Innovation financing and the role of relationship lending for SMEs

Suboptimal information transmission can arise from the trade-off in transferring the scientific and technological content of innovative projects. On the one hand, better signals reduce informational asymmetries and lower firms’ costs of funding. On the other, in line with neo-schumpeterian models of creative destruction, a full disclosure of the project may increase the likelihood of being replaced on the monopolistic market generated by the innovation, reducing its flow of future expected returns.

Dynamic firms face sizable sunk costs linked to skilled workers, researchers, engineers and scientists who cannot be fired and hired without a consistent loss in human capital and accumulated knowledge.

Coherent results are also found by Atanassov et al. (2007) who show the advantage of equity financing, relative to bank debt financing, in developing innovations for large and listed U.S. firms.

In other words, \(\mu _{it} \hbox{RL}_{it}\) is the sum of the baseline effect and its interaction with firm size: \(\mu _{it} \hbox{RL}_{it}=\mu _{1} \hbox{RL}_{it}+\mu _{2} (\hbox{RL}_{it}\times \hbox{size}_{it-1})\rightarrow \mu _{it}=\mu _{1}+(\mu _{2}\times \hbox{size}_{it-1})\).

Indeed, the model requires prior parameter restrictions (typically \(\theta \varphi =0\)) to be logically consistent. Imposing \(\varphi =0\) simplifies the system into a recursive bivariate model that does not leave room for any feedback effect of innovation on the financial status.

In particular, the structural propensity is computed as the predicted probability of the following probit model:where \({1\!\!1}R \& D\) is a dummy variable identifying R&D projects, \(\lambda _{t}\) are time controls and \(\varOmega _i\) is a vector of fixed effects for firms’ belonging strata (identified by 4 size classes, 110 geographical provinces, and 12 industries). Firms are grouped into four size classes (depending on 25th, median, and 75th percentile of the size distribution), 110 provinces, and 12 industries. Their interaction provides a set of \(4\times 110\times 12=5,280\) dummy variables associated with the different strata in the sample.

The fitting property of this first-stage regression (pseudo-\(R^2=0.16\)) ensures a good power of the instrumenting set. Moreover, since the determinants of \(\widehat{{1\!\!1}R \& D}_{i,t-1}\) are controlled for in the FC equation, \(\varphi _{\widehat{{1\!\!1}R \& D}_{i,t-1}}\) should be correctly interpreted as the direction of the feedback effect of firms’ dynamism on the financial status.

R&D and innovation share common characteristics and the leading causes of financing constraints (Hall 2002), allowing to catch this additional channel.

Although models that fully exploit the panel structure of the data have the great advantage to control for firm-specific idiosyncratic components, they require variation across time of the binary dependent variable. Given the high persistence and state dependence of both the innovation propensity and the FC status, all these models produce an excessive reduction in the sample and lead to a selection bias due to the empirical approach itself. Explanations about the persistence of innovation are mainly based on the cumulative nature of learning processes (Rosenberg 1976), “success-breeds-success” (according to which succeeding in innovation increases generated cash flows that may be devoted to finance further innovations, Stoneman and David 1986) and on strategies of innovation smoothing. To address concerns of unobserved heterogeneity, I also include regressors’ mean in both equations. Results are qualitatively similar to the main specification.

Organizational–managerial innovations are defined as “the implementation of new organizational or managerial methods in the firms’ business practices, workplace organization or external relationships”.

The original question asks to quantify firms’ difficulties in accessing external credit on a scale from one to ten. The definition employed considers as “grave difficulties” values (strictly) greater than seven. Results are however robust to the choice of different thresholds.

“Hold up” is related to banks’ ability to extract rents from the firm, resulting into inefficient investment choices. “Soft-budget-constraint” problems concern banks’ incentive to refinance some of the ex-post inefficient projects.

The price to pay to include measures of soft information into the FC equation is a relevant reduction in the sample (roughly 9,900 firms). Since the question on the lender identification has been introduced only in the last wave of the survey, I restrict the analysis to firms interviewed in 2011, exploiting also observations in the previous waves when repeated interviews were available (this is done to avoid an excessive sample reduction).

Notice that this approach requires the assumption of stability of the firm-bank relationships over time. Several considerations support this hypothesis. First of all, in a system dominated by SMEs, firms do not usually have the reputation needed to get credit from a new financial institution and they have to rely on prolonged relationships (Diamond 1991). This issue is even more relevant in times of crisis characterized by increased opaqueness of less structured companies. Moreover, evidence from Italy indicates that firms attempted to broaden the range of financial sources rather than substitute one bank with another (D’Auria et al. 1999). Given the low diffusion of multiple banks in the sample (15 %), the potential issue of “bank-switcher firms” should be negligible.

The survey does not contain information about the length of the relationship or the share of the main bank on total banking debt, typically considered as alternative measures of relationship lending.

