Using data from the Community Innovation Survey for Belgium in two consecutive periods, this
paper explores the relationship between firm-level innovation activities and the propensity to start
Product and process innovation and the decision to export:
firm-level evidence for Belgium
Ilke Van Beveren and Hylke Vandenbussche
D I S C U S S I O N P A P E R
Center for Operations Research
Voie du Roman Pays, 34
CORE DISCUSSION PAPER
Product and process innovation and the decision to export:
firm-level evidence for Belgium
Ilke VAN BEVEREN and Hylke VANDENBUSSCHE
Using data from the Community Innovation Survey for Belgium in two consecutive periods, this
paper explores the relationship between firm-level innovation activities and the propensity to start
exporting. To measure innovation, we include indicators of both innovative effort (R&D
activities) as well as innovative output (product and process innovation). Our results suggest that
the combination of product and process innovation, rather than either of the two in isolation,
increases a firm?s probability to enter the export market. After controlling for potential
endogeneity of the innovation activities, only firms with a sufficiently high probability to start
exporting engage in product and process innovation prior to their entry on the export market,
pointing to the importance of self-selection into innovation.
Keywords: exports, product innovation, process innovation, self-selection, firm heterogeneity.
JEL Classification: D24, F14, L25, O31, O33
Katholieke Universiteit Leuven, LICOS and Lessuis, Antwerpen, Belgium.
Université catholique de Louvain, CORE, B-1348 Louvain-la-Neuve, Belgium and KULeuven, LICOS, B-
3000 Leuven, Belgium. E-mail: firstname.lastname@example.org. This author is also member of ECORE,
the association between CORE and ECARES.
We are thankful to Bee Aw, Mark Roberts, Bruno Cassiman, Stijn Vanormelingen, Davide Castellani,
Maurizio Zanardi and Marton Csillag for helpful discussions and suggestions. Participants of the International
Economics Group Seminar Series at UCL, the International Trade Workshop in Brussels and the European
Trade Study Group conference in Rome provided useful comments. We would also like to thank Manu
Monard, Peter Teirlinck and the CFS-STAT Commission for allowing us to access the CIS data for Belgium,
for answering questions related to the data and for their hospitality during visits there. While we are thankful
to all of them, none of them are responsible for any errors we have made.
This paper presents research results of the Belgian Program on Interuniversity Poles of Attraction initiated by
the Belgian State, Prime Minister's Office, Science Policy Programming. The scientific responsibility is
assumed by the authors.
There is a large and growing body of literature dealing with the link between firms? decision
to export and their productivity. The seminal works of Bernard and Jensen (1999) and Melitz
(2003) have shown that only the more productive firms self-select into exporting, since only
firms with an efficiency level above a certain threshold, are able to overcome the fixed costs
associated with entry on the export market. This literature typically assumes that firms?
productivity is a random, exogenous draw from a Pareto distribution.
More recent contributions to the literature (e.g. Bustos 2005; Yeaple 2005) have sought to
endogenize firm-level productivity, hence allowing for the possibility that firms can influence
their own efficiency level, rather than simply observing it in each consecutive period. One of
the ways in which firms can increase their productivity, is through innovation activities. In the
theoretical framework of Yeaple (2005), firms have the possibility to adopt either a high-
technology, low unit cost or low-technology, high unit cost production process. The low unit
cost technology entails a higher fixed cost of technology adoption. In the presence of fixed
costs to enter the export market, only those firms that adopt the low unit cost technology will
be able to start exporting.
In response to these developments, several authors have explored the relationship between
firms? innovation activities and their propensity to engage in exports. However, thus far, the
empirical results on the link between innovation activities and the firm?s export decision have
been mixed. Moreover, results seem to depend on which innovation measures are used.
Specifically, both Aw, Roberts and Winston (2007), as well as Cassiman and Martinez-Ros
(2007) fail to find a significant link between firm-level R&D (innovative effort) and the
probability of firms to start exporting, using firm-level data on manufacturing firms in Taiwan
and Spain respectively.
When innovation output measures are considered, the link between innovation and firms?
propensity to export appears to yield stronger, but mixed results. While Caldera (2009) and
Cassiman and Martinez-Ros (2007) both use the Spanish ESEE data set, their analysis yields
different findings. In particular, Cassiman and Martinez-Ros (2007) identify product
innovation, but not process innovation, as a driver of firm-level export propensity; while
Caldera (2009) finds both product and process innovation to matter, although the impact of
product innovation is higher than that of process innovation. Damijan, Kostevc and Polanec
(2008) on the other hand, using data on the Slovenian manufacturing sector and applying
matching techniques to account for the endogeneity of the innovation activities, find no
evidence that product or process innovation acts as a significant driver of export propensity at
the firm level. They do provide evidence that firms engage significantly more in process
innovation after entering the export market. Finally, Becker and Egger (2007), who apply
matching techniques to German survey data, find that firms introducing a product and process
innovation simultaneously increase their propensity to export by about ten percentage points.
Product innovation acts as a significant driver of firms? export propensity when introduced in
isolation, but process innovation does not.
Recent theoretical models incorporating productivity improving investment in a Melitz (2003)
type of model like Bustos (2008) and Bas and Ledezma (2008), show that productivity and
investment are strategic complements i.e. more productive firms invest more. While these
models confirm the complementarity between innovation and exporting, due to their static
nature they do not solve the ambiguity surrounding the causality direction. The exception is
Constantini & Melitz (2007) that explore the relationship in a dynamic setting and find that
only firms that anticipate export market entry engage in prior productivity improving
investment. The empirical results in this paper are in line with this theoretical finding.
