and sample selection.21 Briefly, in estimating the exporting-productivity relationship,
selection bias occurs because participants in export markets may posses certain
characteristics such that they would achieve better performance vis-à-vis non-
participants even if they did not enter export markets, and this productivity gain is
correlated with the decision to participate in the global market. This will mean that
standard estimation techniques lead to biased results. Thus the essential problem at the
core of evaluating the effect of exporting is to obtain an estimate of the unobserved
counterfactual that is not biased because of any simultaneous relationship between the
decision to export and the gains from exporting.
There are several approaches that attempt to eliminate the bias that arises from self-
selection (cf. Blundell et. al., 2005). The first considered here is matching. Essentially,
this involves matching every exporting firm with another firm that has (very) similar
characteristics but does not export. Essentially, under the matching assumption
exporters and non-exporters have the same (observable) attributes that impact on
productivity (and the probability of exporting). Thus the non-exporting, matched sub-
group constitutes the correct counterfactual for the missing information on the
outcomes that exporters would have experienced, on average, if they had not exported.
There are a number of issues associated with the matching process, including the need
for a rich dataset that includes all relevant variables that impact on productivity and
all variables that impact on whether the firm exports or not. Matching is done on the
set of selection criteria, so that any selection on unobservables is assumed to be trivial
and does not affect outcomes in the absence of exporting.22, 23 Here, we have adopted
21 Standard evaluation problems are discussed in Heckman (2000), Moffitt (2004), and Heckman and
Navarro-Lozano (2004).
22 Typically firms that export which are not ‘supported’ by firms from the non-export population are
dropped, which can reduce significantly the size of the export sub-group included in any analysis. So
where there is little common support between the treated and non-treated comparators, matching breaks
down.
23 Another issue is that by definition, matching assumes that the effect for the average export firm is the
same as the effect for the marginal firm (the ‘treatment on the treated’ effect equals the unconditional
16