CI it is not possible to interpret? Or is reporting the total indirect and specific indirect effects suffcient? Thank you so much again! Monica 


Dear Professor Muthén, Thank you very much for your answer! I already considered this, I was just hoping that I could condense my model and I overlooked something...Then I will try to calculate the estimates by hand, taking into consideration the bias of the logistic function. Kind regards, Zsofia Ignacz 


Monica: Confidence intervals are given for total effects as are standard errors. 

Jiyeon So posted on Wednesday, June 20, 2012  1:19 am



Hi Prof. Muthen, I have a path model that looks like this: A > B > C > D. If we define standardized coefficientsfor each path as: a = A > B path b = B > C path c = C > D path If the direct effect of A on D was nonsignificant, the total effect of A on D is: (a)*(b)*(c) [the product of the three path coeffieints] Is this product still standardized beta? If not, what should this product (indirect effect) be labelled as? I am having difficulty interpreting my result since multiplicative product of the three paths was .01 and I'm not sure if that should be interpreted as a path coefficient (or is it the amount of variance explained as R square is in regression contexxt?). I would very much appreciate your feedback! Thank you very much. 


I would not use the standardized coefficients. I would multiply the raw coefficients and standardize by a and d. So multiply by the standard deviation of a and divide by the standard deviation of d. This is a regression coefficient not an Rsquare. 

Bridget posted on Monday, July 02, 2012  7:57 am



Dear Prof. Muthen, we estimated a path model with one dummycoded predictor variable (X), one continuous mediator (M), and two continuous criterion variables (Y, Z). We used a biasbased bootstrapping procedure (in MPlus Version 5.2) and found indirect, but no direct or total effects from X on Y and from X on Z. We would like to know why we have significant indirect, but no total effects. We know that total effects can be suppressed when indirect and direct effects are opposite in sign. This is, however, not the case in our analysis (the total effects are greater than the corresponding indirect effects). We noticed that the standard errors of the direct and total effects (from X on Y/Z) are much greater than the standard errors of the indirect effects and would be pleased if you could help us understand why this is so. 1.Why are the standard errors of the direct effects higher than the standard errors of the indirect effects? 2.How is the standard error of the total effect (X>Y) estimated in MPlus? In what way does the standard error of the direct effect (X>Y) influence the estimation of the standard error of the total effect? 3. By comparison: How is the standard error of the indirect effect estimated in MPlus? 4. Do you have any other ideas on how to explain why we have significant indirect, but no total effects? Many thanks in advance and best regards! 


Please send your output and license number to support@statmodel.com. Standard errors of the direct effects are estimated using the Delta method. 

Jen posted on Friday, March 08, 2013  1:06 pm



Hi Linda, Hope you are well! My coauthors and I have been puzzling over a similar situation to that above, where the indirect effect is significant but the total effect (which is larger) is nonsignificant due to a much larger SE. I wondered if there was an explanation for this or if, based on your request for the output above, this might indicate an issue with our models? I am happy to send the output and license number if relevant, but wanted to see if this was actually an issue or if there might be some explanation first! Thank you! 


This can happen. It is not indicative of a problem. 


Dear Professor Muthén and Muthén: How do you recommend I address this error message from Mplus: ALL INDIRECT EFFECTS CANNOT BE COMPUTED BECAUSE THE ABSOLUTE VALUE OF AN EIGENVALUE OF THE REGRESSION MATRIX IS GREATER THAN OR EQUAL TO 1. All I actually need in my observed variables structural equation model are the total effects of all the predictors (x and y variables) on one other yvariable of interest. I don't need the indirect effects. I note that my model is a nonrecursive model (y6 >y7 and y7>y6). Thank you! 


This is explained here https://books.google.com/books?id=9ACs50RjacC&lpg=PA34&ots=lVvDZze0J&dq=%22The%20stability%20index%20%20is%20the%20largest%20eigenvalue%20of%20the&pg=PA33#v=onepage&q=%22The%20stability%20index%20%20is%20the%20largest%20eigenvalue%20of%20the&f=false I would recommend you look for a model modification that resolves the problem. 

Back to top 