Message/Author 

Anonymous posted on Monday, May 24, 2004  10:48 am



I'm now using the new version 3 program and I've started receiving the warning message below. This is true even when using programs that previously ran fine. Is this a new message and how serious is it when nothing else in the model looks strange? Thanks. THE MODEL ESTIMATION TERMINATED NORMALLY WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE F3. 


This message is serious and it is new. In most cases, you will find a negative residual variance or a correlation greater than one. This can also be caused by dependencies in your data. These conditions result in inadmissable models. We added the error message because we have found that negative residual variances and correlations greater than one are sometimes ignored. You can ask for TECH4 in the OUTPUT command to see the model estimated correlations among the latent variables. 

Dustin posted on Tuesday, November 02, 2004  1:27 pm



I am having some difficulty with a negative residual variance within a simple measurement model. The negative residual is not statistically significant suggesting that it includes "0," but I am worried about potential criticisms from reviewers about fixing the residual to zero. Can you offer any guidance in this regard. Below is the output for the analysis: Analysis: Type = missing h1; MODEL: pun by punr punt; rem by remr remt cu ; fer by pd fear3; OUTPUT: sampstat ; MAXIMUM LOGLIKELIHOOD VALUE FOR THE UNRESTRICTED (H1) MODEL IS 1692.554 THE MODEL ESTIMATION TERMINATED NORMALLY WARNING: THE RESIDUAL COVARIANCE MATRIX (THETA) IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE FEAR3. TESTS OF MODEL FIT ChiSquare Test of Model Fit Value 11.096 Degrees of Freedom 11 PValue 0.4352 ChiSquare Test of Model Fit for the Baseline Model Value 475.114 Degrees of Freedom 21 PValue 0.0000 CFI/TLI CFI 1.000 TLI 1.000 Loglikelihood H0 Value 1698.102 H1 Value 1692.554 Information Criteria Number of Free Parameters 24 Akaike (AIC) 3444.203 Bayesian (BIC) 3519.035 SampleSize Adjusted BIC 3443.047 (n* = (n + 2) / 24) RMSEA (Root Mean Square Error Of Approximation) Estimate 0.007 90 Percent C.I. 0.000 0.082 Probability RMSEA <= .05 0.742 SRMR (Standardized Root Mean Square Residual) Value 0.036 MODEL RESULTS Estimates S.E. Est./S.E. PUN BY PUNR 1.000 0.000 0.000 PUNT 1.030 0.089 11.614 REM BY REMR 1.000 0.000 0.000 REMT 1.490 0.166 8.997 CU 2.408 0.324 7.432 FER BY PD 1.000 0.000 0.000 FEAR3 16.210 8.906 1.820 REM WITH PUN 0.282 0.057 4.936 FER WITH PUN 0.090 0.054 1.654 REM 0.040 0.029 1.374 Intercepts FEAR3 16.353 0.345 47.382 PD 2.672 0.058 46.440 PUNR 2.485 0.077 32.459 PUNT 2.720 0.077 35.246 REMR 3.165 0.062 51.159 REMT 2.729 0.069 39.382 CU 5.563 0.168 33.156 Variances PUN 0.790 0.119 6.618 REM 0.314 0.065 4.826 FER 0.092 0.058 1.583 Residual Variances FEAR3 4.393 12.575 0.349 PD 0.460 0.069 6.627 PUNR 0.189 0.060 3.134 PUNT 0.157 0.062 2.513 REMR 0.325 0.043 7.519 REMT 0.106 0.057 1.855 CU 2.881 0.355 8.120 35.246 2.720 2.727 REMR 3.165 0.062 51.159 3.165 3.959 REMT 2.729 0.069 39.382 2.729 3.047 CU 5.563 0.168 33.156 5.563 2.566 Variances PUN 0.790 0.119 6.618 1.000 1.000 REM 0.314 0.065 4.826 1.000 1.000 FER 0.092 0.058 1.583 1.000 1.000 Residual Variances FEAR3 4.393 12.575 0.349 4.393 0.221 PD 0.460 0.069 6.627 0.460 0.833 PUNR 0.189 0.060 3.134 0.189 0.193 PUNT 0.157 0.062 2.513 0.157 0.158 REMR 0.325 0.043 7.519 0.325 0.509 REMT 0.106 0.057 1.855 0.106 0.132 CU 2.881 0.355 8.120 2.881 0.613 RSQUARE Observed Variable RSquare FEAR3 Undefined 0.12208E+01 PD 0.167 PUNR 0.807 PUNT 0.842 REMR 0.491 REMT 0.868 CU 0.387 


FER has only two indicators and one indicator has a negative residual variance. You can fix the residual variance to zero but you might want to rethink that factor. With only two indicators, it is not identified unless it is part of a larger model. 

Dustin posted on Wednesday, November 03, 2004  9:47 am



Any ideas on what is causing the negative residual variance for the latent factor? The fer construct is part of a larger model so I could fix the residual to be zero, but I was wondering why the loading for the indicator fear3 is so much higher than pd for the factor. Is this something I can explore, and if so, how? 

bmuthen posted on Wednesday, November 03, 2004  11:00 am



This may be an indication that the model is not fitting as well as it should. For example, consider the 2 indicators of fer (y1, and y2, say)and a 3rd variable (z say). If (1) y1 and y2 correlate moderately and (2) y1 correlates highly with z while (3) y2 correlates much less with z, then the factor model loadings cannot both be high because it violates (1) and the y1 loading needs to be higher than the y2 loading to satisfy (2) and (3). It may then happen that the y1 loading needs to be so big that there is nothing left for the residual variance and it turns negative. Although one could try to make y1 and z correlate in a new way, not only through the factor, that may lead to a complicated model. And, my example really points to y1 and y2 not acting as indicators of a single factor visavis z. 


