Message/Author 

Student 09 posted on Monday, January 16, 2012  4:33 am



Hi using the multiple group analysis in Mplus, I would like to test whether an indirect effect differs between two subsamples (group a & group b). For each group, I requested the indirect effect by using the model constraint command. MODEL: y2 ON y1 (p1a); y1 ON x (p2a); MODEL CONSTRAINT: NEW (ind_1); Ind_1 = p1a*p2a; MODEL group b: y2 ON y1 (p1b); y1 ON x (p2b); MODEL CONSTRAINT: NEW (ind_2); Ind_2 = p1b*p2b; This syntax yields the two indirect effects (ind_1 & ind_2). But now I wonder how to test their difference – how would I constrain these parameters to be equal, given that they come from two different groups? 


You should use a Model Constraint where you define their difference  this will give you the test as a z score: MODEL CONSTRAINT: NEW (ind_1 ind_2 diff); ind_1 = p1a*p2a; ind_2 = p1b*p2b; diff = ind_1ind_2; 


Hello, Can I run a moderated mediation analysis as a twogroup model, employing the MODEL INDIRECT command? In other words, can MODEL INDIRECT be used in conjunction with a grouping variable, allowing for tests of moderated mediation? Thanks! 


You can run moderated mediation using multiple group analysis but you can't use MODEL INDIRECT. You will need to use MODEL CONSTRAINT to compare indirect effects across groups. MODEL INDIRECT does not have this function. 


How do I test whether multiple indirect effects differ between 2 groups? The model includes paths from 2 Xs (ext2, stigma2) through a single mediator (ai_t) to 2 DV's (VDATE, PDATE). I used the model constraint language as per Bengt's response above to create the 4indirect diffs... MODEL CHILD: ai_t on ext_pt2 (a1a); ai_t on STIGMA2 (a2a); PDATE on ai_t (b1a); VDATE on ai_t (b2a); MODEL ADOLESCENT: ai_t on ext_pt2 (a1b); ai_t on STIGMA2 (a2b); PDATE on ai_t (b1b); VDATE on ai_t (b2b); MODEL CONSTRAINT: NEW (ind_1 ind_2 diff); ind_1 = a1a*b1a; ind_2 = a1b*b1b; diff = ind_1ind_2; NEW (ind_3 ind_4 diff); ind_3 = a2a*b1a; ind_4= a2b*b1b; diff = ind_3ind_4; NEW (ind_5 ind_6 diff); ind_5 = a1a*b2a; ind_6 = a1b*b2b; diff = ind_5ind_6; NEW (ind_7 ind_8 diff); ind_7 = a2a*b2a; ind_8 = a2b*b2b; diff = ind_7ind_8; Mplus will run a single constraint test but when I include more than one I get an error message saying that I've entered an already used parameter. Although true, each indirect is comprised of unique pairs of parameters. Is there a way to simultaneous test differences in the 4 indirect effects? 


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


Hello, In using the MODEL CONSTRAINT option to define new parameters for one's model, does this increase the number of parameters estimated and hence affect degrees of freedom? A fiveknownclass model in which I used MODEL CONSTRAINT to define an a*b indirect path (to test for mediation) had different degrees of freedom than I expected, and I wondered if this is because I added parameters to the model through using this command? Thanks! 


If you put a constraint on a parameter in MODEL CONSTRAINT, this can affect the degrees of freedom. If you define a new parameter, this should not. 


Thanks, Linda. Can I label parameters that are classspecific estimates in a knownclass model, then use those classspecific labels to constrain parameters within classes? For examplewould this work? !Whites %cg#1% [gender] (gen_m_w); gender (gen_v_w); !Chinese() !%cg#2% [gender] (gen_m_cn); gender (gen_v_cn); !MODEL CONSTRAINT: gen_v_w = gen_m_w*(1gen_m_w); gen_v_cn = gen_m_cn*(1gen_m_cn); 


Yes, you can label classspecific parameters. 


