«Tobias Schmidt (Deutsche Bundesbank) Wolfgang Sofka (Department of Organisation and Strategy, CentER, CIR, Tilburg University) Discussion Paper ...»
Hence, we add the number of a firm’s other banking contacts of a firm as a control variable to the model. Relationship length between a bank and its client has also been found to alter lending decisions (Herrera & Minetti, 2007). The overwhelming majority of the firms in our sample (86%) never change their main banking relationship during the five- year observation period. Hence, an exact measure of relationship length appears inappropriate. Conversely, we add a dummy variable indicating whether the firm has changed its main bank in the preceding three-year period. We proxy firm’s internal cash flow through the return on sales from the previous year and its overall creditworthiness through its credit score. Creditreform determines the credit score as an index based on a proprietary formula which places heavy penalties on negative events such as delayed payments to suppliers or insolvency. All indices are standardized to ensure comparability. In addition, firms that are traded on the stock exchange have broader access to external financing. We control for this effect by adding a dummy variable for whether the firm is incorporated based on stock market shares.
Other factors may influence firm R&D decisions. Firms can suffer from liabilities of size or newness in their resource acquisition (e.g. Shinkle & Kriauciunas, 2009). We add firm size (number of employees in logs) and firm age (number of years since founding) to the model.
Firms can also be part of a larger group and draw from its resources. Hence, we add a separate dummy variable for such cases. Finally, we capture remaining industry differences by including industry dummy variables. These follow grouped two-digit industry classes as suggested by OECD and Eurostat. The resulting groups are low-tech, medium low-tech, medium high-tech and high-tech manufacturing as well as low knowledge intensive and knowledge intensive services (see Appendix A for full details). Similarly, we add an additional control dummy for a firm’s location in eastern Germany which, a result of reunification, has been found to provide geographical opportunities and challenges different from those in westen Germany as (e.g. Czarnitzki, 2005).
4.2.3 Estimation model A logical inference from our theoretical reasoning is that some firms may not be able to invest in R&D at all, i.e. their R&D investment equals zero. Hence, a technique is required that takes into account that the dependent variable being censored at zero. We estimate censored panel regression models. In particular, we estimate random effects Tobit models.
Fixed effects tobit models are only beginning to emerge and existing approaches have been criticized for delivering inconsistent estimates as well as being overly demanding on assumed data and variation (Cameron & Trivedi, 2005; Grimpe & Kaiser, 2010). The inconsistency stems primarily from the finite nature of empirical samples. Non-linear, fixed-effects models suffer especially from inconsistency issues because estimates are more likely to be influenced by incidental parameters (Heckman, 1987; Neyman & Scott, 1948). Inconsistencies can be assumed to be reduced if the sample encompasses more than eight time periods, but random effects estimators are more commonly used (Cameron & Trivedi, 2005). Given our data availabilities, we opt for a random effects model. We run several model specifications and include the independent variables of interest stepwise.
We inspect the dataset for multicollinearity based on correlations and variance inflation factors and find no evidence by any conventionally applied standard (e.g. Chatterjee & Hadi, 2006). The mean variance inflation factor equals 1.63 with the highest individual variance inflation factor equaling 3.94 (see Appendix B for full details). Table 1 provides descriptive statistics for the sample as well as a comprehensive overview of variable observation levels and scales.
5.1 Main bank specialization and market share Table 2 presents the estimation results of the tobit models testing hypotheses 1 and 2. Model 1 contains only control variables and can serve as a benchmark for all other models.
Significant effects remain stable across models and the quality of model fit increases (log likelihood and Chi squared test).
Insert Table 2 about here
Model 2 includes the main effects of bank specialization and industry market share. A bank’s market share in the focal firm’s industry has a positive and highly significant effect on R&D investment. Hence, hypothesis 1 is supported. The effect of bank specialization is negative and highly significant, lending support to hypothesis 2. We calculate effect sizes based on a one standard derivation difference from the average in main bank industry specialization and market share. The former reduces firm R&D intensity by 14%, the latter increases it by 6%. Hence, the effects are not just significantly different from zero but also have a sizeable impact on firm R&D. This result reinforces the theoretical logic that both portfolio and information externality theory can inform predictions of firm R&D investment through the main bank’s client portfolio. Diversification in firms’ main bank portfolio allows more firm R&D spending while the increasing industry specialization within a bank’s portfolio allow less. We test a simultaneous relationship by adding a multiplicative interaction term between bank specialization and market share in model 3. Interestingly, there is no immediate relationship between the two variables beyond the main effects. As a result, neither mitigating nor reinforcing effects can be found and hypothesis 3 has to be rejected.
