Anonymous posted on Sunday, April 14, 2002 - 4:14 pm
I ran a growth mixture model with four time-points data. The model was running fine with two classes. However, when I wanted to correlate the covariates, I got some error messages like the following:
"The following variable in a latent class regression ON statement is not an x-variable:MALE. *** WARNING in Model command The following variable in a latent class egression ON statement is not an x-variable: EDUC"
Here is my Mplus program: DATA: FILE IS BDIW0_3.dat; VARIABLE: NAMES ARE BLACK MALE AGE EDUC Y1-Y4 M1-M4; MISSING = ALL (-9); USEVAR = M1-M4 y1-y4 BLACK MALE EDUC AGE2; CLASSES=C(2); DEFINE: AGE2=(AGE-37.4); ANALYSIS: TYPE = mixture MISSING; ITERATIONS = 2500; MODEL: %overall% i BY M1-M4@1; s BY M1@0M2@1M3@2M4@3; m1 on y1; m2 on y2; m3 on y3; m4 on y4; [M1-M4@0 i*25 s*-1]; i with s; firstname.lastname@example.org s*-1; male with educ; i s on BLACK MALE EDUC AGE2; c#1 on black male educ age2;
%C#1% [i*10 s*0]; i with s*6.276;
OUTPUT: STANDARDIZED tech7; !tech4 sampstat; SAVEDATA: File is LCM_0.prb; Save is CPROB;
Your help will be appreciated.
bmuthen posted on Monday, April 15, 2002 - 8:22 am
Short answer: Don't say "male with educ".
Longer answer: These are "x variables" and you don't model them, but they are correlated by default. Mentioning them turns them into variables that have modeled relations, which is not allowed for variables influencing class membership.
Anonymous posted on Wednesday, April 17, 2002 - 10:44 am
Thank you very much for your prompt reply! I was just trying to test the covariances among the covariates in growth mixture modeling. For simplicity, I just specified one covariance: "Male with Educ," which specifies the covariance between variables Male and Educ in the model. It has no other meanings.
Variables Black, Male, Educ, and Age are hypothesized to affect both the growth factors and class membership. They are also hypothesized to be correlated with each other. I am glad to know that these variables are correlated by default in growth mixture modeling. Is there any option to print the correlations as well as their significance test information in output? And is it possible to uncorrelated, in necessary, some of the x variables (e.g., x1 with x2@0;)?
In addition, another set of outcome measures (y1-y4) in my model were treated as time-varying predictors of the outcomes m1-m4 in the model. I also want to test how y1-y4 were correlated with variables Black, Male, Educ, and Age. However, it was not allowed to specify these correlations in the program while regressions were allowed. I am wondering if the variables y1-y4 are also correlated with other "x variables" (e.g., Black, Male, Educ, and Age) by default? If so, is it possible to see them in output to check which correlations are significant?
Thanks for your help.
bmuthen posted on Thursday, April 18, 2002 - 12:08 pm
You can study the correlations among any of your variables by asking for analysis type = basic.
You cannot impose zero correlations among x variables that predict class membership. For other x variables you can however have x's uncorrelated.
Time-varying predictors are by default correlated with time-invariant predictors (black, male), but not if the time-varying predictors are also dependent variables. I raise this last issue since you refer to y1-y4 as outcome measures. If by that you mean that y1-y4 are both time-varying covariates for m1-m4 AND y1-y4 are themselves predicted by other variables then y1-y4 cannot be correlated but regressed on their predictors.
I am running a 5-class GMM model where one class is specified a priori as a "zero" class of nonusers of a substance. I also include predictors of class membership and ask for their variance to avoid listwide deletion. I receive an error message regarding a "reciprocal interaction."
USEVAR ARE POT8 POT9 POT10 POT12 POT96 POT99 POT02 POT05 MALE FRL567 BLACK ASIAN NATIVE POSFAM57 POSSCH57 ZASFR57 FAMMJ57 ZPRMJ7 BD8;