We explore this link between innovation and firms? export propensity further using data from
a detailed innovation survey for Belgium . Belgium is a small and open economy with more
than 80% of its GDP resulting from exporting. In view of the importance of exporting for the
economy as a whole, it therefore offers a good opportunity to study the link between
innovation and exporting. The Belgian innovation survey that we use covers two consecutive
periods 1998-2000 and 2002-2004. The relatively short nature of the panel is compensated by
detailed information on firms? innovation characteristics, reporting data on firms? innovative
effort (R&D investment) as well as their innovative output (product and process innovation).
The survey also records other firm-level variables, such as export status and export intensity.
We combine the innovation data with firm-level company accounts data.
Recent efforts to incorporate investment in a Melitz & Ottaviano (2008) type of model even get the opposite
result i.e. that lowly productive firms invest more, which further enhances the ambiguity surrounding the relation
between innovation and exporting (Mayneris, 2009).
The Belgian innovation survey data are collected by a national agency (BELSPO) and are part of the
Community Innovation Survey (CIS) database. We use CIS 3 (1998-2000) and CIS 4 (2002-2004). The
innovation survey cover a period of two years. Firms are asked to report innovation activities between the
beginning and end of this period. All financial and accounting information, such as the value of sales, export
intensity and total amount spent on R&D pertains only to the last year in the data (2000 and 2004).
A growing number of recent papers have used data on innovative output (rather than only on R&D inputs) to
analyze the innovation decisions of firms. Examples include Castellani and Zanfei (2006, 2007), Damijan et al.
(2008), Griffith, Mairesse and Peters (2006) and Mairesse (2004).
In contrast to previous papers we correct for three different types of endogeneity issues that
can arise when analyzing the link between firms? innovation activities and their propensity to
export. First, since firms typically make their innovation and export decisions simultaneously,
a simultaneity bias emerges. Second, since exporting activities tend to exhibit persistence over
time (Aw et al. 2007), a causality bias arises when past exporting history is not properly
controlled for. Finally, to the extent that firms can anticipate entry into the export market and
their innovative efforts are driven by this future prospect, the introduction of new innovations
is endogenous to firms? export decision (Costantini and Melitz 2007). This anticipation effect
is the third source of endogeneity when analyzing the link between firm-level innovation and
exporting activities .
To account for the simultaneity of firm-level innovation and exporting decisions and to rule
out past exporting history, we limit the sample used in the empirical analysis in two ways.
First, we focus on firms that have answered the innovation survey in two consecutive periods.
This will allow us to use lagged (initial) innovation and other firm-level characteristics as
potential determinants of firms? propensity to export. This limits the sample to 600 firms.
Second, to control for the causality bias and the persistence of firm-level exports (see for
instance Aw et al. 2007), we focus our attention on firms that started exporting in 2004
(Starters) and compare these to a control group of firms that did not export in either period
Similar to previous research, we fail to find a link between firm-level internal (or external)
R&D and the exporting decision. Moreover, when we add product and process innovations
simultaneously as determinants of the firms? exporting decision, our findings are similar to
those obtained by Cassiman and Martinez-Ros (2007), i.e. product innovation, but not process
innovation, acts as a significant driver of firms? entry on the export market. However,
inspection of the data reveals that more than fifty percent of innovating firms in our sample
introduce a product and process innovation simultaneously, rather than one of the two in
Looking more closely at our data we find that a substantial share i.e. 48% of firms in our
sample that introduced a product innovation between 1998 and 2000 also introduced a process
innovation. Similarly, 58 % of all firms that have introduced a process innovation during the
Costantini and Melitz (2007) analyze the joint entry, exit, export and innovation decisions of firms confronted
with trade liberalization in a dynamic setting. They find that the anticipation of upcoming liberalization can
induce firms to innovate prior to their entry on the export market.
An alternative interpretation of the anticipation effect is that it is similar to a simultaneity effect with a lag.
same period, simultaneously introduced a product innovation. When we account for the high
correlation between firms that engage in both product and process innovation, our results
suggest that it is not so much product or process innovation in isolation, but rather the
combination of the two, which drives firms into the export market. Process innovation is
usually associated with an improvement in cost efficiency, while product innovation is
associated more with an improvement in the quality of a product. When we apply this
interpretation, we can infer that the results that we find suggest that firms engage in a
combination of both cost reduction and quality improvement in the run up to their entering
Finally, to account for the anticipation effect we use an instrumental variable approach. As
instruments for product and process innovation we use firm-level innovation inputs (internal
and external R&D) which are highly correlated with output innovation measures but appear to
have no direct impact on the exporting decision as discussed above. Another instrument we
use for the innovation decision is the level of training activities at the firm-level which
appears highly correlated with product and process innovation but not correlated with
exporting. After controlling for endogeneity of the innovation activities, we find no evidence
that firms engaging in product and/or process innovation are more likely to enter the export
market. These results suggest that only firms with a sufficiently high probability to start
exporting will engage in product and process innovation prior to their entry on the export
market, pointing to the importance of self-selection into innovation activities of firms
anticipating entry in the export market in the future. One interpretation for this result is that
firms anticipating exports expect tougher competition abroad and as a result engage in cost
reducing investment (process innovation) and quality upgrading (product innovation) prior to
exporting. Another explanation is that when a firm starts exporting it increases its market size
which makes the investment in cost reduction and product improvement more worthwhile i.e.
it raises the return to investment as theoretically shown by Lileeva and Trefler (2007). For
Belgian firms this is particularly relevant since they have a relatively small home market.
Hence exporting offers an important potential for increasing the market they serve.
Theoretically, our results are in line with the model by Constantini and Melitz (2007)
stressing the importance of an anticipation effect in explaining the evolution of firm
productivity. Recent industry studies also seem to confirm firms? quality upgrading in
anticipation of entering the export market (Iacovone & Smarzynska Javorcik 2008).