I am estimating a SEM which includes (among other things) two exogenous variables: RACE and INCOME. In my sample, RACE and INCOME are correlated and I would like to include a WITH statement to model this correlation. However, when I do, I get this error message. *** FATAL ERROR VARIABLE TRACE1 CAUSES A SINGULAR WEIGHT MATRIX PART. THIS MAY BE DUE TO THE VARIABLE BEING DICHOTOMOUS BUT DECLARED AS CONTINUOUS. RESPECIFY THE VARIABLE AS CATEGORICAL. However, when I put TRACE1 in the CATEGORCAL statement I get this message. *** ERROR in Variable command CATEGORICAL option is used for dependent variables only. TRACE1 is not a dependent variable. Am I missing something here? Is the correlation included by default and the WITH Statement not necessary? 


Exogenous variables should not be inlcuded on the CATEGORICAL list. The scale of categorical variables does not affect model estimation. Correlations among observed exogenous variables should not be included in the MODEL command. The model is estimated conditioned on the x variables as in regular regression. If you want to know the correlation between you x variables, use the sample statistics from TYPE=BASIC or another program. Think about regular regression where y is regressed on x1 and x2. The parameters in the model are an intercept and two slopes, one for x1 and one for x2, and a residual variance for y. The is no parameter for the correlation of x1 and x2. 

Martie posted on Friday, March 18, 2005  9:41 am



Correlations of exogenous variables in the PHI matrix only apply to latent variables? 


Do you mean the psi matrix? If so, the psi matrix is the matrix of variances and covariances of latent variables. There is some confusion in the field because with continuous observed variables, it is the same if you include the covariances of the variables in the model or not. So the practice has been not to distinguish between exogenous and endogenous variables in this case. This has made it confusing for those transitioning to other outcome types and other models. 

Anonymous posted on Saturday, July 09, 2005  9:47 am



whatever I try with LISREL 8.7, as soon as I try to run it, I get "THIS MODEL DOES NOT CONVERGE" I don't know what the problem is. Is there something wrong with my dataset? is there some kind of troubleshootingmanual for this? 


You can look at the suggestions in the Mplus User's Guide for things to do when models don't converge. 

Annonymous posted on Thursday, January 19, 2006  1:05 pm



I have been conducting SEM with a categorical outcome variable (ACCOMP) that has 4 levels, and have been able to run the analyses sucessfully. however I have been trying to run ACCOMP as a binary variable  changing nothing else in the model  and this error message came up: *** ERROR One or more variables have a variance of zero. Check your data and format statement. I noticed that the other categorical variables also had a variance of 0.00, but only ACCOMP had asterixes beside it. Any suggestion as to what the problem might be? 


You would need to send your input, data, output, and license number to support@statmodel.com. 


Hi: My model runs fine (fits poorly, but runs fine) when my control variables are not included, but gives the message: THE MODEL ESTIMATION TERMINATED NORMALLY. WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE ADHERE. (The control variables are listed in the "adhere ON MALE..." line.) discrim BY b09d01a b09d01b b09d01c f10b04 f10b05 f10b06; distrust BY g09c12 g09c13 g09c14 g09c15 g09c16 g09c17 g09c18; psybur BY g45a54 g45a55 g45a58 g45a59 g45a61 g45a62 g45a63 g45a64; acdiff BY g45a67 Rg45a69 g45a71 g45a73; schdiff BY g45a51 g45a52 g45a60; arteff BY g45a39 g45a41; adhere BY rFORGOT rSKIP rLESS PERFECT g45a47; psybur acdiff arteff discrim distrust adhere ON MINORITY; adhere ON MALE b12age03 EDUC b12inc05 MEDICAID HETERO R_IDU R_MSW viral_ct g08ovr01 g08ovr02 b11g03 b11g05 b11dru08 g45_a14 b09ssp04; arteff distrust adhere ON discrim; psybur arteff adhere ON distrust; adhere ON psybur acdiff schdiff arteff; ADHERE does have a negative residual variance, which causes this error. What might be the problem? I checked for correlation > .85 among the control variables, but didn't find any. There are two correlations, not among the control variables, that approach .85: RG45A69G45A39 (0.839) and RG45A69G45A41 (0.849). However, there was no problems with singularity before the control variables were included. I know there are a lot of control variables, but they are both theoretically and empirically associated with my outcome. Would using factors (2 emerge from analysis) instead of the observed variables be an acceptable substitution? Or someone above suggested fixing the residual to zero? Thanks! 


I would check the model without the covariates ("control variables") to see if the ahdhere factor has a significant variance and to see that the model fits the data well. If yes on both, I would make sure the model with covariates fits well. For example, it looks like adhere has many more covariates than the other factors  are those extra covariates really not related to the other factors? If they are, the model is misspecified. 


Thanks for your response  you were right about the misspecification, and I made those changes. Without covariates, ADHERE has a residual variance estimate = .054, critical value = 5.10 but the fit is poor (n = 1653; X2 = 219.961, df = 13, p = .000; CFI = .770, TLI = .788, RMSEA = .098, WRMR = 3.170). One problem was that one of my covariates had substantial (15%) missing data. When that variable was dropped, however, ADHERE remained npd and the fit of the model (with covariates) was still not that good (n = 1909; X2 = 107.154, df = 14, p = .00; CFI and TLI = .858, RMSEA = .059, WRMR = 2.348). Any other suggestions? If it makes a difference, the model incorporated weights and used the WLMSV estimator. 


You should send your input, data, output, and license number to support@statmodel.com so we can see exactly what you are doing. 


Following up on your earlier posting about the error message: "WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE X"  I am getting this error, but do not find either a negative residual variance or a correlation greater than one in relation to the variable in question or any other variable in the model. Is there anything else I should look for? Thank you in advance, Inna 


If you do not see either a negative variance or residual variance or a correlation of one or greater, you must have a dependency among some of your variables. 


