Who Benefits from Longer Lending Relationships? An Analysis on European SMEs

Abstract

This paper empirically investigates the impact of lending relationships duration on SMEs financial stability. Our research hypothesis is that the balance between benefits and costs of longer bank‐firm ties may be different depending on the degree of firms' financial health. Using a large sample of European manufacturing SMEs that excludes firms that have defaulted and those with less than ten employees, we find that the overall positive effect of enduring lending relationships tends to progressively increase for more stable firms, being greater when the main bank operates nearby the firm. Our findings, yet, are conditional on firm survival and may not be generalized to the smallest of firms.

Notes

8. Indeed, an analogous scenario emerges from the whole literature focusing on the implications of lending relationships on firms' financing and performance. Incidentally, and focusing on the empirical field, we recall that close lending relationships have been found beneficial for firms—especially for those more affected by asymmetric information problems—as close bank‐firm ties: increase credit availability (Angelini, Di Salvo, and Ferri 1998; Berger and Udell 1995; Bharath et al. 2011; Chakraborty and Hu 2006; Cole 1998; Elsas and Krahnen 1998; Harhoff and Korting 1998; Hernandez‐Canovas and Martínez‐Solano 2010; Kano et al. 2011; Petersen and Rajan 1994), reduce loan interest rates (Berger and Udell 1995; Brick and Palia 2007; Harhoff and Korting 1998), provide stronger protection against the interest rate cycle (Ferri and Messori 2000), decrease firms' probability to pledge collateral (Brick and Palia 2007; Harhoff and Korting 1998; Jimenez, Salasa, and Saurina 2006), promote firms' product and process innovations (Benfratello, Schiantarelli, and Sembenelli 2008; Giannetti 2012; Herrera and Minetti 2007), reduce firms' dependence on trade debt (Petersen and Rajan 1994), foster firms' foreign direct investment (De Bonis, Ferri, and Rotondi 2010). Other studies, however, found that close lending relationships are negatively, or unclearly, related to loans interest rates (Blackwell and Winters 1997; D'Auria, Foglia, and Marullo Reedtz 1999; Degryse and Van Cayseele 2000; Hernandez‐Canovas and Martínez‐Solano 2010; Kano et al. 2011; Petersen and Rajan 1994; Stein 2011), may induce banks to avoid financing risky long‐term investment projects, even though profitable (Weinstein and Yafeh 1998), increase collateral requirements (Ono and Uesugi 2009), tend to lower firms' profitability (Montoriol Garriga 2006), and to hamper the growth of small firms (Gambini and Zazzaro 2013).

9. Close bank‐firm ties might also lead to a monopoly power of firms, so that the hold‐up problem may be reciprocal. Indeed, banks devote several specific assets to preserve a lending relationship (human, organizational, physical), which could make costly to end the lending contract, even when firms' prospects appear deteriorated (Longhofer and Santos 2000). Since the relation‐specific resources employed by banks increase with firm's size, the circumstance that firms might capture banks is more likely as the former are larger (Gambini and Zazzaro 2009; Hoshi, Kashyap, and Scharfstein 1990; Peek and Rosengren 2005).

10. We do not need to trim the distribution of our dependent variable, as we will estimate several quantile regressions separately modeling the relationship under investigation for the tails (as well as for the other segments) of the Z‐score distribution. Therefore, outliers in the response distribution cannot exert undue influence on the majority of the data (see next sub‐section for a depiction of the other quantile regression properties). Nevertheless, when we do eliminate the first and the last percentile of the Z‐score variable, our results are confirmed.

11. There are numerous measures of financial health at an individual firm level. Broadly speaking, it is possible to distinguish two categories of indicators, one based on accounting data, the other on market data. Since our analysis is not confined to quoted firms, we have to discard the second category. Among the indicators left, our choice has been determined by the availability of data. Indeed, since we lack information on firm bankruptcy, we cannot employ logit or probit models to predict firm failure. For the same reason, we cannot adopt multiple discriminant analysis to retrieve the most appropriate Z‐score model à la Altman (Altman 1968; Altman, Hartzell, and Peck 1995; Altman and Hotchkiss 2006). Analogously, the lack of data on some financial ratios (such as retained profits to total assets) prevents us from applying Altman Z‐scores originally developed for different time periods, countries and industries. Incidentally, such an application could imply a significant decline in the accuracy of the analysis. In the robustness checks subsection, we will change our dependent variable making use of the information available.

