How to Do xtabond2

David Roodman
Research Fellow
Center for Global Development
 xtabond2 in a nutshell
 Firstado version in 11/03, Mata version in 11/05.
 Extends built-in xtabond, to do system GMM,
Windmeijer correction, revamped syntax
 Estimators designed for
– Small-T, large-N panels
– One dependent variable
– Dynamic
– Linear
– Regressors endogenous and predetermined
– Fixed individual effects
– Arbitrary autocorrelation and het. within panels
– General application
 Outline of paper
Introduction to linear GMM
Motivation and design of
difference and system GMM
xtabond2 syntax
 Black box problem
 Canned & sophisticated procedure
 Dangers in hidden sophistication
– finite sample ≠ asymptotic
 Users should understand motivation and
limits of estimator
 Linear GMM in one slide
 Instrument vector z such that E[z ] 0
 # instruments > # parameters so can’t have E N [z ]  N1 Z 'Eˆ 0
 Want to “minimize” N1 Z 'Eˆ in some sense
 In what sense? By a pos-semi-def. quad. form given by A:
'
1 ˆ  1 ˆ   1 'ˆ  1 ˆ'
E N z  A Z 'E N  Z 'E  A Z E   E ZAZ E
'ˆ

N A N  N  N
β
 Given A, minimizing leads to Aˆ  X '
ZAZ '
X 1
X '
ZAZ '
Y
 Always unbiased, but which A is efficient? Answer: A should weight
moments z i' E inversely with their variances and
covariances:
 '
A EGMM Var Z E X, Z  Z Var E X, ZZ Z ΩZ
1 ' 1 ' 1

 To make feasible, choose arbitrary proxy for Ω, call it H. Do GMM

(one-step). Use residuals to make robust sandwich estimator of
Z'ΩZ  1 . Rerun. Two-step is feasible, theoretically efficient.
 Linear GMM and 2SLS

 
βˆ EGMM  X ' Z Z ' ΩZ 
1
ZX '
 X ZZ ΩZ Z Y
1
' ' 1 '

 If Ω  I, reduces to
2

 ' ' 1 '


βˆ 2 SLS  X ZZ Z  Z X X ' ZZ ' Z  Z ' Y
1 1

 If errors i.i.d., efficient GMM is 2SLS

 If not, 2SLS inefficient
 Linear GMM in another slide
(Holtz-Eakin, Newey, and Rosen 1988)
(1) Y Xβ  E
OLS inconsistent: EX E 0
'

' ' '

(2) Take Z-moments: Z Y  Z Xβ  Z E
OLS consistent E X '
ZZ' E 0 
but inefficient Var Z' E Z' ΩZ not scalar 
Left-multiply by Z ΩZ  :
' 1
2

Z ΩZ
' 1
2
Z Y Z ΩZ 
' ' 1
2
Z Xβ  Z ΩZ 
' ' 1
2
Z 'E
1 1 1

Let X Z ΩZ  Z X, Y Z ΩZ  Z Y, E Z ΩZ  Z E

* '  ' ' * '  ' * ' 
2 2 2

* * *
(3) Y X β  E
OLS efficient Var E *
 I 
OLS on (3) = GLS on (2) = GMM on (1)
GMM = GLS on Z-moments
 Difference and system GMM
Basic model:
yit yi ,t  1  x it' β  it
 it i   it
Ei  E it  Ei it  0

Conceptual starting point: OLS

Problem: Dynamic Panel Bias (Nickell 1981)
 Fixed effects in disturbance term make y endogenous
i ,t  1

o Example: Indonesia
 A problem of short panels
 Individual dummies (=Within Groups) don’t help
o Transformed y endogenous, as are deeper lags
i ,t  1
 Partial solution: OLS in differences

yit yi ,t  1  x it' β   it

 Purges fixed effects, doesn’t spread endogeneity much
 Transformed yi ,t  1 still becomes endogenous since the
yi ,t  1 in y i ,t  1  y i ,t  1  y i ,t  2 correlates with the  i ,t  1 in
 it  it   i ,t  1
 But deeper lags exogenous if no AR(), offering
instruments
 Problem: Other endogeneity
 Differencing eliminates endogeneity to fixed
effects error component. But
o y now endogenous to 
i ,t  1 it

o Other predetermined variables become

endogenous in same way
o Still other variables may be endogenous
from the start
 For general application, assume no perfect
instruments waiting in the wings
Solution: Instrument with lags (2SLS)
(Anderson and Hsiao 1981)
 Assuming no AR() in  , natural
it

instruments for y are y and y

i ,t  1 i ,t  2 i ,t  2

 Both mathematically related to y i ,t  1

 y seems preferable: available at t = 3

i ,t  2

 Again, small T influences

 Do same for other endogenous variables
 Problem: Inefficiency
 Deeper lags available as instruments
o But reduce sample in 2SLS
o Problem for short panels
 In differences, errors not i.i.d.
o  and  mathematically correlated
it i ,t  1

o 2SLS not efficient

Solution: GMM & GMM-style instruments
(Holtz-Eakin, Newey, and Rosen 1988)
 Use many lags, replacing missing with zero
 Generate separate instrument for each lag and time
period instrumented
 0 0 0 0 0 0 
 .   0 0 0 0 0 0 
 .  
   y i1 0 0 0 0 0 
 y i1   .
   0 yi 2 y i1 0 0 0 
IV-style:    GMM-style:  0 0 0 yi3 yi 2 y i1 
 y i ,T  2   
          
