Using Stata to Replicate Table 4 in Bond (2002)
These notes refer to using Stata/SE 9.1, in March 2006.
Preliminaries
Open the dataset usbal89.
The main variables are:
id - firm identifier
year - year
y - log sales
n - log employment
k - log capital stock
Other variables have been derived from these. E.g.
y_1 - first lag of y
yk - (y - k), log of sales-capital ratio
Set panel format
tsset id year , yearly
Pooled OLS (OLS levels); Table 4, column (i)
xi: regress y n l.n k l.k l.y i.year , robust cluster(id)
Within Groups; Table 4, column (ii)
xi: xtreg y n l.n k l.k l.y i.year , fe robust cluster(id)
or
xi: areg y n n_1 k k_1 y_1 i.year , absorb(id) robust cluster(id)
Here the coefficients are identical to those in Table 4, column (i), since this is a balanced
panel.
First-Differenced GMM; Table 4, column (iii)
xi: xtabond2 y n l.n k l.k l.y i.year , gmm(y n k, lag(2 .)) iv(i.year) robust noleveleq
First-Differenced GMM; Table 4, column (iv)
xi: xtabond2 y n l.n k l.k l.y i.year , gmm(y n k, lag(3 .)) iv(i.year) robust noleveleq
This omits the levels of the variables dated t-2 from the set of instruments. The serial
correlation tests reported in Table 4, column(iv) are slightly different from those
produced by Stata.
System GMM; Table 4, column (v)
xi: xtabond2 y n l.n k l.k l.y i.year , gmm(y n k, lag(2 .)) iv(i.year, equation(level)) robust
h(1)
The noleveleq option is not specified. This uses a “system” combining equations in first-
differences with equations in levels.
The h(1) option uses 2SLS as the one-step estimator. This was also the case in the
Blundell-Bond (2000) production function estimates, which are reproduced in Table 4.
This is not the one-step weight matrix used in DPD98 for Gauss, or the DPD package in
PC Give and OX. The h(2) option in Stata uses the same one-step weight matrix as these
programs. Neither of these is the default option in Stata, which corresponds to the h(3)
option. See help xtabond2 for further details.
The gmm(y n k, lag(2 .)) option uses the lagged levels of y, n and k dated t-2 and earlier
as instruments for the equations in first-differences; and (correspondingly) the lagged
first-differences of y, n and k dated t-1 (only) as instruments for the equations in levels.
This is the default specification of gmm-style instruments for the levels equations. Other
options are available; see help xtabond2 for further details.
The iv(i.year, equation(level)) option uses the year dummies as instruments for the
equations in levels only. This treatment is specific to year dummies, and ensures that the
correct number of moment conditions of the form E[uit - ct] = 0 are used. For other
strictly exogenous variables used as iv-style instruments, the equation(level) restriction
would not normally be used.
The results are similar but not identical to those in Table 4, column (v).
System GMM; Table 4, column (vi)
xi: xtabond2 y n l.n k l.k l.y i.year , gmm(y n k, lag(3 .)) iv(i.year, equation(level)) robust
h(1)
The gmm(y n k, lag(3 .)) option uses the lagged levels of y, n and k dated t-3 and earlier
as instruments for the equations in first-differences; and (correspondingly) the lagged
first-differences of y, n and k dated t-2 (only) as instruments for the equations in levels.
The results are similar but not identical to those in Table 4, column (vi).
To obtain the Difference-Sargan test in column (vi):
g dsar = 75.80 - 53.66
g df = 55 - 40
g pval = chi2tail(df, dsar)
su pval
This gives a p-value of 0.104.
To test the “common factor” restrictions in column (vi):
testnl (_b[l.y]*_b[n] = -_b[l.n]) (_b[l.y]*_b[k] = -_b[l.k])
This gives a p-value of 0.796.
A useful feature of xtabond2 is that different assumptions can be made about the validity
of different instruments. For example, suppose we do not want to use lagged first-
differences of capital as instruments for the equations in levels, but we do want to use
lagged first-differences of sales and employment. This requires two separate uses of the
gmm(.) option:
xi: xtabond2 y n l.n k l.k l.y i.year, gmm(k, lag(3 .) equation(diff))
gmm(y n, lag(3 .)) iv(i.year, equation(level)) robust h(1)
The gmm(k, lag(3 .) equation(diff)) option specifies the use of lagged levels of k dated t-3
and earlier as instruments for the equations in first-differences; with no lagged
differences of k used as instruments for the equations in levels.
This is particularly useful when we expect first-differences of some but not all of the
variables to be uncorrelated with the individual-specific effects.
