I am quite new in SEM and perhaps my inquiry is trivial to you. I am trying to analyze a model (N-519)with 8 observed variables (path analysis) of which 5 are continuous, one is Likert 1-10 and 2 are categorical. The Likert variable and the two categoricals are DVs. Further, the continuous variables depart from mv normality. I am quite confused about which method of estimation to use. Any suggestions? Is there any literature on the subject?
You have two estimator choices in Mplus -- weighted least squares (WLSMV) or maximum likelihood (MLR). You may find the following articles helpful:
Muthén, B. & Kaplan D. (1985). A comparison of some methodologies for the factor analysis of non-normal Likert variables. British Journal of Mathematical and Statistical Psychology, 38, 171-189.
Muthén, B. & Kaplan D. (1992). A comparison of some methodologies for the factor analysis of non-normal Likert variables: A note on the size of the model. British Journal of Mathematical and Statistical Psychology, 45, 19-30.
See also Web Note 4 and other references that are available on the website.
Hi Linda and Bengt, I have two latent variables in an SEM, each of which has categorical indicator variables with the same four-point scales. Is it possible to interpret significant results of predictor variables on the criterion factors in terms of odds ratios since the latent variable is based on observed variables with the same four point scale. For instance, can I say, for each unit change in the predictor, the odds of risk beliefs increase by such and such (exponentiating the beta log odds)?
The regression of a continuous factor on a covariate is a simple linear regression and is interpreted as such. It is the regression of the categorical factor indicator on the factor that is a probit or logistic regression.
Linda, I also ran a test for an indirect effect, which was significant. I will report the effect plus confidence interval. However, is the parameter estimate a log odds beta if the outcome variable is categorical? I just want to be sure how to interpret the results for my paper.
Jungeun Lee posted on Friday, August 03, 2007 - 1:16 pm
I have a SEM model in which all observed variables are categorical. I used 'Categorical' option for this model and estimated it with Weighted least squares estimator. I am not quite sure about how to interpret coefficents. Following is a short version of my mplus input for the model. In parenthesis, I added my current thought about how these coefficents should be interpreted. Am I on track???
MODEL: (Probit)f1 by zpotgodr zpotplrr zpotharr; (linear regression)f2 on f1; (probit)ncmopotr on f2; (linear regression)f4 on ncmopotr;
I am running a path model with a combination of categorical and continuous indicators. I am running the same model on five different groups (one with the groups pooled, the other on four different groups separately). I have two main questions:
1) Although I use the same variables in with each group, I get starkly different sample statistics (means/thresholds/intercepts) for each group. For example, in one group I get a set of means that seems to make sense (e.g. 42 for a continuous variable with a range of 20-60, .42 for a binary ordered categorical variable ranging 0-1). But in another group, the same variables will have means of 4, and 2.6 for example. Why is this the case? Could this be causing convergence problems?
2) It seems like models with combinations of categorical and continuos variables have poorer model fit as measured by RMSEA compared with models having only continuous variables. Is this the case?
I have a relatively complex SEM model in which endogenous variables are a mixture of continuous, categorical, AND count variables. I can't seem to find an estimator that can simultaneously handle all three types of data. I am running version 3.1 (I know I need to upgrade!). Can MPLUS handle this situation? If so, what estimator would be able to handle a complex model with all three types of data? Thank you for your time.
You can do this in the current version of Mplus using maximum likelihood estimation. Note that numerical integration is required and each factor represents one dimension of integration.
Cecily Na posted on Wednesday, December 01, 2010 - 4:52 pm
Dear Linda, Can a factor contain indicators in different scales? For example a factor is drug use, the three indictors are 1)frequency of drug, 2)age at first time drug use, 3)whether or not IV users (dichonomous)? I guess I should use WLSMV estimation in this case? How would I interpret the paths between the factor and different indicators? Especially between the factor and the dichonomous indicator? Thanks a lot!
Factor indicators can be measured on different scales. The scale determines the type of regression coefficient. For continous, it is linear. For categorical, it is probit with WLSMV and logistic with ML unless the probit link is used.
mari posted on Monday, February 14, 2011 - 1:37 pm
I have two follow-up questions to the posting from [Jungeun Lee posted on Friday, August 03, 2007 - 1:16 pm]. I have a similar model in which all observed variables are binary. I used WLSMV for my ESEM model.
1) In the printed output under "model results", does the estimates of "factor by indicator" mean factor loadings or probit regression coefficients?
2) You confirmed that the estimates of "factor on covariate" are linear regression cofficients, and the estimate of "distal outcome on factor" is a probit regression coefficient(if distal outcome is binary). My ESEM model also have a path from covariate (binary) to distal outcome (binary). In the output, does the estimate of "distal outcome on covariate" also mean a probit regression coefficient (instead of a logistic regression)?
I am sorry for such a beginner's question. Many thanks in advance.
Following up the questions above (posted on 2/14/2011), I have three questions for my ESEM model with WLSMV. I have two EFA factors for 20 items, three covariates, and one distal outcome. All observed variables are dichotomous.
Q1: I am wondering if a path coefficient(e.g., 1.2)from a covariate to a EFA factor can be interpreted like one unit change in x increases y by 1.2?
Q2: one covariate is gender in my model. I am wondering if I should test measurement invariance before interpreting the path coefficient from gender to a factor.
Q3: If Q2 is yes, how about the path from a factor to distal outcome. My distal outcome is drug use (yes/no). In this case, do I need to test measurement invariance?
Q1. ESEM factors (typically) have the metric set so that their variances are 1. So 1.2 means that when x changes 1 unit, the factor changes 1.2 SDs.
Q2. That's always a good idea to make sure you are talking about the same factor for the two genders.
Q3. Yes for the same reason as Q2.
mari posted on Wednesday, April 13, 2011 - 7:42 am
As a follow-up question to your answer about Q3 above, I am wondering how to test measurement invariance for EFA factors in this ESEM model. I am learning multiple group (MG) analysis to test measurement invariance, but it seems that MG analysis cannot handle efa factors. Since all 20 items load on both factors, I believe that MG is not an option for me.
Then, how can I test measurement invariance for the path from efa factors to distal outcome? Am I missing something? I would appreciate any guidance. Thank you.
mari posted on Thursday, April 14, 2011 - 11:34 am
Thank you, Linda! After I tried ex.5.27, I got two more questions.
Q1. When I used "type=imputation" for 20 imputed data sets, the output did not print model fit information. When I used the same syntax to one of the 20 data sets without type=imputation, it printed model fit info. I wonder if model fit info cannot be computed when using multiply imputed data sets.
Q2. Even when using one data set, any MG models with commands "model g2" did not run. The following is input excerpts and errors:
------------------------------------- GROUPING is mrjfq3dyb (0 = g1 1 = g2);
Model: people disorder by stepsab2 - balcoab2 (*1); [people disorder @ 0];
Model g2: [stepsab2 - balcoab2];
*** ERROR The following MODEL statements are ignored: * Statements in Group G2:
*** ERROR in MODEL command EFA factors in the same set as PEOPLE must have all fixed or free means. Problem with: [ PEOPLE ]
-------------------------------------- When I did not have group-specific commands, the models ran well. I am wondering how I can resolve this problem.
Hi, I am trying to run an sem model with one binary dependent variable (u1) and two independent contiunous latent variables (f5 and f9). f5 contains f1-f4, f9 contains f6-f8. I have written the syntax as follows:
variable: names are x1-x39 u1; categorical is u1;
model: f1 by x1-x4; f2 by x5-x9; f3 by x10-x15; f4 by x16-x21; f5 by f1-f4;
f6 by x22-x25; f7 by x26-x30; f7 by x31-36; f8 by x37-x39; f9 by f6-f8;
u1 on f5 f9;
is this right?or i have to add anything else? note: the CFA for f5 and f9 has a good fit. Many thanks,
2) I would only correct for measurement error in a single-indicator model if you have very good information about the reliability and it is not high. You don't need to do that just to have the interaction.
1) If my exogenous variable (see ECWC above) is categorical (Likert scale 1, 2, 3), do I need to list it as categorical in my syntax? In the MPlus user manual it states to only list dependent variables as categorical.
Only dependent variables go on the CATEGORICAL list. In regression, covariates are either binary or continuous and in both cases are treated as continuous. You can treat your variable as continuous or create a set of dummy variables.
Thank you for your help. When I run the model (see syntax below) with the ML/MLR estimator I get the warning below. This warning goes away when I use the MLF estimator - can the results with the MLF estimator be trusted?
VARIABLE: NAMES ARE ECWC ECWCD LOVER REJECT LIVED SEXPRE SEXCOPE DSI2000 DSI2007 TOTVIC TOTCHVIC SSPIC SSPIA; USEVARIABLES ARE ECWC SEXPRE SEXCOPE DSI2007 TOTCHVIC SSPIC; MISSING ARE ALL (-9.00); CATEGORICAL ARE DSI2007 SEXPRE SEXCOPE! ONLY LIST DEPENDENT INDICATORS AS CATEGORICAL ( ANALYSIS: ESTIMATOR IS MLR; TYPE = RANDOM; ITERATIONS = 1000; CONVERGENCE = 0.00005; H1ITERATIONS = 500; H1CONVERGENCE = 0.0001; COVERAGE = 0.10; MODEL: SSR BY SEXPRE SEXCOPE; ! SEXUAL SELF REGULATION; PEDO BY SSPIC DSI2007 TOTCHVIC; !PEDOPHILIA; ECWCXSSR | ECWC XWITH SSR; ! COMPUTING INTERACTION TERM; SSR WITH ECWC; PEDO ON ECWC SSR ECWCXSSR;
WARNING: THE MODEL ESTIMATION HAS REACHED A SADDLE POINT OR A POINT WHERE THE OBSERVED AND THE EXPECTED INFORMATION MATRICES DO NOT MATCH. AN ADJUSTMENT TO THE ESTIMATION OF THE INFORMATION MATRIX HAS BEEN MADE.THE CONDITION NUMBER IS -0.245D-03.THE PROBLEM MAY ALSO BE RESOLVED BY DECREASING THE VALUE OF THE MCONVERGENCE OR LOGCRITERION OPTIONS OR BY CHANGING THE STARTING VALUES OR BY INCREASING THE NUMBER OF INTEGRATION POINTS OR BY USING THE MLF ESTIMATOR.
Chi-square and related fit statistics are not available if means, variances, and covariances are not sufficient statistics for model estimation. Difference testing of nested models can be done using -2 times the loglikelihood difference which is distributed as chi-square.