the propensity score matching approach where we first estimate a model to identify
the probability of exporting (i.e. the propensity score) using the following (random
effects panel) probit model:
P(Export it = 1)=φ(lnLPit-1,ln Ageit-1, Intangit-1, Sizeit-1, Industryit,Regionit) (3)
where Export is coded 1 if the firm exported at any time during 1996-2004; LP is
labour productivity; Age is the age of the firm; Intang is coded 1 if the firm has non-
zero intangible assets24 ; Size represents a set of dummy variables that indicate
whether the firm belongs to one of the following 4 size bands: 10-19, 20-49, 50-199
or 200+ employees; and Industry and Region are dummy variables indicating each
industry sub-group or Government Office region. Following Girma et. al. (2004), if
Pi is the propensity score of exporting for firm i at time t, we then use the propensity
score matching procedure available in STATA 9 to find the closest match (using the
“nearest-neighbour” approach) for each exporting firm in terms of the propensity
scores from the sub-group of non-exporting firms, i.e.:
∣P∙ - Pj = min {Pi - Pj} (4)
i j k∉{Export k =0} i j
Having obtained a matched sample, we estimate a multivariate model using the
matched data to test the learning-by-exporting hypothesis. This combination of
matching and parametric estimation is argued to improve the results obtained from
non-experimental evaluation study (e.g. Blundell and Costa Dias, 2000), as other
impacts on the outcome variable are explicitly controlled for.
average treatment effect). Heckman and Navarro-Lozano (2004) argue that this is an unattractive
implication.
24 Here these non-monetary assets usually refer to corporate intellectual property (e.g. patents,
copyrights, trademarks, etc.), innovative activities, advertising, goodwill, brand recognition and similar
intangible assets. There is sufficient ambiguity of exactly what should be included as intangible assets
(and issues over how to measure such assets - see, for example, Webster and Jensen, 2006)) that we
have chosen to use a dummy variable rather than the actual monetary amount reported in FAME.
17