We focus on the case of France. Robin and Schubert (2013) provide an interesting comparative narrative on the French context in terms of public research institutes. France, historically, has been characterized by a centralized and mission-oriented science and technology policy, where missions are defined at the central level and implemented by national, publicly funded research centres such as CNRS, INRA, INRIA and INSERM. Interestingly, for the purposes of our work, until 2006 (under the Act of Law of 12 July, 1999) the National Research Agency (ANR) did not provide explicit public financial support for collaborative research projects involving public research organisations and private firms (Robin and Schubert 2013). Because our empirical investigation is based on data from the French CIS4,Footnote 5 our empirical findings unravel the tendency for firms to cooperate for innovation before and regardless of the introduction of the 2006 tax credit regulation.
The CIS is conducted in EU countries, using a EUROSTAT harmonized questionnaire based on the OECD Oslo Manual (OECD 2005). The French CIS4 was launched in 2005. It targeted a representative sample of firms with more than 10 employees, in non-agricultural sectors; 25,000 firms were interviewed, a response rate of around 86%.Footnote 6 The majority of French CIS4 variables cover the 3-year period 2002–2004; some information on firm structure (e.g., employment, turnover) refer to the initial and final years of the period, and innovation expenditure and outcome variables refer to 2004.
Although we rely only on cross-sectional data, the French CIS4 questionnaire includes a wealth of information covering firms’ structure and location, innovation inputs, outputs and outcomes, and—most importantly for our analysis—innovation barriers and cooperation with external partners. The last two items are important because they allow us to retrieve information on both cooperation and barriers in a period (i.e. 2002–2004) when perception of obstacles was less likely to be affected by the (probably unobservable) confounding and, thus, biasing, factors related to the last global economic crisis.
In what follows, we focus on manufacturing firms only. Given the filtered structure of the CIS questionnaire and in order to have complete information for all the firms included in our sample, we restrict our sample to innovative firms with no missing values for the variables employed (see Mairesse and Mohnen (2010) for a discussion on the opportunities and constraints from the CIS filtering).Footnote 7 Our working sample includes 3825 firms.
We exploit a twofold econometric strategy. First, we investigate whether and how specific barriers linked to cost, market and knowledge factors are related to specific types of cooperation. To this end, we estimate a series of probit models as follows:
$$ Cooperation_{i} = a + b_{1} \varvec{X}_{i} + b_{2} \varvec{Barriers}_{i} + \varepsilon_{i} . $$
where Cooperation denotes the dependent variable the firm’s cooperation agreements with different types of partners; X is a vector of appropriate control variables; Barriers is a vector of the variables synthesizing specific types of obstacles to innovation perceived by the firm; and ɛ is the error term. A further, augmented specification includes the interaction terms between pairs of barriers in order to obtain a preliminary sense of the potential influence of the joint perception of barriers on the propensity to cooperate.
Our dependent variables refer to the firm’s engagement in innovation cooperation. We employ a general binary variable, COOP, which captures whether the firm is engaged in formal cooperation with any type of partner. We distinguish between cooperation with firms (COOPFIRM) and research organizations (COOPORG). As mentioned in Sect. 1, the choice of partner might be dictated by different incentives, such as cost and risk sharing or outsourcing of information on technologies.
Key explanatory variables include a set of dummies that indicate perception of obstacles to innovation. These binary variables are built on the 4-point likert scale items included in the CIS questionnaire, and take the value 1 if the firm reports high relevance for the influence of at least one item related to: costs (COST) (lack of internal funds, lack of external funds, or excessive cost related to innovation); demand and market structure (MKT) (demand uncertainty or market structure dominated by incumbent large firms); knowledge (KNOW) (lack of skilled personnel, lack of information on markets, or lack of information on technologies). It is important to remember that the barrier variables synthesize perception of their importance for innovation-active firms (i.e., firms that have engaged in innovation): in this respect, according to the classification proposed in D’Este et al. (2012) and used in other studies, we focus on the revealed rather than the deterring barriers, which are the hampering factors encountered in the production of innovations, rather than obstacles that deter firms from engaging in innovation activities (D’Este et al. 2008, 2012; Pellegrino and Savona 2013; Hölzl and Janger 2013, 2014).
Building on these variables, we create a set of interactions to provide a preliminary picture of whether and how different types of barriers are complements influencing the cooperation propensity. To this end, we constructed the interactive variables COST*MKT, COST*KNOW and MKT*KNOW, which we add to the baseline specification to address the first two research questions.
The control variables in X aim to reduce potential omitted variable bias. First, we control for firm size, measured as the logarithm of employment. Size can be related to cooperation since large firms are more likely to adopt a combined strategy of internal and external knowledge acquisition (Veugelers and Cassiman 1999) due to their critical mass, resources and likely higher capacity to manage cooperation agreements effectively (Belderbos et al. 2004; Segarra-Blasco and Arauzo-Carod 2008).
We control also for technological capability, by including a dummy variable (i.e. R&D Continuous), which captures continuous engagement in R&D investment (e.g., presence of a dedicated department). This may exert a positive effect on the propensity to cooperate (e.g. Colombo and Garrone 1996; Cassiman and Veugelers 2002; Belderbos et al. 2004). Indeed, persistent and sustained engagement in R&D reflects higher absorptive capacity (Cohen and Levinthal 1989), higher capacity to take advantage from cooperation and thus to recognize its strategic value.
We control for the effect of public funding (R&D Funding) on the probability to cooperate since innovation policy programmes may explicitly require firms to cooperate (e.g. in the case of collaborative R&D subsidies). Public support can change the strategic behaviours of beneficiaries and how they conduct their R&D and, eventually, might lead to increased cooperation with external partners (Marzucchi et al. 2015).
Two dummies for whether firms belong to a national group (GROUP) or to a transnational corporation (TNC) are included. Firms belonging to a group, either national or international, are expected to be more likely to cooperate. When looking for partners, they benefit from the power and prestige of the wider group. In addition, firms belonging to foreign groups may need to establish cooperation agreements in order to acquire specific local knowledge and capabilities, for instance, related to local markets requirements (e.g. Tether 2002). Moreover, they may exploit intra-group communication channels and knowledge pools to gather more information about potential partners, create easier contacts and more easily tap into knowledge from the interacting firms or organizations (Tether 2002; Mohnen and Hoareau 2003).
We include firms’ engagement in international markets (EXPORT) which might exert an effect on cooperation. On the one hand, exporting firms may revert to cooperation to acquire capabilities and maintain their competitiveness. On the other hand, when faced with a strong competitive environment, exporting firms may be induced to protect their know-how (Cassiman and Veugelers 2002) and may reduce cooperation in order to minimize knowledge leakages.
Finally, we include two sets of dummies. The first is the firm’s regional location (NUTS 2).Footnote 8 It accounts for regional heterogeneity in terms of availability of cooperating partners and structural, institutional and social aspects, which might affect the propensity for cooperation (for an interesting take on this issue, see D’Este et al. 2013). Institutional features can provide different grounds for cooperation activities because of the different formal and informal instruments at the regional/local level to stimulate cooperative activities between firms and other institutional actors such as universities and research centres (Robin and Schubert 2013). Also, some regions are more industrialized than others, providing firms with a large ‘reservoir’ of partners to choose for cooperation activities. Moreover, the cognitive distance among cooperating partners may be shorter if partners are located within the same regional borders, making it easier for the same partners to cooperate. Overall, each region has idiosyncratic specificities that can influence the propensity of embedded firms to cooperate. The second is a set of NACE 2-digit dummies to control for sector specificities.
Our working variables are presented in Table 1, with descriptive statistics reported in Table 2. Table 3 presents the correlation matrix, which does not suggest relevant collinearity issues.
Table 1 List of variables
Table 2 Descriptive statistics
Table 3 Correlation matrix
The second part of our empirical analysis addresses the third research question, that is, whether there is a super- or sub-modular effect of the perception of barriers on cooperation activities. This requires a series of complementarity tests (Antonioli et al. 2013; Hottenrott et al. 2014) which allow us to exploit the formalization of complementarities for discrete cases (Mohnen and Röller 2005) examining the presence of complementarity (super-modularity) or substitutability (sub-modularity) among barriers that do not emerge directly from the analysis of simple interaction terms. In order to implement the tests we consider the ‘cooperation function’ of firm j (COOPj) as the firm’s objective function and focus on two types of barriers at a time that might affect the firm’s cooperation function, \( b^{{\prime }} \) and \( b^{\prime \prime } \):Footnote 9
$$ COOP_{j} = COOP_{j} (b^{\prime } ,b^{\prime \prime } ,\theta_{j} )\forall j. $$
Each firm j faces a combination of the two barriers, \( (b^{\prime } ,b^{\prime \prime } ) \) and a set of endogenous and exogenous controls θj, including the remaining barrier.
Complementarity between the two different barriers can be analysed by testing whether \( COOP_{j} (b^{\prime } ,b^{\prime \prime } ,\theta_{j} ) \) is super-modular in \( b^{{\prime }} \) and \( b^{\prime \prime } \). Our aim is to derive a set of inequalities to be tested in the empirical analysis.
Each firm might be in one of the four following states of the world: facing both b
’ and b
’’; neither of the two; or one, but not the other one; and vice versa. This leads to four consequent elements in the set B (forming a lattice):
$$ B = \left\{ {\{ 00\} ,\{ 01\} ,\{ 10\} ,\{ 11\} } \right\}. $$
It is possible to demonstrate that \( b^{{\prime }} \) and \( b^{\prime \prime } \) are complements and, hence, COOP
j
is super-modular if and only if:
$$ COOP_{j} (11,\theta_{j} ) + COOP_{j} (00,\theta_{j} ) \ge COOP_{j} (10,\theta_{j} ) + COOP_{j} (01,\theta_{j} ), $$
Or
$$ COOP_{j} (11,\theta_{j} ) - COOP_{j} (00,\theta_{j} ) \ge \left\lfloor {COOP_{j} (10,\theta_{j} ) - COOP_{j} (00,\theta_{j} )} \right\rfloor + \left\lfloor {COOP_{j} (01,\theta_{j} ) - COOP_{j} (00,\theta_{j} )} \right\rfloor $$
This second inequality clearly shows the interpretation of the super-modularity, or complementarity between two barriers. If the inequality holds, it means that the gain in the propensity to cooperate (increase in the probability to cooperate) that the firm achieves by moving from a state of the world characterized by the absence of relevant barriers (0,0) to a state of the world in which both barriers are perceived as relevant (1,1), is higher than the sum of the gains in the propensity to cooperate (increases in the probability to cooperate) obtained by moving from a state of the world (0,0) to those in which only one barrier is perceived as relevant (1,0) and (0,1).
In order to test for complementarities or substitution effects we operationalize the methodological framework in two steps.
In the first step we set up the ‘Cooperation function’, which can be specified as follows, using two types of barriers, e.g. COST and MKT,Footnote 10 to define the states of the world, while controlling for both the third barrier (e.g., KNOW) and the set of control variables defined above:
$$ \begin{aligned} \left[ {\text{COOP}} \right]_{\text{i}} &= {\text{b}}_{{0{\text{i}}}} \left[ {\text{Controls}} \right] + {\mathbf{a}}{\text{KNOW}} \\ &\quad + {\mathbf{b}}_{{1{\text{i}}}} \left[ {{\text{COST}}\left( 1 \right)/{\text{MKT}}\left( 1 \right)} \right] \\ &\quad + {\mathbf{b}}_{{2{\text{i}}}} \left[ {{\text{COST}}\left( 1 \right)/{\text{MKT}}\left( 0 \right)} \right] \\ &\quad + {\mathbf{b}}_{{3{\text{i}}}} \left[ {{\text{COST}}\left( 0 \right)/{\text{MKT}}\left( 1 \right)} \right] \\ &\quad + {\mathbf{b}}_{{4{\text{i}}}} \left[ {{\text{COST}}\left( 0 \right)/{\text{MKT}}\left( 0 \right)} \right] + {\text{u}}_{\text{i}} \\ \end{aligned} $$
Both the cooperation variable (COOP) and the two types of cooperation (COOPORG and COOPFIRM) are dummy variables: therefore, we run a set of probit regressions, excluding the constant term, since we are interested in the marginal effects associated with all the four states of the world b1, b2, b3 and b4. It is important to stress that while we focus on the complementarity between two types of barriers (e.g. COST and MKT), we control for a third type of obstacle (e.g. KNOW). Specifically, the marginal effects associated with the four states of the world used in the complementarity test are computed setting at 0, 1, the mean value and excluding the third barrier. This allows us to infer whether the results of the complementarity between two barriers typologies test hold for the different values of the third type of obstacle.
Having retrieved the marginal effects using the probit estimates, the next step is to implement a set of Wald tests, which allow us to test the following linear restriction on the state-of-the-world-dummies’ marginal effects: b1 + b4 = b2 + b3 where b1 is associated with the (1,1) state of the world; b2 is associated with the (1,0) state of the world; b3 is associated with the (0,1) state of the world and b4 is associated with the (0,0) state of the world.
The Wald tests are distributed as a χ2 with one degree of freedom since we are testing a single linear restriction at a time. Given that we are interested in the following inequalities, b1 + b4 − b2 − b3 ≥ 0; b1 + b4 − b2 − b3 ≤ 0, and since each Wald test has one degree of freedom, we can apply the appropriate procedure for the p value adjustment in testing inequalities.Footnote 11 Moreover, as a further robustness check, we combine the results of the ‘adjusted’ Wald test with the resulting sign of the linear combination of the coefficients.
By looking at the joint sets of results, we can infer whether rejection of the Wald test null hypothesis allows us to identify complementarity or substitutability between barriers: on the one hand, if b1 + b4 − b2 − b3 ≥ 0 and the Wald test leads us to reject the null, we can argue that we are in presence of super-modularity and, hence, of complementary barriers; on the other hand, we infer sub-modularity if b1 + b4 − b2 − b3 ≤ 0 and the Wald test null is also rejected.