Model identification PreviousNext
Mplus Discussion > Categorical Data Modeling >
Message/Author
 Anonymous posted on Tuesday, December 02, 2003 - 9:22 pm
This is the code I've made for behavioral genetics model on item level for categorical data.
The variables have 3 levels each.
It could be run, but still have the error message that this model may not be identified.
I have 5 variables for each pair so 10 variables.
10*11/2=55 data points.
5*3=15 different coefficients, and because equal variance is not constrained, 10 error variances. total 25 estimeates.
I don't know why this model is not identified.
Could someone can answer what was wrong with my program?
Thank you.

Mplus VERSION 2.14
MUTHEN & MUTHEN
12/02/2003 10:39 PM

INPUT INSTRUCTIONS

Title: Exploratory factor analysis Mplus for political cons.
Data: File is
'U:\SEM\mplus\punch.dat';
Format is (6X,F1.0,F1.0,50F1.0/8X,50F1.0);
Type is INDIVIDUAL;

Variable: Names are sex zyg
att1 att2 att3 att4 att5 att6 att7 att8 att9 att10
att11 att12 att13 att14 att15 att16 att17 att18 att19 att20
att21 att22 att23 att24 att25 att26 att27 att28 att29 att30
att31 att32 att33 att34 att35 att36 att37 att38 att39 att40
att41 att42 att43 att44 att45 att46 att47 att48 att49 att50

t2att1 t2att2 t2att3 t2att4 t2att5 t2att6 t2att7 t2att8 t2att9 t2att10
t2att11 t2att12 t2att13 t2att14 t2att15 t2att16 t2att17 t2att18 t2att19
t2att20
t2att21 t2att22 t2att23 t2att24 t2att25 t2att26 t2att27 t2att28 t2att29
t2att30
t2att31 t2att32 t2att33 t2att34 t2att35 t2att36 t2att37 t2att38 t2att39
t2att40
t2att41 t2att42 t2att43 t2att44 t2att45 t2att46 t2att47 t2att48 t2att49
t2att50;
categorical=
att6 att7 att16 att17 att23 t2att6 t2att7 t2att16
t2att17 t2att23;
useobservations=zyg eq 1 or zyg eq 3;

grouping is zyg(1=mzf, 3=dzf);
usevariables att6 att7 att16 att17 att23
t2att6 t2att7 t2att16 t2att17 t2att23 zyg;
model:
e1 by att6*(1)
att7*(2)
att16*(3)
att17*(4)
att23*(5);
e2 by t2att6*(1)
t2att7*(2)
t2att16*(3)
t2att17*(4)
t2att23*(5);
a1 by att6*(6)
att7*(7)
att16*(8)
att17*(9)
att23*(10);
a2 by t2att6*(6)
t2att7*(7)
t2att16*(8)
t2att17*(9)
t2att23*(10);
c by att6*(11)
att7*(12)
att16*(13)
att17*(14)
att23*(15)
t2att6*(11)
t2att7*(12)
t2att16*(13)
t2att17*(14)
t2att23*(15);

e1 e2 a1 a2 c@1;

model mzf:
a1 with a2@1;

model dzf:
a1 with a2@.5;

analysis:
matrix=covariance;
type=mgroup;
estimator=wlsmv;


INPUT READING TERMINATED NORMALLY


Exploratory factor analysis Mplus for political cons.

SUMMARY OF ANALYSIS

Number of groups 2
Number of observations
Group MZF 1232
Group DZF 747

Number of y-variables 10
Number of x-variables 0
Number of continuous latent variables 5

Observed variables in the analysis
ATT6 ATT7 ATT16 ATT17 ATT23 T2ATT6
T2ATT7 T2ATT16 T2ATT17 T2ATT23

Grouping variable ZYG

Categorical variables
ATT6 ATT7 ATT16 ATT17 ATT23 T2ATT6
T2ATT7 T2ATT16 T2ATT17 T2ATT23

Continuous latent variables in the analysis
E1 E2 A1 A2 C


Estimator WLSMV
Maximum number of iterations 1000
Convergence criterion 0.500D-04
Parameterization DELTA

Input data file(s)
U:\SEM\mplus\punch.dat

Input data format
(6X,F1.0,F1.0,50F1.0,/,8X,50F1.0)


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 64.
 bmuthen posted on Wednesday, December 03, 2003 - 6:06 am
To match the standard ACE model, the e1, e2 factors should be specified as uncorrelated and uncorrelated with the a and c factors, since the factors are correlated by default in Mplus. Note, however, that with categorical outcomes, the e-related variances are not identified. With continuous indicators, the e factors are really residuals. You have categorical indicators and residual variances are not separately identifiable parameters. You can handle this in 2 ways in Mplus. First, using the default Delta parameterization, fixing delta @1 for each item for both groups. Or, equivalently, using the Theta parameterization, fixing theta @1 for each item in both groups. An article by Carol Prescott is on its way out in Behavioral Genetics which will be posted this month. You may also contact her for further information.
 BMuthen posted on Wednesday, December 03, 2003 - 6:31 am
Just to clarify, with categorical indicators the e factors should be deleted from the model.
 Silvia Bacci posted on Friday, March 02, 2007 - 3:02 am
Dear authors,
I have a problem with a second order CFA (really similare to the example 5.6 in Mplus Giude).
I have 6 ordinal variables, with 4 categories each one, so we have 21 DF.
I want to estimate a model with three latent variables, two of them measured by three indicators each one, and one second order factor.

I used this input:

USEVARIABLES ARE
GIUD_COM GIUD_DIA GIUD_VAL
GIUD_STR GIUD_ST1 GIUD_ORG;
CATEGORICAL ARE
GIUD_COM GIUD_DIA GIUD_VAL
GIUD_STR GIUD_ST1 GIUD_ORG;
MISSING ARE all(999);
ANALYSIS:
ESTIMATOR IS WLSMV;
ITERATIONS = 50000;
CONVERGENCE = 0.0005;
MODEL:
DOCENTI by GIUD_COM GIUD_DIA GIUD_VAL;
STRUTT by GIUD_STR GIUD_ST1 GIUD_ORG;
GIUD by DOCENTI STRUTT;

The output says:
INPUT READING TERMINATED NORMALLY

and then:
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 8.

First af all, how can I understand wich is parameter 8?
Then, why the model is not identified?
By the way, I tried also the model with the only two latent factor (first order) and it works.
How can I solve the problem?

Best regards,
Silvia
 Linda K. Muthen posted on Friday, March 02, 2007 - 6:30 am
You can find out what parameter 8 is by asking for TECH1 in the OUTPUT command. Your second order factor has only two factor indicators. Such a model is not identified.
 Jaime Derringer posted on Thursday, October 11, 2007 - 10:03 am
Hello,

I am attempting a multi-group categorical CFA with thresholds with model input as follows:

MODEL:
f1 BY tob alc mar leg ill; !reference=m2
MODEL m1:
f1 BY tob alc mar leg ill;
[tob$1];
[alc$1];
[mar$1];
[leg$1];
[ill$1];
{tob@1};
{alc@1};
{mar@1};
{leg@1};
{ill@1};

...repeated for four additional groups (3 ages, divided by sex)
I consistently get an error message that the model may not be identified (so no standard errors or model fit tests are given) due to an error in parameter 21. Output tells me (I think) that 21 is alpha, the start value (currently 0) for the mean of the latent trait in group 2 (label=m1). Is there a way to change this start value? (Or am I even interpreting this error correctly? =)

Thank you,
Jaime
 Linda K. Muthen posted on Thursday, October 11, 2007 - 10:10 am
I think the problem is that when you mention the factor indicator tob in the group-specific MODEL command, it is freed causing the model not to be identified. If this does not solve the problem, please send your input, data, output, and license number to support@statmodel.com.
 Sara Jaffee posted on Monday, April 19, 2010 - 2:22 am
I am testing a measurement model with 15 indicators, all of which are 3-category variables. I am using WLSMV. I am getting an error message that says the model is not identified and I cannot see any obvious reason why. The message says:

THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL.
PROBLEM INVOLVING PARAMETER 64.

THE CONDITION NUMBER IS 0.973D-17.

Parameter 64 is the PSI matrix for the variance of the latent factor.

I'd be grateful for any insight into this problem.
 Linda K. Muthen posted on Monday, April 19, 2010 - 6:24 am
Please send your output and license number to supprot@statmodel.com.
 Kai Savi posted on Wednesday, November 03, 2010 - 1:55 pm
Hello, I am testing a model for measurement invariance across time. I am running a multi-group categorical model with the following syntax:

DATA: File (2004) = c:\Kai\bps2004.csv;
File (2006) = c:\Kai\bps2006.csv;
VARIABLE: Names are
ID EMPLMT RLV PELL IMPTB IMPTC IMPTE IMPTF IMPTI FREQA FREQB
FREQD FREQC FREQE FREQF FREQG HIDEGEX GPA PROUT AFFORD JOBHOUR
COMSERV WTA000 ATTENDA ATTENDB ATTENDC ATTENDD ATTENDE ATTENDF
ATTENDG;
Missing are all (-3);
Usevariables FREQC FREQD FREQE FREQF FREQG COMSERV;
Categorical are FREQC FREQD FREQE FREQF FREQG COMSERV;
idvariable = ID;
Weight = WTA000;
MODEL: contv by FREQC FREQD FREQE FREQF FREQG COMSERV;
MODEL 2004: contv@0;
{FREQC-COMSERV*.5}
MODEL 2006: contv@0;
{FREQC-COMSERV*.5}
OUTPUT: STANDARDIZED SAMPSTAT;

I receive the following error:
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.

Parameter 22 is DELTA for model 2004. Why is there an error for 2004 but not 2006? I am new to Mplus, and would appreciate any input you may have.
 Linda K. Muthen posted on Wednesday, November 03, 2010 - 3:04 pm
Scale factors must be fixed at one in one of the groups for the model to be identified. See the Topic 2 course handout on the website under multiple group analysis to see the models we suggest for testing measurement invariance of categorical outcomes.
 Kristine Khachatryan posted on Tuesday, October 11, 2011 - 5:22 am
I am running a multi-group CFA with categorical indicators studying measurement invariance. In my model I have four factors. The first factor is measured by 5 indicators (y1-y5), the second one with 2 indicators (y6-y7). For the third and fourth factors I have (unfortunately) only one indicator for each (y8; y9). So, for my model to be identified I have to fix the residual variance of y8 and y9 to 0, am I right? To test for the configural invariance I fixed the factor means to 0 and the scale factors to 1 in both groups and allow loadings and tresholds to vary across groups.
Here is my input:
MODEL:
f1 by y1-y5;
f2 by y6-y7;
f3 by y8;
y8@0;
f4 by y9;
y9@0;
[f1@0 f2@0 f3@0];
{y1-y9@1},
MODEL west:
f1 by y2-y5;
[y1$1-y9$1];

Doing this I got an error message saying that scale factors for categorical outcoms can only be specified using PARAMETERIZATION =DELTA with estimators WLS, WLSM, or WLSMV. But, as far I understood, with the delta parameterization scale factors are not allowed to be parameters in the
model.
I would very much appricate if you could help me to solve this identification problem and help me with the syntax.
 Linda K. Muthen posted on Tuesday, October 11, 2011 - 5:37 pm
Scale factors are allowed with the Delta parametrization. Residual variances are allowed with the Theta parametrization. It sounds like you have PARAMETRIZATION=THETA; in the analysis command. You should remove it.

Having one indicator with residual variance of zero is the same as working with the observed variable. I would do that.
 Andres Fandino-Losada posted on Wednesday, February 15, 2012 - 11:06 am
Dear Drs. M & M.
I am working on a path model with categorical ordinal (observed) dependent variables.
How does the "identification counting rule" work with ordinal variables?
Each ordinal variable should be consider as a set of dummy variables (i.e. number of levels - 1)?
Thank you.
Andres.
 Linda K. Muthen posted on Wednesday, February 15, 2012 - 11:12 am
It works the same way as for continuous variables. An ordinal variable is one variable unless you create a set of dummy variables and use those.
 Lyndsey Gott posted on Tuesday, June 19, 2012 - 2:01 pm
Hello,

I am trying to conduct a CFA involving 13 factors, with the ultimate goal of generating correlations between the factors (also need corresponding p-values).

I received 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 390.

THE CONDITION NUMBER IS -0.733D-16.

1. I requested TECH1 output to investigate what parameter 390 is. The only time 390 appears is beneath PSI. The 6th edition user manual tells me that the psi matrix contains the variances and covariances of the continuous latent variables and that the both the rows and columns represent the continuous latent variables in the model.

How am I to interpret this with regards to resolving the problem preventing standard errors (and subsequently, p-values) from being computed?

2. What is more appropriate to use when creating a correlation matrix: The MODEL RESULTS containing WITH comparisons between factors or the TECH4 estimated correlation matrix for the latent variables? Upon comparison, the correlation values generated by these two analyses are different. And from other CFA analyses I performed successfully, the WITH correlation section (and not the TECH4 output) includes the two-tailed p-values I need for my summary matrix.

Thank you so much in advance.

Lyndsey Gott
 Linda K. Muthen posted on Wednesday, June 20, 2012 - 10:46 am
1. Please send the output and your license number to support@statmodel.com.

2. TECH4.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: