Firm exports, innovation and the regional dimension in Spain

ABSTRACT Firm exports, innovation and the regional dimension in Spain. Regional Studies. Firm-level data are used to estimate the effect of product and process innovations on firms’ exports in each Spanish NUTS-2 region. Results show that the effect of innovation on exports is far from regionally uniform. The gap in the propensity to export between innovative and non-innovative firms, conditional to other sources of firm heterogeneity, is shown to be particularly wide in regions with a high extensive margin of exports. An immediate implication is that policies aiming at stimulating innovation, which are likely to be effective in increasing the number of exporting firms, will not exert the same effect in all regions.


Introduction
Different studies have provided empirical support for the existence of a close link between firms' exports and productivity and other sources of firm heterogeneity (Bernard, Eaton, Jensen, and Kortum, 2003;Bernard and Jensen, 2004;Melitz, 2003). Some of these studies have compared the export performance of innovative and non-innovative firms, concluding in favour of a significant positive correlation between innovation and exports (Aw, Roberts, and Winston, 2007;Becker and Egger, 2013;Bratti and Felice, 2012;Caldera, 2010).
Nonetheless, all the studies that have analysed the link between innovation and firm exports so far have neglected the role of space, despite several studies using aggregated regional data having shown sharp disparities across regions in exports that are linked in one way or another to some regional characteristics (Gil, Llorca, and Serrano, 2008;Nicolini, 2003). Similarly, the still small number of firm-level studies that have recognized the potential role played by regional factors have concluded that the export activity of similar firms in different regions varies depending on the characteristics of the region in which they are located (Farole and Winkler, 2014;Rodríguez-Pose, Tselio, Winkler, and Farole, 2013). However, none of these studies have stressed the role of innovation and in particular studied whether the regional differences in firms' propensity to innovate are behind those observed in their export activity.
Against this background this paper hypothesizes that differences across regions in firms' innovation performance account for an important portion of the territorial disparities observed in export activity. More precisely, the main objective of the paper is to show that, in addition to the well-known disparities in the amount of innovation, regions also differ in the effect that innovation has on firms' propensity to export. From the evidence supporting the causal effect of innovation on firms' exports (Basile, 2001;Becker and Egger, 2013;Brancati, Marrocu, that policies favouring innovation will also be effective in enhancing firms' export performance. However, the immediate implication of the hypothesis in this paper is that a similar increase in innovation activity induced by a policy intervention would not necessarily have the same effect on export performance in all regions. Hence, this potential source of heterogeneity should be taken into account when the intervention is designed. To the authors' knowledge, regional heterogeneity in the effect of innovation on exports has not been considered in the literature so far. In this regard it is worthwhile noting that accounting for regional heterogeneity is far less feasible if the analysis of the relationship between innovation and exports is performed using aggregated data at the regional level. In addition, using the macro-level approach to studying the effect of innovation on exports would not allow researchers to control for important sources of firm heterogeneity (such as size, sector of activity, and productivity), increasing the chance of falling into an ecological fallacy.
The contribution of innovation to regional disparities in firms' export performance is analysed using a representative sample of manufacturing firms in each of the Spanish NUTS2 regions.
To be clear, in the first stage, the effect of product and process innovation on the probability of exporting for the firms in each Spanish region is estimated using a bivariate probit model, which aims to control for endogeneity of the measures of innovation. In the second stage, the estimated coefficients for each region are combined with the sample values of firm characteristics to compute counterfactual average propensities to export in each region in a counterfactual scenario for the propensity to innovate in products and processes. A comparison of actual and counterfactual regional export propensities provides an intuitive assessment of the impact of regional differences in innovation on those observed in export performance. local markets. Therefore, exporting makes firms more prone to innovate. The empirical literature investigating the innovation-exports link has provided strong support for the selfselection hypothesis (e.g. Basile, 2001;Becker and Egger, 2013;Cassiman et al., 2010). The evidence that empirically supports the learning-by-exporting explanation is somewhat weaker, although some recent studies have pointed out that addressing the issue of endogeneity appropriately leads to not rejecting the causal effect of exports on innovation (Aw et al., 2007;Bratti and Felice, 2012;Bustos, 2011;Love and Ganotakis, 2013). In any case the strong evidence on the interplay between firms' innovation and their internationalization (with exports as one of its components) led Altomonte, Aquilante, Békés, and Ottaviano (2013) to suggest better coordination of internationalization and innovation policies.