Our analysis contributes to the existing literature in a number of important ways. First, unlike
most of the existing empirical work , we take the correlation between product and process
innovation explicitly into account in the empirical analysis, hence allowing for potential
complementarities between the two innovation types. Furthermore, our results point to the
importance of accounting for all potential sources of endogeneity of firm-level innovation in
the exporting decision. After accounting for the three types of endogeneity outlined above i.e.
simultaneity bias, causality bias and anticipation effect, we find that only firms that anticipate
export market entry engage in innovation activities.
The policy implications arising from our empirical analysis appear important. A more general
interpretation of our results is that trade liberalization induces innovation activities. Previous
literature has shown that when governments engage in trade liberalization, trade costs are
reduced which leads to a larger number of exporters. The additional insight arising from this
paper is that firms prepare themselves for this entry into export market by simultaneously
engaging in cost reduction and quality improvement.
The rest of the paper is structured as follows. Section 2 reviews the relevant literature, while
section 3 discusses the data and reveals interesting empirical facts. Section 4 introduces the
empirical model and sections 5 and 6 presents the empirical results. The final section
concludes and formulates relevant policy recommendations.
2. LITERATURE REVIEW
While the literature on the relationship between firms? participation on export markets and
their productivity abounds , until recently, it remained largely silent on the sources of the
productivity advantages associated with firms? entry on export markets. Following theoretical
models of entry and exit (e.g. Jovanovic 1982; Hopenhayn 1992; Melitz 2003), researchers
have long continued to assume that the productivity advantage that enabled firms to start
exporting (or start producing) was exogenous in nature, hence not determined by any firm-
From this early literature dealing with the relationship between exports and productivity, a
dual relationship emerges, whereby firms exogenously self-select into the export market (i.e.
Becker and Egger (2007) are a notable exception. However their analysis includes both starters on the export
market and continuing exporters, i.e. they do not control for past exporting history in their analysis.
For reviews on this extensive literature, we refer to Greenaway and Kneller (2007) and Wagner (2007).
their productivity is higher than the minimum efficiency level required to enter export
markets) and, once they start exporting, have the potential to further increase their
productivity through learning effects.
Recently, however, efforts have been made to endogenize firm heterogeneity, allowing firms
to engage in productivity-enhancing activities prior to engaging in international markets.
Important theoretical contributions in this field include Bustos (2005) and Yeaple (2005).
Unlike earlier models of firm dynamics (e.g. Jovanovic 1982), in Yeaple?s model firms are
born identical. After being born, they have the possibility to adopt a high-technology, low unit
cost production technology, or a low-technology high unit cost technology. In the presence of
fixed costs associated with both technology adoption and exporting, the model shows that
only those firms adopting the low unit cost technology are able to start exporting. In related
work, Costantini and Melitz (2007) analyze the joint entry, exit export and innovation
decisions of firms in response to or in anticipation of trade liberalization. Their findings point
to the importance of taking the timing and speed of trade liberalization into account when
analyzing firms? export and innovation decisions. In particular, they find that anticipation of
upcoming trade liberalization and a slow liberalization process can motivate firms to innovate
ahead of export market entry.
From these different strands of the literature, three different hypotheses concerning the link
between exporting and productivity emerge (Alvarez and Lopez 2005). Apart from exogenous
self-selection and learning effects, firms have the possibility to engage in investments aimed
specifically at raising their productivity prior to entry on export markets (Conscious self-
selection). Existing empirical evidence has thusfar been interpreted in favor of the exogenous
self-selection hypothesis although the evidence is observationally equivalent to conscious
self-selection. For learning effects the results tend to be mixed.
Recently, several empirical papers aim to provide more direct evidence on the conscious self-
selection hypothesis, investigating the link between firms? export propensity and a number of
firm-level investments or decisions: training and R&D (Aw et al. 2007), product and process
innovation (Damijan et al. 2008) and physical investment (Alvarez and Lopez 2005; Iacovone
and Smarzynska Javorcik 2008). A common feature all these papers share is that they
investigate to what extent certain (investment) activities of firms increase their propensity to
engage in exports. Furthermore, all of the studies cited provide evidence on the
complementary nature of these investment activities and firms? export propensity.
When investigating the link between firm-level innovation activities and its propensity to
(start) export(ing), two types of innovation measures have been used in the literature.
Specifically, either innovation input measures, usually expressed as the ratio of R&D over
sales or as a dummy variable indicating whether firms engage in R&D, or innovation output
measures, typically expressed as dummy variables representing whether firms have
introduced a product or process innovation; are used as measures of firm-level innovation
activities. As was already noted in the introduction, the impact of firms? innovation activities
on their export propensity are mixed and seem to depend on the type of measures used.
Aw, Roberts and Winston (2007) explore the link between firm-level R&D, training,
productivity and exports using data on the Taiwanese electronics sector. Their findings
suggest that R&D and exporting are not complementary activities, but they have a
complementary effect on firm-level productivity. These results seem to imply that the
combination of exporting and R&D increases productivity more than the sum of both
conducted in isolation. Cassiman and Martinez-Ros (2007) find similar results for the Spanish
manufacturing sector, i.e. firms engaging in R&D investment do not exhibit a significantly
higher export propensity.
While research spending of firms can be considered a reasonable proxy of firm-level
innovative output in the absence of information on the actual innovations firms have
introduced, there are several drawbacks associated with the use of R&D spending, which is
essentially an input in the innovation production function , as a measure of firm-level
innovation. First, not all innovation efforts actually lead to the introduction of product or
process innovations, i.e. it is possible that firms? efforts to innovate fail for some reason, in
which case using R&D rather than actual innovations leads to an overestimation of firms?
innovative activities. Second, it is not unlikely that there is a considerable time lag between
firms? investment in R&D and the actual introduction of an innovation to the market, in which
case the timing of the R&D and innovation decisions do not match, leading to an
overestimation of innovation in some years and an underestimation in later years, when the
level of R&D spending is lower and innovative output is higher.
