[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: Xtabond2 and large samples

From   "Bromiley, Philip" <[email protected]>
To   <[email protected]>
Subject   st: Xtabond2 and large samples
Date   Sun, 16 Sep 2007 10:18:15 -0700

I'm running xtabond2 with a large sample (55000 observations).  I'm
running things like:

global v2 Aeqdebt  R11_eqdebt R12_eqdebt R13_eqdebt R14_eqdebt
Ashortlong R11_shortlong  R12_shortlong R13_shortlong R14_shortlong
Ainv_sales R11_inv_sales R12_inv_sales R13_inv_sales R14_inv_sales
Acapexp_sales R11_capexp_sales R12_capexp_sales R13_capexp_sales
R14_capexp_sales  Arec_sales     R11_rec_sales R12_rec_sales
R13_rec_sales R14_rec_sales 

xi  i.year

xtabond2    L(0/1).Aroa L(1/4).($v2)   L(0/3).Aindroa  q2 q3  q4
R11_indroa R12_indroa R13_indroa R14_indroa _Iyear_2000-_Iyear_2004 ,
gmm (L.Aroa , lag( 2 7))   ///
gmm ( $v2, lag(2 7)   )   ///
iv( L4.(Aindroa q2 q3  q4 R11_indroa R12_indroa R13_indroa R14_indroa
		indcapsales indrecsales indinvsales indcashsales
indeqdebt  ))  ///
iv( _Iyear_2000-_Iyear_2004)   ///
noconstant   robust twostep  orth 

I end up with 800 to 1500 instruments.  Until I lag the gmm's and iv's
so far back that I get extremely high variance parameter estimates, I
tend to get statistically significant Hansen statistics.  The serial
correlation tests generally are fine - statistically significant in
first and insignificant in second lag.  Some of these statistically
significant Hansen statistics are on variables that are clearly
exogenous (including the year dummies).

The predicted errors from the estimation are not normal.  

I'm wondering:

(i) Hansen tests for normality of a function.  While I don't see that
Hansen (1982) explicitly assumes normality of the observation-specific
errors, in practical terms, could non-normality of the
observation-specific errors hurt the test?

(ii) Would the large sample size result in rejections of normality
(i.e., significant Hansen) even if the deviations are modest?  This is a
standard problem with such tests where a large enough sample size will
let you reject the null hypothesis for modest deviations from the
hypothesized distribution.
(iii) Am I screwing up the specification in some obvious way?  I've
tried numerous ways to specify the model but the problem remains.


Philip Bromiley
Dean's Professor of Strategic Management 
Merage School of Business University of California, Irvine 
Irvine, CA 92697-3125
(949) 824-6657
(949) 725-2898 (fax)

*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index