Although evidence based on the analysis of regional exports has revealed sizeable territorial differences in the amount of trade flows, consideration of the regional dimension in the study of firms' export activity remains scarce. As indicated by Farole and Winkler (2014), location has typically been reduced to a control dummy in firm-level analyses. Regional dummies would capture differences across regions in exports originated by geography and disparities across locations in some socio-economic features and in the endowment and quality of the infrastructures that facilitate exports (see for instance Márquez-Ramos, 2016;Naudé and Gries, 2009;Nicolini, 2003). In fact, it is reasonable to think that these factors are behind the regionally differentiated impact of some trade agreements and policies designed to stimulate exports (Coughlin and Wall, 2003;Gil et al., 2008). In any case this evidence provided by studies using aggregated regional data on exports and their determinants advises on the importance of including the regional dimension in studies analysing export performance at the firm level. In this regard some of the most recent contributions to the literature on firms' export performance have replaced the set of regional dummies with proxies for the geography and characteristics of the region in which each firm is located.
A group of papers explored the influence of agglomeration economies on firm exports under the assumption that market and non-market interactions of other local firms reduce the costs of exporting. Although Bernard and Jensen (2004) did not find evidence on the effect of agglomeration spillovers on firms' export propensity in their study for the US, the most recent evidence has suggested a positive external effect (Anderson and Weiss, 2012;Koenig, 2009;Koenig, Mayneris, and Poncet, 2010). In addition to agglomeration, another group of recent papers has included in the analysis of firm exports the effect of the business environment and the institutional setting of the region in which the firm is located. They have assumed that not only firm characteristics and agglomeration economies affect firm export performance.
Regional features, such as the endowment of education and infrastructures, the investment climate, and the quality of local institutions, may also exert an impact on firms' decision to export (Farole and Winkler, 2014;Mukim, 2012;Rodríguez-Pose et al., 2013). 2 These recent studies include measures of agglomeration and regional endowments in addition to firm characteristics in empirical models aiming to explain firms' export activity, 3 assuming that regional determinants shift export performance conditional on firm characteristics but neglect any influence that regional factors may have on the effect of firm characteristics on exports. In other words they impose similar effects of firms' characteristics regardless of their location. In the case of interest to our study, it should be noted first that most studies incorporating the regional dimension have not included innovation as a source of firm heterogeneity. Secondly, the approach followed in the above-mentioned studies (adding regional-level variables to the specification of firm-level export activity) does not account for the region-specific effect of innovation on exports, since it imposes the same impact on all firms regardless of the regional context. Conversely, in this study we estimate a separate effect of innovation in each region. This allows us to assess the soundness of the study's hypothesis, that is, the extent to which the observed differences across regions in firms' innovation activity account for regional disparities in export performance, taking into consideration the fact that the effect of innovation on firm exports is likely to vary across regions.

The Data Set
The study of the impact of firms' innovation on export performance in each region demands data on proxies for these two magnitudes, and for other sources of firm heterogeneity, from a sample that needs to be representative of the population of firms in each region. In Spain such data are available from the Innovation in Companies Survey (ICS) undertaken by the Spanish Statistical Office. The ICS provides detailed information on technological and nontechnological innovations following a methodology based on the OECD Oslo Manual.
Interestingly, the ICS also provides information on firm performance, including sales abroad and total sales, as well as some characteristics such as firms' size and sector of activity.
The sample in the ICS is representative of firms with 10 or more employees in all branches of activity. However, this paper focuses on firms in the manufacturing sector, since, as in most of the previous literature, it is assumed that they are the ones producing tradable goods, at least in a much larger proportion than firms in other sectors of activity. The information corresponding to the 2005 ICS wave was exploited. The choice of that year took into account the availability of the firm-level data required for the analysis and the fact that the phenomenon under study was not contaminated by the turbulence caused by the great recession. 4 Unfortunately, the ICS did not allow us to track each firm in the sample over several years, which means that the information could not be treated as a panel data set. 5 However, the ICS sample was designed to guarantee representativeness at the regional level.
