CFA all binary Variables
Message/Author
 Maria Rueda posted on Friday, April 29, 2011 - 9:42 am
Hi,
I'm performing CFA with all binary variables. Which time of estimator do you recommend me to use ML or WLSMV? there is an specific estimator I should not use?

Maria
 Linda K. Muthen posted on Friday, April 29, 2011 - 10:19 am
You can use either. If you have several factors, WLSMV is best because with ML each factor with binary factor indicators requires one dimension of integration. If you want to include residual covariances between factor indicators, WLSMV is also best because with ML each residual covariance requires one dimension of integration. Models using more than four dimensions of integration are not recommended.
 Scott Smith posted on Monday, October 14, 2013 - 11:25 am
Can you further explain what dimensions of integration are? Does this mean that I shouldn't run more than four factors within one CFA model at a time?
 Linda K. Muthen posted on Monday, October 14, 2013 - 12:05 pm
Numerical integration is discueed on pages 470-473 of the Version 7 user's guide.
 Hugo posted on Tuesday, June 24, 2014 - 3:20 am
Hi,

I'm performing CFA with all binary variables. I have a problem becouse I have to set to 1 the variances of the three latent variables (all latent variables) to estimate the model, and I think that this is the reason to obtain the same results in Model results and Standardized estimates.

I need the results of the STDYX output for this model and I have 7 standarized estimates greater than 1, e.g x4 has a estimate=1.141 and S.E.=0.177

I´m wondering if you know some formula to fix the standardized estimates in the STDYX and obtain values less than 1.

 Linda K. Muthen posted on Tuesday, June 24, 2014 - 8:02 am
Standardized coefficients can be greater than one. See the FAQ on our website.
 Danica Cruz posted on Sunday, August 03, 2014 - 6:53 pm
I read in the Mplus User's Guide that it's possible to have combinations of continuous, binary and categorical indicators in a measurement model. Is it possible to have these combinations on the same factor? For example a binary variable and two categorical variables (one with 4 levels and one with 5 levels)? If so, how would the scaling work? Thank you.
 Linda K. Muthen posted on Monday, August 04, 2014 - 11:56 am
Yes, you can have combinations. You can set the metric of the factor by having all factor loadings free and fix the factor variance to one.
 Jorge Fernando Pereira Sinval posted on Monday, March 07, 2016 - 11:45 am
Dear Professor Muthén.

I'm working with an instrument with 10 dimensions (151 binary items). Some of the items belong to more than one factor, what can I do? I tried doing a common CFA for binary items, after 7 hours of working the output said "NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED". What can I do?

Thank you.
 Linda K. Muthen posted on Monday, March 07, 2016 - 2:37 pm
Allow the items to load on more than one factor, for example,

f1 BY y1 y2 y3 y4;
f2 BY y5 y6 y7 y2;
 Jorge Fernando Pereira Sinval posted on Monday, March 07, 2016 - 3:14 pm
Dear Professor.

Yes, I did that, I put in each factor all items that the original author proposed. But I always get that message. For example:

MODEL:

DS BY Item1 Item3 Item4 Item6 Item7 Item9 Item15 Item20 Item22 Item27
Item31 Item33 Item36 Item42 Item53 Item55 Item56 Item57 Item58 Item60
Item64 Item66 Item67 Item69 Item72 Item73 Item74 Item77 Item79 Item81
Item84 Item85 Item88 Item93 Item102 Item103 Item104 Item107 Item112
Item113 Item114 Item115 Item117 Item122 Item125 Item127 Item129 Item131
Item133 Item135 Item140 Item142 Item145 Item147 Item148 Item150 Item151
Item152 Item154 Item155 Item158; !62

OV BY Item5 Item9 Item12 Item20 Item23 Item24 Item29 Item30 Item31
Item33 Item41 Item48 Item53 Item55 Item56 Item58 Item67 Item71 Item72
Item77 Item80 Item84 Item85 Item93 Item94 Item102 Item103 Item107
Item111 Item115 Item122 Item123 Item125 Item138 Item148 Item149 Item152
Item157

Some items are in four factors.

Should I write any specific command on the input file?

Again, thank you.
 Bengt O. Muthen posted on Monday, March 07, 2016 - 5:50 pm
Try an EFA. Maybe the CFA is too poorly fitting.

If you have convergence problems send output to Support along with your license number.
 Jorge Fernando Pereira Sinval posted on Monday, March 07, 2016 - 6:07 pm
Dear Professor.

Thank you, I will try the EFA. In the CFA can I increase the number of interactions?

Once again, thank you.
 Anton Dominicson  posted on Tuesday, August 23, 2016 - 12:19 pm
Hi. I am trying to do a monte carlo simulation to determine an appropriate sample size for the estimation of a CFA model with a single latent variable measured by 8 binary variables. However, Im having trouble finding a way to do the CFA simulation with all binary variables.

Using the example mcex5.2 as a template (montecarlo example for a 2 factor CFA, all binary variables) I tried the following:

montecarlo:
names = u1-u8;
generate = u1-u8 (1);
categorical = u1-u8;
nobs = 300;
nreps = 10000;
model montecarlo:
f1 by u1@1 u2-u8*.8;
u1-u8*.2;
f1*.8;
model:
f1 by u1@1 u2-u8*.8;
u1-u8*.2;
f1*.8;
output:
tech9;

The result is the following:

Variances for categorical outcomes can only be specified using PARAMETERIZATION=THETA with estimators WLS, WLSM, or WLSMV.

After adding ANALYSIS: PARAMETERIZATION=THETA, I get these error messages for each replication:

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

Is there a problem with my use of the mplus language? Or is there something inviable about what I want to do?

Also, I understand that the simulation can be carried out with ML/MLR, but I don’t know about the pertinence of using numerical integration.
 Bengt O. Muthen posted on Tuesday, August 23, 2016 - 2:21 pm
Note from mcex5.2 that the Model command drops the line with residual variances that is used in Model Population.

Residual variances with binary outcomes are only identified with multiple groups or time points.

Note also that your parameters values in Model Population don't make the latent response variable variances add to 1 as is assumed for the default Delta parameterization of WLSMV. With factor loading 0.8, factor variance 0.8, and residual variance 0.2, the latent response variable variance is 0.64*0.8 + 0.2 which is not 1. This makes the estimates to come out in the wrong metric.

ML/MLR would be fine here - then you don't mention the residual variances in Model Population either.
 Anton Dominicson  posted on Wednesday, August 24, 2016 - 9:23 pm
I understand now, and the simulation ran appropriately. Thank you very much for your help.