themselves when introduced as right-hand-side variables in (5) although they are
generally important as prior determinants of the likelihood of exporting. Our second
approach is to include the sample selectivity correction terms λ0 and λ1 in Equation
(5) so as to control directly for the correlation between the error terms in Equations (3)
and (5). This is labelled the ‘control function’ model in our results. Lastly, we use the
propensity score matching procedure to obtain a matched sample of exporters and
non-exporters based on Equation (4), and this matched sample is used when
estimating Equation (5).
The full set of results from estimating Equation (5) are reported in Table 3, based on
long-run estimates (but including the lagged coefficient on the dependent variable in
order to assess how long it takes to converge on the long-run equilibrium reported)
and diagnostic tests for each industry sub-group. In most cases, the models estimated
are satisfactory in terms of autocorrelation (cf. the AR(1) and AR(2) test statistics)
and the adequacy of the instrument set used (cf. the Hansen test results).30 Here we
concentrate on the variables linked to ‘learning-by-exporting’, but it is interesting to
note that our results show that increasing returns-to-scale generally were present for
all sub-groups (across the 16 industries examined, the average sum of the output
elasticities was 1.14 for those firms that had always exported, followed by a value of
1.13 for those moving into exporting; the average RTS for firms never exporting was
the lowest at 1.02).
30 Similar results (in terms of diagnostic statistics and often parameter estimates) are obtained when the
‘control function’ model and the matching model are estimated. Detailed results for these two models
are not reported in this paper but they are available upon request.
22