Social Capital, Trusting, and Trustworthiness: Evidence from Peer-to-peer Lending

How does social capital affect trust? Evidence from a Chinese peer-to-peer lending platform shows regional social capital affects the trustee’s trustworthiness and the trustor’s trust propensity. Ceteris paribus, borrowers from higher social capital regions receive larger bid from individual lenders, have higher funding success, larger loan size, and lower default rates, especially for low-quality borrowers. Lenders from higher social capital regions take higher risks and have higher default rates, especially for inexperienced lenders. Cross-regional transactions are most (least) likely to be realized between parties from high (low) social capital regions.


I. Introduction
Trust, defined as the willingness that a trustor voluntarily places resource at the disposal of the trustee with expectation of a fair payoff, is fundamental to finance and economic growth. A considerable body of work highlights social capital (SC) stock as an important antecedent of trust (Arrow (1973), Knack and Keefer (1997), Guiso, Sapienza, andZingales (2004) (2008)). However, the channels through which SC affects trust are unclear. Moreover, the link between society-level SC and micro-level economic transactions has a conceptual gap, which is highlighted when trading partners come from different SC environments.
To examine the impact of SC on trust, we draw from the extant trust literature, which distinguishes trustworthiness from generalized trust (Colquitt, Scott, and Lepine (2007)). On the one hand, trustworthiness relates to the objective characteristics (e.g., integrity, competence) of a trustee (Ang, Cheng, and Wu (2015), Hasan, Hoi, Wu andZhang (2017a, 2017b)). Generalized trust, on the other hand, refers to the subjective belief of a trustor on the likelihood that a potential trading partner will act honestly (Hong, Kubik, and Stein (2005), El-Attar and Poschke (2011)).