Specifically, it contains samples that represent the population of firms in each of the NUTS II regions in Spain. They correspond to the 17 autonomous communities, which are historical, geographical, and administrative regions with a high level of political and financial autonomy, including the promotion of trade and innovation within their boundaries.
It should be noted that the size of the manufacturing sector varies between the Spanish regions. This means that the number of manufacturing firms included in the sample varies markedly among regions: from a maximum of 3118 in Catalonia -the region with the largest manufacturing sector -to a minimum of only slightly above 200 firms in the Canary and the Balearic Islands -regions specializing in tourism with scarce manufacturing activities (see last column of Table 1). In any case it should be mentioned that the size of the entire sample, 14078 manufacturing firms, is large enough to guarantee the quality of the estimates provided in the following sections.

Definition of the Main Variables
As mentioned above, the ICS 2005 includes the volume of total sales and sales abroad in the surveyed period. This information was used to compute a binary variable for the firm decision on exporting, which equals 1 if the firm exported in 2005 and 0 otherwise. The sample average of the binary export variable in each region is an estimate of its extensive margin of exports (i.e. the share of exporting firms in the region).
As for innovation, the ICS includes detailed information on the inputs and outputs of the innovation process. Following the arguments in the previous literature, measures of outputs rather than measures of inputs were selected. Among the available output measures, we also followed the innovation internationalization literature (e.g. Becker and Egger, 2013) in selecting product and process innovations implemented by firms and distinguished between the two types because of their expected differentiated impact on exports. Accordingly, a dummy variable was defined for product innovation that equals 1 if the firm implemented some product innovation in the last two years and 0 otherwise. Similarly, the dummy variable for process innovation equals 1 if the firm implemented some process innovation in the last two years and 0 otherwise. Finally, a dummy variable that accounts for the implementation of innovation regardless of the type was computed. This innovation (prod/proc) variable equals 1 if the firm implemented product and/or process innovations and 0 if it did not implement any kind of innovation.
Beyond export and innovation activities, the ICS includes useful information to control for other sources of firm heterogeneity in the empirical analysis. In particular, data on the total sales and the number of employees were combined to compute a simple measure of firms' labour productivity. 6 To mitigate the issue of endogeneity for this variable, the values for 2003, which are also reported in the survey, were used. 7 In addition, several dummy variables were defined to account for other sources of firm heterogeneity: size (10 to 49, 50 to 249, and 250 and more employees), branch of activity (NACE 2 digits), being part of an enterprise group (not part of a group, part of a national group, and part of an international group), and the NUTS2 region where the firm is located (17 regions).

Exports and Innovation in the Spanish Regions. Descriptive Analysis
The extensive margin of exports for manufacturing firms in each of the Spanish regions and in the country as a whole is reproduced in the first column of Table 1. The figures indicate that slightly above 50% of the Spanish manufacturing firms with at least 10 employees exported in 2005, which is in line with the evidence reported elsewhere (e.g. Barba Navaretti et al., 2010). They also reveal sharp regional disparities in the propensity to export, which ranges from about 35% in Andalusia, Asturias, Cantabria, Castile La Mancha, and Extremadura to above 60% in Catalonia and the Basque Country.
Table 1 also provides the share of manufacturing firms that innovate in products, in processes, and in at least one of these two types of innovation. In the entire country, just 32.4% and 38.1% of firms reported that they implemented some product and process innovation, respectively. The share rises to almost half of the firms under the softer criterion of reporting at least one of the two types of innovation. The data also show that regions differ markedly in firms' propensity to innovate. The share of firms reporting innovations is larger in the most advanced regions (Catalonia, Madrid, and the Basque Country) than in the traditionally lagging regions (Extremadura, Castile La Mancha, and Andalusia) -not to mention the low values for the share of innovative firms in the island regions. A comparison of the regional figures for the extensive margin of exports with those for the share of innovative firms suggests a connection between the two magnitudes. The share of exporting firms is larger in regions where innovation is more abundant; the opposite also holds true. Actually, the correlation coefficient for the regional figures of the extensive margin and the share of firms that innovate in products, in processes, and in at least one of them is, respectively, 0.82, 0.84, and 0.86. 8 < TABLE 1 NEAR HERE > In sum, the results from the descriptive analysis in Table 1 confirm that i) the regional disparities in export activity are sizeable and ii) the differences across regions in the propensity to innovate in products and processes are linked with the disparities observed in firms' export activity. In the following sections, we investigate these issues further by controlling for other sources of firm heterogeneity that are far from being distributed homogeneously across regions.

