Culture and Multiple Firm–Bank Relationships: A Matter of Secrecy and Trust?

These include, among several other firm decisions and outcomes, debt maturity choice (Zheng et al. 2012), cost of debt (Chui et al. 2016), syndicated loans’ terms (Giannetti and Yafeh 2012), and trade credit provisions (El Ghoul and Zheng 2016).

An unpublished conference paper, by Pasiouras et al. (2018), examined the relationship between individual national culture dimensions and multiple firm-bank relationships. The present study is substantially different in the sense that it: (i) uses a different database (SAFE in the present study versus OSIRIS in the past study); (ii) uses a different sample of firms (large firms in the past study versus primarily SMEs in the present study); (iii) coves different countries (mostly developing countries in the past study versus Eurozone countries in the present study); (iv) examines a different firm-bank relationship (bank advisors in the past study versus bank lenders in the present study); and (v) it develops a very different conceptual model, leading to the empirical investigation of a different research question (culture of secrecy in the present study versus individual dimensions of national culture in the past study).

In fact, Guiso et al. (2006) highlight that culture entered the economic discourse through the concept of trust. However, they also argue that trust is not just an inherited cultural variable. Therefore, for the purposes of the present study, we distinguish between national culture—in particular secrecy—and trust. It should be mentioned here that the literature suggests that culture might also influence corruption (Jha and Panda 2017), and most importantly corruption in bank lending (Zheng et al. 2013; El Ghoul et al. 2016), that could somehow relate to our work. This is because corruption in bank lending will have implications for the arrangements between the loan officer and the loan applicant (Zheng et al. 2013), and subsequently for the firm’s access to bank credit (El Ghoul et al. 2016).

As discussed in Kysucky and Norden (2016), relationship lending may create benefits for borrowers by reducing information asymmetries. In a meta-analysis of 101 studies they find that long-lasting, exclusive, and synergy-creating bank relationships are associated with higher credit volume and lower loan rates.

Campbell (1979) also points out the role of secrecy by arguing that When the manager identifies projects which are expected to yield supernormal risk-adjusted profits, his assessment of the project is conditional upon his strategy remaining secret. He is preluded, therefore, from openly revealing his information to the market (p. 915). While the discussion of Campbell (1979) is centered around the decision to finance a project from private versus public sources, we outline in the main text how secrecy also relates to decisions about the establishment of single versus multiple banking relationships.

Although a few studies examine the role of trust in banking relationships, the association between trust and the number of banks has received limited attention in the literature. Harhoff and Korting (1998) consider the role of trust in a study that examines the costs and collateral requirements in bank lending extended to German SMEs. Hernandez-Canovas and Martinez-Solano (2010) examine the impact of trust on having access to financing and on borrowing costs in Spain. Howorth and Moro (2012) investigate the impact of trusting relationships on interest rates in Italy. In another study that uses Italian data, Moro and Fink (2013) examine whether the perceptions of the loan manager for the trustworthiness of an SME’s management can enhance its access to credit.

The importance of trust for information sharing has been documented in various studies in the context of other economic transactions and relationships such as those of a supply chain. For example, Özer et al. (2011), (i) determine the conditions under which trust is crucial in forecast information sharing between a manufacturer and a supplier; (ii) how trust is affected by changes in the supply chain environment; and (iii) how trust affects related operational decisions. Additionally, Özer et al. (2014) examine the differences in trust and trustworthiness between China and the United States in the context of strategic information sharing behavior in a cross-country supply chain. Finally, Beer et al. (2018) show that buyers benefit from working with trustworthy suppliers.

This survey was conducted between 12 March and 18 April 2018, and it was made available on 4 June 2018. The reference period is the previous 6 months—i.e., October 2017 to March 2018. As outlined in the methodological information that accompanies the survey (European Central Bank 2018): (i) the SAFE is run every six months on a given set of questions and in a limited number of euro area countries (i.e., the ECB Round). In each of the ECB rounds, the smallest countries in the euro area (currently Estonia, Cyprus, Latvia, Lithuania, Luxembourg, Malta and Slovenia) were excluded from the sample by the ECB, because they represent less than 3% of the total number of employees in the euro area; (ii) the interviews in the SAFE are predominantly conducted by telephone (using computer-assisted telephone interviewing or CATI); (iii) the interviewees in each company are considered to be a top-level executives (general manager, financial director or chief accountant); (iv) the companies in the sample are selected randomly from the Dun & Bradstreet business register. In some countries, where the Dun & Bradstreet register was not sufficient or not available, other (mainly local) sources have also been used in the past; and (v) the sample is stratified by country, enterprise size class and economic activity. Furthermore, as outlined in the sample questionnaire that is made available at the website of the European Central Bank, the questionnaire consists of sections that include questions about the: (i) general characteristics of the enterprise; (ii) general information on the type and situation of the enterprise; (iii) financing of the enterprise; (iv) availability of finance and market conditions. Finally, it should be noted here that there also exists a more comprehensive survey in terms of country coverage (i.e., all EU countries plus some neighboring countries) performed by the ECB in cooperation with the European Commission (i.e., the Common Round). This is conducted once a year; however, information about the number of banks per firm, that is our dependent variable, was not available in either the 17th wave (2017H1) or the 19th wave (2018H1) that were Common Round surveys. Hence, the present study is restricted to 12 euro area countries that form the universe of countries included in SAFE 2017H2.

The European Social Survey (ESS) is an academically driven cross-national survey conducted across Europe. The ESS was established in 2001 at the National Centre for Social Research (now NatCen Social Research) while as of 2003 the ESS Headquarters are hosted by the City, University of London (UK). On 30 November 2013, the ESS was awarded European Research Infrastructure Consortium (ERIC). Every two years, face-to-face interviews are conducted with newly selected, cross-sectional samples to measure the attitudes, beliefs and behavior patterns of diverse populations in European countries. The survey involves strict random probability sampling, a minimum target response rate of 70% and rigorous translation protocols. The hour-long face-to-face interview includes questions on a variety of core topics like, among others: (i) media and social trust; (ii) politics; (iii) subjective well-being; and (iv) human values. Country-level aggregated data like the ones that we use in the present study are available online at: https://​nesstar.​ess.​nsd.​uib.​no/​webview.

Hofstede Insights was created in 2017 from a merger between itim International and The Hofstede Centre. The data are available at: https://​www.​hofstede-insights.​com/​country-comparison

The firms that we lose due to missing information for each variable are as follows: 31 (FOREIGN), 9 (AGE), 44 (INTEXP_CH), 121 (PROFIT_CH), 159 (DEBTAS_CH), 71 (more than one variable). These variables are discussed in section “Firm-level Control Variables”.

The composition of the sample used in the regressions for this baseline specification is as follows: Austria (561), Belgium (548), Finland (420), France (1075), Germany (1009), Greece (534), Ireland (384), Italy (1117), Netherlands (620), Portugal (594), Slovakia (323), and Spain (1073).

Although not referring to corporate disclosures, an example from Hofstede et al. (2010) might be of some relevance here as it relates to secrecy and the disclosure of information. They mention that in large-power-distance cultures scandals involving persons in power are expected, and so is the fact that these scandals will be covered (p. 77), as well as that people read relatively few newspapers (p. 77).

For example, in the case of France, the responses are as follows: 0 (5.60%), 1 (3.50%), 2 (7.50%), 3 (12.40%), 4 (12.90%, 5 (27.30%), 6 (11.40%), 7 (10.90%), 8 (6.60%), 9 (1.20%), 10 (0.70%). Thus, the trust index for France is calculated as (0 × 0.56%) + (1 × 3.50%) + (2 × 7.50%) + (3 × 12.40%) + (4 × 12.90%) + (5 × 27.30%) + (6 × 11.40%) + (7 × 10.90%) + (8 × 6.60%) + (9 × 1.20%) + (10 × 0.70%) = 4.59.

Our approach is consistent with the one used in some other studies, such as Hernandez-Canovas and Koeter-Kant (2010). In further analysis, we replace these indicators with dummy variables. The main results remain the same.

BANK_CONC corresponds to the assets of three largest commercial banks as a share of total commercial banking assets, whereas BANK_LIQUID is captured by the ratio of liquid assets to deposits and short-term funding.

Cameron and Trivedi (2013), mention that an indication of the magnitude of overdispersion or underdispersion can be obtained simply by comparing the sample mean and variance of the dependent count variable. As a rule of thumb, they suggest that if the sample variance is more than twice the sample mean, then data are likely to remain overdispersed after the inclusion of regressors. They also highlight that this is particularly so for cross-section data (like the ones of this study), for which regressors usually explain less than half the variation. With an average 2.821 bank relationships per firm and a variance of 7.860, the variance-mean ratio in our sample equals 2.786. Thus, the variance-mean ratio exceeds 2, and while the overdispersion may not be severe, we opt for a negative binomial model that does not impose equidispersion. In further analysis, we use Poisson regression as a robustness test. We also employ a Poisson model with endogenous covariates to account for potential endogeneity. In all the cases, the main results hold.

There are also 111 firms (out of the 8258 firms in the baseline specification with SECRECY, and the 7401 ones in the case of TRUST) reporting between 11 and 90 bank relationships. Most of the banks falling in this group have between 11 and 15 banks (75 firms), another 26 firms have between 16 and 25 banks, and the remaining 9 firms have between 30 and 90 banks. The 75th percentile is equal to 3 banks. In unreported regressions we re-estimate the baseline specification while excluding firms that report: (i) zero bank relationships; (ii) more than 3 banks (i.e., 4 to 90); and both (iii) zero banks and more than 3 banks. In all the cases the results hold. To conserve space, the estimations are available from the authors upon request. As we discuss in the text, in further analysis we also estimated our specification with a quantile regression for counts using the approach of Machado and Santos-Silva (2005).

The coefficients of the control variables that are not shown in the tables are available from the authors upon request.

When we include all three indicators in the regression for TRUST (column 4), we observe a switch in the sign of CONT_ENFOR and INSOLV_REGS. One potential reason is the high correlation between TRUST and INSOLV_REGS (0.68) combined with the moderate correlations between the three indicators as well as with the other control variables.

In unreported regressions, we re-estimate the specification in Column (3) of Table 5, while replacing FINMARK_DEV by the individual components of (i) venture capital financing, and (ii) financing through local equity and bond markets. They both enter the regressions with a negative and statistically significant coefficient, confirming the hypothesis of substitution between banking and alternative means of financing. The main results thus hold again.

To conserve space, we do not tabulate all these regressions. However, they are available from the authors upon request.

We re-estimate the specification in Column (1) of Table 6 using the actual percentage of the loans held with the main bank instead of the categorical indicator LOAN_MAIN for a somewhat reduced sample for which this information is available (N = 7291 in the case of SECRECY and N = 6536 in the case of TRUST). The main results remain the same.

Our main results hold when we re-estimate the specification in Column (3) of Table 6 with the actual percentage of firm exports instead of the categorical indicator EXPORTS, for a somewhat reduced sample for which this information is available (N = 7732 in the case of SECRECY and N = 6908 in the case of TRUST).

SAFE classifies the firms into the four major economic activities (industry, construction, trade and other services) based on the one-digit level of the European NACE classification (according to Rev. 2). The procedure, described in the methodological information of the Survey (European Central Bank, 2018), is as follows. Enterprises from mining and quarrying (B); manufacturing (C); and electricity, gas, steam, and air conditioning supply (D); and water supply, sewerage, waste management, and remediation activities (E); are combined under “Industry”. “Construction” refers to construction (F). “Trade” includes wholesale and retail trade; repair of motor vehicles, motorcycles and personal and household goods (G). “Services” includes: enterprises in transport and storage (H); accommodation and food service activities (I); information and communication (J); real estate activities (L); professional, scientific and technical activities (M); administrative and support service activities (N); arts, entertainment and recreation (R); and other service activities (S). SAFE excludes the following activities from the sample: agriculture, forestry and fishing (A); financial and insurance activities (K); public administration and defence, compulsory social security (O); education (P); human health and social work activities (Q); activities of households as employers; undifferentiated goods- and services-producing activities of households for own use (T), activities of extra-territorial organizations and bodies (U); holding companies (NACE 64.20) and private non-profit institutions.

These results are not reported but they are available from the authors upon request.

We do not include all the variables because the number of observations falls by nearly 40% (compared to the baseline specification), and at the same time, the model eithers fails to converge in the case of SECRECY or several variables are dropped due to collinearity in the case of TRUST (including TRUST). Thus, in addition to the firm-specific variables and the macroeconomic conditions, we include the variables for financial development and getting credit, which have been used in most past studies. Additionally, the financial development indicator that we use is a composite index that captures several dimensions of the market, including perceptions on bank soundness.

Given the nature of the variables described in the sections “Indicators of Secrecy and Trust”, “Firm-level control variables”, and “Country-level control variables”, outliers should not be (in general) an issue. The only firm-level variable that is actually continuous is the firm’s number of years of relationship with the main bank (LENGTH_RELATION), and many of the country-level variables are indices that fall in a given range. Using a graph box approach in Stata, in general confirms this expectation and we detect only one potential outlier in the case of GDP growth and inflation and some outliers in the case of LENGTH_RELATION. Nonetheless, to lessen any potential concerns, we cap (winsorize) LENGTH_RELATION and all the country-level variables (including SECRECY and TRUST) at the 5th and 95th percentiles and we re-estimate the baseline specification of Table 3 and the extended one of Table 7. The results remain the same and we do not tabulate them to conserve space. They are available from the authors upon request.

We acknowledge at this point, that the suitability of the instruments is a concern in many cases, and our study is no exception to this. Therefore, we do not claim that these are the best or only possible instruments. However, such instruments are grounded in the empirical literature on culture and trust, and they are also theoretically justified as they could explain a culture of secrecy and trust. Furthermore, when using the Hansen’s J test for overidentifying restrictions, we fail to reject the null hypothesis that the instruments are valid ones, and the F-statistic of the first stage regression appears to be very high. Consequently, even though one can never completely rule out endogeneity, we believe that these results mitigate concerns about omitted variable bias driving our findings.

A detailed discussion of the relationship between religion and the various dimensions of culture is out of the scope of our study. However, we provide some examples in the text. More detailed discussions are available in Nash and Patel (2019), Hofstede et al. (2010), and House et al. (2004).

Ashraf and Galor (2013), calculate this measure by applying the coefficients obtained from the regression of expected heterozygosity on migratory distance at the ethnic group level, using the worldwide sample of 53 ethnic groups from the HGDP-CEPH Human Genome Diversity Cell Line Panel.

Ashraf and Galor (2013) and Arbatli et al. (2020), document a negative relationship. Additionally, Arbatli et al. (2020) mention that interpersonal population diversity, as determined predominantly during the exodus of humans from Africa tens of thousands of years ago, has contributed significantly to the emergence, prevalence, recurrence, and severity of intrasocietal conflicts. They discuss several mechanisms that explain their findings, along with various advantages of the genetic diversity indicator compared to other diversity metrics like ethnolinguistic fractionalization and polarization. However, it should be mentioned that many others outline that the relationship between ethnic fractionalization and trust is not necessarily negative (Dinesen and Sønderskov 2018). For example, Uslaner (2010) argues that integrated and diverse neighborhoods will lead to higher levels of trust, if people also have diverse social networks. Similarly, in the case of genetics, there could a positive relationship between genetic heterogeneity and trust. For example, under the parasite stress theory of values and sociality, genetic diversity might have positive implications for trust. In more detail, Thornhill and Fincher (2014) mention in their book that: low parasite stress evokes a value system of individualism that includes anti-authoritarianism, and tolerance, validity and trust of out-groups; a willingness to interact with, support, and empathize with different others” (p. 269). Empirical evidence also suggests that lower historical disease burden and infectious diseases prevalence increase trust and social capital (Le 2013; Varnum 2014).

The model is also known as an exponential conditional mean model with endogenous regressors. We rely on a two-step GMM estimator. In this case, the GMM obtains parameter estimates based on the initial weight matrix, computes a new weight matrix based on those estimates, and then re-estimates the parameters based on that weight matrix. The parameter estimates produced by GMM estimators make the sample-moment conditions as true as possible given the data. We estimate the model under the assumption of additive error terms. As before, to allow for heteroskedasticity, we use robust standard errors. Further information on the GMM estimators implemented in IV Poisson is available in Mullahy (1997), Cameron and Trivedi (2013), and Windmeijer and Santos-Silva (1997).

The results for the baseline specification lead to the same conclusions as the ones of the extended specification, and they are available from the authors upon request.

Hofstede et al. (2010) point out that in societies characterized as masculine, the emotional roles of the two genders are clearly distinct. Men are supposed to be assertive, tough, and focused on material success, whereas women are supposed to be more modest, tender, and concerned with the quality of life. In contrast, in societies that are characterized as feminine, the emotional gender roles overlap.

The six countries included in these regressions are: France, Germany, Ireland, Italy, Netherlands, and Spain. The baseline specification has 5278 observations, while the extended specification has 4054 observations.

The fieldwork for the 2017 Edelman Trust Barometer was conducted between October 13 and November 16, 2016, taking into account about 1000 adult respondents per country.

For example, in the case of France, the responses are as follows: Distrust a lot (15.1%), Distrust (30%), Neither trust nor distrust (21.8%), Trust (29.2%), Trust a lot (4%) Thus, the trust index for financial institutions for France is calculated as (0 × 15.1%) + (1 × 30%) + (2 × 21.8%) + (3 × 29.2%) + (4 × 4%) = 1.77.

The negative binomial model does not converge when we estimate the extended specification with the industry dummy variables. When we drop the industry dummy ACT_TRADE, the model converges, and we confirm the results for TRUST_FINAN. Thus, in this case, the results in column (6) of Table 10 are without the industry dummy ACT_TRADE. Nonetheless, the model converges, and we obtain the same results when we use the alternative estimators (e.g., OLS, Poisson, Logit, Tobit, Quantile regression for count) to estimate the extended specification with all three industry dummies. These estimations are available upon request.

For example, the correlation of FINMARK_DEV with TRUST is 0.72 (p = 0.00) and the one with SECRECY is − 0.48 (p = 0.00). TRUST is also moderately correlated with GET_CREDIT (r = 0.43, p = 0.00). Finally, there are correlations between the control variables like for example FINMARK_DEV and GET_CREDIT (r = 0.51, p = 00), FINMARK_DEV and INFLAT (r = 0.49, p = 0.00), and GDPCAP_GR and INFLAT (r = − 0.51, p = 0.00).