I want to run a path analysis with two dichotomous dependent variables. My path analysis is fairly simple, but I also want to control for a number of variables that are not directly in the analysis. Usually, in SEM I would partial these variables out of my correlation matrix and then run the analysis on the partial matrix. However, I know that in Mplus I must use the raw data for categorical dvs. Is there any way to still control for these variables without adding them all to my path model?
bmuthen posted on Thursday, October 31, 2002 - 10:29 pm
I am not clear on how you partial variables out of a correlation matrix - is that something a regression approach can do? If so, this approach could be applied to raw data for categorical dvs.
What I was doing to control for variables in the past was to run my correlation matrix on the variables of interest, while partialling out the control variables. This results in a partial correlation matrix, where the bivariate correlations are a reflection of the correlation between the two variables after the variance that is due to a correlation with a third variable has been removed. I would then use that correlation matrix to run the path analysis. However, my impression from the Mplus manual is that with a categorical dv, I have to give the program raw data. Can I use the raw data(with my categorical DVs) and ask Mplus to use the partial correlations of my variables of interest?
Otherwise, my understanding of how to run control variables in path analysis is this. I have to add the control variables to the model and estimate their variances, and then covary them with all of the other variables. This results in a path model with a lot of extra paths. That is why I would prefer the former approach that I describe.
bmuthen posted on Friday, November 01, 2002 - 3:19 pm
It sounds like your correlation partialing-out procedure (CPO) is the same as what would be obtained by having the "third variable" as a covariate in a regression or path analysis. I may be wrong, but I suspect that the CPO procedure distorts the standard errors in the final path analysis, so I would vote for avoding this approach. The CPO approach could not be used with categorical dv's because even if you got the adjusted correlations, you would not have the corresponding correct weight matrix to get the correct s.e.'s and chi square.
Anonymous posted on Thursday, March 13, 2003 - 12:17 pm
I am estimating a model with one observed dichotomous outcome, but with one of the predictors being the outcome of a regression equation.
In the first equation, I estimated y* = f(y1, x1, x2,x3) & y1 = f(w1, w2, and w3). In the second equation, I estimated y* = f(y1, x1, x2,x3, x4) & y1 = f(w1, w2, and w3), adding in additional predictors for the dichotomous outcome.
What was strange was that the coefficients for w1-w3 changed drastically. If I changed y* to a continuous variable, the coefficients for w1-w3 remained constant. I assume that this change in coefficients is a function of the estimator (WLS), but the dramatic changes in the coefficients (one went from 0.3 to -1.1) was disconcerting.
No, it cannot. The reason for this is that a mediating variable must be able to be treated as binary or continuous in its role as a indepndent variable in the mediation and a nominal variable can't be treated that way.
Anonymous posted on Thursday, June 02, 2005 - 7:35 pm
Hi, I have two endogenous variables, one is binary y1 (0/1) and the other is nominal y2 (three categories, 0/1/2). Three questions:
1. I did the following specifications but MPLUS cannot execute my commands. What's wrong with my procedure?
2. When I use y2$1 and y2$2, I want to make y2=0 as the comparison baseline but I am not sure if I am doing it correctly.
3. In MODEL INDIRECT, can I use the following command?
y1 IND y2$1 y2$2 x1;
Please see my procedure:
... USEVARIABLES ARE ... CATEGORICAL IS y1; NOMINAL IS y2; ANALYSIS: PARAMETERIZATION=THETA; TYPE=MEANSTRUCTURE; MODEL: y1 ON y2$1 y2$2 x1 x2...; y2 ON x1 ... MODEL INDIRECT: y1 IND y2 x1;
I want to run a Multinominal logistic regression (Example 3.6) with one trichonomous dependent variable (Group 1,2, and 3) and five continuous covariates. Some like in Example 3.6 I want to take the alternative specification of dependent Group:
This is not how the categories are reordered. You would need to use the DEFINE command to change the numbers for the categories.
Cecily Na posted on Friday, December 10, 2010 - 7:48 pm
Dear Linda, The outcome variable in my SEM model is a dichotomous variable STD (yes/no). In addition to defining STD as a categorical variable, and use WLSMV estimator, is there anything else I need to do? For example, the predictors of STD are criminal (ordinal) and druguse (a continuous latent factor with categorical indicators v1 v2 ) My partial syntax are the following:
categorical variables are criminal v1 v2 STD druguse BY v1 v2; STD ON druguse; STD ON criminal;
The "ON" command is a logistic regression? How do I interpret the path values?
The CATEGORICAL list is for dependent variables only. Predictors must be binary or continuous. You can treat criminal as continuous or create dummy variables for it. The regression coefficients using WLSMV are probit not logistic. A good book to consult on probit regression is:
Long, S. (1997). Regression models for categorical and limited dependent variables. Thousand Oaks: Sage.
Dear Prof. Muthen, I am estimating a Longitudinal Cross-Lagged Design for Structural Equation Models. In this model I regress all control variables of the first wave on their second wave counterparts. But some of those control variables are dichotomous. When I identify those control variables of the second wave as categorical my model does not convergence. Do I need to indicate them as categorical? Or is this not required for those control variables? Thank you very much in advance, Kind regards, Caroline
Any variable only on the right-hand side of ON should not be placed on the CATEGORICAL list.
dvl posted on Wednesday, August 21, 2013 - 1:51 pm
I am performing a path model but one of my endogenous variables is an ordinal categorical variable with 4 categories. When I use WLSMV as estimator, mplus defines the categorical dependent variable as a continuous latent variable. I only get one parameter for each of my independent variables in the regression on the ordered categorical dependent variable. For example: my dependent categorical variable is "how often do you pray": Never, once a week, more than once a week, everday. So if I get for example an effect of -0,15 for age, how should I interpret it? Is that with a reference category? And if so, how do I know which category is the reference category?
Is it possible to do a path analysis with a nominal endogenous variable? I'd like to get a separate path coŽfficiŽnt for each category of my dependent (in comparison to the reference category). If so, could you give me some quick instructions where I could find information on this topic?