Linda, Thanks for your reply. Could the problem be caused by an indicator variable that is part of the same latent factor and is binary but is being treated as continuous? That is, I have a latent factor (among many) that has four indicators and one of them is binary. When I run the measurement model for the factor and treat that variable as binary invoking the WLSMV estimator, all is fine. But when I run a larger model with this factor associated with other factors and covariates and treat this variable as binary, the model does not converge (even with iterations set to 6000). If I comment out the CATEGORICAL line and treat the indicator as continuous, the model converges. So, I've been letting the binary indicator get treated as continuos (although it is not symmetric). Could this be causing a problem with one of the other indicators in the same latent factor? Independent of the "(PSI) IS NOT POSITIVE DEFINITE" issue, does this approach give me invalid results? Or is it a reasonable approximation given that this is one variable in about 50? Alternatively, I could use the MLR estimator with numerican integration. What would be the best option? Thanks again, Inna 


I don't think this has anything to do with treating the variable as categorical versus continuous. One thing that can cause convergence problems when there are a combination of continuous and categorical outcomes is when the variances of the continuous outcomes are large. We recommend rescaling to keep them between one and ten. When psi is not positive definite, the results are not valid. 

Susan Scott posted on Thursday, September 14, 2006  7:53 am



I am running a SEM with 5 imputed datasets. In the Tech4 output, I get a message for each dataset: Errors for replication with data file H:\qol\SEM_chked\MPd\t2e1.txt: THE MODEL ESTIMATION TERMINATED NORMALLY What does the first message "Errors for replication..." mean? Thank you, Susan Scott 

Thuy Nguyen posted on Thursday, September 14, 2006  8:56 am



This isn't part of TECH4 output. It should be part of the TECH9 output that shows model convergence messages for each data set. If you get "THE MODEL ESTIMATION TERMINATED NORMALLY", then everything is okay for that data set. 

Susan Scott posted on Thursday, September 14, 2006  12:38 pm



My apologies  you are correct, it is part of the Tech9 output. So I can ignore this statement? It seems to appear with every run. What does it mean? Susan Scott 

Thuy Nguyen posted on Thursday, September 14, 2006  1:27 pm



If the message under each data set is MODEL ESTIMATION TERMINATED NORMALLY, then you can ignore it. But if you run into a problem where a particular data set does not produce model convergence, then a different message will appear and this will help to figure out which data set had a problem and what the problem is. This applies to multiple imputation and external Monte Carlo. 

Susan Scott posted on Friday, September 15, 2006  11:00 am



For 4 of my 5 imputed datasets, I see MODEL ESTIMATION TERMINATED NORMALLY, but dataset #4 does not converge: Errors for replication with data file H:\qol\SEM_chked\MPd\t2e4.txt: NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED. But I get the same 'Errors for replication...' message. This tells me the noncovergence is in dataset t2e4.txt. Can it tell me anything else? Thank you, Susan Scott 

Thuy Nguyen posted on Monday, September 18, 2006  3:34 pm



The "Errors for replication.." is just a heading for each data set. The actual error message for that data set follows. In this case, you get "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED." You should try setting the model up as a regular model with this data set to see if increasing the number of iterations will help and see if there are more clues as to why the model does not converge with this data set. 

Jeremy Miles posted on Wednesday, October 11, 2006  12:42 pm



Hi, I'm getting a weird error running Mplus version 4, doing a CFA, with MLR estimation, and missing data. I've got 8 latent variables and the fit is just about OK. Then I add some variables and correlate them with the factors, and I get an error: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NONPOSITIVE DEFINITE FIRSTORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER 0.262D16. PROBLEM INVOLVING PARAMETER 201. Parameter 201 is the last one  it's the variance of one of the added variables. But it doesn't matter which variables I enter (from the ones I want to enter) I get the same error, and it's always parameter 201  which is always the variance of the last variable. If I change the model, and use different variables, even a completely different set, I get the same error, with the variance last variable in the model. Which seems weird to me, but I might be missing something. 


Can you send your input, data, output, and license number to support@statmodel.com. 


Hi  I'm running Mplus 4.2 doing a path analysis with MLR estimation and missing data. i'm getting the following error message: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NONPOSITIVE DEFINITE FIRSTORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS 0.305D10. PROBLEM INVOLVING PARAMETER 136. When i orginally ran this model, I had two additional variables, and the model ran without errors, even though it included the variable that the current error message refers to (school achievement). what does this error message mean, and is there anyway to address the problem? Thanks! aprile 


Please send your input, data, output, and license number to support@statmodel.com. I need more information to diagnose the problem. Please include TECH1 in the OUTPUT command. 


Hi, I have a few questions after I did an SEM calculation: I have 17 categorical dependent variables, where 3 are independent and 14 dependent. The estimator WLSMV and theta parameterization was used. I did two different calculations one with the help of the M+ language generator (SEM). An output was delivered without erreo message. However, I also wanted to calculate the regression coefficients between all our variables. Therefore, I added ON statements. Then, regression coefficients came out well, but the following error message was displayed: THE DEGREES OF FREEDOM FOR THIS MODEL ARE NEGATIVE. THE MODEL IS NOT IDENTIFIED. NO CHISQUARE TEST IS AVAILABLE. CHECK YOUR MODEL. and consequently no CHIsquare value for the model was calculated. Questions: 1) Why this error message when adding the ON statement? 2) Can we use the CHIsquare value earlier obtained for the model with ON? Also, what does thresholds under Model results mean? Thanks! Lollo 


Your model is not identified because there is not enough information in the data to estimate all of the parameters that you have specified. You cannot use the chisquare from the identified model for the not identified model. 

tommy lake posted on Saturday, April 28, 2007  12:47 am



Dear Linda, I was running an SEM model and got this error message: *** ERROR Unexpected end of file reached in data file. What does this error message mean, and is there anyway to address the problem? Thanks! Tommy 


This means that the amount of data in the data set and the amount of data expected by the number of variables specified in the NAMES statement don't match. It is often caused by blanks in the data set when free format is used. If you can't figure it out, please send your input, data, output, and license number to support@statmodel.com. 


