I have conducted mediation analysis with INDIRECT using bootstrapping and I am interested in whether the significant mediation holds in the sub-groups. Can I use INDIRECT in multiple group analysis? I have done the multiple group testing up to the structural level per the UG recommendations. Thanks, Sue
We are testing mediation models and we would like to see if there are sex differences on the IND pathway (multi-group analyses).
When searching on the Mplus forum discussion, we found this solution:
'If you have the estimate of the mediated effect and its SE for each of the 2 groups, you can simply use those numbers to create the approximately normal test variable:
(e1 - e2)/(se(e1-e2)),
where the denominator is sqrt(var(e1-e2)), where var(e1-e2) is var(e1) + var (e2), where var(e) is the square of the SE(e).'
Our questions are: (1) Is there another way to conclude about differences between groups on the mediated effect (i.e., Mplus command)? (2) If the formula above is the only way to conclude about differences between groups, is there a reference we could use?
You can use MODEL TEST to compute a Wald test or you can do difference testing of the model where the parameter of interest in free across groups compared to a model where the parameter is held equal across groups. Difference testing is discussed on pages 434-435 of the Version 6 Mplus User's Guide.
I have a question regarding an error I get when i try to run the following model: INPUT INSTRUCTIONS
data: File is P.csv; variable: Names are P I M; Usevariables are P I M; Categorical are I M; Analysis: Type=general; bootstrap=5000; Model: I ON M P; M ON P; Model indirect: I IND P;
*** ERROR The length of the data field exceeds the 40-character limit for free-formatted data. Error at record #: 1, field #: 1 *** ERROR The number of observations is 0. Check your data and format statement. Data file: P.csv
Dear Linda, thanks a lot for your fast response, I figured it out and Mplus is now working like a dream ;-)
Please, can you advice me how to report the results below (A mediation-analysis with one dichotomous variable) And I need to report confidence intervals for the bootstrap analysis. I am a bit in a doubt how to report it. I haves search APA, but without luck. Can you help me???
CONFIDENCE INTERVALS OF TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT, AND DIRECT EFFECTS
Thanks a lot, so to get it clear (sorry, but I am new in the realm of mediation-analysis ;-)
My mediator is the MAS-variable (dichotomous), so I report the point estimate -0.082 and 95% CI [-0.151, -0.024], right (see above). And then the S.E. Est./S.E. as well for the MAS variable? and that will be it?
I am wondering whether it is possible to do mediation analysis in Mplus with a categorical variable consisting of 4 subcategories? I have done it with a dichotomous variable, but would be very excited if it is possible with a 4-category mediator variable as well????
Dear Linda, Thanks a lot, it sure looks very interesting and promising for future research possibilities...
I have looked into the article and the appendix as well, and I can see that there is syntax for a 3-categorical mediator, but can you in anyway help me with a syntax, if my mediator consist of 4 categories? I would be very grateful since I have some very interesting psychological data where it would be superb with this opportunity.
I'm using all observed variables in a path analysis with multiple groups (2 groups). I have the following in the model: m1 m2 on x; y1 y2 on m1 m2 x;
model indirect: y1 ind x; y2 ind x;
model lowses: y1 on x (a1); m1 on x (b1); y1 on m1 (c1); y1 on x (a3); m2 on x (b3); y1 on m2 (c3);
model highses: y1 on x (a2); m1 on x (b2); y1 on m1 (c2); y1 on x (a4); m2 on x (b4); y1 on m2 (c4);
model constraint: new (f g); f=(b1*c1)+(b3*c3); g=(b2*c2)+(b4*c4); model test: 0=f-g;
However, y1 and m1 are based on similar measures (across time) and y2 and m2 are based on similar measures (across time). Should I run 2 models (one for y1 and m1 and one for y2 and m2) or keep the analysis as one (as above)? If I run it as above, I need to run one model constraint for predicting the indirect effects for y1 across the 2 groups and another for predicting the indirect effects for y2, right? Thanks!
Thanks so much for your response! These boards as well as the manual help so much.
I was trying to re-run my model using Bootstrapping and CINTERVAL in the output. I know you said that MODEL TEST cannot be used and so 0 = f - g needs to be brought into the MODEL CONSTRAINT field. However, when I do this, I get an error of convergence: NO CONVERGENCE. SERIOUS PROBLEMS IN ITERATIONS. ESTIMATED COVARIANCE MATRIX NON-INVERTIBLE. CHECK YOUR STARTING VALUES. Do you have any advice?
Hi, I was wondering if it is computionally right to test with MODEL INDIRECT a model in which a latent dimension has an effect on a DV via a second latent dimension of which the DV is one of the observed variable. From a tehoretical perspective it make sense, but I am not sure if I cant translate it into SEM, that is:
f1 BY x1 x2 x3; f2 BY z1 z2 dv1;
f2 on f1 (7); dv1 on f1 (8);
MODEL INDIRECT: dv1 on f2 f1;
In addition I wanted to test with a multi-group approche whether the mediation model holds in different group. Thus I constrained the relation between variables to be equal across groups.
