

Two group LGM with categorical data 

Message/Author 

Daniel posted on Monday, April 05, 2004  1:55 pm



Hi, I am running a two group model and am testing for group difference on a fixed effect (e.g., slope on gender). I am using WLS as my estimator for the chisquare difference test. However, when I constrain paths to equality, I keep on getting a 2 df difference when I should only get a 1 df difference, isn't that correct? Is there something I'm missing? In other words, shouldn't the degrees of freedom only differ by 1 for constraining one path to equality across groups? 

Daniel posted on Monday, April 05, 2004  2:24 pm



Never mind, I found the problem. I was a dope! I didn't realize I used one of the same numbers in parens to constrain my paths to equality as was used for threshold constraining previously. Sorry to bother you. 


I am trying to run a multiple group latent growth curve analysis with a categorical (in this case binary) outcome. First I used the following sytanx... USEVAR ARE H1SU1 H2SU1 H3TO130; CATEGORICAL ARE H1SU1 H2SU1 H3TO130; WEIGHT IS gswgt1; CLUSTER IS psuscidn; GROUPING IS smy_dc (0 = hetero 1 = SMY); ANALYSIS: TYPE = MEANSTRUCTURE Complex; ESTIMATOR = MLR; MODEL: Isuicide BY H1SU1@1 H2SU1@1 H3TO130@1; Lsuicide BY H1SU1@0 H2SU1@1 H3TO130@6.5; [H1SU1@0 H2SU1@0 H3TO130@0] [Isuicide]; [Lsuicide]; Model Hetero: [Isuicide]; [Lsuicide]; Model SMY: [Isuicide]; [Lsuicide]; OUTPUT: tech4 sampstat; And got the following error: *** ERROR in Analysis command ALGORITHM = INTEGRATION is not available for multiple group analysis. Try using the KNOWNCLASS option for TYPE = MIXTURE. I then switched to a default estimator instead of MLR, and I switched to theta paramaterization, and got the following error: *** ERROR Cluster ID cannot appear in more than one group. Problem with cluster ID: 371 There is not a problem with the cluster variable. Before and after I got this error warning, I used these same data and the cluster command in my MG LGM with continiuos outcomes and it worked fine. Second, I'm not sure what this error means... by "group" I assume it does not mean the two groups I am analyzing since it is fine if individuals from these different groups are in the same cluster ID. Then, just to peel back the onion layers, I got rid of the cluster command and dropped the "complex" from the analysis command, and received the following error: *** ERROR The following MODEL statements are ignored: * Statements in the GENERAL group: [ H1SU1 ] [ H2SU1 ] [ H3TO130 ] Perhaps the strategy I'm using now is a close to correct and just needs a few changes, or perhaps I am way off. Either way, can you suggest a way to conduct multiple group LGM with a continous outcome. Also, ideally I'd rather do it in "meanstructure" opposed to "mixture" because I am more familiar with analyses using meanstructure, and I don't think it is even possible for me to carryout my planned subsequent analyses when using mixture. Thanks! 


If you use the weighted least squares estimator (WLSMV), you can use the GROUPING option with categorical outcomes. 


As suggested I used WLSMV as the estimator, and the model "ran" without an error. However, the output lists no estimates or standard errors since I have no free parameters. I should have 2 free parameters since, according to the tech 1 output, I am estimating less parameters (16) then my limit (18 since I have a 3 observed variables for each of my 2 groups). Here is my syntax: USEVAR ARE H1SU1 H2SU1 H3SU1; CATEGORICAL ARE H1SU1 H2SU1 H3SU1; WEIGHT IS gswgt1; GROUPING IS smy_dc (0 = hetero 1 = SMY); ANALYSIS: TYPE = MEANSTRUCTURE; ESTIMATOR = WLSMV; PARAMETERIZATION = THETA; MODEL: Isuicide BY H1SU1@1 H2SU1@1 H3SU1@1; Lsuicide BY H1SU1@0 H2SU1@1 H3SU1@6.5; [H1SU1$1@0 H2SU1$1@0 H3SU1$1@0] [Isuicide]; [Lsuicide]; Model Hetero: [Isuicide]; [Lsuicide]; H1SU1; H2SU1; H3SU1; Model SMY: [Isuicide]; [Lsuicide]; OUTPUT: tech4 sampstat; Here is my tech 1 output: TECHNICAL 1 OUTPUT PARAMETER SPECIFICATION FOR HETERO TAU H1SU1$1 H2SU1$1 H3SU1$1 ________ ________ ________ 1 0 0 0 NU H1SU1 H2SU1 H3SU1 ________ ________ ________ 1 0 0 0 LAMBDA ISUICIDE LSUICIDE ________ ________ H1SU1 0 0 H2SU1 0 0 H3SU1 0 0 THETA H1SU1 H2SU1 H3SU1 ________ ________ ________ H1SU1 1 H2SU1 0 2 H3SU1 0 0 3 ALPHA ISUICIDE LSUICIDE ________ ________ 1 4 5 BETA ISUICIDE LSUICIDE ________ ________ ISUICIDE 0 0 LSUICIDE 0 0 PSI ISUICIDE LSUICIDE ________ ________ ISUICIDE 6 LSUICIDE 7 8 PARAMETER SPECIFICATION FOR SMY TAU H1SU1$1 H2SU1$1 H3SU1$1 ________ ________ ________ 1 0 0 0 NU H1SU1 H2SU1 H3SU1 ________ ________ ________ 1 0 0 0 LAMBDA ISUICIDE LSUICIDE ________ ________ H1SU1 0 0 H2SU1 0 0 H3SU1 0 0 THETA H1SU1 H2SU1 H3SU1 ________ ________ ________ H1SU1 9 H2SU1 0 10 H3SU1 0 0 11 ALPHA ISUICIDE LSUICIDE ________ ________ 1 12 13 BETA ISUICIDE LSUICIDE ________ ________ ISUICIDE 0 0 LSUICIDE 0 0 PSI ISUICIDE LSUICIDE ________ ________ ISUICIDE 14 LSUICIDE 15 16 What I am I missing? I assume that whatever it is, my problem is related to either the estimator I'm using (WLSMV) or the fact I'm using categorical outcomes. Thanks 


This is an Mplus support question. Please send your input, data, output, and license number to support@statmodel.com. 

gibbon lab posted on Wednesday, October 12, 2011  1:57 pm



Hi Professor Muthen, For longitudinal binary outcomes (1/0), if the proportion of 1's increases along study waves and assume 1 does not change back to 0, (e.g., suppose the outcome is yes/no smoking and no one quit smoking during the study), is latent growth model still applicable in this case? Thanks. 


Answered under another thread. 

Back to top 

