export shares separately, and also variables indicating the export share of firms that do
not import their inputs and import share of firms that do not exports (column 5).
One other important and significant control is the share of foreign ownership that, as
expected, is positively correlated to the firm’s performance.
Yet, identifying the relationship between productivity and trade practices though the
variation across plants can introduce a bias. In fact the foreign network index could be
correlated with omitted plant characteristics that affect productivity. Under the
hypothesis that these characteristics are time invariant, it is possible to control for
unobserved firm heterogeneity with fixed effect estimates. This estimator identifies the
impact of the variable of interest relying on the within-firm time variation. Such
estimates are reported in column (6) and (7) where is shown how the coefficient on the
IE index remains positive and statistically significant.
To further test the robustness of our findings in column (8) we also introduce among the
regressors, the lagged value of TFP index assuming that firm’s productivity follows a
Markov process. This inclusion introduces however a bias that we correct trough the
Arellano-Bond dynamic panel estimator in column (9) (Arellano and Bond, 1991). The
coefficient of interest maintains both significance and sign even when import and export
shares are introduced as controls (column (10)). This latter estimator has also the
advantage of permitting to address more general endogeneity issues. In fact we introduce
in the GMM instruments matrix also the lagged values36 of the IE index to overcome the
endogeneity between the level of productivity and the value of the index. However the
use of lagged values of the variables to control for endogeneity leads to a significant
decline in the number of observations which does not permit to draw very definite
conclusions from the analysis. The same happens when using traditional instrumental
variables estimators such as the one reported in columns (11) and (12). The first case
corresponds to the two stages least squares estimator with first and second lag of the IE
index used as instruments. Column (12) instead displays the two-step instrumental
variables GMM estimates37 obtained with the same instruments. Nonetheless, in both
cases the IE index shows a positive and statistically significant coefficient and the tests
on the validity of the instrument confirm that they are uncorrelated with the error term38.
36 Starting from t-2
37 The efficiency gains of this estimator relative to the traditional instrumental variable two step estimator
derive from the use of the optimal weighting matrix that generates efficient estimates of the coefficients as
well as consistent estimates of the standard errors in presence of heteroskedasticity.
38 therefore first and second lags are valid instruments for the IE index.
20