Greetings - I'm curious as to whether a solution was discovered for the problem reported here Friday, January 20, 2012 - 8:55 am. I have a model of multiple group mediation, and when the testing statement is brought into the model constraint statement, I get the NO CONVERGENCE. SERIOUS PROBLEMS IN ITERATIONS. ESTIMATED COVARIANCE MATRIX NON-INVERTIBLE.CHECK YOUR STARTING VALUES. error messages.
Dear Dr. Muthen, Thank you for your help. I tried to run the model as you suggested but I am still getting the same error. I tried this alternative model and it did run: MODEL CONSTRAINT: NEW(a1a2 b1b2 c1c2 diff diff1 diff2); a1a2=a1*a2; b1b2=b1*b2; c1c2=c1*c2; diff=a1a2-b1b2; diff1=b1b2-c1c2; diff2=c1c2-a1a2;
The output gave me these new parameters: New/Additional Parameters A1A2 -0.023 0.011 -2.079 0.038 B1B2 -0.008 0.006 -1.300 0.193 C1C2 -0.032 0.021 -1.549 0.121 DIFF -0.014 0.012 -1.162 0.245 DIFF1 0.024 0.021 1.105 0.269 DIFF2 -0.009 0.023 -0.405 0.685
Is there any other way of testing whether these are significantly different from each other?
One follow up question, i ran two models one with the direct paths constrained and one with them unconstrained. There does not seem to be significant change in model fit. If the direct paths are the same across group, couldn't I just assume that the indirect paths are the same as well?
I ran a multiple group path analysis with two groups. The chi squared difference test between the unrestrained and the restrained models was not significant. However, the patterns of mediation are different between groups. I have three mediators, so for example in one group the mediation is significant through two mediators but in the other group it's only significant through one. Are these differences significant? Or are they not since the models are not different overall? Any advice you have on how to interpret that would be appreciated! Thank you!
I am currently running a multiple group mediation model to test whether the indirect paths significantly differ across gender.
Here is the model I am specifying between my latent variables. I want to run one model (below) with freely estimated paths across gender and compare this to the fit of a model with equated indirect paths across gender (e.g. boys a1 and girls a1).
MODEL: socanx10 ON scdiff; socanx13 ON scdiff;
ovvic ON scdiff(a1); relvic ON scdiff(b1);
socanx10 ON ovvic(c1); socanx10 ON relvic(d1); socanx13 ON ovvic(e1); socanx13 ON relvic(f1);
MODEL girl: socanx10 ON scdiff; socanx13 ON scdiff;
ovvic ON scdiff(a2); relvic ON scdiff(b2);
socanx10 ON ovvic(c2); socanx10 ON relvic(d2); socanx13 ON ovvic(e2); socanx13 ON relvic(f2);
MODEL constraint: NEW (ind_go ind_go2 ind_gr ind_gr2 ind_bo ind_bo2 ind_br ind_br2);
I ran a multiple group path analysis with two groups. I tested whether the indirect effects were significantly different using likelihood ratio tests and they were not. In the output I see the indirect effects separated by group, so I'm not sure which I should report. Should I run the analysis without the grouping variable and report those indirect effects?
I've been running a multiple group path analysis on a complex survey data. The BRR are the replicates weights used in the dataset. Since there are some missing, I used the default estimator declaring a multivariate normality assumption. I've tested also the difference between boys and girls on specific indirect effects by using MODEL CONSTRAINT. Checking the output, the parameter relative to the difference was statistically significant. I kindly ask whether the computation of the standard error of this parameter takes into account the sampling design, considering the BRR weights.
Dear Dr Muthen, We are testing mediation models (two mediators) and we would like to see if there are cohort differences on the IND pathway (multi-group analyses with three groups). We have a dichotomous outcome. According to your commend earlier in this thread you could test the difference of indirect effects between groups with model constraint in two ways. First, obtain indirect effects in model constraint: a1b1 = c1 * a2; aa1bb1 = cc1 * aa2; model test: a1b1=aa1bb1; We do this for all indirect effects. This works fine. However, it is not possible to use the bootstrapping command in combination with model test. Is this a concern at all? The second way is to define this command in MODEL constraint: 0 = a1b1 - aa1bb1; We can use bootstrapping. However, we are getting this error: THE DEGREES OF FREEDOM FOR THIS MODEL ARE NEGATIVE. THE MODEL IS NOT IDENTIFIED. NO CHI-SQUARE TEST IS AVAILABLE. CHECK YOUR MODEL. NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED. Can we conclude that the constraint does not work and therefore the bootstrapped indirect effects are not the same? Or does this mean that there is a problem with our model? How else could we obtain a significance test of bootstrapped indirect effects across cohorts? Thanks a lot in advance!
