| Message/Author |
|
|
|
|
I’m using latent class regression on a 2-class solution with 6 indicators. For the regression, I am fixing the conditional probabilities for the indicator thresholds at the starting values and entering groups of covariates in the OVERALL statement using C#1 ON X1 X2 etc. I get the following error: ONE OR MORE MULTINOMIAL LOGIT PARAMETERS WERE FIXED TO AVOID SINGULARITY … The odd thing is that when I enter the first four covariates, all regression coefficients are fixed. In the second group (now 9 total), all are again fixed. However, when I enter the last two and a new parameter from Model Constraint: New, all 12 parameters are estimated. Since I’m fixing the latent class indicator thresholds, the classes don’t change in makeup, only in size (78 obs. are dropped in the first LCR and 120 are dropped in the 2nd and 3rd). I can’t figure out why adding more covariates would make the coefficients estimable.
Thanks for any insight you can give-
Michelle |
|
|
|
|
| I don't understand why you fix the indicator thresholds at the starting values. Perhaps you want to use the R3STEP option. |
|
|
|
|
Sorry; I should've been more clear--I'm using this approach so I can add a new parameter in Model Constraint, since I understand that command cannot include variables specified only as auxiliary variables. It seems like this approach is the one described by Masyn (http://www.depts.ttu.edu/immap/Masyn_LCAWorkshop_Dec2013_TTU.pdf, slide 119). My main consideration is to be able to create the linear combination parameter.
Is fixing the thresholds this way what Mplus is doing with Auxiliary (R) or (R3STEP)? |
|
|
|
|
I don't see on slide 119 anything related to what you are saying. The slide has the heading:
ALTERNATIVES TO DISTAL-AS-INDICATOR
I don't know what you mean by adding a new parameter.
Are you considering covariates or distal outcomes? |
|
|
|
|
I was referring to the part that says "Modified 1-step fixes all measurement parameters (e.g., item thresholds) at their estimated values from the unconditional model."
I am using covariates and talking about doing what we had discussed in another thread at http://www.statmodel.com/discussion/messages/22/19041.html. I am coding it as below; is this incorrect? The goal is to estimate the linear combination coefficient with standard error.
[omitted]
Usevar are
year1 educ1 ethnic1
lon1 rlf1 dep1 socanx1 selfest1
cximfr1 cxsocfr1
cxmofr1 cxbrofr1 cxofffr1
xtage1
olfc1 olfs1
bivat1 ;
Categorical are
cximfr1 cxsocfr1 cxmofr1 cxbrofr1 cxofffr1 bivat1;
[omitted]
Model:
%OVERALL%
C#1 ON
YEAR1
EDUC1
ETHNIC1
LON1
RLF1
DEP1
SOCANX1
SELFEST1
XTAGE1
OLFC1(p1)
OLFS1(p2) ;
[omitted]
Model constraint:
NEW(lincom1);
lincom1=p1+p2; |
|
|
|
|
| I can't see anything wrong in this setup. |
|
|
|
|
| Great! So back to the initial question--I'm not sure what could be going on that would make the coefficients estimable when all coefficients are included, but force them to be fixed when the first set and again the second set are added. Shouldn't the additional covariates make it harder to estimate parameters and not easier? |
|
|
|
|
| I don't know the answer to that in your specific setting. The general reason for the coefficients to be fixed is too little variation in the x variables within classes. |
|
|
| Back to top |