I've run an sem with 3 latent factors and I get a negative residual variance for one of my indicators, although it is nonsignificant. This indicator is correlated with another indicator of the same latent factor (r=.71); might that be the cause of the negative residual variance? I see in previous posts that you suggest setting neg residual variance to zero if nonsignificant. How do I do that? Also, are there other ways to address neg. res. variance? Thanks 


I don't think this would cause but you can see by eliminating the covariance from the model. To set a residual variance of y to zero, say: y@0; You can think about changing the model. 

Maja Cambry posted on Thursday, June 28, 2007  6:35 am



I'm running an sem with 5 continuous latent vars, 29 dependent variables and 4 independent variables as controls/ covariates. My sample of 368 is drawn from a large dataset with a complex survey design, thus I'm using stratification, cluster, and weight options. I obtain the following error messages. Not sure what these mean, how serious, and what to do about them. Any suggestions? Thanks! WARNING: THE VARIANCE CONTRIBUTION FROM A STRATUM WITH A SINGLE CLUSTER (PSU) IS BASED ON THE DIFFERENCE BETWEEN THE SINGLE CLUSTER VALUE AND THE OVERALL CLUSTER MEAN. THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NONPOSITIVE DEFINITE FIRSTORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS 0.991D17. PROBLEM INVOLVING PARAMETER 74. THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS. 


The first message is informational. It is desirable to have more clusters than model parameters. We are making your aware that you don't. We believe that the results are acceptable but this has not been studied. 


This is the error message from a path analysis. ERROR MESSAGE BELOW: THE MODEL ESTIMATION TERMINATED NORMALLY THE CHISQUARE DIFFERENCE TEST COULD NOT BE COMPUTED BECAUSE THE H0 MODEL IS NOT NESTED IN THE H1 MODEL. MODEL BELOW IS THE FIRST MODEL RUN: MODEL: s on mddpar ; dep on s mddpar ; SAVEDATA: DIFFTEST IS DIFFTOT.DAT; MODEL BELOW IS THE SECOND MODEL RUN: ANALYSIS: DIFFTEST=DIFFTOT.DAT; MODEL: s on mddpar (1) ; dep on s (2) mddpar ; I have run taken this approach before with the only difference the people in the sample and have gotten no error msg . This error msg seems to come and go what could be the causes? I have tried changing names of data sets deleting saved data sets from prior runs changing the variables (e.g., not using total depression but adolescent MDD as the outcome). It seems to be an issue with the program that I am unaware of. Do you have any suggestions? Thanks . FYI I think it is really great you do this. 


Please send your input, data, output, and license number to support@statmodel.com. I need more information to answer your question. 


Hello, I'm getting the following message when I use WLSMV as the estimator. I do not get this message when I use WLS estimation. THE MODEL ESTIMATION TERMINATED NORMALLY THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING PARAMETER 70. THE CONDITION NUMBER IS 0.315D17. Why am I getting this message and what should I do now? 


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


Hi, I'm getting a warning that the latent variable covariance matrix is not positive definite and the problem involves variable SUB3. SUB3 does not have either a negative residual nor any correlations greater than or equal to one. From the TECH4 Output it has correlations (and covariances as we fixed the variance of each latent variable to equal 1) of .285, .217 and .813 with the 3 other latent variables in the model. Must there be a dependency involving SUB3 or is it possible that the negative correlation with one of the other latents would be causing the not positive definite warning? IF there must be a dependency, do you have any suggestions for figuring out what the dependency is arising from? Thanks! 


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


Dear Dr. Muthen, How bad is the following error? THE STANDARD ERRORS FOR H1 ESTIMATED SAMPLE STATISTICS COULD NOT BE COMPUTED. THIS MAY BE DUE TO LOW COVARIANCE COVERAGE. THE ROBUST CHISQUARE COULD NOT BE COMPUTED. THE MODEL ESTIMATION TERMINATED NORMALLY. Does this affect the estimation of the growth factors i and s? I get a nice plot, but I am wondering how trustworthy the solution is. Thanks in advance. 


The H1 model is the unrestricted model that is used to compute chisquare. From the information that you give, your results should be fine. If you want a more definitive answer, send your output and license number to support@statmodel.com. 


So this error is only a problem if you want to use the chisquare, and it does not affect parameter estimates? I am glad to hear that my results should be fine. Thank you very much. 


Yes. 


I have a model with one binary variable x1, six manifest variables y1y6 and two latent variables f7 and f8. My input statement is: variable: names are x1 y1y12; model: f1 by y1; f2 by y2; f3 by y3; f4 by y4; f5 by y5; f6 by y6; f7 by y7y9; f8 by y10y12; f1 on x1; f2 on f1; f3 on f2 f1; f5 on f3 f2 f1 f7 f8; f6 on f5 f2; f1@1; f2@1; f3@1; f4@1; f5@1; f6@1; f7 with f8@0; analysis: estimator = mlr; output: standardized; modindices; residual; fscoefficient; tech4; I am getting the following message in the output: WARNING: THE RESIDUAL COVARIANCE MATRIX (THETA) IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR AN OBSERVED VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO OBSERVED VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO OBSERVED VARIABLES. CHECK THE RESULTS SECTION FOR MORE INFORMATION.PROBLEM INVOLVING VARIABLE Y1. I tried some model modifications  but there is always this error message. Is there any error in my model specification? Thank you. 


For a single indicator, you do not fix the factor variance to one. You fix the residual variance of the factor indicator to zero. With one indicator, it is the same if you do this or just use the observed factor indicator in the model. I suggest doing this. For further questions of this type, please send your input, data, output, and license number to support@statmodel.com. 


Hi, I keep receiving a warning message "error in variable command, duplicatre variable on NAMES list" [then indicates name of variables). When I go to the names list the variable is only listed once. I have been using the same format in other SEM and did not received this warning. Any assistance provided would be appreciated. 


