Jen Bailey posted on Tuesday, April 06, 2004 - 11:36 am
Dear Dr.s Muthen: I am running an SEM with non-normally distributed indicators on latent factors (indicators are frequency of substance use). I would like to run the analysis using ML to see, for curiosity's sake, how big a problem it is to use ML in this situation. I was hoping to output a file containing residuals, so that I could examine their distribution. I've looked in the new manual and in existing conversations at statmodel.com, but haven't found how to do this. Is it possible? Does Mplus 3 offer some other way of checking whether residual distributional assumptions are met?
bmuthen posted on Tuesday, April 06, 2004 - 12:01 pm
Mplus currently does not output individual residuals. To check effects of using linear models with non-normal indicators you can instead compare regular linear modeling with ML to (1) linear modeling with MLM, (2) non-linear modeling treating the outcomes as ordered categorical, (3) non-linear modeling treating the outcomes as censored, or (4) non-linear modeling using a non-parametric mixture approach. The last 2 approaches are available only in Version 3.
Jen Bailey posted on Tuesday, April 06, 2004 - 1:06 pm
Thanks for your response! I'll check out the methods you suggest.
Jen Bailey posted on Wednesday, May 19, 2004 - 11:11 am
I'm working on an SEM with ordered categorical indicators for most of my latent variables (I also have 2 latents with continuous indicators). I'm using the WLSMV estimator and Delta parameterization. In my output, I see that I have a Heywood case. I have fixed the factor loading for the problematic indicator to 1, but the Heywood problem persists. The variable in question is dichotomous, with about 94% of cases having a score of 1, and the other indicators of the factor are trichotomous.
I'm not sure what the implications of this are. Because I'm using Delta parameterization, I can't fix the error variance to 0. Do I need to switch to Theta parameterization? Or should I not worry about the Heywood case since I've fixed the factor loading to 1? Your thoughts would be greatly appreciated.
bmuthen posted on Wednesday, May 19, 2004 - 11:22 am
You can fix the factor loading to a somewhat smaller value until you get away from the Heywoood case if that doesn't significantly worsen the fit. You can stay in the Delta parameterization.
Dustin posted on Thursday, July 01, 2004 - 11:18 am
I am getting the following warning for a strucutral model using a binary manifest variable ("ET") to predict a continuous latent variable.
WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE ET.
Is this something I should be concecerned about. All parameters within the model are within reasonable values and the overall model fit is good.
David Bard posted on Tuesday, April 18, 2006 - 11:15 pm
Where do I find more detail on how M+ calculates the residual covariance matrix (using residual option) under various estimators? Could you briefly explain why this matrix does not equal the difference between observed and estimated covariances when using MLMV?
The residual covariance is the difference between the observed and estimated covariance matrices for all estimators. If you send your input, data, output, and license number to firstname.lastname@example.org, I can see what is happening.
It depends on the scale of the dependent variable and for categorical outcomes whether there are covariates in the model. For continuous outcomes, the values are for covariances. For categorical outcomes in a model without covariates, the values are for correlations. For categorical outcomes in a model with covariates, the values are for residual correlations. The residuals are not standardized.
Jon Elhai posted on Saturday, January 13, 2007 - 3:16 pm
Drs. Muthen, To follow up on this email thread from several months ago... I have a couple of questions... 1) Regarding the "Residuals for Covariances/Correlations/Residual Correlations" in the OUTPUT: RESIDUAL for Mplus - are these also known as "covariance residuals" or "fitted residuals"?
2) You mentioned in the email thread that these numbers in this output are "unstandardized residuals." It seems that these correspond in my data with the unstandardized residuals that I obtained in AMOS. How can they be interpreted - is this what Kline's SEM book says that numbers greater than .10 (absolute value) indicate that the model doesn't explain the observed correlation well?
I'm not familiar with the terms covariance residuals and fitted residuals. The residuals in Mplus are the difference between the sample values and the model estimated values. They are not standardized so no absolute value would apply. You would look for large differences relative to the scale on which the variables are measured.
jemila seid posted on Wednesday, December 03, 2008 - 12:20 pm
Dear Dr. Muten,
I am wondering if can get individual residuals (residuals from fitting individual growth curves) from Mplus version 5.
J.W. posted on Wednesday, March 11, 2009 - 1:03 pm
In Mplus output, both Beta and Gamma parameter specifications are all shown in the BETA matrix in Tech1 Output (they are separate parameter matrixes in LISREL). In addition, the slope coefficient of regressing an endogenous indicator (e.g., y1) on an exogenous indicator (e.g., x1) in an extended MIMIC model testing DIF should also be a Gamma parameter (Kaplan, 2000, p.73). But it is also included in the BETA matrix in Tech1 Output. Is there a way to separate the BETA and GAMMA matrixes in Tech1 output? Thanks a lot for your help.
Mplus uses only the Beta matrix for regression coefficients. This has no impact on the model estimates.
Steve posted on Wednesday, June 26, 2013 - 1:47 pm
I am relatively new to Mplus.
I am needing to obtain the correlation residuals matrix - that is, the difference between the actual sample correlation matrix and the implied correlation matrix by the model (as is often recommended to find specific areas of misfit).
I have used another program before which provided both matrices and then I needed to calculate by hand. Is there way to get this information somehow in Mplus?
[I just purchased the mixture add-on thinking I could obtain this information from one of the more advanced Tech outputs.]
If it doesn't come out when your request RESIDUAL, I don't think we give that when analyzing a sample covariance matrix (as opposed to a sample corr matrix for EFA). And I'm not sure I would prefer that approach. Instead we give standardized residuals - and my preferred way to understand model misspecification, namely modification indices. In fact, a primary motivation for developing modification indices by Sorbom was that basing model modification on residuals can lead to the wrong model.
Sorry, I'm a bit confused about the RESIDUAL output. I'm conducting an equity analysis and in order to do so, I need the unstandardized residual variance produced by multiple linear regression. I'd like to use Mplus to deal with missing data appropriately.
Mplus gives "***" for the variance of the residuals in the unstandardized portion, and Mplus only provides numerical estimates for the standardized variance of the residuals.
Is there any way to see the unstandardized variance of the residuals, perhaps in one of the graphs? As Mplus does not save individual-level residuals, I can't save them and then compute this value.
Is there any way to ask Mplus to output the full covariance matrix with variances on the diagonal. By this I mean some way to format the output. The current version of MPlus outputs the lower half of the matrix only. Programs used to compute the CVI and other cross-validation indices (Cudeck and Browne, 1992 and MacCallum 1994) need the full matrix. Naturally, we can prepare the full matrix outside of MPlus using SAS PROC IML, however, it would save time. Also, have you considered at least providing a single sample CVI in MPlus which can be computed from the "F" discrepancy function based on F(S;Sigma_hat).