Hi Dr. Muthen, Similar to Student09 in Jan 2012 of this thread, I am interested in testing whether an indirect effect differs between two groups. On Jan 16 2012, you wrote: You should use a Model Constraint where you define their difference  this will give you the test as a z score: MODEL CONSTRAINT: NEW (ind_1 ind_2 diff); ind_1 = p1a*p2a; ind_2 = p1b*p2b; diff = ind_1ind_2; I ran my model with this syntax, but I want to be sure that I am looking in the right place in the output and interpreting correctly. My understanding is that I should be looking at the third and fourth columns of the model results section, under "New/additional parameters." This section of my output looks like this: New/Additional Parameters INDIRECT 0.003 0.024 0.144 0.885 INDIRECT 0.001 0.065 0.010 0.992 DIFF 0.003 0.069 0.039 0.969 My understanding is that for DIFF, Est./S.E = 0.039 and p = .969, meaning that the population means are equal across groups. This would mean that the indirect effect is not moderated by the variable I used to group (weight status). Am I interpreting this correctly? Thank you! 


That's correct. One cannot reject that the two indirect effects are equal. 


Thank you for your response. As a follow up question related to my model, I found partial measurement invariance across gender in my initial invariance testing (at the weak level). Once freeing one parcel loading on my independent latent variable, the chi sq diff tests held for weak and strong invariance. My understanding is that I would need to free this loading across gender groups in my structural model. However, I am already specifying two groups for my moderated mediation analyses. Can I also specify two separate groups by gender to account for that partial measurement invariance? If so, what would that syntax look like? And how would that change my interpretation of model fit and the diff score created to determine moderation? Thank you! 


Sounds like you would then have 4 groups (2 x 2). In which case you can e.g. specify genderinvariance of the slopes involved in the moderated mediation. Just label the parameters according to the equalities you want to specify. 


Is there a way to determine probability of finding group differences in a multiple group analysis, posthoc? I ran a moderated mediation model using a multiple group analysis and did not find significant moderation. This may be because one group only had 29 participants in it, compared to 95 in the other. I am wondering if I can determine the probability of detecting group differences, given the numbers that I had? Thank you! 


You can do a Monte Carlo simulation  see the UG chapter 12. This would give you an idea of the power to reject equality. 

Shiny posted on Tuesday, August 26, 2014  7:48 am



I d like to follow up with this discussion. I understand that when difference test Shows insignificant p value, it means the indirect effects are equal. Is the difference test in mplus a heterogenity test in its nature? I checked some statistical mateirals. Some researchers used a relatively subjective approach, comparing the Mediation sizes (the coefficients) or simply look at whether the indirect effects are full v.s parital v.s non effects in different groups. They sometimes skipped the difference test. or they consider the Group difference existed (e.g., when full Mediation in one Group and partial in another), although in the difference test result p is not significant. I would love to get a comment from Dr. Muthen. 


Q1. I don't know what a heterogeneity test is in this context. I think it is good to carry out a test. 

ri ri posted on Saturday, September 06, 2014  12:24 am



To compare two indirect effects, I ran the difference test. In one Group I have a Mediation effect, in the other Group, there is no Mediation effect. However, difference test showed no difference, because p>.05. I wonder how could that be? New/Additional Parameters ind1 0.657 0.311 2.113 0.035 ind2 0.567 0.397 1.427 0.153 diff 0.090 0.505 0.179 0.858 I guess the difference of the Parameters is small (.090), thus it gives a big p value? Can I Report the indirect effects differ across groups? Does it only make sense, to conduct a difference test when indirect effects exist in both Groups? Maybe in my case, when in one Group there is no indrect effect, I do not Need a difference test? 


I think it is natural that if ind2 is not significantly different from zero, the ind2ind1 difference is not significantly different from zero  the latter distance is smaller. I would report all 3 findings. 

ri ri posted on Saturday, September 06, 2014  11:05 am



Thank you for your answer! I will Report all the findings as you suggested. But can I still say, there is a moderated Mediation because the Mediation effects are not the same in two Groups. may I ask the statistical Background of the ind1  ind2? I think I read somewhere this is a z score test?I would like to also indicate in my paper what the difference test means in mplus. Thanks! 


If ind1 and ind2 are not significantly different you can't say that you have moderated mediation. The estimate/SE for "diff" is a Zscore. 

ri ri posted on Friday, September 19, 2014  10:10 am



Thank you for the helpful Information. I am now summarizing the Analysis and Tools in my paper and want to add one sentence about difference test provided by mplus. can I say: we used a zscore difference test in Mplus to test whether indirect effects vary across Groups. Or is there a more precise term to describe this difference test? It is good to get your expert view! Thanks! 


It looks like you used MODEL CONSTRAINT to test the difference. The test of the difference is a zscore. There is nothing more to say. 


