I'm a mplus novice and have a simple question. Under the categorical outcome analyses examples on your website, you have a path analysis example. In the analysis command, you don't specify "logistic". How then do I interpret the regression coefficients in this model. For example, for y8 on y5, the estimate is .246. I assume that this in not in logodds. Could I use the ouput to determine what the coefficient would be in logodds? Thank you.
With weighted least squares estimation and categorical outcomes, the regression coefficient is a probit regression coefficient. With maximum likelihood estimation and categorical outcomes, the regression coefficient is a logistic regression coefficient or a log odds. Version 3 allows both estimators with categorical outcomes.
I am running mediational models using path analysis with continuous and binary predictors and continuous and binary outcomes. The estimates are WLSMV using the THETA matrix. I am also using the TYPE=GENERAL COMPLEX because I need to use the CLUSTER function (I am looking at students within schools). Which estimates are appropriate to report?--the StdYX? I am assuming the two binary outcomes are probit estimates?
Thank You, Peggy Tonkin
BMuthen posted on Saturday, January 22, 2005 - 3:34 pm
The regression coefficients for binary dependent variables with WLSMV are probit regression coefficients. In line with regular regression, I would report raw coefficients as well as StdYX coefficients except when binary covariates are involved.
Peter Martin posted on Wednesday, September 28, 2005 - 4:59 am
Within a path model, how do I interpret the coefficient of a path where X is ordinal (but not binary) to a Y that may be either binary, ordinal, or continuous?
The scale of y determines the type of regression that is estimated. The scale of the exogenous x variable is not as issue in estimation. x variables can be either binary or continuous. With a binary or ordinal y variable, WLSMV estimates probit regresison coefficients. With a continuous y variable, WLSMV estimates a simple linear regression coefficent.
Peter Martin posted on Thursday, September 29, 2005 - 1:55 am
Thanks, Linda. So does this mean that an ordinal X is treated as if it was on interval scale? E.g. in a linear regression, the coefficient would state the increase in Y given an increase of 1 rank in X?
Or does the procedure use the latent variable that underlies X (this latent variable would be estimated, because the X has also paths leading to it)?
melissa posted on Thursday, July 12, 2007 - 8:30 am
I am running a SEM in which:
One endogenous latent variable is indicated by three dichotomous variables.
Another endogenous latent variable is indicated by two continuous and one dichotomous variable.
I have specified the categorical variables in the input and am using the wlsmv estimator.
Here are my questions: 1. Are the estimates related to the first mentioned latent variable (with all three dichotomous indicators) interpreted as probit estimates? 2. How are coefficients related to the second latent variable interpreted? 3. I am currently reporting both B's and StdYX's in my tables. I have the standardized coefficients labeled as Beta's. Is this inappropriate given the above mentioned latent variables? (I have other latent variables that do indeed include only continuous indicators).
The scale of the dependent variable determines the type of regression coefficient. For categorical factor indicators and WLSMV, probit regression coefficients are estimated. For continuous indicators and WLSMV, linear regression coefficients are estimated.
The labels don't depend on the variables being categorical or continuous. In both cases, the parameter estimates are regression coefficients.
Preeti posted on Saturday, December 13, 2008 - 2:29 pm
Hello. I am running an SEM model with a categorical outcome using the probit function. I have both latent and observed predictors. The reviewers would like effect sizes on my parameters. Could you please advise me on what would be the appropriate effect size to use and how to calculate it?
You have to decide what effect size is relevant here. Effect size is typically a difference in means under 2 different covariate conditions such a treatment/control, divided by the SD.
Sarah Ryan posted on Tuesday, September 20, 2011 - 3:08 pm
My model involves: 4 binary and 1 continuous control covariate (x1-x5) 2 observed exogenous predictors (z1-z2) 3 latent exogenous predictors (L1-L3) 1 latent mediator (LM1) 1 observed ordinal outcome (y)
I have a few questions related to coeff. interpretation.
1) I read elsewhere that the interp. of stdzd. probit is not as straightforward as with linear coeff.- Is this simply due to awareness of when to use STDYX and STDY, as well as how there is a diff b/w a unit change in continuous x versus a change in category (binary x)?
2) One of my controls shares a large and large and signif assoc w/ my mediator such that in the full model the expected large and signif assoc b/w the control and the outcome is negative and signif. Can I interpret that to mean that once one controls for the rel. b/w the control and LM1, the remaining variance in the control no longer shares the formerly assumed relationship with y?
3) With an ordinal outcome, does a positive beta indicate that an increase in x is associated with an increase in the probability of moving from one category to the next?
1) Standardizing probit/logit with respect to the covariate variance is no different from linear regression with continuous outcomes. For instance, you don't want to do it for a binary covariate. As for the DV, you don't really need to standardize wrt the binary outcome (or rather its latent response variable counterpart).
2) I'm unclear about this question. Did the control->y relationship go from positive to negative once the mediator LM1 was introduced?
3) Think of it as the latent response variable increasing and when it does the probability of a lower category goes down and a higher category goes up. However, a middle category probability first goes up but with further covariate increase then goes down, favoring a higher category. It is easiest to see the effect in a graph.
Sarah Ryan posted on Wednesday, September 21, 2011 - 8:29 am
Regarding 2) above: Yes, you understand correctly. LM1->control is strong and positive, y->LM1 is strong and positive, y->control is small to moderate and negative.
Further, if I model the paths between this control and the two indicators of the latent mediator with which the control shares a direct relationship (indicated by M.I.'s), the sign of the control/indicator relationship is also negative and the standardized y->LM1 now just over 1; y->control remains neg, but grows in magnitude.
I've been doing some reading on suppression effects, but I am not sure that is what is going on. I've also done all the things I can think of to do to assess multicollinearity effects. Collinearity diagnostics with all of the measured varbs. in the model were okay. All of the latent variable correlations are below .6 except for that b/w the mediator and outcome. None of the bivariate correlations are above .5, and most are quite a bit less. Because I'm using three waves of data, it is not logistically possible for the predictors, mediator, and outcome to be measuring the same thing (nor is it theoretically possible). My standard errors range from .01 to .07 (however, the sample size is about 5000- when I run the baseline with the second group (N=1000), the SE's are larger (.03-.14))