12. The dummy indicating proximity with the main bank is available only for the last year (2009) of the EFIGE survey. Before imputing the value of PROXIMITY to the years 2001–2008, we verify whether in those years there existed a relationship with the main bank. We deduce the existence of such a relationship from a positive value of duration. Thus, when DURAT is greater than zero we proceed imputing the end of the survey value of PROXIMITY; if DURAT is missing, we consider also PROXIMITY as missing.

13. On the role of bank‐firm physical proximity with respect to loans availability and credit conditions see, also, Petersen and Rajan (2002), Carling and Lundberg (2002), Degryse and Ongena (2005), DeYoung, Glennon, and Nigro (2006).

14. The literature investigating the determinants of firms' insolvency (or failure) suggests the relevance of a rich set of firm level characteristics along with industry and macroeconomic factors, as well as institutional and cultural drivers (see, for instance, Bhattacharjee and Han 2000; Bhattacharjee et al. 2009; Bottazzi et al. 2011; Mihet 2012; Mishra 2013). Since several of these factors may be highly correlated, to select our control variables we proceed as follows. First, we consider all variables suggested by previous studies for which data are available. To give an example, the firm age is often considered among the conditioning variables of our dependent variable. However, the EFIGE database provides information on the age bracket to which each firm belongs only for the last year of the survey. Thus, we cannot resort to an imputation process, as we did for DURAT. Second, we discard those explanatory variable that are highly correlated with at least one other variable (for instance we do not employ a standard measure of trade openness, as it is found highly correlated with VAR_RGDP). Finally, as robustness checks, we re‐estimate model 1 by substituting some regressors with correlated indicators. To exemplify, the GDP growth regressor is replaced with the rate of unemployment in column 2 of Table 3 and with the (log of the) GDP in column 3 of the same table (see the “Robustness” section for further details).

15. As a matter of fact, while OLS estimates are retrieved from the minimization of the sum of squared residuals (difference between an observed value and its fitted value), quantile coefficients are obtained by minimizing a weighted sum of absolute residuals, with weights depending on whether the observation lies above or below the fitted line (for a detailed illustration of the quantile regression approach, we refer to Koenker 2001).

16. The bootstrap method (Efron 1979) does not make assumptions on the distribution of the response variable and consists in extracting—with replacements—a large number of samples (of size n) from the observed sample. These resamples will be randomly different from the original sample and will be used to get parameter estimates, as well as variance and covariance estimates. In particular, the covariance matrix will also provide the covariance of the estimated coefficients of the same regressor across distinct quantiles, thus one can verify the equivalence of the marginal impact of a covariate at different quantiles.

17. The absolute value of the DURAT parameter slightly increases. Nevertheless, such a differential impact appears limited (if compared to that registered when the Z‐score indicator is employed as dependent variable), and not statistically significant when we test the equivalence of the coefficient of DURAT across the quantiles.

18. On the supply side, the main bank may decide to either continue or interrupt lending when the customer risk‐taking increases. On the demand side, rather than establishing information‐intensive relationships with a main bank, insolvent firms might decide to resort to transaction‐oriented lending, trying to dissimulate their risk profile since banks accessing less soft information are less able to distinguish good from bad borrowers (Bolton et al. 2013).

19. To verify the significance of the above mentioned sum for each quantile, we compute the following standard error: $$\text{error}:\widehat{\sigma }\sqrt{\mathrm{var}\left({\widehat{\beta }}_{\text{DURAT}}\right)+\mathrm{var}\left({\widehat{\beta }}_{\text{INTE}}\right)+2\text{COV}\left({\widehat{\beta }}_{\text{DURAT}},{\widehat{\beta }}_{\text{INTE}}\right).}$$ .

Additional information

Notes on contributors

Mariarosaria Agostino

Mariarosaria Agostino is Senior Lecturer of Economics in the Department of Economics, Statistics and Finance at the University of Calabria.

Francesco Trivieri

Francesco Trivieri is Associate Professor of Economics in the Department of Economics, Statistics and Finance at the University of Calabria.