Chi-square difference test
Message/Author
 Carol posted on Wednesday, March 01, 2006 - 3:29 pm
Hi,

I am running a cholesky model using twin pairs. I'm trying to compare the full model to a submodel. I'm using MLMV as the estimator and I tried to use the DIFTEST option, but I keep getting the following error:

THE CHI-SQUARE COMPUTATION COULD NOT BE COMPLETED BECAUSE OF A SINGULAR MATRIX.

Can you please speculate as to what might be causing the error and how to get around it?

I should add that I also get a warning that one or more matrices are not positive definite.

Thank you,
Carol Van Hulle
 Linda K. Muthen posted on Wednesday, March 01, 2006 - 4:51 pm
This is an indication that the data and model are not compatible. I would need to see your input, data, output, and license number to say any more as this is a data dependent problem.
 Ken Sakurai posted on Friday, July 09, 2010 - 12:38 am
Hi Drs. Muthen

I need to compare and see if 2 path coefficients in a structural equation differs.

Let's say I have the following model.

A by A1-A3;
B by B1-B3;
C by C1-C3;
C on A B;

To compare if the size of path between A-C and B-C differs, should I compare the chi-square value of the above model against the chi-square value from the model below?

A by A1-A3;
B by B1-B3;
C by C1-C3;
C on A B(1);

This seems somewhat different from a multiple group analysis wherein one compares the difference in one path regression as a function of a grouping factor. In my case, I am comparing two different path coefficients.

 Linda K. Muthen posted on Friday, July 09, 2010 - 10:10 am
You can use difference testing to do this or MODEL TEST, for example,

A by A1-A3;
B by B1-B3;
C by C1-C3;
C on A (p1)
B (p2);
MODEL TEST:
0 = p1 - p2;
 Ken Sakurai posted on Saturday, July 10, 2010 - 11:18 pm
Hi Dr. Linda Muthen,

This is a follow up question to the above thread. I did run "MODEL TEST: 0=p1-p2;" like you suggested.

After that, I wasn't able to find the output that addresses the 0=p1-p2 model test. Was I supposed to request a specific output?

My question really is what happens after I run the analysis you suggested. Thank you,
Ken
 Linda K. Muthen posted on Sunday, July 11, 2010 - 6:33 am
You will find it with the fit statistics under the following heading:

Wald Test of Parameter Constraints
 shumail paracha posted on Wednesday, April 03, 2013 - 5:34 am
AIC and BIC values are also considered good for the comparison of two models....mplus has already given these values then why there is a need for calculating their values from log likelihood....are these values not effective for MLM? and please suggest me what is the correct formula of calculating their values if we can't consider AIC and Bic values given by output...
 Linda K. Muthen posted on Wednesday, April 03, 2013 - 10:54 am
There is a discussion in multilevel modeling about whether to use the number of clusters or the number of observations to compute AIC and BIC. Mplus uses the number of observations. If you want the number of clusters, you need to compute AIC and BIC by hand.
 Chilla Lee posted on Wednesday, November 27, 2013 - 4:01 am
Hi,

What should I do if I want to compare 3 coefficients at once?

It doesn't generate outcome if I write syntax like;

MODEL TEST:
0 = p1 - p2
0 = p2 - p3
0 - p1 - p3

Is it only way to make 3 runs in order to detect the difference and find out where the differences came from?

(I found that mplus runs the syntax if i write 2 lines of 0 = p1-p2 and 0=p2-p3, but i cannot find where the difference came from just like ANOVA in univariate analysis)
 Linda K. Muthen posted on Wednesday, November 27, 2013 - 11:28 am
MODEL TEST:
0 = p1 - p2;
0 = p1 - p3;

is a joint test of the three contrasts.

MODEL TEST:
0 = p1 - p2;

is a test of the difference between p1 and p2 being zero.
 Hadar Nesher Shoshan  posted on Wednesday, August 29, 2018 - 12:46 pm
Dear Linda,
5 years after the last question, I would like please to add a question to the topic. If the wald test is not significant, does it mean that the difference is not 0?
 Bengt O. Muthen posted on Wednesday, August 29, 2018 - 1:21 pm
It means that the difference is not significantly different from zero.