R&D expenditure (as a share of total sales) and a dummy variable are also used as alternative measures.

The MET survey also contains information on informal connections. Firms are classified into “stand alone” and belonging to networks, depending on the existence of prolonged and relevant relationships with other firms. The complexity of the inter-firm connection (in parentheses) then discriminates between “simple” and “advanced” networks.

Since Credit score\(_{it-1}\) is a generated regressor, standard errors in the estimations throughout the paper are obtained with bootstrapping techniques.

\(\hbox {Size}_{it-1}\) and \(\hbox {Age}_{it-1}\) are defined as the log of (one plus) firm number of employees and age, while \(\alpha _t\), \(\alpha _{\mathrm ind}\), and \(\alpha _{\rm{reg}}\) are, respectively, time controls, 2-digit industry controls (12 dummies), and region controls (20 dummies).

About 40 % of the sample in each wave.

Within the industrial classification available, hightech sectors are as follows: chemicals, plastic, means of transportation, engineering, electric, and electronic equipments.

The significance of the estimated correlation coefficient (\(\hat{\rho }\)) provides a further validation of the use of simultaneous estimations, showing the relevance of neglected third-party factors. In this regard, Lollivier (2001) showed that restricting the residuals’ correlation to zero (which is imposed with two distinct probit models) yields to endogeneity problems and biased and inconsistent estimates.

Indeed, neglecting micro firms may lead to underestimate the effects of FC on innovation. This is particularly true for the introduction of new products, when the increased riskiness of the project adds up to the low capability of self-financing, the little tangible assets to pledge as collateral, and the higher opaqueness that characterize very small companies.

Although the magnitude of the coefficient is slightly higher for organizational–managerial innovations, an F test on the difference of the coefficients does not detect any significant difference with respect to the introduction of new products.

A dummy variable identifying firms with “grave difficulties in accessing the credit market”.

One opened issue is whether results are driven by a subset of unsound firms with low-quality projects and reduced financial means. Given the absence of a control group, I exploit firms’ unconditional innovation likelihood to identify the subset of (more) innovative companies. The sample is then split into “Innovative” and “Non-innovative” firms depending on the median value (or 75th percentile) of the predicted probability to innovate from the following probit model: \(Pr(\hbox {Inn}_{it}=1)=\phi \left( \beta _1^\top X_{1,it-1}\right)\). Results hold for both subsets, although they are stronger for more innovative firms.

Although lagged covariates clear problems of simultaneity bias in the main equation, the persistence of relevant attitudes (such as innovation, export, and R&D) may leave residual endogeneity into the estimation. To control for this possibility, the baseline model is augmented with lagged values of innovation. Controlling for previous realizations allows to purge all the persistent behaviors already embedded in Innovation \(_{t-1}\) and to focus on the “pure” effect of each regressor. Although innovation presents a high persistence, the other coefficients are very stable, suggesting that the main results are not affected by endogeneity problems generated by “sticky behaviors.”

This approach is justified by the non-feasibility of id-fixed effects and provides further support to the validity of the baseline regression.

This evidence is consistent with the higher informational asymmetries of high-return projects.

This discrepancy may be due to the very nature of soft forms of innovation, which embed a great variety of improvements and are often adopted by very small companies. An alternative explanation can be found in the higher expected payoff of product and process innovations relative to organizational–managerial improvements.

It is worth reminding that marginal effects are computed at regressors’ mean values. This implies that values in Tables 10 and 12 (ranging from a 1-to-2 %), actually induce substantial effects for micro firms borrowing from distant and big banks (from 3-to-8 % increase in firms’ probability of FC).

Results are qualitatively analogous although not as robust as the main specification.

This aims at controlling for the possible self-selection of banks toward provinces with a higher share of dynamic firms with low financial problems.

The purpose of this robustness check is to control for the possible spurious correlation between the size of the lender bank with other relevant characteristics affecting its lending practices.

The estimation of the credit score is based on a confidential dataset provided by Fiditoscana (a credit-warranty structure operating on market basis and in the allowance of warranties based on public funds), consisting in 3,000 credit ratings assigned by several Italian banks to local firms.

The proportional odds model is based on a multi-equation estimation where coefficients are constrained to be the same across different states of the dependent variable. The high significance of the LR test suggests the violation of this hypothesis and requires switching to a generalized ordered logit model that allows for variations in the beta estimates across states. The advantage of using such a regression (with respect to standard multinomial logit models) is the possibility of imposing constancy for all the covariates that do not violate the proportional odds assumption, having in such a way a more parsimonious model.

This evidence may be explained by the higher informational asymmetries due to the increased opaqueness of SMEs in times of crisis.

This is required by the absence of actual ratings for most of the companies in the original sample.

The accuracy of the model is tested out-of-sample with a bootstrap procedure to avoid standard problems related to over-fitting of in-sample tests.