Dear dr. Muthén, I’m testing a moderated mediation model in which the association between an independent variable IV and a dependent variable DV is mediated by Med, and paths a and c are moderated by Mod. The moderation effect is represented by a product term (Int = IV*Mod). We used the following syntax: MODEL: DV on IV Mod Int; DV on Med (b1); Med on IV (a1); Med on Mod; Med on Int (a3); MODEL CONSTRAINT: new (indlow indmed indhigh indirect difind); !indirect effect at low, medium, and high value of moderator (moderator is centered) indlow=(a1+a3*0.5)*b1; indmed=(a1+a3*0)*b1; !also represents indirect effect regardless of moderation indhigh=(a1+a3*0.5)*b1; !total indirect effect indirect =(a1+a3)*b1; !difference between total indirect effect and indirect effect regardless of moderation difind = indirect  indmed; However, I get two error messages: *** ERROR (A1+A3*0.5) *B1 ^ERROR *** ERROR in MODEL CONSTRAINT command A parameter label has been redeclared in MODEL CONSTRAINT. Problem with: INDMED So there seems to be a problem with the way I define new variables in the MODEL CONSTRAINT command. What am I specifying wrongly here? Thank you for your help! 


Try changing indlow=(a1+a3*0.5)*b1; to indlow=(a1+a3*(0.5))*b1; 


Thank you, that get's rid of the first warning. The second warning still remains, however: *** ERROR in MODEL CONSTRAINT command. A parameter label has been redeclared in MODEL CONSTRAINT. Problem with: INDMED If I remove all the syntax pertaining to INDMED, the warning persists, but then starts blaming another new variable, INDIRECT. 


Then we need to see the output  send to support along with license number. 

ri ri posted on Monday, May 11, 2015  8:50 am



I did a difftest using MODEL CONSTRAINT.Is there a Name for the estimate coefficient? Here is an example of my result: DIFF2 0.117 0.548 0.214 0.830 I want to Report what .0117 represents for. I understand the difftest is a zscore test, but is it possible to get a z value in difftest? Thanks! 


The DIFF2 parameter is what you specified in MODEL CONSTRAINT. The third column is the ztest. 


I am having trouble with conducting a moderated mediation similar to user (student09). My two groups are gender (1 = male 2= female). Could you take look to see what I may be doing incorrectly? variable: names =.... usevariables = Gender...; missing = all(999); grouping = Gender (1=Male 2=Female); analysis: type = general; model: X by... Y2 on Y1(b1_Male) X; Y1 on X (b2_Male); model constraint: New (ind_Male); ind_Male = b1_Male*b2_Male; model: X by... Y2 on Y1(b1_Female) X; Y1 on X (b2_Female); model constraint: New (ind_Female); ind_Male = b1_Female*b2_Female; model test: ind_Male = ind_Female; Below is the error I am getting. *** ERROR in MODEL CONSTRAINT command The following parameter label is ambiguous. Check that the corresponding parameter has not been changed. Parameter label: B1_MALE Why would this parameter be ambiguous? Would I need to define the parameter first? Do I need to instead include the direct effect coefficients from my nongrouping analysis? I am unsure of whether this parameter is "made up" or defined in some way. 


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

Jayoung Gong posted on Wednesday, November 25, 2015  3:44 am



I'd like to know if this syntax is right to investigate my model. X is the predictor, Y is the outcome, M is the mediation, W and Z are the moderations.X>M is moderated by W, and M>Y is moderated by Z. My syntax is like below. ANALYSIS: TYPE = RANDOM; ALGORITHM=INTEGRATION; MODEL: X BY g1g6; W BY i1i2; M BY c1c3; Z BY s1s2; Y BY j1j2; XW  X XWITH W; M ON X(a1) W XW(a3); MZ  M XWITH Z; Y ON M(b1) Z X W XW MZ(b3); MODEL CONSTRAINT: NEW(ind1 ind2); ind1=(a1+a3*1)*(b1+b3*1); ind2=(a1+a3*0)*(b1+b3*0); OUTPUT: SAMPSTAT; TECH8; Is this right for moderated mediation model? 


For observed variables you would not use XWITH but instead use DEFINE to create the product. 

Jayoung Gong posted on Wednesday, November 25, 2015  5:59 pm



Thank you for the answer. in my model, there are all continuous variables. I assume they are all latent variables. (I use "BY" for them.) I'm quite confused if I need to use "DEFINE" instead of "XWITH" in this case. 


