sem question

-----Original Message-----
From: Spencer Graves [mailto:spencer.graves at pdf.com] 
Sent: Sunday, July 16, 2006 7:29 PM
To: Denis Fomchenko
Cc: r-help at stat.math.ethz.ch; John Fox
Subject: Re: [R] sem question



MODEL UNDERIDENTIFIED?

	  I've looked at 'sem' for many years but never found 
that application that seemed to me to require that machinery. 
 However, I know that it's very easy to get models that are 
"underidentified."  One of the simplest cases is the 
classical "errors in x regression" problem:

Observe:
	  X = xi + e.x, e.x~N(0, s2.x)
	  Y = eta + e.y, e.y~N(0, s2.y)
Model:
	  eta = a+b*xi

	  If I'm not mistaken, I believe that it is 
theoretically impossible to estimate a, b, s2.x, and s2.y 
without additional information, like for example the ratio 
between s2.x and s2.y.


LAGS IN BOTH TIME AND SPACE?

	  I've copied John Fox, the 'sem' package author and 
maintainer, on this reply.  He might educate us both on how 
to include lags in both time and space into an 'sem' model.

	  Failing that, are you familiar with Pinheiro and 
Bates (2000) Mixed-Effects Models in S and S-Plus (Springer). 
 This book and the companion 'nlme' packages include 
facilities for linear and nonlinear models in both space and 
time.  The follow-on 'lme4' package and accompanying 'lmer' 
function will also handle non-normal response distributions.  
I'm a firm believer in trying the simple things first, and I 
think the mixed-effects models are simpler than 'sem', though 
Prof. Fox may wish to disabuse me of my ignorance on that point.


MORE HELP?

	  If you would like more from this listserve than just 
this, please submit another post.  When you do, however, 
please include a simple, self contained example to illustrate 
briefly what you want, what you tried, and the deficiencies 
with what you tried, as suggested in the posting guide! 
"www.R-project.org/posting-guide.html".

	  Hope this helps.
	  Spencer Graves 	  	

Denis Fomchenko wrote:

Dear all,

I am trying to estimate simultaneous equation model 

concerning growth in russian regions.

I run the analysis by means of FIML in R sem package.
I am not familiar with SEM yet, but I've just got several 

suitable estimated specifications.

Nevertheless, sometimes R gives the following warning message:

Warning message:
Negative parameter variances.
Model is probably underidentified.
 in: sem.default(ram = ram, S = S, N = N, param.names = pars, 
var.names = vars,

I check for rank condition - all three equations in the 

system are turned out to be exact...

Does anybody know what it means? and how to handle with 

that problem?

P.S.
Do you know any examples of models estimated in SEM by 

means of FIML, incorporating spatial lag on endogenous variable?

Thanks, in advance

Denis Fomchenko
research fellow
Department for Economic Development Problems Institute for 

the Economy 

in Transition 5, Gazetny lane, Moscow 125993, Russia
e-mail: fomchenko at iet.ru
http://www.iet.ru
 

	[[alternative HTML version deleted]]