Regional Impact of Innovation on the Propensity to Export
The main aim of this section is to provide evidence supporting the hypothesis that, having controlled for other sources of firm heterogeneity, the difference in export propensity between innovative and non-innovative firms varies across regions. As initial evidence, we report the extensive margin of exports for innovative and non-innovative firms in each region and the whole of Spain in Table 2. The raw data show that the extensive margin is much higher for innovative firms in all regions. The figures also indicate that the gap between innovative and non-innovative firms varies across regions, which suggests that the impact of implementing innovation on firms' export propensity might well depend on regional characteristics. On the other hand, a comparison of the gap in the extensive margin of exports reveals that it is somewhat wider for product than for process innovations. This is consistent with the premise that having new or substantially improved goods contributes to a greater degree to the firm export opportunities than implementing new or improved technologies of production or methods of delivery.

Empirical Specification
The raw data in Table 2 are informative about the gap in the extensive margin of exports between innovative and non-innovative firms in each region. However, it neglects the effect of other sources of firm heterogeneity on this gap. Hence, in this section we aim to obtain an estimate of the impact of innovation on firms' propensity to export in each region, controlling for the other sources of heterogeneity. We followed Roberts and Tybout (1997) in assuming that firms decide to export if the profits obtained when exporting exceed those obtained when only serving the country market. !" !"# being such a difference in profits made by firm i in region r when exporting, Therefore, the export status for each firm in each region (export ir ), conditional on the loglevel of productivity (Prod ir ) measured two years before and other firm characteristics in X ir , is supposed to depend on the firm innovation status (Inn ir ). Consistent with our main assumption in this paper, the specification in (1) allows for regional differentiated effects of innovation ( ! ), while regional heterogeneity is also allowed in the impact of productivity ( ! ). Under the assumption of normality for the random component, u ir , the estimate of the impact of innovation can be obtained from a probit model: where Φ is the cumulative normal distribution function and X ir includes dummy variables for the size of the firm, the sector of activity, being part of an enterprise group, and the region where the firm is located.
As discussed in section 2, there are arguments against the exogeneity of the measure of innovation in a specification such as that in (2). Since innovation is a discrete variable, the treatment of endogeneity in the context of a (non-linear) probabilistic model is far from simple, as the standard IV approach is not suitable in this case. 9 As in other studies in the literature analysing the link between firm innovation and export status (e.g. Aw et al., 2007; Girma, Görg, and Hanley, 2008), we opted to control for endogeneity by estimating a bivariate probit model. 10 We assumed that firm i in region r implements an innovation if the profit that it obtains by doing so exceeds that of not implementing the innovation: where !" !"" is the difference in profits, Inn ir is the innovation status, Z ir is the set of factors other than productivity affecting that status, and !" is an error term. As with exports, a regionally uniform effect of productivity on innovation is not imposed, but it is allowed to vary across regions ( ! ). Under the normality of the error term, the probit specification for the innovation status is: If Cov !" , !" = 0, that is, the errors of the two decisions are independent, the univariate probit model in (2) can be used to obtain consistent estimates of the impact of innovation on the export status in each region r. Otherwise, the errors of the two processes are related to one another, Cov !" , !" = ≠ 0, and the estimates from the probit model in (2) in isolation will not be consistent. This will be the case if, for instance, unobservable firm characteristics affect simultaneously export and innovation statuses. In such a case, consistent estimates of the impact of innovation on exports, and of the other unknown parameters, can be obtained by estimating the bivariate probit formed by (2) and (4).
As regards identification in the bivariate probit model, it needs to be stated that we included in Z all the variables in X plus a set of instruments for the innovation status of the firm.