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

Jay Irwin posted on Thursday, January 15, 2009  10:11 am



I am conducting a SEM model that includes latent factors and observed covariates. My dependent variable is the Iowa CESD scale (an observed variable) and it appears that something is happening with this variable. I am getting an error about the Psi factor not being positive definite. When running the Tech4 output, I see that the residual variance for the Iowa CESD is negative (est: 14.185, S.E.: 89.009, Est/S.E.: 0.159, Std.: 14.185, StdYX: 0.614. Should I set this to zero in the syntax? If so, how do I go about doing this? Is there any repercussions for doing this on the dependent variable? Thank you! 


Please send your full output and license number to support@statmodel.com. Please use Version 5.2. 


Hi  My colleagues and I are conducting a regression analysis that includes one dependent variable and 22 predictors/covariates. The model includes a weight and a cluster. When I run the model, I receive the following error: THE STANDARD ERRORS FOR H1 ESTIMATED SAMPLE STATISTICS COULD NOT BE COMPUTED. THIS MAY BE DUE TO LOW COVARIANCE COVERAGE. THE ROBUST CHISQUARE COULD NOT BE COMPUTED. THE MODEL ESTIMATION TERMINATED NORMALLY. I have checked my covariance coverage, and values range from 0.67  1.00, so I am unsure what the source of the error is. Do you have any suggestions for us? We are working with restricted use data, so we would not be able to send the data file to you. Thanks! 


It sounds like you are bringing the means, variances, and covariances of the covariates into the model and that making distributional assumptions about these variables is creating problems. 


Hello, I am running a 4 factor SEM model with interaction effects. I am having trouble inputting my data. I double checked my format and it looks fine. Everything works fine if I don't use my last variable. If I use it, I get a message saying that one of my variables has a variance of zero. I checked this on another statistical package and that variable did not have a variance of zero. The rest of the variables read fine  I only have a problem with the last variable. Any suggestions? Thank you in advance for your help. 


Please send the input, data, output, and your license number to support@statmodel.com. 


Hello, I want to estimate a WileyWiley Model with count data. It works fine as long as I don't define my variables as counts (syntax see below). Variable: Names are v1 v231 W231 X231 v231b W231b X231b; Usevariables are v231b W231b X231b; Missing are all (9999) ; Model: f1 by v231b@1; f2 by W231b@1; f3 by X231b@1; f2 on f1; f3 on f2; f3 with f1@0; v231b W231b X231b (1); ANALYSIS: Type is general; OUTPUT: TECH1 SAMPSTAT RESIDUAL STANDARDIZED; But as soon as I add this to Variable part: Count are v231b W231b X231b; it results in the following error: *** ERROR in Model command Variances for count variables are not currently defined. Variance given for: V231B and for variables w231B and x231B as well. The problem is the same when I try to estimate a two indicator model where both indicators are count variables. I am not sure why that is, if I made some misspecification or if it just isn't possible to estimate these models using count variables. I would be very thankful for your help with this. 


Yo need to remove the line that says: v231b W231b X231b (1); In a count variable, the mean and the variance are related  if you know the mean, you know the variance, and your model (currently) tries to estimate both. That's not allowed. Jeremy 


Hello Jeremy, v231b W231b X231b (1); restricts the error variances of the three variables to be equal. I fail to see how that should be a problem? I might not have been clear that the problem is the same for the variables W231b and X231b, I get the exact same error. If I take this restriction out, the model isn't identified anymore, since it is just identified in the current state. 


Count variables do not have variance parameters. Please send your input, data, output, and license number to support@statmodel.com to see what the identification problem is. 


Hi. I´m trying to build a personality model, including latent variables, and I am getting the following error message: WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. Now my problem is this: this message is probably caused by multicollinearity in my data. This is understandable because the data includes the same personality traits from the same people measured some years apart. So multicollinearity is something we would expect and which is desirable. It proves that that personality is stable over time. So knowing that high correlations are something we actually want and hope, can I ignore this error message? Or does it mean that the model doesn´t work? The model is terminated and evaluated normally except for this error message. Can I trust these results despite of the error message, generally speaking, in this kind of desirable high correlation situation, or should something be done about it? Kim 


This message usually should not be ignored. Please send your full output and license number to support@statmodel.com. 

Jeremy Miles posted on Thursday, September 03, 2009  11:14 am



Have you correlated the residuals of the items over time? if you don't do that, the fact that (say) "enjoys parties" at time 1 is more correlated with "enjoys parties" at time 2 than it is with "like telling jokes" at time 2 will force the model to increase the size of the correlations between the latents. Instead, you should account for that by correlating the equivalent items over time. 


Does Mplus currently have the facility to print warnings and errors to file? 


Warnings and errors are printed in the output file. 


I've been having a problem with my Psi matrix being NPD. I think part of the problem is what I'm modeling and am hoping for suggestions. I'm looking at the "effects" of moving on the family context during three developmental periods. I've modeled the three periods using three observed variables. Naturally, some of these variables (and hence the latents) are highly correlated (most not over .5 however). I'm not using a growth model because this isn't growth over time, just contexts over time. I've correlated the latent variables over time and the residuals of the observed variables. Other potential fixes are welcomed. 


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


Hello Linda, I'm trying to build a secondorder but have convergence problems. (NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED.) The model contains two 2ndorder factors g1 and g2 which are respectively measured by four 1storder factors (f1f8). Since the 1storder factors cause the 2ndorder ones, we chose a formative model on this level.The 1storder factors have a reflective measurement model. We have two outcome variables which are observable variables (quanti and qual). To calculate the model we proceeded as follows: In accordance with literature we first calculated a factor anlysis for the 1storder factors, saved the scores (FA1FA8) and used them as indicators for the secondorder factors. Input: MODEL: f1 BY x1x4; f2 BY x5 x6; f3 BY x7 x8; f4 BY x9x11; f5 BY x12x15; f6 BY x16x20; f7 BY x21x23; f8 BY x24x26; g1 BY FA1FA4; g2 BY FA5FA8; g1 ON f1f4; g2 ON f5f8; g2 ON g1; quant ON g1; quant ON g2; quali ON g1; quali ON g2; Do you have any ideas why the model doesn't converge? Does our proceeding on the calculation of secondorder model look reasonable to you? Thank you very much for your response. 