 Result: Arellano-Bond (1991) difference GMM
 Problem: Autocorrelation
 E.g., if  it are AR(1), then yi ,t  2 ~  i ,t  2 ~  i ,t  1 ~  it
 Must assume yi ,t  2 is invalid instrument in i,t
 Solution: Restrict to deeper lags
 If we find AR(l) in  , use lags l + 1 and deeper
it
 Arellano-Bond AR() test
 Expect AR() in  it i   it
 To check for AR(1) in  it , test for AR(2) in eit
 E.g., compare eit  ei ,t  1 and ei ,t  2  ei ,t  3 to detect ei ,t  1 ~ ei ,t  2

 Test statistic for AR(l) in differences:

 e e
i ,t
it i ,t  l

 Normal under null of no AR(l)

 Arellano and Bond calculate its standard deviation
 z test for AR()
 More general than other AR() tests in Stata.
 abar: post-estimation command for regress, ivreg, ivreg2
 Problem: Weak instruments

If y is nearly a random walk, yi ,t  1 is a poor

instrument for yit , mathematical
relationship notwithstanding
Solution: Instead of purging fixed effects,
find instruments orthogonal to them
(Arellano and Bover 1995)
 If E y   stationary, then Eyit i  0
it i

 yi,t  1 uncorrelated with fixed effects, thus with it i 

good instrument in levels (if no AR)
 Make system of difference and levels equations
 Concretely, make a stacked data set, with difference
up top, levels below. Treat as single estimation prob
 Instrument differences with levels and v.v.
 “System GMM” (Blundell and Bond 1998)
 Relationship among moments
(Tue Gorgens)
Ewi1i1  D Ewi1i 2  D Ewi1i 3  D Ewi1i 4 
L L L
Ewi 2i1  Ewi 2i 2  D Ewi 2i 3  D Ewi 2i 4 
L L
Ewi 3i1  Ewi 3i 2  Ewi 3i 3  D Ewi 3i 4 
L
Ewi 4i1  Ewi 4i 2  Ewi 4i 3  Ewi 4i 4 
 Problem: Two-step errors too small
Regression for Arellano-Bond (1991) column (a1), Table 4
Arellano-Bond dynamic panel-data estimation, one-step difference GMM results
------------------------------------------------------------------------------

| Robust
| Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
n |
L1. | .6862261 .1445943 4.75 0.000 .4003376 .9721147
L2. | -.0853582 .0560155 -1.52 0.130 -.1961109 .0253944
w |
--. | -.6078208 .1782055 -3.41 0.001 -.9601647 -.2554769
L1. | .3926237 .1679931 2.34 0.021 .0604714 .7247759
k |
--. | .3568456 .0590203 6.05 0.000 .240152 .4735392

Arellano-Bond dynamic panel-data estimation, two-step difference GMM results

------------------------------------------------------------------------------
| Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
n |
L1. | .6287089 .0904543 6.95 0.000 .4498646 .8075531
L2. | -.0651882 .0265009 -2.46 0.015 -.1175852 -.0127912
w |
--. | -.5257597 .0537692 -9.78 0.000 -.6320709 -.4194485
L1. | .3112899 .0940116 3.31 0.001 .1254122 .4971675
k |
--. | .2783619 .0449083 6.20 0.000 .1895702 .3671537
 Problem, cont’d

 Problem appears to be one of overfitting

o Efficient GMM deemphasizes moments
with high variance (high second moments)
o Feasible efficient GMM in small samples
may deemphasize outliers (high first
moments)
o Spurious precision
 Solution: finite-sample correction
(Windmeijer 2005)
o One-step estimate: β̂1  f Y (conditioning on X, Z)
o One-step residuals used to construct Ω̂ :
   
1 1
ˆβ  X Z Z Ω
' ' ' ˆ Z  1 Z ' Y  g ( Y, Ω
ˆ Z Z X  X ' Z Z 'Ω ˆ )  g (Y, f (Y))
2
 
 
o Standard estimate of Var βˆ 2 treats Ω̂ as constant,
observed, precise—despite dependence on random Y
o Taylor expansion of g around true β :