Finding appropriate instruments is a challenge, as one has to ensure that they do not affect firms' exports in a direct way, but only through their effect on innovation (exclusion restrictions). In our empirical exercise, we used the type of instruments suggested in the recent literature aiming to identify a causal effect of innovation on firm exports. In particular, as in Becker and Egger (2013) and Lachenmaier and Wößmann (2006), we used information on obstacles/impediments to firms' innovation.
To be more precise, the ICS questionnaire includes four possible answers ("very important", "important", "not very important", and "not important at all") to the question concerning the importance assigned by a firm to the lack of its own capital in the decision to innovate. Firms also report the extent to which the fact that the innovation expense is too high is an obstacle to their innovative efforts (the same four possible responses are also available for this item). For each of these two questions, we generated a set of binary variables for each possible response, that is to say, four binary variables for the lack of equity capital and another four for too high innovation expenses. 11 Although the ICS includes information on other obstacles/impediments to innovation, we decided not to use them as instruments, since they are likely to exert a direct influence on firm exports (e.g. a lack of skilled workers and a lack of access to external capital). 12 On the other hand, consistent with Altomonte et al. (2013), we exploit the variation across sectors (NACE 2 digits) and firm size (small, medium, and large) in the share of firms that benefited from innovation subsidies in each region. The ICS includes information on whether the firm received public financial support for innovation activities in the period 2003-2005.
However, we decided not to use the firm-level information on innovation subsidies, given that we cannot ensure that the selection of recipient firms was not based on performance (which would influence exports in 2005). Instead, we used the information on subsidies to compute the share of subsidized firms in the same sector and size category. This measure correlates with firm innovation, while it may be assumed to be exogenous to the export status in our sample as long as the group of firms that might have influenced the set-up of R&D promotion policies (prior to 2005) did not coincide completely with the group of firms in the sample for which we measured the export status in 2005. 13 It is noteworthy that, in the empirical exercise, the variable for the measure of innovation was interacted with the set of region dummy variables to obtain separate estimates for the effect of innovation on exports. Therefore, we also included as instruments the interaction between the regional dummies and the set of instruments defined above (see Aghion, Howitt, and Mayer-Foulkes, 2005;Wooldridge, 2002).

Results
The marginal effects calculated from the coefficient estimates of the bivariate probit model are used to measure the impact associated with innovation and the other firm characteristics.
The results for the full set of characteristics are shown in Table A.9 of the online Appendix.
This table reports the estimates of the marginal effects from the specification that imposes a common effect of innovation and productivity in all regions (the first three columns of results) and from the one in which the effect of innovation and productivity is allowed to vary across regions. It also includes basic statistics for assessing the adequacy of the bivariate probit model. In both cases a set of regional fixed effects is included to control for the impact of contextual factors affecting the propensity to export of the firms in each region. More precisely, they capture the differences across regions in firms' export performance due to geography, the endowment of infrastructures (hard and soft), and any other influence of the socio-economic environment (e.g. Márquez-Ramos, 2016;Nicolini, 2003). Alternatively, estimates were obtained using proxies for some of these regional factors. A summary of these results is reported in Tables A.10 to A.12 of the Appendix. Since there are no significant differences in the estimated marginal effect of innovation for each region between the specification including the fixed effects and that including the regional variables, detailed results and the corresponding discussion are provided only for the former. In brief, the estimates of the marginal effects for all the characteristics are in line with those reported in the literature: a positive and significant effect of innovation and productivity in all regions, a higher probability of exporting in medium and large firms, and an increase in the probability of firms being in a group, although only when it is an international one. The results also confirm that there are significant differences across industries in the probability of exporting, whereas the differences across regions are only significant when imposing the same effect of innovation and productivity in all regions. This suggests that regional variability in the effect of these two magnitudes accounts for most of regional disparities in the propensity to export, conditional on the observed firm characteristics. On the other hand, the correlation of the errors of the exports and innovation equations (ρ) is highly significant, which supports the bivariate probit model as the preferred specification. Finally, the coefficients are highly jointly significant. 