Modification Indices for Categorical ... PreviousNext
Mplus Discussion > Categorical Data Modeling >
Message/Author
 Linda K. Muthen posted on Thursday, November 18, 1999 - 2:50 pm
TECH2 in OUTPUT provides parameter derivatives. The derivatives for constrained parameters (fixed or equal) can be used as modification indices although they have not been scaled to a chi-square metric. The relative size of the derivatives rather than their absolute value should be considered. When all variables are categorical, the derivatives can be examined to determine which parameters are most likely to improve model fit.
 Michael Bohlig posted on Saturday, July 07, 2001 - 12:43 am
Hello. I would like to verify that I am using the derivatives correctly to determine which parameters may help improve model fit. I have been looking at the magnitude of the absolute value of the derivatives as an indicator of potential parameters to free. Is this correct? If the sign is important, how would I interpret negative versus positive derivatives?
Thanks for your assistance.
 Linda K. Muthen posted on Sunday, July 08, 2001 - 9:23 am
The sign of the derivatives is the opposite of the sign that the parameter will take when it is freed if it was fixed at zero. Derivatives are not scale free so they cannot be compared across variables of different scales.
 Anonymous posted on Wednesday, August 27, 2003 - 2:38 pm
I have four questions about the use of TECH2 in Mplus.

1. Is it the case that TECH2 derivatives can only be used when "forward fitting" a model -- in other words, in a strategy where one starts with a restricted model and then successively frees parameters to achieve a better fit ?

2. How should the TECH2 derivatives for parameters already included in the model be interpreted ? Do large parameters suggest free terms that one might want to consider restricting to zero ?

3. What do you suggest should be considered a large TECH2 derivative in Mplus ?

4. Mplus often gives derivatives that don't seem to have any readily discernable meaning -- such as the derivative of the coefficient between the indicator of a CFA indicator with respect to an outcome variable with respect to a lambda. Does this make sense ?
 Linda K. Muthen posted on Wednesday, August 27, 2003 - 7:10 pm
1. Yes.
2. If the parameters are free, the derivatives should be zero. For fixed parameters, the absolute value of the derivative are not meaningful.
3. This can only be determined by freeing a parameter and seeing how large the chi-square decrease is.
4. Derivatives are given for all fixed parameters. It is up to the researcher to decide which parameters are meaninigful.
 Anonymous posted on Tuesday, November 30, 2004 - 2:58 pm
Hello,
I'm doing a two group CFA. In my nested model, I have both factor laodings and respective thresholds constrained to be equal across the two groups. I would like to use derivatives to see which factor loading/threshold should be freed to improve the chi square diff test. Could you please let me know what derivatives (Theta, Tau, etc) would give me the proper information? In freeng parameters, should I first look at factor loadings to be freed or the threshold?
 Linda K. Muthen posted on Tuesday, November 30, 2004 - 5:28 pm
In Version 3, modification indices are available for models with categorical outcomes. You don't need to look at derivatives. If you wanted to look at derivatives, factor loadings are found in lambda and thresholds in tau.
 Laura A. Szalacha posted on Friday, March 03, 2006 - 8:19 am
Sorry if this is too elementary - but, when I have my intervention parameter fixed it is positive and the model fits, when I free the parameter, it is negative. The rest of the model remains much the same.
Any help would be appreciated.
Thank you,
 Linda K. Muthen posted on Friday, March 03, 2006 - 9:34 am
It is unclear what you mean when you say that you have your intervention parameter fixed and it is positive. Have you fixed it at a positive value? The parameter estimates says that you have a negative intervention effect.
 Lois Downey posted on Friday, January 05, 2007 - 11:53 am
I notice that for CFAs with categorical outcomes, modification indices do not consider the addition of correlated residuals until one has actually specified at least one correlated residual in the model. Is there a way to see the Lagrange Multiplier test results associated with adding "WITH" relationships when the current model includes none?
 Linda K. Muthen posted on Sunday, January 07, 2007 - 9:01 am
No, you need to include them set at zero to open that matrix. For example,

u1 WITH u2@0;
 Kathy posted on Wednesday, April 16, 2008 - 8:18 am
I have a two group (categorical data) mgfa and I have a MI between two questionnaire items for one group which is : sd3 with sd19 = 157.13. Is this MI indicating that the error terms are correlated between these two questionnaire items? Is it appropriate to allow this correlation for one group and not the other? How would I do this in Mplus?
Would I:
model zz:
sd3 with sd19@0
 Linda K. Muthen posted on Wednesday, April 16, 2008 - 9:17 am
The modification index says that if you free the residual covariance between the two items in that group, chi-square will drop approximately 157.13. I wonder why this residual covariance is important for model fit in only one group. A residual covariance is a measurement parameter and this could be an indication of measurement non-invariance. It could represent a minor factor. I wonder if you did an EFA in each group separately as a first step in your analysis to determine if a model with the same number of factors fits best in each group.

