The long-run parameter estimates that refer to the impact of ‘learning-by-exporting’
for the IV, control function and matching models are shown in Table 4. There are 3
sets of estimates that consider whether post-entry exporting improves productivity:
firstly, there are the terms that show whether firms new to exporting have the
expected pattern of significant, positive estimates in t and t + 1 (cf. the EXPentry
variables); second, we measure the TFP impacts for those firms leaving exporting
expecting that (if learning-by-exporting is prevalent) there should be significant,
negative effects in t and t+1 for firms that exit overseas markets (cf. the EXPexit
variables); lastly, we also allow for the effect on TFP of those that both enter and
leave export markets, with the expectation of significant, positive estimates in t and
t + 1 (cf. the EXPboth variables).
The results show that generally all three approaches to controlling for selectivity
effects produce broadly similar results. The sample selectivity terms (λ0 and λ1 ) are
generally insignificant, suggesting that the IV model has adequately controlled for
potential selectivity bias. The matching approach results in substantial reductions in
the number of observations available in those industries where exporters are in the
minority, and the loss of exporters without ‘common support’ in those sectors where
the majority of firms do export,31 but the parameter estimates obtained are generally
not too different to those obtained using the standard IV approach.
31 We use the ‘pstest’ procedure available in STATA 9 to inspect the extent of covariate balancing after
matching (see Leuven and Sianesi, 2003, for details of this test). In all cases the matched exporter and
non-exporter sub-groups have the same mean propensity scores, and there is always a 100% reduction
in ‘bias’ with respect to the values of propensity scores in the matched sample.
27