Sorry, I missed that you had X, M, Y latent. Yes, your input looks correct. 

Ofer Eldad posted on Sunday, May 27, 2018  12:27 pm



Hi Dr. Muthen, In moderated mediation, can I test the significant of the difference between two indirect effects through simple subtraction when the outcome is dichotomous in crossclassified data? Meaning, can I still use the following? diff = ind1  ind2; In general I am a bit confused as to how to calculate terms of interest in mplus given a dichotomous outcome. For example, when calculating indirect effect, should I go with ind = a*b; or is it more appropriate to approach it as an odds ratio ind = exp(a*b); Many thanks, ~Ofer 


Binary Y in mediation modeling needs special considerations. See our Topic 11 Short Course handout and YouTube video at http://www.statmodel.com/course_materials.shtml 

Ofer Eldad posted on Monday, June 04, 2018  9:11 pm



Thanks for the response. I saw the lecture and read more about using counterfactuallydefined indirect effect. Am I correct that I cannot find the indirect effect through multiple mediators for binary outcome through "model indirect"? Is there an alternative way to do this? Thanks. 

Ofer Eldad posted on Monday, June 04, 2018  9:22 pm



In addition, any suggestion on how to approach the problem when the data is crossclassified? That is, any way to compute the counterfactuallydefined indirect effect for cross classified data with a binary outcome? 


9:11 post: Q1: right Q2: How to do it in Mplus is shown in the article Nguyen, T.Q., WebbVargas, Y., Koning, I.K. & Stuart, E.A. (2016). Causal mediation analysis with a binary outcome and multiple continuous or ordinal mediators: Simulations and application to an alcohol intervention. Structural Equation Modeling: A Multidisciplinary Journal, 23:3, 368383 DOI: 10.1080/10705511.2015.1062730 


If you are using a crossclassified model, the potential outcome based approach has to be conditioned on both clustering effects. That is because the random effects would influence the computation of the indirect effect. You can use the factor scores if you decide to go that route but you will have to compute this for every pair of clusters from the two crossednesting levels. You might consider using the simpler approach from Chapter 2 in the Regression and Mediation Analysis book. 

DavidBoyda posted on Monday, April 06, 2020  12:53 pm



Hello, regarding the LOOP command, if my moderator has a small range 17, should i really be specifying a range that is larger as seen in the regression with interaction video? MODEL CONSTRAINT: LOOP(x,1,1,0.1); Plot(effect); effect=b1+b2*x; 


Which slide are you referring to here? 

DavidBoyda posted on Tuesday, April 07, 2020  12:40 am



This video here with Marten: https://youtu.be/cNDmABvCz3I?t=800 This is my plot. https://tinyurl.com/wgz23j4 


The video uses LOOP(x,1,1,0.1); where 1 to +1 is the relevant range for that example. If your moderator has the range 1 to 7, you should use those values. 


Dear Drs Muthen, I am conducting a moderation analysis which involves interpreting the New/Additional Parameters section of the output. I am conducting an analysis where the outcome is ordered categorical, and the predictor and moderator are continuous. My output is: SIMP_LO 2.693 0.602 4.473 0.000 SIMP_MED 2.488 0.561 4.435 0.000 SIMP_HI 2.284 0.527 4.334 0.000 I wondered what the best way would be to interpret this output. For example, are the estimates from this output in the format of odds ratios because of the ordered categorical outcome? Thank you in advance for your help. Best wishes, Stefania 


If you use ML where Link=Logit is the default, these estimates are in logit metric. Looks like you are expressing simple slopes are 3 moderator values. You can use Model Constraint to exponentiate these and get odds ratios for each of the 3. 


Dear Drs. Muthen, Thank you for your quick response, and apologies, I did not clarify. The outcomes have more than two groups so it is a probit regression that I have conducted. So I think the estimates are in probit metric? Is there a way to use Model Constraint to generate probabilities for these scores? Thank you in advance. Best wishes, Stefania 


Your output segment Summary of Analysis will tell you if you used probit or logit. Both can be used for categorical outcomes with more than 2 categories. The Plot command can be used to view the estimated probabilities. See also chapter 5 in our RMA book: http://www.statmodel.com/Mplus_Book.shtml 

Back to top 