Message/Author |
|
|
Hi: I have a simple path model (3 IVs, 3 DVs) but for some reason MPlus is calculating intercepts for the DVs without my asking. note: x3 should have no DVs associated with it. Code is: x1; x2; x3; y1; y2; y3; x1 with x2 x3; x2 with x3; y1 on x1 x2; y2 on x1 x2 y1; y3 on x1 x2 y1 y2; This should have 24 free parameters and 18 estimated, leaving 6 df. However, MPlus is estimating intercepts for y1-y3, leaving only 3 df. How can I tell MPlus NOT to estimate those intercepts? |
|
|
Apologies. The above post was incorrect. The model has 6 variables, so there are (6*7)/2 = 21 unique elements. So 3 degrees of freedom is actually correct because I am requesting 18 parameter estimates. Even though the df is correct, it still SAYS it is estimating 24 parameters (not 18). This is because it is estimating means (for all IVS) and intercepts (for all DVS). Given that the degrees of freedom are correct, is this something to worry about? |
|
|
Is there a reason you are mentioning the variances and covariances of the observed exogenous variables in the MODEL command. The model is estimated conditioned on these variables. Their means, variances, and covariances are not model parameters. For some estimators you can say MODEL=NOMEANSTRUCTURE in the ANALYSIS command. Note that having unstructured means, intercepts, and thresholds does not affect model fit. |
|
|
I was taught by my institution to include the variances and covariances of IVs in the model command. This is inclusion / exclusion of this in the model command has not historically mattered, so I typically include them. If this is not impacting model fit, then that works for me. Thank you! |
|
|
Having means, variances, and covariances of observed exogenous variables in the model has no impact only when all observed dependent variables are continuous and there is no missing data. In all other cases, it changes the model. In regression, a model is estimated conditioned on the observed exogenous variables so no distributional assumptions are made about them. Their means, variances, and covariances are not model parameters. |
|
|
Thank you. One additional question: When the variances and covariances were removed and the "MODEL=NOMEANSTRUCTURE" option was used, it still calculated intercepts and means and gave me the following error: *** WARNING in ANALYSIS command MODEL=NOMEANSTRUCTURE is not allowed in conjuction with INFORMATION=OBSERVED. Request for MODEL=NOMEANSTRUCTURE will be ignored. Is this because I am using observed variables? |
|
|
No, this is the information matrx. You would need to change this to INFOMRATION=EXPECTED. I think you should simply leave the meanstructure in as it does not change model fit. |
|
|
Ah, that makes sense. Thanks for the helpful feedback. |
|
|
This solved the issue. Given that there was missing data, I removed the mean structure. No estimates really changed, however. If you have a reference for more information, I will happily follow up. Thanks again, |
|
|
Hello, I am conducting a path analysis using all continuous variables, bootstrapped standard errors, and indirect effects. In my output I noticed that one of my intercepts is negative and nonsignificant. How should I interpret this finding, and is this a cause for concern? Thank you for your time! |
|
|
Intercepts are not usually interpreted. I would not be concerned. |
|
|
Dear Dr. Muthen Do I understand correctly that I can use the MODEL=NOMEANSTRUCTURE to exclude the intercepts from a path model? What additional changes do I have to make in the Mplus command? Thank you for your help! |
|
|
In some cases you can do this. A model with unstructured means is the same as a model without means so it is not necessary to exclude them. |
|
|
Follow up question: I would like to exclude the intercepts to run regressions through the origin (as can be done in SPSS with the command /origin). I would like to do that as I want to test the mediated effect of four experimental groups on behavior. As I want to test all the groups against each other, I do not want to use three coding variables representing the four groups but I would like to enter the four groups as four independent variables. This can easily be done within SPSS by exactly using the command “/origin”, so that the regression is calculated through the origin and the four groups can be represented by an independent variable each. I would like to do the same now within Mplus to run a mediation model (and not only the direct effect which I tested in SPSS). I hoped that I can do that by applying the command MODEL=NOMEANSTRUCTURE. Unfortunately, this is not possible and I get continuously the following warning: WARNING: THE SAMPLE COVARIANCE OF THE INDEPENDENT VARIABLES IS SINGULAR. PROBLEM INVOLVING VARIABLE E8PCHW5. NO CONVERGENCE. SERIOUS PROBLEMS IN ITERATIONS. CHECK YOUR DATA, STARTING VALUES AND MODEL. Is there an option to run regression through the mean within Mplus? Can I use the MODEL=NOMEANSTRUCTURE but then my question is: why do I get this warning. Thank you so much for your help! |
|
|
This should work - in Mplus you would fix the intercept to zero. Please send your data, input, output and license number to support. |
|
|
This is exactly what I do not know how to do: fix the intercept to zero. In the SPSS handbook I only found references reg intercepts fixed to zero by default. But how can I do it purposely? |
|
|
If you have a regression y ON x; you would fix the intercept simply by saying [y@0]; If you continue to have problems, send to support. |
|
Majel Baker posted on Wednesday, August 09, 2017 - 4:22 pm
|
|
|
Hello Drs. Muthen, I am estimating a model with 8 indicators and 5 latent factors, only some of which are correlated with each other. This is a total of 31 parameters (6 variances, 10 paths, 7 latent factor covariances, 8 error variances). With 8 indicators, I'm provided 36 vars/covars, so my df should be 36 - 31 = 5 df. But I see Mplus is estimating the means of the indicators; this puts me up to 31 + 8 = 39 parameters, and my model becomes under-identified. I get the error message: " NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED." How can I tell Mplus not to estimate the indicators' means? Using NOMEANSTRCTURE and INFORMATION=OBSERVED have not helped it converge. Here is my code. Thank you for your help. DATA: FILE IS xy.csv; LISTWISE=ON; VARIABLE: NAMES ARE RUT1 SCT1 PCT1 DAHT1 RUD SCD PCD DAHD; USEVARIABLES ARE RUT1 SCT1 PCT1 DAHT1 RUD SCD PCD DAHD; MISSING ARE ALL (-999); ANALYSIS: TYPE IS general; ESTIMATOR = ML; ITERATIONS = 20000; MODEL=NOMEANSTRUCTURE; INFORMATION=EXPECTED; MODEL: Rum BY RUT1 RUD; CS BY SCT1 SCD; PC BY PCT1 PCD; DAH BY DAHT1 DAHD; Retro BY RUT1 SCT1 PCT1 DAHT1; Daily BY RUD SCD PCD DAHD; Rum WITH CS PC DAH; CS WITH PC DAH; PC WITH DAH; Retro WITH Daily; Rum CS PC DAH WITH Daily @0; Rum CS PC DAH WITH Retro @0; |
|
|
When Mplus estimates mean parameters it also used sample information on the means so the degrees of freedom should not change (unless your model has a mean structure). Regarding the non-convergence, you can look at the (non-converged) estimates that are printed to see if something unexpected is seen. If that doesn't help, you can send your output to Support along with your license number. |
|
Majel Baker posted on Thursday, August 10, 2017 - 3:23 pm
|
|
|
Thanks, Dr. Muthen. My model does not have a mean structure, as far as I'm aware of. Looking at the non-converged estimates, running it with NOMEANSTRUCTURE looks like it does take out estimating the intercepts, because I don't see them in the non-converged estimates, indicated by [] notation, and I don't see them [] anymore. The output is indeed showing 31 parameters--factor loadings (BY), covariances (WITH), and error variances and latent variances (*). So, I still can't understand why I'm getting "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED." I'll send my output and license to support@statmodel.com. |
|
|
Note that non-convergence does not have to do with non-identification. |
|
Back to top |