Long-lasting bank relationships and growth of firms

In a similar vein, Houston and James (2001) study the investment behavior of bank-dependent firms and conclude that, in the USA, “banks are unwilling to finance relatively large capital expenditures and thus bank-dependent firms must rely more on internally generated funds for these types of expenditures” (p. 349); see also Yafeh and Yosha (2003), who show that, in Japan, bank-dependent firms invest less in research and development (R&D) than independent firms.

See Becchetti and Trovato (2002), Carpenter and Petersen (2002), and Fagiolo and Luzzi (2006).

See Degryse and Ongena (2001), Fok et al. (2004), Herrera and Minetti (2007), Montoriol Garriga (2006), and Alessandrini et al. (2009, 2010).

Evidence that longer relationships with a lender increase credit availability for small firms is provided by Angelini et al. (1998) and Harhoff and Körting (1998) for European countries, and by Petersen and Rajan (1994), Berger and Udell (1995), Cole (1998), Chakraborty and Hu (2006), and Bharath et al. (2007, 2011) for the USA.

This expression is due to Williamson (1976, 1985).

Consistent with the hypothesis that relational banks capture small firms, many studies concerning countries other than the USA have found that interest rates and collateral on small business loans increase with the length of the lending relationship and that the usage rates of credit lines decrease for older customers (Angelini et al. 1998; Degryse and Van Cayseele 2000; Hernandez and Martinez 2006; Ono and Uesugi 2005; Barone et al. 2010; Ogawa et al. 2007; Grunert and Norden 2011; Jiménez et al. 2009; Ioannidou and Ongena 2010). In addition, the profitability of small firms tends to be lower when they maintain exclusive relationships with few banks (Montoriol Garriga 2006). By contrast, Degryse and Ongena (2001) and Fok et al. (2004) found that exclusive relationships with banks improve performance of large publicly listed firms.

Sakai et al. (2005), analyzing a large panel of small Japanese firms, found that firms that defaulted during the sample period pay higher interest rates than survivors regardless of their age, thus indicating that banks do not have incentives to support and subsidize small troubled borrowers. Similarly, Brunner and Krahnen (2008) showed that, in Germany, the probability of a pool of banks forming to revitalize distressed small–medium enterprises is greater when the firm debt is evenly distributed among a large number of banks. By contrast, evidence on the capture of relationship lenders by large firms has emerged for the Japanese economy, especially during the crisis of the 1990s. Hoshi et al. (1990) found that firms under financial distress experience a relatively modest contraction in their investments and sales if they belong to industrial groups including their main bank (keiretsu). Kawai et al. (1996) reported that interest rate premia paid by firms at the time of financial distress are significantly lower if they rely on the same largest bank lenders for at least 10 years. Peek and Rosengren (2005) showed that large troubled firms with strong bank ties are more likely to obtain additional loans than other firms.

It is worth noting that the surveyed firms are not necessarily customers of the UniCredit bank (to be exact, the survey does not provide any information on the identities of the lending banks). This survey has been widely used in banking literature; amongst others, see Herrera and Minetti (2007), Benfratello et al. (2008), Alessandrini et al. (2009, 2010), and Presbitero and Zazzaro (2011), to which we refer for further details on the survey.

By restricting the analysis to the period after the introduction of the euro, we limit the distortionary effects on balance-sheet entries of price adjustment that occurred during the currency changeover. However, to check the robustness of our results, we reproduced all regression exercises by considering also the seventh survey for the period 1995–1997. On the whole, estimation results are qualitatively unchanged; those for the employment growth rate (our preferred measure of firm growth) are statistically more significant, while, not surprisingly, those for asset growth are slightly less significant (results available from the authors upon request).

The sample is stratified by size (distinguishing five classes according to number of employees), geographical macro-areas (Center–North and South), and industrial macrosectors (according to the Pavitt classification).

A detailed description of variables used in the analysis and their sources is reported in Table 2.

Following Evans (1987), our regression specification (1) derives from the firm growth relationship \( Size_{it} = \left[ {G_{it} \left( {Size_{it} ,\,Age_{it} ,\,\sum\limits_{j} {X_{jit} ,\,{{Main{\text{-}}bank}}\,\,{{relationship}}\,\,{{length}}_{it} } } \right)\,} \right]^{d} \left( {Size_{it} } \right)\left( {\nu_{it} } \right) \), where d denotes the growth time interval and ν is a lognormal error term. Alternatively, in the empirical literature on corporate growth rates, a logarithmic specification of Gibrat’s law has been used (Geroski 2005). Even though the span of time of our analysis is short, the growth rate specification allows us to overcome possible problems of nonstationarity of size series.

Typically, empirical models for Gibrat’s law also include Size and Age squares and the interaction term between Size and Age. In our sample, however, these terms are generally insignificant and increase standard errors for Size and Age, whereas their inclusion leaves the sign and magnitude of other covariates (in particular, of Main-bank relationship length) virtually unchanged (results available on request from the authors).