Several authors have taken these drawbacks into account and rely on measures of firm-level
innovation output rather than inputs to investigate the link between firm-level innovation and
export propensity. Becker and Egger (2007) use German survey data, Caldera (2009) and
Cassiman and Martinez-Ros (2007) use data for Spain and Damijan et al. (2008) use survey
See for instance Mairesse and Mohnen (2002).
data for Slovenia to explore the relationship between firm-level innovative output, measured
as the introduction of product and process innovations and firms? propensity to (start)
export(ing). While these papers share a common purpose and in the case of Caldera (2009)
and Cassiman and Martinez-Ros (2009) also use the same data set, some differences between
them, both in terms of sample selection, methodology and empirical results are worth noting
As shown by Aw et al. (2007) firms? exporting status is characterized by a high persistence, a
finding that is also consistent with the prediction of Melitz?s model that entry into the export
market leads to the incurrence of a fixed cost, which cannot be recovered. Given these
preliminaries, it is not unlikely that firms? initial entry versus its continued presence on the
export market have different determinants. Iacovone and Smarzynska Javorcik (2008)
document, for a sample of Mexican manufacturing firms, an increase in physical investment
prior to the introduction of a domestic variety on the export market, but only for new
exporters. For firms with prior export experience, no such increase was recorded. These
findings point to the importance of taking firms? prior export experience into account in the
For this reason, Damijan et al. (2008) focus only on first-time exporters when investigating
the impact of firm-level innovation activities on firms? propensity to export. While Caldera
(2009) and Cassiman and Martinez-Ros (2007) both use the full sample of exporters (starters
and firms with export experience) in their analysis, they both perform a number of robustness
checks to account for prior experience in exporting. Specifically, Caldera (2009) estimates a
dynamic model as a robustness check and Cassiman and Martinez-Ros (2007) repeat their
analysis using only starters on the export market versus a control group of non-exporters. In
both cases, the main findings are robust to these alternative specifications. Becker and Egger
(2007) on the other hand, focus on the full sample of firms and do not differentiate between
first-time exporters and continuing exporters. In the empirical analysis below, we follow
Damijan et al. (2008) by focusing only on starters on the export market and a control group of
In terms of the methodologies used, the four papers cited above can be divided in two groups.
Caldera (2009) and Cassiman and Martinez-Ros (2007) both use a probit model to investigate
the relationship between firm-level innovation and export status. To control for unobserved
firm heterogeneity, they add random effects to the baseline specification. Apart from a
number of control variables, both papers add lagged innovation status for product and process
innovation as independent variables. This allows them to control for the simultaneity of the
export and innovation decisions. However, while selection on prior export status (i.e. using
only starters on the export market) and the use of lagged firm characteristics avoids the
pitfalls of persistence in exports and of a simultaneity bias resulting from the timing of the
innovation and export decisions, this does not rule out the existence of feedback effects,
rendering firm-level innovation endogenous in the export decision framework.
Specifically, if firms have some prior knowledge of their prospects on the export market, they
are likely to make their innovation decisions with this prospect in mind. In other words, to the
extent that firms can anticipate their entry on the export market and if their innovation efforts
are driven by this expectation, product and process innovation cannot be considered
exogenous in the export decision. To take this anticipation effect into account, Caldera (2009)
and Cassiman and Martinez-Ros (2007) estimate, in addition to their baseline model, several
instrumental variables (IV) regressions. Caldera relies on a linear probability framework and
uses firm-level funding for innovation as an instrument, while Cassiman and Martinez-Ros
rely on IV probit estimation.
Becker and Egger (2007) and Damijan et al. (2008) take a more direct approach to account for
the potential endogeneity of the innovation decision in the firm?s exporting decision, both
papers apply matching estimators. Becker and Egger (2007) focus on the causal link going
from innovation to exporting; while Damijan et al. (2008) look at the bi-directional causal
impact. As was noted in the introduction, Becker and Egger (2007) are among the first to take
the correlation between firm-level product and process innovation explicitly into account. In
their matching analysis, they distinguish between four types of firms: i) firms that did not
introduce a product or process innovation, ii) firms that introduced a product, but not a
process innovation, iii) firms that introduced a process, but not a product innovation and iv)
firms that introduced both a product and process innovation. However, their analysis includes
both continuing exporters and starters, i.e. they do not control for past exporting history in
In terms of the empirical results, the existing literature remains inconclusive. Caldera (2009),
Cassiman and Martinez-Ros (2007) and Becker and Egger (2007) find that the introduction of
a product innovation results in an increase in firms? export propensity. On the other hand,
Note that matching does not necessarily deal with all kinds endogeneity. Unobservable heterogeneity
surrounding the treated firms may still persist if the variables in the first order probit are not perfectly correlated
with the unobservables.
Damijan et al. (2008) find no significant impact of product innovation on the export
propensity of Slovenian firms. For process innovation, the findings of Becker and Egger
(2007), Cassiman and Martinez-Ros (2007) and Damijan et al. (2008) suggest that process
innovation does not increase firms? export propensity. Caldera (2009) on the other hand
reports a positive and significant impact of process innovation on the probability of firms to
export, a finding that is robust to several endogeneity controls.
From a theoretical point of view, there are reasons to expect that product innovation and not
process innovation drives firms into exporting. Klepper (1996) analyzes the patterns of exit,
entry, innovation and growth over the product life cycle. His findings indicate that firms are
more likely to conduct product innovations in the beginning of their life cycle (prior to
exporting), while they are more likely to focus on process innovations during the later stages
of their life cycle. This pattern is in line with the product life cycle as put forth by Vernon
(1966), where firms first introduce a product innovation on the domestic market, after which
they start exporting that product. Rationalization of the production process, for instance
through process innovations aimed at improving the efficiency of production only takes place
at a later stage. As noted by Cassiman and Martinez-Ros (2007), process innovation is also
likely to become more attractive to the firm once production volumes are large and
competition is mounting.
However, to the extent that the introduction of a process innovation makes the firm more
productive, process innovations can help firms to attain the minimum efficiency level needed
to enter the export market in a profitable way.
3. DATA AND EMPIRICAL FACTS
To investigate the relationship between a firm?s innovation activities and its probability to
start exporting, we use data from an innovation survey for Belgium, obtained from the Belspo
(2006). The survey is conducted every four years, the data we use are for the years 2000 and
2004. The population for each survey is selected on the basis of the full population of Belgian
firms, registered at the National Office for Social Security at the end of the period considered
(2000 and 2004). Of these, all firms with at least ten employees are selected. The full sample
of firms in 2000 amounts to 2,100 firms; while for 2004 data are available for 3,322 firms. To
ensure that the sample is representative for the full population of firms, additional sampling is
performed in each period. The final sample of firms is representative for the population in
terms of sector distribution, size classes and regional distribution (Teirlinck 2005).
The survey data contain detailed information on firms? innovation activities, as well as some
general information, such as firm?s export intensity in 2000 and 2004. The survey further has
information on both innovative efforts of the firm (internal and external R&D) as well as on
its innovative output. For innovative output, a distinction is made between a product
innovation, defined as a new or significantly improved good or service that is new to the
market or new to the firm; and process innovation, which concerns new or significantly
improved methods of production, logistics, etc .
Apart from export intensity and innovation characteristics, the questionnaire contains
information on firm-level sales and employment and identifies foreign affiliates of
multinational firms. In order to obtain additional information required to calculate the
productivity of firms in the sample, we merge the survey data with firm-level annual accounts
information, obtained from the Belfirst database (BvDEP 2006).
As was already noted in the introduction, we restrict the sample used in the empirical analysis
in two important ways. First, we limit attention to those firms that have replied to the
questionnaire in two consecutive periods. Since sampling for the innovation survey is
performed independently in each period, the overlap between the two periods is limited to 600
firms (i.e. these firms have responded to both questionnaires). Reducing the sample in this
way allows us to use (four-year) lagged innovation and firm-level characteristics in the
empirical analysis, and hence to avoid a simultaneity bias resulting from the fact that firm-
level innovation and export decisions are taken at the same point in time.
Second, to control for past exporting history, we restrict the sample to two types of firms: i)
firms that start exporting in 2004 (i.e. they did not export in 2000) and ii) a control group of
firms that did not export in either period (Never exporters). 97 firms in the sample start
exporting in 2004, while 92 firms did not export in 2000 or 2004. Hence, the total sample size
amounts to 189 firms .
For the definitions of all variables used in the empirical analysis, we refer to Appendix A.
The representativeness of this sample is no longer guaranteed but includes all firms for which all information
necessary for our analysis could be found.
Table 1 summarizes the sector distribution specifically for our sample. As can be seen in the
table, the sample covers all sectors of the economy. Apart from the number of firms, table 1
also lists the number of non-exporters and starters in each of the sectors considered.
Ever since the seminal work by Bernard and Jensen (1995), many empirical papers have
documented the differences between exporters and non-exporters in terms of several firm
characteristics, such as size, productivity, etc. (see for instance De Loecker 2007 for Slovenia
or Mûuls and Pisu 2009 for Belgium). In a similar vein, Table 2 reports summary statistics
(mean and standard deviation) of a number of firm-level characteristics, separately for non-
exporters and starters on the export market. However, unlike the papers cited above, Table 2
looks at the difference between exporters and non-exporters prior to their potential entry on
the export market.
As was noted in the previous section, we will include lagged firm-level (innovation)
characteristics in the empirical analysis in order to avoid a simultaneity bias, resulting from
the fact that firms? innovation and export decisions, as well as decisions related to the
allocation of inputs and outputs are taken at the same point in time. By including (four-year)
lagged firm characteristics, we aim to control for these simultaneity issues. Analogously to the
empirical analysis, Table 2 therefore reports lagged firm-level characteristics.
Table 2 shows that exporters are larger and more productive already four years prior to
engaging on the export market. These differences are statistically significant. In the empirical
analysis below, we take these differences into account, in addition to industry dummies to
control for differences across sectors.
Table 3 summarizes the innovation characteristics of the sample. Similar to Table 2, the table
distinguishes between non-exporters and starters on the export market. The values reported in
the table refer to the number of firms engaging in a particular innovation activity, the
percentages are calculated with respect to the total number of non-exporters or starters,
reported in the first row of the table. Several interesting facts emerge from Table 3.
Firm size is defined using employment data. Similar to Aw et al. (2007), total factor productivity is calculated
using the index number methodology. While this methodology has a number of drawbacks, i.e. constant returns
to scale and perfect competition are assumed and no allowance is made for unobservable factors, unlike
parametric estimation, it does not assume a homogeneous production technology for all firms in a particular
sector (Van Biesebroeck 2007). Variables are defined in Appendix A.
First, comparing the last two columns in Table 3, it is clear that firms that will start exporting
in 2004, already exert greater innovative effort in 2000 compared to non-exporters in both
periods. For internal and external R&D, the differences between the two groups are relatively
small. About 30 percent of the starters engage in internal R&D in 2000, compared to 26
percent for the non-exporters. For external R&D, the relevant figures are 13 and 8 percent
respectively. For innovative output however, the differences between the two groups are
much larger. While 58 percent of the starters introduced a product innovation in 2000, only 33
percent of the non-exporters did. Similarly, 49 percent of the starters introduced a process
innovation in 2000, compared to 26 percent for the non-exporters.
