Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: GLLAMM -- soprobit link


From   Stas Kolenikov <skolenik@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: GLLAMM -- soprobit link
Date   Fri, 17 Feb 2012 09:26:55 -0500

My first reaction would be to check identification of your model (and
-gllamm.hlp- does note that scaled probits are difficult to deal
with). What I would be more worried about is the fact that your model
specification assumes the same thresholds for all items -- is that
appropriate? If not, you would want to concentrate on creating
item-specific thresholds first, and then try to move towards a more
flexible model on the raters' side. Modeling thresholds with
-thresh()- option usually requires -noconstant- to avoid problems in
identifying the shift parameters.

On Fri, Feb 17, 2012 at 7:05 AM, George Bouliotis
<g.bouliotis@bham.ac.uk> wrote:
> Dear Statalist members,
>
> I am using a dataset of 30 rates evaluating 34 items in an attempt to implement a polytomous ordered probit model with rater-specific measurement errors and thresholds, as it is described in the section 7.12 , book "Multilevel and Longitudinal Modeling Using Stata; Hesketh and Skrondal".
>
> Unfortunately, and despite numerous model modifications (e.g. nip(40), matrices, initial values etc)  , the Newton Raphson algorithm fails to converge although adaptive quadrature is reasonably stable. The usual error message "flat or discontinuous range" emerges any time I use the link "soprobit" whereas full convergence is achieved with "oprobit" links. Please see below some Stata output.
>
> My questions are:
> 1.      Is there any other trick/way for achieving full convergence? Would it make any difference if I omit a particular rater or raters from the model? Note that I do not have missing values.
> 2.      Is there any other way for taking rater-specific thresholds with GLLAMM or other command without using "soprobit"?
> 3.      Finally, could I base inference on only adaptive quadrature convergence?
>
> Thank you in advance
> George
>
>
> Dr. George Bouliotis
>
> Research Fellow in Biostatistics
> MRC-MHTMR (Midland Hub for Trials Methodology Research)
> University of Birmingham
>
>
> ----------------MODEL SPECIFICATION WITH GLLAMM
> gllamm   cat rtr1-rtr29, i(item) l(oprob) f(binom) adapt
> estimates store GLLAMM1
> **COND.NO: 20.61
>
> gllamm   cat rtr1-rtr29, i(item) l(oprob) f(binom)
> estimates store GLLAMM3
> **COND.NO: 26.68
>
> gllamm   cat rtr1-rtr29, i(item) l(oprob) f(binom) adapt iter(0)
> estimates store GLLAMM5
> ***COND.NO: 28.12
> *********Adaptive quadrature has not converged
>
>
> ***********
> ***********
> ***********
>
>
> gllamm   cat rtr1-rtr29, i(item) l(soprob) f(binom) adapt
> estimates store GLLAMM2
> **COND.NO:1191.92
>
> gllamm   cat rtr1-rtr29, i(item) l(soprob) f(binom)
> estimates store GLLAMM4
> **COND.NO: 4706.49
>
> gllamm   cat rtr1-rtr29, i(item) l(soprob) f(binom)  from(M1) skip
> estimates store GLLAMM6
> ***Condition Number = 3273.9524
>
>
> ***********
> *********** HETEROSCEDASTIC ERRORS
> ***********
>
>
> eq het: rtr1-rtr16 rtr18-rtr30 //measurement error
> gllamm   cat rtr1-rtr16 rtr18-rtr30, i(item) l(soprob) ///
> fam(binom) adapt s(het) nip(20) iterate(20) from(M3) skip init
>
> estimates store GLLAMM8
> ***NO convergence for adaptive ---Con number: NO
>
>
>
>
> ***********
> *********** THRESHOLDS
> ***********
> eq thr: rtr1-rtr29 //thresholds
> gllamm   cat rtr1-rtr29, i(item) l(oprob)  thres(thr) from(M3) skip  adapt
> ***Adaptive converged , Newton Raphson NO convergence ---Con number: NO
>
> eq thr: rtr1-rtr16 rtr18-rtr30  //thresholds RATER17 ref-guy
> gllamm   cat rtr1-rtr16 rtr18-rtr30 , i(item) l(oprob) f(binom) thres(thr) adapt init
> ***Adaptive converged , Newton Raphson NO convergence ---Con number: NO
>
> eq thr: rtr1-rtr16 rtr18-rtr30  //thresholds RATER17 ref-guy
> gllamm   cat rtr1-rtr16 rtr18-rtr30 , i(item) l(oprob) f(binom) thres(thr) adapt init
> ***Adaptive converged , Newton Raphson NO convergence ---Con number: NO
>
>
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/



-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index