You should use the firstorder factors as indicators of the secondorder factors and not use factors scores. See Slide 160 of the Topic 1 course handout for the proper setup for a secondorder factor model. 


Thank you very much for your response Linda. Do you see any possibility to measure the secondorder factors in a formative way? Since we assume that the 1storder factors cause the 2ndorder factors, we would like to use a formative measurement model on this level and don't want to use the BYstatement. Is this possible in mplus? How could we avoid convergence problems? Thanks again. 


If you want g1 and g2 to be influenced by the f factors you should simply drop the statements g1 BY f... and g2 BY f... and instead define the g factors by saying: g1 BY; g1@1; g2 BY; g2@1; Keep all the other statements. I'm not sure this model is identified, but you can see. You may have to fix one slope per g factor. 


I am running the following constructs STIGMA BY SOCIAL CONCEAL DISCUSS FEAR AUTO; PARENT BY PRAISE IRRESP UNORGNIZ PRNT_COG STIGMA ATTEN HYPER; FAMILY BY COHESION TRAUMA ENGMENT UNITY PARENT; Where each one of these variables (such as Social or Conceal, etc.) are factors themselves. The message I got was Std errors could not be computed because problem with parameter 291, which is PSI for Family What can I do to get std errors? 


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


Greetings Drs. Muthen, I am running a two group CFA with categorical outcomes. When I ran the unconstrained baseline solution I got the error copied below. However, when I ran each group separately the models ran as expected. I am using version 6. Parameter 37, which is identified below, is the last parameter in the model, and if I remove the variable impacted, the same error occurs with what is then the last parameter. Thank you for your guidance, Ben THE MODEL ESTIMATION TERMINATED NORMALLY THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING PARAMETER 37. 


It's hard to say without seeing the output. Please send it and your license number to support@statmodel.com. 


Ben, the loading of the first variable in the first group is automatically constrained to 1.00. This is not true of the second or subsequent groups, and so the variance of the latent variable is not identified. You can fix this with an "@1" on the first loading in the second group. 


This is one possibility but not the only reason that you might get the identification message. If this is not your problem, please send your output and license number. 


Hi! I am fitting a simple prediction model with one latent construct /predictor at time 1 predicting two latent constructs / outcomes at time 2, controlling for time 1 levels of the outcomes. The structural model works well, but everyt ime I try to add a covariate (such as gender, age, mom's employment status), I automatically get "a nonpositive definite firstorder derivative product matrix". Any suggestions? 


I would need to see the output. Please send it and your license number to support@statmodel.com. 


Hi, I'm new to the website, so I've accidentally posted this question elsewhere. It makes sense for it to be "here" though so I'm apologetically posting it twice. I received the following error message: *** ERROR in MODEL command Covariances for categorical, censored, count or nominal variables with other observed variables are not defined. Problem with the statement: L246_002 WITH L03X5002 *** ERROR in MODEL command Covariances for categorical, censored, count or nominal variables with other observed variables are not defined. Problem with the statement: L246_006 WITH L03X5006 *** ERROR The following MODEL statements are ignored: * Statements in the GENERAL group: L246_002 WITH L03X5002 L246_006 WITH L03X5006 These are binary variables, which I listed in the "categorical are" statement. Because it is the same subject filling out the questionnaire at two different time points, it conceptually makes sense to have the errors be correlated. What do I need to include in my syntax so that these variables will be "defined"? Thank you. 


It sounds like you are using the CATEGORICAL option and maximum likelihood estimation. Each residual covariance requires one dimension of integration in this case. A model with more than four dimensions of integration can be computationally demanding and is not recommended. You can use the BY option to specify a residual covariance, for example, f BY L246_006@1 L03X5006; f@1; [f@0]; The residual covariance can be found as the factor loading for L03X5006. I suggest using the default estimator of WLSMV where you can specify residual covariances using the WITH option. 


Thank you. I originally used MLR as my estimator (because of the categorical variables). If I switch to WLSMV, then I can stick with my "with" statements? Sidenote: I was wanting to include a Latent Variable interaction (xwith), but I wasn't sure how switching from "type is general" to "type is random" would affect things. Can I do "type is random" and use WLSMV? I'm working on my SEM project (hopefully basis for an article) so I am likely asking some basic questions. My apologies in advance. 


Yes, with categorical dependent variables one advantage of weighted least squares over maximum likelihood is the ability to easily estimate residual covariances. XWITH can be done only with maximum likelihood. 


Dear Professor, I am having some difficulty with this error "THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. I check tech4 and the problem involving a variable "vitt" that is correlated with another variable "perp"1.065. This variables were asses with the same instrument, that allow to measured the victimization and the perpetration and for this reason the two variables are correlated. I don't know how to solve the problem. Could you help? tech4 STAB FUNZ PERP VITT STAB 1.000 FUNZ 0.360 1.000 PERP 0.364 0.333 1.000 VITT 0.191 0.391 1.065 1.000 the measurement model is: model: stab by sta1 sta2; funz by lav fam int; perp by per1 perp2 perp3 perp4; vitt by vit1 vit2 vit3 vit4; vitt with perp; output: ChiSquare Test of Model Fit 205.328 df 59 p.000 scaling correction factor MLM 1.303 CFI 0.901; TLI 0.87; RMSEA 0.067 SRMR 0.046 thank you very much, alice 


When two factors correlate greater than one, the model is not admissible. You also have two factors that correlate one. This means they are statistically indistinguishable. I would suggest looking at an EFA to see how many factors are found for this set of observed variables. 