Second, as can be seen in the last row of Table 3, many firms introduce a product and process
innovation simultaneously. Within the group of non-exporters, this is the case for 10 firms
(accounting for about 11 percent of the number of non-exporters), while for the starters on the
export market, this is true for 32 firms (or 33 percent of the number of starters). Hence, it is
clear that firms, and particularly those firms that will start exporting in 2004, often carry out
product and process innovations simultaneously rather than in isolation. Within the group of
starters, 57 percent of all firms that introduced a product innovation simultaneously
introduced a process innovation and 67 percent of the firms engaging in process innovation
simultaneously engaged in product innovation. For the group of non-exporters, the relevant
percentages are 33 and 42 percent respectively. The correlation between the two variables
amounts to 0.4428. The overlap between these two different types of innovation will be taken
into account in the empirical analysis below.
4. EMPIRICAL MODEL
In order to investigate to what extent firm-level innovation activities increase firms? export
propensity, we estimate the following empirical model:
Size Firm-level employment in 2000;
TFP Total factor productivity in 2000;
INN Innovation characteristic, differs depending on specification;
I Sector dummy.
The dependent variable in  is equal to one if the firm starts exporting in 2004 and zero
otherwise. As noted before, our sample is limited to those firms that start exporting in 2004
and firms that did not export in both periods. Since we only have access to two consecutive
periods for the innovation survey, the use of initial characteristics in  implies that we can
only include one year of data in the regression (2004). The year 2000 is used to define the
lagged characteristics. Total Factor Productivity (TFP) is defined according to the index
number methodology (Törnqvist index). An important advantage of this method is that it
allows for heterogeneity in production technology across firms. As a robustness check we also
verify our results using labor productivity defined as net value added per employee. For the
definitions of the variables used in the empirical analysis, we refer to Appendix A.
We will include both innovative input and output measures in . All innovation measures
are defined as dummy variables, indicating whether the firm has engaged in a particular
activity or not. We use two input indicators, referring to whether the firm has engaged in
internal or external R&D in 2000 and two output indicators, referring to whether the firm
has introduced a product or process innovation in 2000. As is illustrated in Table 4, the
correlations between the different innovation variables are generally high. Only the
correlations between the two output measures and external R&D are lower than 0.40, in all
other cases, the values are larger than 0.40. As argued before, we will take this high
correlation into account in the empirical analysis.
Specifically, in order to avoid multicollinearity issues, which might result in the
insignificance of some of the variables caused by the high correlation between them, we
include only one innovation measure at a time. Moreover, to take the large degree of overlap
between product and process innovation into account, we will further distinguish between
Aw et al. (2007) note that R&D intensity, unlike the discrete choice to engage in R&D, is more likely to be
driven by firm-specific unobservable factors and noise. Therefore, although the innovation survey data report
data on firm-level expenditures on internal and external R&D, we follow Aw et al. (2007) and rely on dummy
variables, indicating whether the firm has actively engaged in R&D, to measure firms? innovation activities.
This choice is consistent with the existence of a fixed setup cost of R&D, such that once this cost has been
incurred, the amount of R&D actually spent matters less.
firms that have only introduced a product innovation, only a process innovation or both
simultaneously. This will allow us to investigate to what extent the simultaneous introduction
of a product and process innovation offers an advantage to the firm in terms of its export
market prospects. In what follows, the results of the baseline specification given by  will be
5. EMPIRICAL RESULTS
Table 5 reports the regression results for the baseline specification given by equation . All
regressions in the table include a full set of industry fixed effects to control for differences
across sectors. Each of the four columns in the table includes a different innovation measure.
In the first two columns, input measures are added, while the last two columns report results
using innovation output measures. All the values reported in Table 5 are marginal effects,
defined as the marginal probability change at the mean of the independent variables (discrete
change from 0 to 1 for dummy variables), standard errors are reported between brackets.
Results in Table 5 show that productivity has a positive and significant influence on firms?
propensity to start exporting. This result is in line with the theoretical and empirical self-
selection literature (Melitz 2003; Muûls and Pisu 2009), i.e. only the more productive firms
are able to enter the export market. Although Table 2 indicated that firms that start exporting
are (on average) larger than their non-exporting counterparts, firm size is only (marginally)
significant in Table 5. These results suggest that the differences between starters and non-
exporters in terms of their size is mainly due to differences across sectors and not so much to
differences within a sector.
Furthermore, in line with the results obtained by Aw et al. (2007) and Cassiman and
Martinez-Ros (2007), our results suggest that firm-level investments in R&D (internal or
external) do not result in a higher propensity to export in the next period. The last two
columns of Table 5 show the results of estimating , but now including innovation output
rather than input measures. We include product and process innovation separately here,
Sectors are grouped as in Table 1.
without taking into account that many firms introduce both innovations simultaneously. The
results suggest that both product and process innovation (irrespective of whether they were
introduced in isolation or simultaneously) have a significantly positive impact on firms?
propensity to start exporting.
Specifically, the magnitude of the marginal effects implies that firms that introduce a product
innovation increase their probability to start exporting by 22 percentage points, compared to
19 percentage points for process innovation. These findings are in line with those reported by
Caldera (2009) for Spain. Using a similar empirical framework , she finds that firms
introducing a product innovation increase their export propensity by 16 percentage points.
Firms introducing a process innovation exhibit a 7 percentage points increase in their
probability to export, which is somewhat lower than in our case.
To determine to what extent the correlation between product and process innovation leads to
serious multicollinearity issues, the first column of Table 6 reports the results of the baseline
specification, which now includes both innovation output variables, i.e. product and process
innovation are both added as independent variables in the regression. When both innovation
variables are taken into account simultaneously, only product innovation emerges as a
significant determinant of firms? export propensity. These results are in line with results
reported by Cassiman and Martinez-Ros (2007) who also report a positive and significant
effect for product innovation, but not for process innovation on firms? export propensity.
However, given the high correlation between the two innovation output measures and the fact
that they both act as significant drivers of firms? probability to enter the export market, it can
be argued that the insignificance of the process innovation variable does not reflect its true
impact. Moreover, while including the innovation measures one by one avoids the
multicollinearity issues discussed above, it fails to take into account potential
complementarities between firms? product and process innovation in shaping their future
export prospects. As was already noted in Section 3, 49 percent of all firms that introduced a
product innovation in 2000 simultaneously introduced a process innovation. Similarly, 58
percent of all process innovators were also product innovators in 2000.
Caldera additionally adds random effects to the baseline specification. Since we only have data for two time
periods and we add lagged firm characteristics in the regression, we cannot estimate a random effects probit
To take this high correlation into account, Table 6 distinguishes between four types of firms:
i) non-innovators (the baseline), ii) firms that only introduced a product innovation in 2000,
iii) firms that only introduced a process innovation in 2000 and iv) firms that introduced both
a product and process innovation simultaneously. Since these categories are mutually
exclusive (a firm is never part of more than one of the four groups), we avoid potential
multicollinearity issues. Moreover, by accounting explicitly for the fact that some firms
introduce a product and process innovation at the same time, we are able to determine to what
extent both innovation activities have complementary effects on firms? export propensity.
Results of the baseline model, but now including three rather than two innovation output
measures, are reported in the second column of Table 6. Again, all regressions include sector
dummies. In the regressions where only one innovation measure is included (only product,
only process innovation or both), we omit all other categories from the sample, e.g. if the
variable Onlyprod is included in the regression, we omit all firms that introduce either a
process innovation in isolation, or a product and process innovation. This implies that in all
regressions the baseline category (the control group) are all non-innovating firms. Similar to
the results for the non-innovation characteristics reported in Table 5, total factor productivity
emerges as a significant driver of firms? export propensity, while firm size is insignificant.
For the innovation measures, results suggest that it is the simultaneous introduction of a
product and process innovation, and not so much either of the two in isolation, that drives
firms into exporting. Firms introducing a product or process innovation in isolation, exhibit
no significant increase in their probability to start exporting.
This finding is in line with findings of Becker and Egger (2007) for Germany, who also find
that the simultaneous introduction of a process and product innovation has a large impact on
firms? export propensity. However, while Becker and Egger (2007) additionally find a
positive and significant impact of product innovation in isolation (though not for process
innovation), this is not the case here. Product or process innovations conducted in isolation
exert no significant impact on the probability of firms to start exporting.
6. ACCOUNTING FOR ANTICIPATION EFFECTS
When using labor productivity instead of TFP all results go through be it that labor productivity itself is no
If firms can anticipate entry on the export market and their innovation activities are driven by
this prospect , innovation cannot be considered exogenous in the analysis reported above. To
control for this potential endogeneity, we will report several instrumental variable estimations
for the innovation output measures. We choose to rely on two-stage least squares
(instrumental variables or IV) regression to estimate the causal impact of firm-level
innovation activities on its export propensity for two reasons. First, unlike linear IV models,
non-linear IV estimation requires fairly strong assumptions, i.e. the error terms in the first and
second stage need to be identically normally distributed and both stages need to be correctly
specified for consistent estimation (Carrasco 1998). Moreover, standard IV probit estimation
procedures require the endogenous variable to be continuous (i.e. the first estimation stage is
linear), yielding inconsistent standard errors for endogenous dummy variables. We therefore
follow Caldera (2009) and rely on two-stage least squares regression to investigate the causal
impact of firm-level innovation activities on its export propensity.
As a first step, we estimate the preferred model of Table 6 (column II), including the three
dummies representing whether the firm introduced a product or process innovation in
isolation or the two of them simultaneously, but now using OLS (i.e. we estimate a Linear
Probability Model or LPM). This will allow us to determine to what extent the LPM results
are comparable to the probit results reported in Table 6. The results for the innovation
measures are similar to the ones obtained with the probit model. Again we find that only those
firms that introduce a product and process innovation simultaneously exhibit a significant
increase in their probability to enter the export market.
While the coefficient on productivity is lower for the Linear Probability Model in Table 7, it
is still positive and significant. The next three columns of Table 7 report results of applying an
instrumental variables approach (IV) in the LPM. We account for the endogeneity of firms?
innovation activities by instrumenting. Generally, instruments need to satisfy two
requirements (Greene 2008). First, they cannot have a direct impact on the dependent variable
(i.e. on the probability to start exporting). Second, they need to be correlated with the
endogenous regressor, conditional on all other covariates. Since there are three endogenous
regressors in Table 6, we need at least three instruments.
The prospect of export market entry can be driven by anticipated trade liberalization as in Costantini and
Melitz (2007) or by firm-level considerations.
For instance, in Stata 10, IV Probit estimation can be achieved using the ivprobit command provided the
endogenous regressors are continuous.
The insignificance of the internal and external R&D dummy in Table 5 (i.e. they have no
direct impact on the probability to start exporting), combined with the fact that internal and
external R&D are essentially the inputs for the innovation outcomes (the endogenous
variables), suggests they might be good instruments. Additionally, it is likely that firm-level
on-the-job training activities, on which we have information from the Belfirst database
(BvDEP 2006) are correlated with firm-level innovation activities and in particular process
innovation, since new production processes need to be executed and therefore introduced to
employees and workers. While firm-level training (which is measured using a dummy
variable) does not feature in Table 5 and 6, we ran an auxiliary regression to ensure that
training is not directly related to firms? propensity to start exporting.