14 Given the particular interest of this paper, Table 3 reproduces the estimated marginal effects of innovation for each region and the country as a whole. For Spain as a whole, the probability of exporting for firms that declared that they innovated in products is 35 percentage points (pp) higher than for similar firms that did not innovate. The size of the effect is similar for process innovation and for the measure that accounts for both types of innovations. Interestingly, the results confirm the existence of substantial disparities across regions in the impact of innovation. The estimated increase in the probability of exporting associated with product innovation is above 40 pp in Aragon and La Rioja, while on the opposite side, apart from the island regions, the smallest impact is shown by firms in Asturias and Catalonia. Regional disparities are also observed in the impact of process innovation, with the largest marginal effect in Murcia, Aragon, and Galicia and the smallest in the two island regions and in Asturias and Andalusia. Accordingly, regional disparities are also observed when the two types of innovation are combined in a single variable, as shown by the estimated marginal effects in the last column of Table 3. 15   < TABLE 3 NEAR HERE > Beyond the particular estimate of the effect of innovation in each region, we can say that, as a general rule, the marginal effect of innovation is greater in regions with high extensive margins of exports, while the opposite is also true. Actually, the correlation coefficient for the extensive margin in each region and the corresponding marginal effect of product innovation, process innovation, and the combination of the two innovations is, respectively, 0.66, 0.69, and 0.72. Therefore, this evidence suggests that firms in regions with a high extensive margin of exports tend to obtain a higher payoff from innovations in terms of export status.

Counterfactual Analysis
As the final step, in this section we discuss the results of a simple counterfactual exercise.
Using sample values for labour productivity and the other firm characteristics in X, and the estimates of the corresponding parameters for each region r obtained in section 4, we computed the counterfactual extensive margin for each Spanish region in an alternative scenario for the share of innovative firms. More precisely, a counterfactual extensive margin of exports for region r was obtained by averaging the marginal predicted probability of exporting for each firm i in region r, using the estimate of the parameters from the bivariate probit model in region r: where ! !"#$% is the counterfactual extensive margin of region r, the bar over the expression on the LHS denotes the sample average, the ^ over the coefficients represents the estimates from the bivariate probit discussed in section 4, Prod r is the vector of values of the labour productivity in each firm in region r, X r is the matrix with observations of the other characteristics of firms in region r, and Φ is the cumulative normal distribution function as in (2). The marginal predicted probability of exporting for each firm in region r was computed by assigning a common counterfactual propensity to innovate, ! !"#$% , instead of the actual value for the firm's innovation status, Inn ir . Concretely, the results were obtained using the share of innovative firms observed in Catalonia as the common-to-all-firms counterfactual propensity to innovate. As indicated in section 3 (see Table 1), Catalonia is the region with the largest share of innovative firms, therefore being an appropriate target for the other regions.
Comparing the counterfactual export margins with the actual values allows us to assess the expected impact of increasing firms' propensity to innovate in each region to the level shown by an average manufacturing firm in Catalonia. 16 Note that the expected impact for a region will depend both on its distance to the innovation propensity target and on the particular effect of innovation on export activity in that region.
The differences between the counterfactual and the actual margins of exports in each region and in the entire country are displayed in Table 4 Islands is significantly negative (-0.70, -0.88, and -0.88 for product, process, and both innovations, respectively). This result leads us to conclude that, other things being equal, increasing the innovation propensity, particularly in products, would contribute to narrowing the regional gap in the proportion of exporting firms.
The analysis so far has not taken into account the country or area of destination of the firms' exports. It could be the case that the effect of innovation on exports varies between markets, depending on certain characteristics (distance, preference for technologically advanced products, etc.). Unfortunately, the information on export destinations in the ICS data set is rather limited. Still, it is possible to know whether the firm sold products in other EU countries and in countries outside the EU. Hence, estimates of the marginal effect of innovation for each region were obtained for the case of exports to the EU and exports to non-EU countries. The results are reported in Table A.13 of the Appendix. It can be observed that, with few exceptions, there are no sizeable differences in the impact of innovation between the two types of destinations. Accordingly, the differences in the change in the extensive margin of exports (to EU countries in one case and to non-EU countries in the other) induced by innovation are minimal (Table A.14 of the Appendix). On a priori grounds, it could be thought that innovation should be more important for non-EU destinations (for instance to compensate for the higher trade costs of markets outside the common trade area). However, the importance of the Latin American and Moroccan markets for a substantial number of Spanish exporting firms must be kept in mind. In the case of the former destination, the greater distance is compensated for by the lower costs associated with the common language and cultural ties. As for Morocco, it is worthwhile noting that it is a closer destination than most of the EU countries and that different trade agreements favour trade between Spain and this country. Last, the income per capita in these destinations is still rather low in comparison with most of the EU countries, which acts against the demand for new and more technologically advanced products produced by innovative firms.