Another relevant measure of firm size commonly used in the literature on Gibrat’s law is turnover (amongst others, Audretsch et al. 2004; Fagiolo and Luzzi 2006). However, as Weinzimmer et al. (1998) argue, which is the appropriate measure of firm growth depends upon the aim of the analysis. As we are interested in testing whether long-lasting bank relationships encourage managers to invest resources in the growth of the firm they direct, and since changes in turnover may be determined by many external causes independent of deliberate investments in the firm’s growth (such as strong market demand or failure of competitors), we consider total assets and (especially) employment as the appropriate measures of firm growth for our analysis. However, for robustness, we re-estimated all regressions by measuring growth in terms of turnover, reaching results qualitatively almost indistinguishable from those found for the growth of assets (available upon request).

Statistics on the OECD Employment Protection Law Index (OECD 2004) concerning the difficulty of dismissal indicate that Italy is one of the countries with the strictest labor protection laws among OECD members, unlike Common Law countries, which have the least restrictive legislation.

This definition of small firms has been replaced by Recommendation 2003/361/EC (May 2003), by which the Commission raised the total asset threshold to 10 million euros.

All the exercises presented in the following sections have also been run omitting firms changing in size status during the sample period, and results are robust in sign, magnitude, and significance of coefficient estimates.

In our samples, the correlation coefficients between return on assets and the growth of employment and total assets range between 0.08 and 0.13 and are always statistically significant at the 99% level. In addition, the average rates of employment and total assets growth for the subsample of firms with return on assets above the median (or the mean) are statistically higher with 99% level of significance (similar results hold when we classify the firms’ health status on the basis of the ratio between pretax operating income and interest expenses).

According to the survival selection bias, the negative impact of initial firm size on their growth rate could be due to the fact that large, inefficient firms are more likely to survive than small, inefficient firms, which are therefore not observed and are excluded from the sample.

It is worth noting that violation of Gibrat’s law does not stem from the presence of M&A (Ijiri and Simon 1977).

New branch entrants and New branch incumbents consider the absolute number of new branches opened in a province. For robustness we also considered instruments where new branches are normalized, alternatively, to the population or total branches in the province (Herrera and Minetti 2007; Alessandrini et al. 2009, 2010). Estimated coefficients are substantially unaltered, while the F-statistics on excluded instruments indicate a worsening of their significance.

Boot and Thakor (2000), Dell’Arriccia and Marquez (2004), and Hauswald and Marquez (2006) offer classical arguments for a positive correlation between competition and relationship lending, while Presbitero and Zazzaro (2011) present evidence for Italy.

A similar indication of weak instrument problems for the length of the bank relationship is provided by the study of Herrera and Minetti (2007) on the likelihood of firms adopting innovation. However, as they use nonlinear TSCML regressions for which tests for weak instruments are not available, they do not provide weak-instrument-robust inference.

Inference on coefficients of Main-bank relationship length based on the Anderson–Rubin (AR) and the LM tests provide similar results (see Table 6). The only differences worth noting with regard to the AR tests are: (i) the nonsignificant impact of Main-bank relationship length for asset growth of medium–large enterprises; and (ii) the rejection at the 10.4% level of significance of a null correlation between Main-bank relationship length and the average conditional asset growth rate for the whole sample.

However, results of IV regressions remain practically identical if we compute standard errors robust to heteroskedasticity and abandon weak-instrument-robust inference.

The change of sign for the coefficient of Age in IV estimations is consistent with results found by Herrera and Minetti (2007). For the sake of robustness, as in the OLS specifications 3, 4, 7, 8, 11, 12, IV regressions have also been run adding the six further controls described in Subsect. 5.1.2, both instrumenting and noninstrumenting the Number of Banks variable. Results on our main variable do not change, especially the negative impact of relationship length on the growth of small firms.

Coefficients at the 73rd and 85th percentiles are not significant at the 10% level, but at a slightly higher level, while coefficients at the last two percentiles are very large in magnitude but statistically insignificant.

In this case, no problem of iterative convergence arises and we obtain point estimates for all the percentiles.

The results reported in Table 7 refer to the generalized version of the Hausman test obtained by using seemingly unrelated estimations (i.e., using the “suest” command of Stata).

The 1-year probability of default is drawn from the RiskCalc™ Italy model developed by Moody’s KMV (Dwyer et al. 2004). The RiskCalc™ model for 1-year risk of default combines firm’s financial statement ratios concerning profitability, leverage, debt coverage, growth, liquidity, activity ratios, and size. We thank Toni Riti of UniCredit for kindly providing us with the Default risk variable.

It is worth noting that, when we include Roa, Default risk, and Credit rationing in the least-squares regression, they are not significantly associated either with \( Growth_{1it} \) or with \( Growth_{2it} \).

It is worth noting that the insignificance of Mill’s ratio also corroborates the idea that the occupational-changing and occupational-intensity choices are mutually independent.