Nominal variables
Message/Author
 freek bucx posted on Thursday, March 15, 2007 - 5:48 am
Dear all,

I want to use a nominal variable in my model, both as a dependent and an independent variable. This variable has four categories, of which one is the reference category.

In the syntax, I indicate that my variables are nominal (command: NOMINAL ARE); in the models I use VARIABLE1#1, VARIABLE1#2 etc. to indicate to which categories I am referring.

However, I get the following error message: "An nominal variable may not appear on the right-hand side of an ON statement".

My question: how can I use my nominal variables, both as a dependent and an independent variable, in my SEM analyses?

Thanks for all your answers and suggestions,

Freek Bucx.
 Bengt O. Muthen posted on Thursday, March 15, 2007 - 5:00 pm
Independent observed variables ("x" variables in Mplus terminology) should not be declared nominal, but should instead be broken up into a set of dummy variables appearing on the RHS of ON. I assume that your independent nominal variable is not in turn predicted by another variable, in which case something different needs to be done.
 freek bucx posted on Monday, March 19, 2007 - 4:18 am
Hi,

Thanks for your answer. The latter is however the case: my independent (nominal) variable (with four categories) is also predicted itself by other variables. How can I fix this in Mplus?

Freek Bucx.
 Linda K. Muthen posted on Monday, March 19, 2007 - 10:31 am
The only way to do this is to create a categorical latent variable that is equal to the nominal observed variable and use the categorical latent variable in the regression. For example, if you have a model:

x -> u -> y

where u is a nominal variable, after you create a categorical latent variable c which is equal to u, you will have

x -> c -> y

where c is a categorical latent variable with four classes. The relationship between x and c is the multinomial regression of c on x and the relationship between c and y is found in the class-varying means of y.
 ylam posted on Monday, October 28, 2013 - 9:34 am
Hi,
I wanna to create a latent variable (L) which is measured by C (4 level categorical variable) and N (nominal variable as indicated by 0 & 1) to reflect the demo characteristics, where IV is independent variable, M is mediator and DV is dependent variable.

L by C N;
IV on L;
D on M IV;
M on IV L;

nominal = N#1 N#2;

however, error message prompt for unknown nominal variable, how can I fix this problem?

And, If an indicator belongs to two latent factors, how to interpret this structure? How about if two indicators of two latent factors are correlated?

Thank you very much.
 Linda K. Muthen posted on Monday, October 28, 2013 - 10:14 am
You would say

NOMINAL = n;

However, if it is binary, just put it on the CATEGORICAL list with the other categorical variables.

Factor indicators can used for more than one factor and factors can be correlated.
 Jinseok Kim posted on Sunday, December 08, 2013 - 7:47 am
Hi, I am trying to estimate a model in which a nominal variable (u) is predicted by other variables and the same nominal variable (u) is acting as an independent variable of another variable.

Earlier in this thread, you mentioned that, in such a case, I will have to create a categorical latent variable, or a latent class (c) that is equivalent to the nominal variable (u). Would you please give us an example of mplus syntax that may do the job?

Jinseok
 Bengt O. Muthen posted on Sunday, December 08, 2013 - 10:41 am
That's an advanced analysis described in section 8 in

Muthén, B. (2011). Applications of causally defined direct and indirect effects in mediation analysis using SEM in Mplus.

which is on our website. It has Mplus setups.
 Andy Daniel posted on Monday, January 13, 2014 - 5:24 am
Hi,

I'm struggling with a problem concerning dummy variables in a SEM derived from a nominal variable.

The way I was used to include nominal variables as predictors in a Regression/SEM Model was splitting the nominal variable into k-1 dummys with the remaining categeory as reference.

I recently learned from a book, that including dichotomous variables into a regression model is highly problematic.

According to this book this is because the standard deviation is a direct function of the variable's skewness, which results in unreliable standardized coefficients and level of significance. Thus the interpretation of effects would be hard.

Is there a way to handle this problem in Mplus?

Thanks a lot for your help!!
 Bengt O. Muthen posted on Monday, January 13, 2014 - 8:41 am
I don't see that there is a problem here. You don't want to standardize with respect to the variance of a dummy variable because you are not interested in a standard deviation change in the variable but in the (raw) change from 0 to 1.
 Andy Daniel posted on Tuesday, January 14, 2014 - 5:40 am
Ok thanks for the quick answer.

Under this circumstances, would it be possible to compare other path coefficients (e.g. between two metric variables) with the path coefficient of the dichotomous variable on a metric variable (like: effect a is stronger than effect b)?

And is the level of signifiance of the coefficients trustworthy? Will it be influenced/interpretable?
 Bengt O. Muthen posted on Wednesday, January 15, 2014 - 11:07 am
I would not compare coefficients for binary covariates with those of continuous covariates. The SE for a standardized coefficient is trustworthy.
 Andy Daniel posted on Friday, January 17, 2014 - 2:07 am
Thanks a lot for this clarification!
 Edwin Wouters posted on Tuesday, September 13, 2016 - 6:29 am
Dear Sir or Madam,

When estimating a model with Nominal outcomes, which results should one present in an article? STDYX, Model results or Odds ratios?

 Bengt O. Muthen posted on Tuesday, September 13, 2016 - 9:53 am
I would present the logit estimates that are printed and also odds ratios. Note that odds ratios for continuous x's refer to a one unit change and you may instead use a one SD change. This would be obtained if you standardize the x's. See also Chapter 5 in our new book including graphical displays.
 Edwin Wouters posted on Friday, September 16, 2016 - 3:09 am
Dear Dr. Muthen,

Thank you for your response. If I want to present the odds ratios, should I then look at the p-values of the models results? or these of the STDYX results?

Kind regards
 Bengt O. Muthen posted on Friday, September 16, 2016 - 11:16 am
If you don't want to work in the original metric of your x variables the simplest approach is to standardize them first. Then report significance of the slope estimates (raw form) and also give the odds ratios we print. See also Chapter 5 of our new book.