To free the parameter you would specify:

model zz:
sd3 with sd19;

The @0 specification means fixed at zero.
 Li Lin posted on Wednesday, April 21, 2010 - 11:13 am
When I ran "model: f1 by y1 - y10; f2 by y11 - y18" where all y's were categorical with 5 categories, I got modification index of "F1 ON F1 / F1 BY F1 999.000". What does this mean? Thanks.
 Linda K. Muthen posted on Wednesday, April 21, 2010 - 1:22 pm
The value 999 indicates the value could not be computed. You should ignore this.
 Alyssa Thomas posted on Wednesday, July 30, 2014 - 7:27 pm
Hi there,
I have a question on obtaining modificiation indices for a categorical outcome. I am using WLSMV as the estimator and so am unable to select the 'modification indices with minimum..." option.
Should I use the TECH2 option to obtain this?
My independent variables are all Likert Scale and only my (one) outcome variable is categorical.
My supervisor has never worked with categorical outcomes so is unable to help.
Thank you in advance for your help
 Linda K. Muthen posted on Thursday, July 31, 2014 - 5:37 am
You should be able to do this unless you are using an older version that does not have this option. Try asking only for MODINDICES.
 Maria posted on Monday, December 07, 2015 - 6:48 pm
Hello Dr. Muthen, I conducted multi-group analyses in my two groups, with WLMVS estimator for categorical variables with the TECH2 option, but I do not get chi-square information to compare my models in the output. I also tried MODINDICES but no still no model fit information is provided. I am also not sure how to interpret derivatives on Lambda, Psi, Nu in the output between models.
 Linda K. Muthen posted on Tuesday, December 08, 2015 - 9:48 am
It sounds like the model did not converge. Please send the output and your license number to support@statmodel.com.
 Michelle Wu posted on Tuesday, February 06, 2018 - 1:59 pm
Hi Dr, Muthen,

I'm conducting analysis with multiple imputed data. When I requested for modification indices, Mplus gave me no output for that. Should I use TECH2 to look at the derivatives?

If so, is it correct that, if the size of derivatives is large for fixed parameters, I should consider to free it. Then what about those with larger sizes of derivatives but are not constrained in the model?
For example, in my measurement model, two items (x1, x2) had large values of derivatives in Theta, should I consider to x1 with x2 ?

Also, I'm confused with how to determine if I should keep or delete any factor loadings by looking at the derivatives.

Thank you very much.
 Bengt O. Muthen posted on Tuesday, February 06, 2018 - 2:08 pm
Please send your output to Support along with your license number.
 Esme Jordaan posted on Wednesday, February 28, 2018 - 5:51 am
I have the following situation:
Observed dependent variables
Continuous: STIGMA PTSD-HIGHLY non-normal ALC-HIGHLY non-normal SELF
Binary categorical:HIV NUMCL
Observed independent variables:
AGE_F EDU PL_B CHH
Continuous latent variables
F7 - 3 INDICATORS

I AM MODELING PATH MODELS AND TRIED USING ML, MLR, WLS, WLSMV ESTIMATORS WITH BIG DIFFERENCES IN PARAMETER ESTIMATES FOR THE ML AND WLS ESTIMATORS. ONE SUCH EXAMPLE IS SELF ON PTSD: ML; RATIO=4.01, MLR; RATIO=3.89, WLSMV: RATIO=4.83, WLS; RATIO=1.50, CHANGE FROM SIGN TO NS.WHICH ESTIMATOR SHOULD I USE?
 Bengt O. Muthen posted on Wednesday, February 28, 2018 - 3:48 pm
Note that with binary variables, ML uses logit as the default while WLSMV uses probit.
 Esme Jordaan posted on Thursday, March 01, 2018 - 12:48 am
My actual question related to the estimation (ML or WLS and their robust counterparts). The model has seven endogenous variables. Two are categorical (binary with 43/57 and 45/55 split). Three are continuous and normal. Two (PTSD and AUDIT score) are continuous and highly non-normal. These variables could also be treated as censored variables.
The path coefficients differ depending on the estimation used.
For ALC on SELF the ratio is 2.24 ML, 2.22 MLR,1.11 WLS and 2.13 WLSMV.
for SELF on PTSD the ratio is 4.01 ML, 3.89 MLR, 1.50 WLS and 4.83 WLSMV.
Please advise on the appropriate method to use.
 Bengt O. Muthen posted on Thursday, March 01, 2018 - 6:31 pm
Send your outputs to Support along with your license number and point out which coefficient differences you are wondering about.
 Bharath Shashanka Katkam posted on Thursday, August 08, 2019 - 6:51 am
Hello Mplus team,

I understand the

1) Derivative as the "Change in the Log-likelihood value" per unit "Change in the Parameter value".

2) Relative Log likelihood as the "Change in the Log-likelihood value" per unit "Change in the Parameter value".

So, if my understanding is correct, why there is a change in the values of the "Derivative & Log-Likelihood" in Mplus Outputs?
 Bengt O. Muthen posted on Thursday, August 08, 2019 - 3:33 pm
1) Please google the definition of derivative.

2) Where do you see "Relative Log likelihood"? In Tech8 we talk about "Relative Change" - is that what you refer to?
 Bharath Shashanka Katkam posted on Thursday, August 08, 2019 - 9:24 pm
1) The derivative is defined as
d/d(x) * ln (x)
which means, the "Change in the Log-likelihood value" [i.e., d ln (x)]
per unit "Change in the Parameter value" [i.e., d(x)].

2) Under the "Convergence Criteria of EM algorithm" there is a term "Relative loglikelihood change"
 Bengt O. Muthen posted on Saturday, August 10, 2019 - 6:35 am
In that context, relative loglikelihood change means

(LL2 - LL1)/LL1

where 2 and 1 are two consecutive iterations.
 Bharath Shashanka Katkam posted on Saturday, August 10, 2019 - 7:56 am
Thank you so much, Sir.
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: