Matt Diemer posted on Wednesday, December 07, 2005 - 8:58 am
I’m writing with questions about the relationships between latent variables in my structural model (the measurement model appeared to fit well) using NELS data. Fit indices are good (CFI .90 and RMSEA < .05).
Based on previous research, I would expect that the “SPD on ParenSup PrincSup Communal;” command would yield stronger relationships between latent indicators than I have obtained. (SPD had a significant relationship on ParenSup but not on PrincSup or Communal, as copied from output below).
Estimates S.E. Est./S.E. Std StdYX SPD ON PARENSUP 0.487 0.246 1.977 0.929 0.929
PRINCSUP 0.096 0.069 1.397 0.145 0.145
COMMUNAL 0.167 0.463 0.362 0.051 0.051
A couple problems/issues that I’m not sure how to address:
1. Large standard errors (or larger than I expected) in the relationships between latent indicators. 2. Whether to use MLR or default (I think WLSMV) as the estimator with complex sample data where one indicator is categorical (ComPart_). Relationships among latent variables are about the same if I do not label ComPart_ as a categorical variable and run MLR. 3. Some other issue I’m missing in my code or in modeling the data? 4. Maybe it’s simply that the relationships among latent variables that I had expected simply do not exist in the data?
TITLE: Test Mplus input syntax for NELS F2 ! Date is 12.6.05 Structural Model code
DATA: FILE IS C:\Documents and Settings\Vaio User\My Documents\Grants \MSU COE Seed Grant\MPlus analyses \11.7.05 Data set all participants for MPlus with STRATID and SCHID.dat; Format is FREE; TYPE IS individual; NOBSERVATIONS are 27394;
weight is cxtwt_; cluster is SCHID; strat is STRATID; categorical are ComPart_; MISSING ARE ALL (99); useobservations cxtwt_ ne 99 and race_ ne 4 and sesband eq 1 and evdost_ eq 0;
! For substantive reasons, analysis is restricted to lower-SES Adolescents of Color ! whom have not dropped out of high school
ANALYSIS: TYPE IS general complex missing h1;
MODEL: Communal by S7B_ S7G_inv C57O_inv T316Oinv; PrincSup by C51A_inv C51B_inv C51C_inv C51D_inv; ParenSup by P52K_tf P49H_ S99H_; SPD by ComPart_ S40F_ S40J_ s33e_; S40F_ with S40J_; SPD on ParenSup PrincSup Communal;
OUTPUT: standardized modindices;
bmuthen posted on Wednesday, December 07, 2005 - 6:26 pm
The CFI of 0.90 is not sufficient - we recommend CFI > 0.95. I would look at each factor at a time to see where the misfit comes from. Standard errors being large depends on many things, including low sample size and low observed-variable correlations.