In hypothesis 4 we develop a theoretical argument for the effects to differing subject to the uncertainty in industry innovation activities depending on how far removed they are from application. The latter is proxied through the importance of scientific knowledge for innovation in an industry, which has a consistently positive and significant effect on firm R&D investment in Table 2. We test hypothesis 4 by splitting the sample along the mean of the industry share of firms that use science as an input in their innovation activities. Table 3 shows the estimation results for both sub-samples. The negative effect of bank specialization remains significant in both samples. However, the significance level drops strongly in the sub-sample with industries that use less scientific knowledge. Conversely, the information externality effect appears to be confined to industries with an above-average use of scientific knowledge. Hence, hypothesis 4 is partially supported. Additionally, we test the significance of the multiplicative interaction effect between bank specialization and market share suggested in hypothesis 3 for the two sub-samples, too. There is no additional significant finding. Models are not reported.
5.2 Signaling effects
We have developed hypotheses on three kinds of signals firms can send to their main banks:
patents, government R&D subsidies and venture capital investments. The main effects of all of these factors are positive and significant. This is fully in line with existing research emphasizing complementarity effects of R&D with existing knowledge stocks embodied in patents (e.g. Cohen & Levinthal, 1989) as well as additionality effects from government R&D subsidies (e.g. Aerts & Schmidt, 2008), and growth-oriented venture capital investments (e.g.
Levitas & McFadyen, 2009). However, the signaling effect that these factors may have on a firm’s main bank is novel. We add separate multiplicative interaction effects with each factor and bank specialization in models 4, 5 and 6. We use separate models for each interaction to avoid potential issues arising from multicollinearity. Table 4 shows the results for these models.
Insert Table 4 about here
All main effects remain stable. Only the interaction effect between a firm’s patent stock and its main bank’s degree of portfolio specialization is positive and significant (Model 4). There are no additional significant interaction effects for firms having attracted a venture capital investor or having received a government R&D subsidy (Model 5 and 6). All of the signals fail to alter the positive effects from information externalities of a bank’s market share. In sum, hypothesis 5 is rejected. Hypothesis 6 is only supported for a firm’s patent stock. Our result are in line with Levitas & McFadyen (2009) who identify a similar positive patent effect for the venture capital market. In conclusion, the proposed signaling effects are limited to reputation effects based on the success of firms’ past patenting success. Hence, the signaling effect is firm-specific and based on past innovation outcomes. Input-oriented signals originating from successfully attracting government R&D subsidies or venture capital fail to alter the risk assessments or information position of banks. We suspect that this is due to the fact that they are general in nature and can be interpreted positively even without in-depth industry experience of a bank.
5.3 Control variables
All models contain an identical set of control variables. Their influence on firm R&D intensity is consistent across all models with regard to significance levels and directions. We have not developed theoretical predictions for any control variables but significant effects should be discussed briefly. First, it is noteworthy that out of all control variables at the bank level only the type of bank has a significant influence on firm R&D investment. An average firm working with a savings or cooperative main bank invests significantly less in R&D. This result supports other studies (Haselmann et al., 2009) which emphasize the inefficiencies in bank decision making induced by political influence through government ownership (Porta et al., 2002; Sapienza, 2004). R&D intensity increases with firm size but decreases with firm age. Similarly, incorporated firms having access to the stock market invest more in R&D.
This provides evidence for the close relationship between overall resource availability for R&D investments (Ahuja, Lampert & Tandon, 2008) as well as R&D as part of a growth strategy for young firms (King & Levine, 1993). Similarly, the negative relationship between return on sales and R&D investment can be interpreted as an investment in generating the potential for future revenue streams in which the positive performance effects of current R&D are, on average, four to five years removed (Pakes & Schankerman, 1984). The significant industry dummies (low tech manufacturing is the reference group) indicate that the R&D investment is a direct reflection of technological opportunities and competitive pressures.
R&D investment increases with the knowledge intensity of the industry in both manufacturing and services. Low knowledge-intensive service sectors such as transportation are the exception reflecting fewer technological opportunities (e.g. Lyons, Chatman & Joyce, 2007).
6 Consistency and sensitivity checks