Concluding Comments
This paper has shown that innovative firms are more prone to export than otherwise similar non-innovative firms in all the Spanish NUTS2 regions. However, the results indicate that the effect of innovation is far from being regionally uniform. The increase in the propensity to export due to innovation has been estimated to be larger in regions where the extensive margin of exports is high; this result is robust to the alternative measures of innovation considered in the analysis.
The evidence on the differentiated regional effect of innovation on firms' exports is a novelty in the literature, since previous studies have added either regional dummies or controls for regional endowments and agglomeration economies but have imposed the same response of exports to the set of firm characteristics. In contrast, in the modelling strategy followed in this paper, the impact of regional factors and agglomeration is captured by the region-specific intercept of the specifications for the export margins, whilst the impact of innovation is allowed to differ between regions as a result of the influence of the particular conditions in each region. Actually, none of the previous studies that included a regional dimension in the analysis of firms' export performance stressed the effect of innovation. Our results for Spain confirm the key role played by a firm's innovative activity and suggest that, when assessing its contribution to stimulating a firm's export propensity, one needs to take into account the characteristics of the region in which the firm is located.
Although the investigation of the sources of the observed regional disparities in the effect of innovation on firms' exports is beyond the scope of this paper, we can speculate that the differences in export sunk costs across the Spanish regions might be causing territorial disparities in the exports' response to innovation. Innovation contributes to raising future firm productivity and/or having more attractive products and thus to making it easier for firms to face the extra costs of exporting. Even under the assumption that firms in all the Spanish regions are similarly effective in translating innovation into higher levels of productivity and competitiveness, it is sensible to think that geography, agglomeration, and certain regional endowments cause differences across regions in sunk costs. As a result, the benefits of innovation allow the entry exporting costs to be covered by firms in some regions but not in others. This argument can explain the greater effect of innovation on the export status estimated for some Spanish regions with a high extensive margin of exports. A deeper study of this hypothesis is part of our future research agenda.
An immediate implication of the evidence in the paper is that policies aiming to stimulate innovation, which are likely to be effective in promoting exports by increasing the number of exporting firms, will not exert the same effect on exports in all the Spanish regions.
Therefore, the ex ante assessment of innovation policies should include the positive expected effect on export performance but take into account the possibility that geography and certain locational endowments can affect the particular impact of these policies in each region. In addition, the results on the effect of innovation on the export status lead us to recommend focusing the effort of direct policies aiming to promote exports preferably on the group of innovative firms in each Spanish region that are not exporting yet. They are the potential candidates to become exporters if the locational disadvantages are compensated for in some way by the effect of these policies.

Notes
1 The data used in this paragraph for all Spanish regions can be found in Tables A.1 and A.2 of the online Appendix.
2 Interestingly, as indicated by an anonymous referee, the studies analysing the role of regional factors in firms' export performance so far are from developing economies. Future studies should investigate whether the response of firm exports to the regional factors is similar in developed countries. 3 Since the focus of this study is the export performance of firms, and because of the space limitations, we do not intentionally summarize the extensive literature on the determinants of innovation that are internal and external to the firm. We refer the interested reader to the reviews of the literature provided by Audretsch and Feldman (2004) and Carlino and Kerr (2015) and to Czarnitzki and Hottenrott (2009) for an analysis of the relative contribution of firm and regional factors. 4 Unfortunately, it was not possible to gain access to the ICS micro-data for the most recent years. In any case it is worth noting that the inspection of the aggregated regional data on export and innovation activity reveals quite stable evolution over time and hence persistent differences across the Spanish regions in both magnitudes (see Tables A.1  provide information based on panel data for a sample of Spanish firms. Although they include information on export activity, innovation, and other sources of firm heterogeneity, they do not guarantee representativeness at the regional level and thus are not useful for the study in this paper. 7 Previous studies have also included the lag of some firm characteristics to mitigate the harmful effect of potential endogeneity (e.g. Bernard and Jensen, 2004;Koenig et al., 2010). 8 More information about the sector-region distribution of the firms in the sample and the main magnitudes under analysis is provided in Tables A.3 to A.8 of the Appendix. 9 See for instance Angrist andPischke (2009) andCameron andTrivedi (2005). 10 We also estimated the impact of innovation on export status by IV using a linear probability model. However, we prefer to report the results from the bivariate probit framework given our interest in simulating the margins of exports for the Spanish regions in an alternative scenario for firms' innovation activity. The bivariate probit model guarantees values for the predicted probabilities that are within the logical range, while the linear probability model does not.
11 Estimates were also computed using a single binary instrument for each type of obstacle, in which the "very important" and "important" responses were grouped into one category and "not very important" and "unimportant" into the other. The results were robust to this alternative definition of the instruments. 12 A detailed discussion of the appropriateness of the several types of obstacles/impediments as instruments of innovation in a firm-level exports equation can be found in Lachenmaier and Wößmann (2006). We do not reproduce their full arguments here for reasons of space. In any case it needs to be stressed that the lack of own capital will not necessarily be exogenous for absolute exports, as they may serve as a funding scheme for innovation. However, similarly to the case in Lachenmaier and Wößmann (2006), since the focus of the analysis is the propensity to export, it can be argued that this will not affect the innovation possibilities, which can be financed by bigger domestic revenues in firms with a low (or null) propensity to export. In any case it should be mentioned that the results are robust to the exclusion of this variable from the list of instruments. 13 In addition to the one based on the share of R&D incentives, Altomonte et al. (2013) defined another instrument based on the average industry-country R&D intensity, measured some years previously. To achieve this, they relied on OECD aggregated data on R&D expenditures and value added. Unfortunately, we could not exploit a similar instrument in our analysis given the lack of data on R&D investments at the industry-region level in Spain.
14 As indicated above, the characteristics of the empirical model used to obtain the results prevent the computation of the standard battery of validity of instruments tests. Hence, we must rely on the arguments discussed above and on some informal evidence to support the validity of the instruments used in this study. In that regard it can be mentioned that a nonsignificant (direct) effect was obtained for the set of instruments in the reduced-form equation for the export propensity using the sample of firms in an industry in which innovation is a quasi-negligible activity, mining and quarrying. Conversely, as expected given the indirect effect exerted through the influence via innovation, the same exercise in the case of the full set of firms in the Spanish regions revealed a significant joint effect of the instruments on the propensity to export. In a similar vein a linear empirical specification was estimated in which the variable of interest was not the propensity to export but the export intensity (the share of exports in the firm's total sales). The results of the tests clearly indicated that the instruments are not weak and that they are orthogonal to the error term of the export equation. In other words they are valid instruments for the export intensity specification. We assume that this could easily be the case as well when the magnitude under analysis is not the export intensity but the export propensity. These results are available from the authors.
coefficient for innovation and productivity in all the regions in the biprobit model points to the convenience of not imposing a common marginal effect of innovation in all regions. 16 It should be noted that this simple counterfactual exercise does not include general equilibrium adjustments that are expected to follow the change in innovation and exports. As already mentioned, it is merely a sort of what if exercise, the aim of which is simply to assess the contribution of differences across regions in firms' innovation to the regional disparities in the extensive export margin.     Notes: The first two groups of columns refer to the contribution of the region in the total figure for Spain of the number of exporting firms and the volume of exports respectively. The last group of columns is the ratio of exports over the GDP of each region and Spain as a whole. Elaborated from the trade data collected by the customs office of the Spanish Tax Agency and elaborated by ICEX. Ministry of Economics, Industry and Competitiveness of Spain.        ( -0.020 -0.018 -0.011 -0.062 -0.074 -0.011 Canary Isl.