βˆ 2  g Y, Ω  
ˆ   g Y, Ω
ˆ ˆ  g Y, Ω
1 
βˆ

ˆ ˆ

  βˆ  β 
1
βˆ β

o “Correction” comes from second term

 Eβˆ1  β  0 so no effect on Eβˆ 2  —no coefficient bias
 Affects variance
Arellano-Bond dynamic panel-data estimation, one-step difference GMM results
------------------------------------------------------------------------------
| Robust
| Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
n |
L1. | .6862261 .1445943 4.75 0.000 .4003376 .9721147
L2. | -.0853582 .0560155 -1.52 0.130 -.1961109 .0253944
w |
--. | -.6078208 .1782055 -3.41 0.001 -.9601647 -.2554769
L1. | .3926237 .1679931 2.34 0.021 .0604714 .7247759
k |
--. | .3568456 .0590203 6.05 0.000 .240152 .4735392

Arellano-Bond dynamic panel-data estimation, two-step difference GMM results

------------------------------------------------------------------------------
| Coef. Std. Err. t P>|t| [95% Conf. Interval ]
-------------+----------------------------------------------------------------
n |
L1. | .6287089 .0904543 6.95 0.000 .4498646 .8075531
L2. | -.0651882 .0265009 -2.46 0.015 -.1175852 -.0127912
w |
--. | -.5257597 .0537692 -9.78 0.000 -.6320709 -.4194485
L1. | .3112899 .0940116 3.31 0.001 .1254122 .4971675
k |
--. | .2783619 .0449083 6.20 0.000 .1895702 .3671537

Arellano-Bond dynamic panel-data estimation, two-step difference GMM results

------------------------------------------------------------------------------
| Corrected
| Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
n |
L1. | .6287089 .1934138 3.25 0.001 .2462954 1.011122
L2. | -.0651882 .0450501 -1.45 0.150 -.1542602 .0238838
w |
--. | -.5257597 .1546107 -3.40 0.001 -.8314524 -.2200669
L1. | .3112899 .2030006 1.53 0.127 -.0900784 .7126582
k |
--. | .2783619 .0728019 3.82 0.000 .1344196 .4223043
 Problem: too many instruments
 In difference and system GMM, # instruments (j) quadratic in T
 Analogy:
– In 2SLS, if j = # of regressors, first-stage R2’s=1.0 and
2SLS=OLS (biased)
– Too many instruments overfit endogenous variables
 '
1
 And # of cross-moments in Var Z E X, Z to be estimated for
efficient GMM quadratic in j—quartic in T!
 '
1
 Estimate of Var Z E X, Z degrades
 Hansen test very weak—p values of 1.000 not uncommon
 Little guidance on how many is too many
 xtabond2 warns if j > N
Solution: consider limiting instruments
 Limit number of lags of variables used as instruments
 Or “collapse” instruments:

 0 0 0 0 0 0   0 0 0 
 0 0 0 0 0 0   0 0 0 
 
 y i1 0 0 0 0 0   y i1 0 0 
. .


 
 0 yi 2 y i1 0 0 0   yi 2 y i1 0 
 0 0 0 yi3 yi 2 y i1   yi3 yi 2 y i1 
   
,

              

y
i
i ,t  2 eˆit 0 for each t 3
 i ,t
y i ,t  2 eit 0.
 xtabond2 syntax
Y X Z
xtabond2 depvar varlist [if exp] [in range]
[, level(#) twostep robust noconstant small noleveleq
artests(#) arlevels h(#) nomata]
ivopt [ivopt ...] gmmopt [gmmopt ...]]

where gmmopt is
“GMM-style”
gmmstyle(varlist [, laglimits(# #) collapse
equation({diff | level | both}) passthru])
and ivopt is
Classic

ivstyle(varlist [, equation({diff | level | both})

passthru mz])
 Examples
 Classic one-step difference GMM with no controls except
time dummies
xi: xtabond2 y L.y i.t, gmm(y, laglim(2 .))
iv(i.t) robust noleveleq

 Equivalents:
xi: xtabond2 y L.y i.t, gmm(L.y, laglim(1 .))
iv(i.t) robust noleveleq
xi: xtabond2 y L.y i.t, gmm(L.y)
iv(i.t) robust noleveleq

 System GMM, two-step, Windmeijer correction,

w1 exogenous, w2 predetermined, w3 exogenous:
xi: xtabond2 y L.y w1 w2 w3 i.t,
gmm(L.y w2 L.w3) iv(i.t w1) two robust
 Examples, cont’d
If conditions imposed only on levels,
difference equation effectively discarded.
Equivalent pairs:
regress n w k
xtabond2 n w k, iv(w k, eq(level)) small

ivreg2 n cap (w = k ys)

xtabond2 n w cap, iv(cap k ys, eq(level))

ivreg2 n cap (w = k ys), cluster(id) gmm

xtabond2 n w cap, iv(cap k ys, eq(level)) two

Or even:
regress n w k
abar, lags(2)
xtabond2 n w k, iv(w k, eq(level)) small arlevel
 Run times for bbest (seconds)
700 MHz PC
xtabond2 ado 57

xtabond2 Mata, favoring space 14

xtabond2 Mata, favoring speed 11

DPD for Ox 3