Thank for you hel me, I try to make two separate model, one for the victimization and one for the perpetration, but now the problem is this: THE RESIDUAL COVARIANCE MATRIX (THETA) IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE STA1. model: funz by lav fam int; vitt by vit1 vit2 vit3 vit4; stab by sta1 sta2; ESTIMATED COVARIANCE MATRIX FOR THE LATENT VARIABLES FUNZ VITT STAB ________ ________ ________ FUNZ 4.858 VITT 3.682 18.587 STAB 0.659 0.579 0.751 What is the problem now?? thank you 


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


Dear Dr. Muthen, Estimating a multigroup multiple indicator latent growth curve model I got: THE STANDARD ERRORS FOR H1 ESTIMATED SAMPLE STATISTICS COULD NOT BE COMPUTED. THIS MAY BE DUE TO LOW COVARIANCE COVERAGE.THE ROBUST CHISQUARE COULD NOT BE COMPUTED.THE MODEL ESTIMATION TERMINATED NORMALLY When I run the model for T group only my model run without any errors. But when I run the model for C group only I got the same error message. Previously you stated that this error is only a problem if you want to use the chisquare, and it does not affect parameter estimates. But what happens if I would like to write up my result for publication. How do I present my results in absence of any model fit indices? Additionally it concerns me that slope estimates in the STDYX solution for C group were not estimated (see below). Should I be concerned about it? Thank you! Group CONTROL Means I 0.000 0.000 999.000 999.000 S 0.006 0.121 0.046 0.963 Group TREATMENT Means I 0.143 0.129 1.104 0.270 S 0.137 0.061 2.256 0.024 STDYX Standardization Group CONTROL Means I 0.000 0.000 999.000 999.000 S 999.000 999.000 999.000 999.000 Group TREATMENT Means I 0.135 0.124 1.090 0.276 S 0.295 0.139 2.132 0.033 


I think you are using MLR. Try ML for the C group. 


Hi. I am running a MVPA model with a complex survey design, using the stratification, cluster and weight variables. I am getting the following error: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NONPOSITIVE DEFINITE FIRSTORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS 0.929D10. PROBLEM INVOLVING PARAMETER 66. THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS MINUS THE NUMBER OF STRATA WITH MORE THAN ONE CLUSTER. Do I just need to worry about the bottom part of the error message or is there 2 issues, one with the nonpositive definite matrix and the other about the sampling design? I have 90 strata and 179 PUSs. Each strata has 2 psu's in it except 1. Any advice you can offer and fixing this would be very helpful! Many thanks! 


The essence of the message is the bottom part. The effect of having more parameters than independent pieces of information has not been studied. You could do a simulation to see what effect this would have. 

Klaus posted on Friday, July 29, 2011  7:35 am



Hi, I work with change score models (MPlus 6.1, type=imputation (25 data sets), ML estimation). To test the significance of a parameter I want to compare a saturated model with a constrained model. The saturated model is estimated without problems but in the constrained one the following error message appears: THE CHISQUARE COULD NOT BE COMPUTED. THIS MAY BE DUE TO AN INSUFFICIENT NUMBER OF IMPUTATIONS OR A LARGE AMOUNT OF MISSING DATA. Could you please give me some advice how to deal with this problem? And also how to get valid information on chisquare of the constrained model? Thank you in advance! 


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


Hi Dr. Muthen, I'm running a SEM and have been receiving the following error message: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING PARAMETER 22. I used TECH1 to determine that parameter 22 was the beta term between two exogenous variables (A and B). I removed one of the terms (B) and still received the same error message, this time for the beta term between A and C. When I receive this error message, I don't receive any measures of model fit (Chi Sq., CFI, etc.). Could you help me interpret this error so that I can get my model to run correctly? Many thanks. 


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


I have a model: f1 by y1y4; Y1 with y2; y3 with y4; f2 by y5@1; y5@16.3; F3 by f1 f2; My questions are: 1) Can I trust the estimates (F3 by f1 f2) from a model where I get nice model fit information but a very low condition number (like 0.765D18) and a warning regarding the trustworthiness of some standard error parameters? 2) Can I trust the estimates from models where no model fit information can be calculated (the standard errors could not be computed)? In both of the above cases I have discovered that I can fix f1@1 and f2@ the (perhaps untrustworthy) calculated estimate from the previous estimation and get a model with good fit and no warnings. This leads me to my last question: 3) Does this mean that the standard error issue is not really a problem? Best, Jörgen 


Your secondorder factor is not identified with only two firstorder factor indicators. The results should not be interpreted. The model is not identified. 

Jörgen Öberg posted on Wednesday, November 09, 2011  8:01 am



Thank you very much. In light of your answer I checked my CFA textbook and I see what you mean. I take it as the first order part has 3 df and the higher order part has 1 df. Mplus reports 2 df, but that would be for the total model then? I still get estimates for the higher order part though  are those just arbitrary? What do Mplus mean by those estimates? The reason for me having only two factors under the higher order one, is that I was hoping to get their common variance in F3. Then I had plans to regress F3 on a bunch of exogenous variables  would that save my unidentification problem? Best, Jörgen 


Because your model is not identified, none of the results should be interpreted. It might help to add covariates as far as identification goes, but it would result in a weak model where parts of the model must borrow from other parts to be identified. 

MJ Kim posted on Thursday, February 02, 2012  9:43 pm



Hello. I have an error message saying Psi is not positive definite and i do have a negative residual variance in TECH4 output. If i needed to change it to zero, how small does it have to be? Thanks, MJ 


It should be both small and not significant. 

MJ Kim posted on Tuesday, February 07, 2012  9:13 am



Thank you, Dr. Muthen. Helps a lot. Have another question. I had actually two scenarios (same subtaxes with a different data), but only one of them had "nonpositive definite" message (which was resolved by setting a negative residual variance to zero) while the other ran okay. Now, i would like to add another path in the model but then i can't due to the identification problem. In a scenario which had "nonpositive definite" error, i can add an additional path and get a fit when i set a negative zero variance to zero. However, i can't in initially okay scenario again due to identification problem (yet, this scenario has residual variance of the same construct very close to zero (.0004)). Do you think i can set residual variance in initially okay scenario to zero as well, add an additional path, and interpret the fit? Or, would this be introducing bias? Thank you so much. MJ 


I don't know what you mean by the identification problem. If you are referring to the nonposdefcov matrix message, that is not the same as nonidentification. And, typically, you don't resolve identification problems by fixing variances at zero. If this doesn't help, you may want to send your materials to Support. 

