Negative residual variances are usually caused by incorrect starting values or an incorrect model. If they are not significant, they can be set to zero. Otherwise, change the model or starting values.
Anonymous posted on Tuesday, October 02, 2001 - 6:03 am
I just need one thing clarified. If my model has a negative residual variance that is not significant, and the standardized value cannot be calculated (999), does this mean that the parameter estimates are definitely invalid and should not be used? I ask this question after having tried setting the residual variance to zero and changing starting values.
It seems that your question has more to it than is being stated in this discussion. Why don't you send your input and data to firstname.lastname@example.org so I can give you a more informed answer related to your exact problem?
What a wonderful forum! I feel very fortunate to have such a renowned expert available to answer my questions!
I'm running a CFA with 24 binary outcomes (true/false responses) and one latent factor using WLS estimation.
Am I correct in my understanding that the chi-square test of model fit probably isn't the best one to use because of problems with non-normal data and that chi-square df with WLS does not represent interpretable information?
Also, I'm not sure how to interpret SRMR with tetrachoric correlations. Is the value of .234 reliable? If so, do you believe it represents a better indication of fit than RMSEA (.028) for this anaysis?
Finally, what do you think would be the best way to compare nested models? Are chi-square difference comparisons appropriate with tetrachoric correlations, or should I use CFI?
Yu, C.Y. (2002). Evaluating cutoff criteria of model fit indices for latent variable models with binary and continuous outcomes. Doctoral dissertation, University of California, Los Angeles.
It can be downloaded from the Mplus website from Mplus Papers. This dissertation examines the behavior of the fit measures you are asking about for categorical outcomes.
I believe your reference to degrees of freedom and weighted least squares estimation refers to the fact that for the WLSMV estimator, the degrees of freedom are not computed in the regular way. This does not make the chi-square untrustworthy. In fact, WLSMV is the Mplus default. I recommend that you use that not WLS. The degrees of freedom for WLS and WLSM are computed in the regular way.
I would compare nested models using chi-square difference testing. I'm not sure how two CFI values can be compared.
I am working with Latent Growth curve models. I have the following warnings:
-THE COVARIANCE COVERAGE FALLS BELOW THE SPECIFIED LIMIT. -THE MISSING DATA EM ALGORITHM WILL NOT BE INITIATED. -CHECK YOUR DATA OR LOWER THE COVARIANCE COVERAGE LIMIT. -THE STANDARD ERRORS FOR THE STANDARDIZED COEFFICIENT -COULD NOT BE COMPUTED DUE TO FAILURE OF THESTANDARD ERROR COMPUTATION FOR THE H1 MODEL
I am conducting a multiple indicator growth curve (4 time points) and I am wondering if there is anyway to be able to estimate the intercept mean (I have attempted constraining other paths and freeing this so the model will still be identified). Also, I am not sure why (forgive me if this is a naive questions), but the STANDARDIZED option wont produce standardized values for my intercept and mean variances (there are no negative residuals in the model). Any help would be greatly appreciated!
You can fix the intercepts of the factor indicators to zero at one time point instead of holding them equal and then free the mean of the intercept growth factor. You don't gain anything by doing this. It is a reparametrization of the model which results in the same fit and the same estimates.
Referring to Linda's post on Sat 24 Nov 2012, she states "You can fix the intercepts of the factor indicators to zero at one time point instead of holding them equal and then free the mean of the intercept growth factor" and then gives the example pertaining to example 6.15.
However, [u11$1@0 u12$1@0 u13$1@0]; refers to the intercepts for the same item constrained to zero at three different time points? Not at one time point?
Or did I misunderstand something?
If I do this in my models, the estimates are the same, but not model fit?
For binary items, there are two choices for the parameterization of a growth model. If you fix the threshold to zero at each time point, you can estimate the intercept and slope growth factor means. If you hold the thresholds equal across time, you can estimate the slope growth factor mean only. The mean of the intercept growth factor is found in the threshold estimate.