[Limdep Nlogit List] problem matrix singular in latent class model

Wed Sep 7 16:12:29 AEST 2022

Dear Greene, 
I ran three types of models (RPL, LC, and LCRPL), for LCRPL, I was reading the different articles, but none of them say what criteria they use to define which parameter should be a random parameter. This is a doubt that I have in LCRPL. I specify my LCRPL model from de results of mixed logit. In this model, I check the standard deviation, and I sow which parameters are significant, and I used this parameter as RPL in my model LCRPL, but I am not sure if this is correct. Another option than I read, is use The  Lagrange  Multiplier  tests  proposed  in  McFadden  and  Train  (2000). 

In the last email, I ask you if LC is better than RP or RP is better than LC, And you ask me about the fit measure. 

Fit measure of RPL, LC LCRPL

Model 
Log likelihood
McFadden R2
AIC
Random Parameter
Observation 
RPL or Mixed logit
-545.31416
.19
1.85

Parameter with Standard Deviation p<0.05= LAND75, LAND25, SILVO, JAG0

LC
-484.69
0.28
1.75

Seven parameters were significant. However, the significance; and signal were the same in the four classes, except for two parameters that have different signal. 
LCRPL
-490.08
0.27
1.88
All parameter
None of parameter were significant
LCRPL
-561.98
0.16
1.96
land75, Silvo, JAG0
Only two parameters were significant in one class.
LCRPL
-494.60
0.26
1.81
Silvo, JAG0
Only one parameters was significant in one class. 

Thank you
Rergards 
Elias 

Enviado desde Correo para Windows

De: William Greene
Enviado: martes, 6 de septiembre de 2022 09:04 a. m.
Para: Elias José Gordillo Chávez
Asunto: Re: [Limdep Nlogit List] problem matrix singular in latent class model

Elias.  Walker is incorrect about only being able to identify two RPs.  However, neither Walker nor I and Hensher were
discussing latent class models.  You can have more than two RPs in a pure mixed logit model.  I am skeptical about
the entire LCRP specification; it seems overspecified.  But, still, it may be identified if there are enough restrictions 
across classes.  As for a better fit with LC than RP, I would want to know more about the fit measure before deciding that
one model fits better than another.  It is possible that you will find your data are more suited to LC than RP, or to RP
than LC.
Regards
Bill Greene

On Tue, Sep 6, 2022 at 1:20 AM Elias José Gordillo Chávez <elias.gordillo at yahoo.com.mx> wrote:
Dear Greene thank you for your advice; I´ve been Reading and have some questions. I was Reading the thesis of Walker 2002, y he gave some recommendations; he said, with three alternatives (my case), the maximum number of random parameters that can be identified is two. But on the other hand, in your article The mixed logit: the state of practice (2002), you talk about the different methods to know which are random parameters. But what happens If I get more two-parameter? Can I run the different models with different parameters and use the AIC para choice as the best model?

I have to comment that I ran a classic latent class model with 4 classes, and I have coefficients more adequate and a better fit than a model RPL, and LCRP. In this case, es possible that my date set is more according to this model than model more complex as LCRP?.

excuse, I know that are a lot question onthe same issues, thank.

Regards
Elías Gordillo 

Enviado desde Correo para Windows

De: William Greene
Enviado: lunes, 5 de septiembre de 2022 07:36 a. m.
Para: Elias José Gordillo Chávez; Limdep and Nlogit Mailing List
Asunto: Re: [Limdep Nlogit List] problem matrix singular in latent class model

Dear Elias. The model is definitely not identified.  The LCRP model would typically be used with one
or a small few RPs.  You have specified the entire model to be both continuously and discretely 
random.  I think this is too much to ask of a finite sample.  I'm not sure even of two classes. You might
want to look at  the two class results.  If the standard errors are *huge* that would be an identification
problem as well.
/B. Greene
/Bill Greene

On Mon, Sep 5, 2022 at 12:10 AM Elias José Gordillo Chávez via Limdep <limdep at mailman.sydney.edu.au> wrote:
Dear Willian Greene 
Thank you very much; I can run the model with my full dataset (876 individuals and three options). But I have a problem I can only run the model with two classes; if I try to use 3 or 4, I get an error message: Error 1027: Models - variance matrix of estimates is singular. 

My question is, is it possible that my data is inappropriate for the model, or am I specifying the model wrong? 

I used the following syntax for the random latent class:

LCRPLOGIT
;LHS=CHOICE
;CHOICES= A,B,C
;RH2=ONE
;RHS=LAND75,LAND25,LAND10,INTENS,SILVO,EXTENS,WP,JAG0,JAG2,JAG9
;RPL
;DRAWS=50
;FCN= LAND75(t),LAND25(t),LAND10(t),INTENS(t),SILVO(t),EXTENS(t),WP(t),JAG0(t),JAG2(t),JAG9(t)
;HALTON
;PDS=6
;PTS=2 
;maxit=100; tlg=.0001
;lcm
U(A)= LAND75*LAND75 + LAND25*LAND25 + LAND10*LAND10 +INTENS*INSTENS + SILVO*SILVO + EXTENS*EXTENS +WP*WP +JAG0*JAG0 +JAG2*JAG2+JAG9*JAG9/
U(B)= LAND75*LAND75 + LAND25*LAND25 + LAND10*LAND10 +INTENS*INSTENS + SILVO*SILVO + EXTENS*EXTENS +WP*WP +JAG0*JAG0 +JAG2*JAG2+JAG9*JAG9/
U(C)= LAND75*LAND75 + LAND25*LAND25 + LAND10*LAND10 +INTENS*INSTENS + SILVO*SILVO + EXTENS*EXTENS +WP*WP +JAG0*JAG0 +JAG2*JAG2+JAG9*JAG9$

Thank you
Regards 
Elias Gordillo 

Enviado desde Correo para Windows

_______________________________________________
Limdep site list
Limdep at mailman.sydney.edu.au
https://protect-au.mimecast.com/s/DC6UCGv0oyCJV3Y17iKOQWb?domain=limdep.itls.usyd.edu.au

-- 
William Greene
Department of Economics, emeritus
Stern School of Business, New York University
44 West 4 St.
New York, NY, 10012
URL: https://protect-au.mimecast.com/s/6Y0uCJyBrGf879rqJtGpKfD?domain=people.stern.nyu.edu
Email: wgreene at stern.nyu.edu
Editor in Chief: Journal of Productivity Analysis

Editor in Chief: Foundations and Trends in Econometrics
Associate Editor: Economics Letters
Associate Editor: Journal of Business and Economic Statistics

-- 
William Greene
Department of Economics, emeritus
Stern School of Business, New York University
44 West 4 St.
New York, NY, 10012
URL: https://protect-au.mimecast.com/s/6Y0uCJyBrGf879rqJtGpKfD?domain=people.stern.nyu.edu
Email: wgreene at stern.nyu.edu
Editor in Chief: Journal of Productivity Analysis

Editor in Chief: Foundations and Trends in Econometrics
Associate Editor: Economics Letters
Associate Editor: Journal of Business and Economic Statistics