st: xtabond2.mata: a bug?

[Roberto <terrymondo@onetel.com>]

Dear all, I very recently passed to Stata9 and therefore I was very 
hopeful in the "speed improvements" of the new code. Unfortunately at 
my first attempt I have encountered an error. I read from the email 
of David Roodman that,  if we got a red message it would probably be a bug
this is what I receive:
xtabond2_mata():  3900  unable to allocate real <tmp>[45984,843]
                 <istmt>:     -  function returned error

I am trying to estimate the following equation following the notation 
suggested to me by the author of the routine:
 xtabond2 lev l.lev mtbv size1 mat1 collateral profitability ndts 
year1-year24, gmm(lev mtbv size1 mat1 collateral profitability ndts, 
lag (2 7) equation(diff))  gmm(lev mtbv size1 mat1 collateral 
profitability ndts, lag(2 2)) iv(year1-year24, equation(level)) 
robust  noc small twostep h(2)

this is a brief summary of the dataset I am using:

       year |      Freq.     Percent        Cum.
------------+-----------------------------------
       1980 |          3        0.03        0.03
       1981 |          6        0.06        0.10
       1982 |          6        0.06        0.16
       1983 |          6        0.06        0.22
       1984 |          9        0.10        0.32
       1985 |         16        0.17        0.49
       1986 |         43        0.46        0.95
       1987 |         75        0.80        1.74
       1988 |        141        1.50        3.24
       1989 |        196        2.08        5.32
       1990 |        234        2.49        7.81
       1991 |        508        5.40       13.20
       1992 |        742        7.88       21.08
       1993 |        802        8.52       29.60
       1994 |        831        8.83       38.42
       1995 |        845        8.97       47.40
       1996 |        850        9.03       56.43
       1997 |        828        8.79       65.22
       1998 |        793        8.42       73.64
       1999 |        693        7.36       81.00
       2000 |        614        6.52       87.52
       2001 |        569        6.04       93.56
       2002 |        524        5.56       99.13
       2003 |         82        0.87      100.00
------------+-----------------------------------
      Total |      9,416      100.00

Last, I have a Pentium 4 3.2Giga with 1 Giga RAM memory.
My Stata9 is fully updated and I am using the most recent version of 
xtabond2: 2.0.0 16Nov2005

I have searched in past Statalist for possible solutions to this 
issue but I had no luck.

Is this a "bug"?. Is there a solution to this issue?

Thank you in advance for your kind attention.
Best Regards
Roberto




--David


David Roodman
Center for Global Development
Research Fellow
1776 Massachusetts Ave. NW
Suite 301
Washington, DC 20036
droodman@cgdev.org
202-416-0723
fax: 413-410-2602

-----Original Message-----
From: Roberto Mura [mailto:fellowsh@tiscali.it]
Sent: Monday, February 14, 2005 8:14 AM
To: David Roodman
Subject: RE: xtabond2

Dear David
first of all allow me to thank you for your kind reply.
I do in fact have the latest version but I still do not seem able to manage.

I have tried this command:

xtabond2 Y l.Y  X1 X2 X3 year1-year11, gmm(Y, lag(3 3)) gmm(X1, lag(2 
3)) gmm(X2 X3, lag(3 3)) equation(diff) gmm(Y, lag(2 2)) gmm(X1, 
lag(1 1)) gmm( X2 X3, lag(2 2)) iv(year1-year11) 
equation(level)  robust small noc twostep

but stata says
year1 dropped because of collinearity.
Building GMM instruments.........
equation(difference) equation(level) invalid.
r(198);

the only way I found to perform a GMM-SYS is

xtabond2 Y l.Y  X1 X2 X3 year1-year11, gmm(Y, lag(3 3)) gmm(X1, lag(2 
3)) gmm(X2 X3, lag(3 3)) iv(year1-year11, eq(level)) robust small noc twostep

Here I do not specify the lag structure of the instruments for the 
equation in levels....see the problem is that, for some reason I 
don't quite get, Stata uses 1 instrument more than DPD in making the system.
I have tried to recreate the matrix instrument in excel just to be 
sure and what I get is the same as DPD provides (84 instruments for 
the GMM-SYS specified above) and so 70 dof in the Sargan test (Y X1 
X2 X3 + 10 year dummies). Stata uses 81 instruments and so 71 dof in 
the hansen stat. More interesting is the fact that, stata uses 5666 
observations while DPD uses 5667.
The dataset has been thoroughly cleaned so there are no missing 
observations anywhere.
Here are the info from Stata

Arellano-Bond dynamic panel-data estimation, two-step system GMM results
------------------------------------------------------------------------------
Group variable: code                            Number of obs      =      5666
Time variable : year                            Number of groups   =       673
Number of instruments = 85                      Obs per group: min =         4
F(13, 672)    =   1943.77                                      avg =      8.42
Prob > F      =     0.000                                      max =        10

Furthermore, now with the new version of the command I cannot seem to 
manage to get exactly the same results even for the GMM-DIFF....

I will stop here because I think I have taken too much of your time already.
I wish you a very nice day and thank you so much for your kind attention.
R

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/