Message/Author 

Xu, Man posted on Friday, June 18, 2010  11:14 am



I plan to fit some SEM models in Mplus (We have ordered an Mplus a while ago, but have not heard back). The measurement model has items that are treated as categorical. Then the factor is predicted by some other observed variables in the data. I was wondering if possible at all to get standard errors for the standardised output. I am partcuarly interested in the path coefficients, so would like to know the standard errors in order to know if the standardised path coefficients are statistically signficant or not. Many thanks! 


Standard errors for standardized estimates are available in most cases. The STANDARDIZED option is not available for models with random slopes defined using the  symbol in conjunction with the ON, AT and XWITH options or for parameters in MODEL CONSTRAINT involving variables named on the CONSTRAINT statement of the VARIABLE command. For weighted least squares estimation when the model has covariates, STDY and standard errors for standardized estimates are not available. For the MUML estimator, STDY and standard errors for standardized estimates are not available. 

Xu, Man posted on Friday, June 18, 2010  11:31 am



Thank you! Yes. I think the default esimator is WLSMV. It is not multilevel or with interaction. It is just SEM with the items forming the depedent variable declared as categorical. Then which estimator would you recommend to use for getting the standard errors for standardised solutions please? 


If you want to use WLSMV and the model has covariates, you would need to compute the standard errors yourself. You could also use maximum likelihood estimation with categorical outcomes where standard errors for standardized coefficients are obtained. The only drawback would be if your model has too many dimensions of integration. 

Sun Kim posted on Tuesday, October 25, 2011  8:23 pm



Hi, I am encoutering the same situation (WLSMV estimator and no standard errors), and I saw from previous posts that you suggested the delta method for computing standard errors but this method using software seems to require transformation function, the means, and variancecovariance matrix. Is this correct, and is there a more simple way to handcalculate the standard errors myself? 


You would need to use the Delta method to compute the standard errors in this case. There is not simpler way to do this. 


Hi, Is it possible to compute standard errors using Mplus output that is produced by SEMs that include latent variable interactions (e.g., that use the "xwith" command)? If so, how can I compute these standard errors? Since TECH4 is unavailable for latent variable interactions using the "xwith" command, I used Bengt's (2012) article entitled LATENT VARIABLE INTERACTIONS to compute standardized parameter estimates; however, I did not in that article see any mention of how to compute the associated standard errors for these estimates. Thank you for any help with this, and sorry if I'm overlooking something obvious here... Best regards, Luke 


You need to define it in MODEL CONSTRAINT. Then you will get a standard error. 


Thank you, Linda! 

Alissa Beath posted on Wednesday, August 13, 2014  11:03 pm



Hi Linda, I am fitting a SEM with all numeric observed variables and one latent variable in which the assumption of multivariate Normality is implausible. I wanted to use the bootstrapping option but find that SEs for the STDYX estimates are not reported, although they are when I do not use the bootstrap. My syntax is below in case that helps. Could you suggest how I could get SEs for standardized estimates or why that is not possible in this case? Many thanks, Alissa Analysis: bootstrap=2000; MODEL: teique on neuro extra agree open consc; avoid on teique; reapp on teique; supp on teique; reapp with supp; distress by stress anxiety; distress on avoid teique reapp supp; 


We don't bootstrap the standard errors of the standardized estimates. 

Vesna Gagic posted on Wednesday, July 26, 2017  5:30 pm



Hi Linda, I found a comment online by Jon Lefcheck (developer of an R package piecewise SEM) that "the standard errors on the coefficients are still a grey area: if you are concerned at all, refer to the SEs and Pvalues of the /un/standardized variables" (https://www.rdocumentation.org/packages/piecewiseSEM/versions/1.2.1/topics/sem.coefs) Would you know why or have any reference for that statement? In my individual models I have high collinearity when using unstandardized coefficients, so it appears to me that interpreting SE and pvalues from standardized coefficients would be more appropriate. 


I think it is clear that the reason the pvalues are different is that the two parameters have different sampling distributions. One may be less normal than the other. I do not agree that unstandardized would be better. I don't think this is an issue of multicollinearity. You can try nonsymmetric confidence intervals. If they don't agree with symmetric confidence intervals, this suggests a nonnormal sampling distribution. 

Back to top 