MODEL constraint: New(diff); diff = a1b1 - aa1bb1;
and look at the bootstrap CI for diff with
Daniel Lee posted on Monday, August 05, 2019 - 3:26 pm
Hi Dr. Muthen,
I am using the WLSMV estimator and bias corrected bootstrap confidence interval to test for group differences (two groups) in the direct and indirect paths in a simple mediation model.
I found significant group differences in the direct path from the mediator to the outcome (y), but did not find group differences between the indirect effects. Of note, the indirect effect was significant in one group, but not the other. I was wondering if this means that there is a significant difference from the mediator to the outcome (M to Y) but that this difference is not robust enough to influence group differences in the indirect effect?
If there are no significant group differences in the indirect effects, which indirect effect should I interpret (e.g., indirect effect from model where the groups are pooled)?
If a and b do not show group differences in a joint 2 df Wald test (Model Test), then use the model with them equal across groups.
Daniel Lee posted on Wednesday, August 07, 2019 - 8:39 am
Hi Dr. Muthen,
Thank you for the response. I have set equality constraints on all model coefficients between the two groups, and only freed the "a" and "b" path (to be unequal between groups). When I did that, the difference in chi-square (2 df) was significant.
Hello, We are aiming to test a mediation model based on latent variables formed from ordinal data (so have used WLSMV). We want to see if there are differences on the IND pathway for two groups (diagnosis/no diagnosis).
Based on what I've read so far, our current syntax is as below:
Y BY Y1 Y2 Y3 Y4 Y5 Y6 Y7 Y8 Y9 Y10 Y11 Y12; X BY X1 X2 X3 X4; M BY M1 M2 M3 M4 M5 M6;
Y ON X M; M ON X;
MODEL INDIRECT: Y IND X;
MODEL diagnosis: Y ON X (a1); Y ON M (b1); M ON X (c1);
MODEL no diagnosis: Y ON X (a2); Y ON M (b2); M ON X (c2);
MODEL constraint: new (D ND DIFF1); D=a1*b1; ND=a2*b2;
OUTPUT: CINTERVAL STAND TECH4;
Is this correct? Could you explain how you would bootstrap this?
I also wanted to ask what the appropriate way would be to test individual pathways within the model between the two groups - would this be DIFFTEST?
We are doing multiple group mediation with a latent continuous determinant, a continuous outcome and a dichotomous mediator (defined as categorical). We have been advised to calculate the path between the determinant en de mediator on a risk difference scale. We have difficulty calculating the risk difference scale in our multiple group model. We use the following formula; MODEL CONSTRAINT: new(im a1_rd a1b); im = SQRT((smok$1+1)^2); a1_rd = (exp(im+a1)/(1+exp(im+a1))) - (exp(im)/(1+exp(im)); a1b = a1_rd*b1; With the ”im = smok$1+1;” code we are attempting to save the intercept. Mplus seems to use the same estimate of 2.0000 each time, which corresponds with none of the intercepts. In addition we get this error; Missing matching right parenthesis. Is there a way to save intercepts of the individual multiple group models (even if they have the same name of smok$1) in order to obtain a risk difference scale for each group individually?
Thank you for this option. I was wondering whether I could ask a follow up question before sending the input. Is there are any other way to calculate indirect effect with a binary mediator that is in line with the counterfactual mediation framework to account for the scale difference between logistic and linear regression in one mediation model? How should I compute this is the model constraint command?
Muthén, B. & Asparouhov T. (2015). Causal effects in mediation modeling: An introduction with applications to latent variables. Structural Equation Modeling: A Multidisciplinary Journal, 22(1), 12-23. DOI:10.1080/10705511.2014.935843 Click here to download the paper.
Thank you very much! Giving smok$1 in the model command works!
I have another follow up question regarding mediation with one continuous latent variable as predictor.
We would like to compare the coefficients of two models: 1) Latent variable (X)--> outcome (Y) without controlling for the mediator (C) 2) Latent variable (X) --> outcome (Y) with controlling for the mediator (C').
However I have noticed that the factor loadings of the latent variable are different across models 1) and 2). In addition, C and C' sometimes behave in an illogical ways suggesting that the mediator is a suppressor while this is clearly not the case.
Is it possible to compare models 1) and 2) (thus C and C’) if there is a latent variable in the models? What would you advise us to do if we want to compere C and C’ in this situation? Is it a problem when they are illogical when the indirect effects are properly calculated?
If 1) and 2) have different factor loadings, this suggests that the model isn't fitting the data well enough in one or both cases. For instance, the correlations between the factor indicators and the mediator may not be well represented by the model.