MJ Kim posted on Wednesday, February 08, 2012  12:32 pm



Thanks, Dr. Mythen. If i be little more specific here... One scenario with Non positive definite (PSI) was resolved by fixing a negative residual variance to zero and the other was initially okay. Now that i want to include another path, i can do so in the scenario with non positive definite (only with negative res variance fixed at 0). When i try to include another path in the other scenario (which was initially okay), i get "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED." and do not get fit output. What's interesting is though in the "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED." output, i see a negative residual variance in the same variable. If i set this res variance to 0 with an additional path, i get a fit output. So, should i increase number of iterations? or is setting a negative res variance to 0 a reasonable approach? Thanks, MJ 


Please send the output(s) and your license number to support@statmodel.com. 

Nidhi Kohli posted on Thursday, March 08, 2012  3:20 pm



I am trying to fit a latent variable path model to a dataset which has n=176. I first started with the measurement portion of the model to see if it fits the data well before I include the structural portion of the model. I get an error message in the output saying that the "THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE." I checked TECH4 and found that the correlation of one particular factor in the model is greater than 1 (r=1.005). Can you please suggest how can I overcome this error message? Thank you. 


This makes the model inadmissible. You need to change your measurement model. I would suggest doing an EFA to see how many factors that suggests. 


Thanks, Linda. I checked the whole model again and made some modifications. I reran the model and got a different error message saying, "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED." I increased the number of iterations to 10000 but still I am getting the same error message. Theoretically speaking, the model should fit the data well based on the research literature. I am struggling to understand why it is not working with the data I have with me. Is there anything else that I can do to make the model work? Thanks again. 


Many times the variables you have in the data set are not valid and reliable measures. This can cause a theoretical model not to fit. It would be impossible for me to say more without seeing an output. This is a question that should be sent to support. 

Dan Merson posted on Sunday, March 11, 2012  8:13 pm



Greetings. We're trying to run a path model with three outcome, seven endogenous, and nine categorical exogenous variables. The outcome and endogenous vars are CFAproduced continuous factor scale scores which range from approximately 4 to 2 with a few missing values (we have listwise=on). We also have a weighting variable that ranges from .39 to 2.51. We're estimating with MLM instead of ML in order to apply the weights and because our variables have problems with normality. When we run the model we get the following error: Weight variable has negative value at observation 110. We're confused because the weight variable is never negative. If we remove case 110 from the data file we get the same error referring to case 109. Considering that the error might refer to the application of the weight variable to one of the outcome or endogenous variables, we troubleshooted by shifting those vars positive by adding 10, at which point the model ran with no errors. However, this type of adjustment affects the results and is obviously not an acceptable solution. Can you tell me what the problem is and what we can do to remedy it? 


It sounds like you are reading your data incorrectly. You may have blanks in the data set or the number of variable names in the NAMES statement is greater than the number of columns in the data set. If you can't see the problem, please send the input, data, and your license number to support@statmodel.com. 


Linda, going back to your message posted on on Friday, March 09, 2012  5:33 pm, may I send the data and the Mplus output to the support team? Thanks. 


Yes. 


Hello, I created a model for my complex data and it appeared to run fine. However, when I added in a set of dummy variables to represent RACE I received the following error message: THE STANDARD ERRORS FOR H1 ESTIMATED SAMPLE STATISTICS COULD NOT BE COMPUTED. THIS MAY BE DUE TO LOW COVARIANCE COVERAGE. THE ROBUST CHISQUARE COULD NOT BE COMPUTED. With this error, I did not receive any fit indices for my model and, of course, there were no S.E.s reported either. I checked and all of the variables have high coverage values (at least .8 but most higher than .9). The dummy variables that I added had complete coverage  of course I only added 4 of the variables to the model because I have 5 categories. How can I address this error and keep RACE in my model? Thanks for your help! 


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


Hi, I am running a piecewise growth model with nonnormal data. Using MLR I get the following message, but the model estimates normally and I get parameter estimates identical to when I use ML. THE STANDARD ERRORS FOR H1 ESTIMATED SAMPLE STATISTICS COULD NOT BE COMPUTED. THIS MAY BE DUE TO LOW COVARIANCE COVERAGE. THE ROBUST CHISQUARE COULD NOT BE COMPUTED. Is it right to assume my chisquare has not been adjusted for nonnormality (and that it doesn't particularly matter if I am not computing chisquare difference tests) but that my standard errors HAVE BEEN adjusted for nonnormality? My concern is that my standard errors are deflated because of nonnormality. MLM and MLMV are not an option because of missing data. Thank you Kara 


The chisquare and standard errors for the H0 results are robust for nonnormality. The problem is with the H1 model. I would be concerned about low coverage. 


Thank you. Yes, I have do have low coverage because of the structure of the data and as a result do not get fit indices such as RMSEA and CFI when using MLR. When you have low coverage and nonnormal data which estimator would you recommend using? Thanks Kara 


MLR. But I would worry about having such low coverage that these problems happen. You may rely to much on statistical modeling assumptions and too little on data. If the coverage is due to planned missingness, that's another story. 


We have a cohortsequential design with participants assessed every 2 years. As a result, when we code time by age, participants only have data every two years. For example, some participants only have data at ages 12,14,and 16 and others only at 13, 15 and 17. If I continue to use MLR how can I assess model fit with this data? Thanks Kara 


You avoid this problem if you approach the analysis as a multiplegroup setup. See UG ex 6.18. 

Back to top 