To investigate to what extent the instruments are sufficiently ?strong?, i.e. are correlated with
the endogenous dummy regressors conditional on all other covariates, we estimate the
baseline model of Table 6 (column II) using the three instruments above. The first-stage
results of the estimation procedure are reported in the Appendix (Table A.1) . From Table
A.1, it can be seen that for each of the three endogenous dummies (Only product innovation,
Only process innovation and Both) at least one of the instruments yields a positive and
significant coefficient. These results confirm our prior that the instruments chosen are indeed
correlated with our endogenous regressors, conditional upon all other covariates.
The last four columns in Table 7 show the results for the innovation output measures, after
accounting for potential endogeneity of firms? innovation activities (i.e. the anticipation
effect). Similarly to Table 6, we distinguish between firms that have introduced a product or
process innovation in isolation and those that have introduced both of them together.
Surprisingly, both size and productivity are insignificant in all three columns.
Results in the last three columns of Table 7 suggest that, after accounting for the potential
endogeneity of the innovation decision, firm-level innovation has no significant impact on
firms? export propensity. While these results are not in line with those of Caldera (2009) and
Cassiman and Martinez-Ros (2007) that both report a positive and significant impact of firms?
innovation activities on its export propensity after accounting for the potential endogeneity of
the innovation measures; they are in line with results reported by Damijan et al. (2008) for
Unreported, but available from the authors upon request.
The second-stage results are reported in Table 7, last column.
Slovenia, who fail to find a significant effect of firm-level innovation on the probability of
firms to enter the export market .
Hence, after controlling for potential endogeneity of the innovation activities in , we find
no evidence that firms engaging in product and/or process innovation are more likely to start
exporting. These results suggest that only firms with a sufficiently high probability to start
exporting will engage in product and process innovation prior to their entry on the export
market, pointing to the importance of self-selection into innovation activities.
To test the validity and strength of our instruments, three test statistics are reported in Table 7.
All test statistics are obtained using the Stata module ivreg2, developed by Baum, Schaffer
and Stillman (2004). The Sargan-Hansen statistic tests for over-identification of the model,
failure to reject the null hypothesis that the model is over-identified indicates that the
instruments are valid. The Kleibergen-Paap statistic on the other hand tests for under-
identification of the model by testing whether the model is of full rank. The null hypothesis
states that the model is under-identified, rejection of the null implies that the model is
identified. Finally, the Anderson-Rubin F-statistic tests whether the first-stage regressors are
jointly significant and whether the model is identified. The Anderson-Rubin test is robust to
the presence of weak instruments. Failure to reject the null hypothesis that the model is
identified indicates that the instruments are valid.
Apart from the last column of Table 7, all test statistics indicate that the instruments used are
indeed valid. The Sargan-Hansen test is never significant , suggesting that the model is
correctly specified. The Kleibergen-Paap test statistic rejects the null hypothesis of under-
identification at the five percent level in all but the last column of Table 7. Finally, the
Anderson-Rubin F-statistic is never significant, suggesting that the model is identified and the
instruments are valid. However, it is worth noting that the Kleibergen-Paap test statistic points
to potential under-identification of the model in the last column of Table 7, where all three
Most of the existing theoretical and empirical work to date focuses on the manufacturing sector. To verify to
what extent our results are confirmed specifically for manufacturing, we have run additional regressions using
only those firms who are active in this sector. While this limits the sample size further to 71 firms, our main
results reported in Table 7 continue to hold. A notable exception is the result on product innovation in the first
column. Specifically, for the manufacturing sector alone, product innovation conducted in isolation (so without a
process innovation) is found to contribute positively and significantly to the probability that firms start
exporting. However, after accounting for the anticipation effect, product and process innovation (conducted in
isolation or simultaneously) are not found to have a significant impact on the firm?s export status. Hence, overall
the main results reported in Table 7 are very similar to those obtained for the manufacturing sector only. Results
specifically for manufacturing are not reported here for brevity, but are available from the authors upon request.
The Sargan-Hansen test requires the model to be over-identified, i.e. there should be more instruments than
endogenous variables. This implies that the test statistic cannot be calculated for the last column of Table 7,
where the number of endogenous regressors equals the number of instruments.
endogenous regressors are included together in the model. Although the other identification
tests do not confirm this result, some caution in the interpretation of our result is warranted.
Future research, ideally based on both a larger sample and including a time dimension, needs
to be undertaken to confirm these results.
This paper contributes to a small, but growing literature that considers firm productivity to be
endogenous, rather than the result of an exogenous draw (Melitz 2003). Empirically, we
analyze the relationship between firm-level innovation the propensity of firms to start
exporting, using data from an innovation survey for Belgium in consecutive periods.
In our empirical analysis, we control for three potential sources of endogeneity: (i)
simultaneity, which is a consequence of the simultaneous character of innovation and export
decisions; (ii) causality, introduced by persistence in exporting activities and (iii) anticipation,
caused by the fact that firms may innovate in anticipation of export market entry that lies
ahead. We account for these sources of bias by using lagged firm-level and innovation
characteristics, by focusing on starters on the export market (versus a control group of non-
exporters) and by applying instrumental variable estimation. A central finding of the analysis
is that it is important to take the potential complementarities between product and process
innovation into account when analyzing firms? propensity to export. Accounting for the fact
that about half of all innovating firms introduce a product and process innovation
simultaneously, our empirical results suggest that it is the combination of product and process
innovation, rather than either of the two in isolation, that is correlated with firms? entry into
the export market.
Correcting further for an anticipation effects by applying an instrumental variables approach,
points to the importa