The level of trust that A (the trustor) places on B (trustee) is a function of B's trustworthiness and
A's generalized trust.
We postulate that regional SC simultaneously affects its trustee's trustworthiness and its trustor's generalized trust. SC is the ability of actors to secure benefits by virtue of membership in social networks (Bourdieu (1985)). Social networks are typically associated with norms that promote coordination, cooperation, and reciprocity for the mutual benefit of members (Coleman (1988), Putnam (1995)). A high SC environment helps spread cooperative norms and civicmindedness (Guiso et al. (2004)), intensifies internal sanctions such as social ostracism (Uhlaner (1989)) and stigmatization (Posner (2000)), and heightens negative moral sentiments associated with opportunistic behaviors (Elster (1989)). Hence, trustors from high SC regions are likely to anticipate cooperative, as opposed to opportunistic, behavior from their counterpart (i.e., trusting), whereas trustees from high SC regions are likely to keep their promises and have low moral hazard (i.e., trustworthy).
We use peer-to-peer lending to test these hypotheses. In the past decade, technological innovations in finance (Fintech) have supported lending between individuals in an online marketplace without the need for financial intermediaries. Owing to the digital and anonymous nature, establishing interpersonal trust is not applicable in online marketplace lending. To overcome the extreme information asymmetry and adverse selection in this market, lenders seek trust signals to help identify a borrower's "type." Unlike financial institutions, individual lenders use representativeness (Kahneman and Tversky (1972)) or even stereotype (Gilbert and Hixon (1991)) to minimize effortful thought processes. In this context, regional SC provides cursory beliefs and generalizations about others (Bottazzi, Da Rin, and Hellmann (2016)). The impact of SC on trust is instantaneous (Durlauf and Fafchamps (2006)), exogenous to each economic transaction, and could be overweighed in probability judgments (Bordalo, Coffman, Gennaioli, and Shleifer (2016)) 1 .
We construct a Chinese provincial SC index to capture the SC environment of lenders and On the impact of SC on "trustworthiness," we show that all else being equal, borrowers from high SC regions receive larger bid from individual lenders, have higher funding success, larger loan size, and lower default rates. The effect is pronounced among "low-quality" (low-educated, non-repeated, and low-score) borrowers. These results are robust to a variety of robustness checks for endogeneity and alternative explanations.
On the impact of SC on "generalized trust," we find borrowers from high SC regions are more likely to become lenders. Conditional on extending loans, lenders from high SC regions bid larger amount and larger fraction of the loan, but incur high default rates. Further evidence shows that their loans to borrowers from low SC regions contribute to inferior performance. These results focused on inexperienced lenders, that is, those who have not encountered any defaults.
Third, on how regional SC affects cross-regional transactions, we show in a two-by-two matrix that 63% of total investments (accounting for 57% of total transactions) are made by lenders in 1 Zingales (2015) comments: "Even within the United States, Americans of Swedish origin are more trusting, more in favor of redistribution, and less thrifty than Americans of Italian origin, in the same way that Swedes are more trusting, more in favor of redistribution, and less thrifty than Italians." 2 Section 4.1 presents the construction of the SC index. Our results are robust to city-level SC measures. See Section 5.3.4 for details.
high SC regions to borrowers in high SC regions. Twenty-one percent of total investments (accounting for 22% of total transactions) are made by lenders in high SC regions to borrowers in low SC regions. Twelve percent of total investments (accounting for 15% of total transactions) are made by lenders in low SC regions to borrowers in high SC regions. Only 4% of total investments (accounting for 5.5% of total transactions) are made by lenders in low SC regions to borrowers in low SC regions. These findings suggest that cross-regional transactions are most (least) likely to be realized between parties from high (low) SC regions, where the aggregate level of trust is highest (lowest).
Our work belongs to the extensive literature on SC and trust. Prior empirical works typically use the word "trust," but they refer to either "trustworthiness" or "generalized trust." For example, to illustrate the impact of SC on trustworthiness, Guiso et al. (2004) show that Italian households in high SC regions have easy access to institutional credit. Hasan et al. (2017a) find that US firms headquartered in high SC counties receive favorable bank loan conditions. Ang et al. (2015) find that foreign firms prefer to invest in Chinese regions where local partners and employees are considered trustworthy. Lin and Pursiainen (2018) find that in equity crowdfunding, entrepreneurs from high SC regions have better campaign outcomes. On the impact of SC on generalized trust, Hong et al. (2005) and Guiso et al. (2008) find that individuals in high SC environments are more likely to participate in stock markets. Bottazzi et al. (2016) find that societal trust positively predicts European venture capital investments but negatively predicts their successful exits. Unlike previous studies, our highly granular data from peer-to-peer lending allow us to observe separately the impact of SC on trustees' trustworthiness and trustors' generalized trust.
This work adds to the growing number of studies on how non-expert lenders process information in a Fintech environment. As Thakor and Merton (2018) note, technology by itself is not substitute for trust. Prior work finds that non-standard soft information provides trust signals for investors to overcome information friction. Data from Prosper, a US-based peer-to-peer lending platform, reveal that borrowers' trustworthy appearance (Duarte, Siegal, and Young (2012)) and online friendship networks (Lin, Prabhala and Viswanathan (2013)) help improve their funding success through the impression of trustworthiness. Herzenstein, Sonenshein, and Dholakia (2011) and Larrimore, Jiang, Larrimore, Markowitz, and Gorski (2011) find that the use of extended narratives, concrete description, and quantitative words contributes to funding success.
Michels (2012) shows that additional unverifiable disclosure is associated with an increase in bidding activity and reduction in the cost of debt. We add to this literature the soft information of SC. We show that lenders' bidding behavior is affected by the SC of their home and that of the borrower. To our best knowledge, this research is the first work on the impact of regional SC in the world's largest debt crowdfunding market, China.
Finally, we contribute to a strand of literature on the role of trust in cross-border transactions by presenting important empirical evidence from peer-to-peer lending 3 . Guiso et al. (2009) show that trade and investment flows are large between countries that exhibit high mutual trust. Bottazzi et al. (2007) provide evidence that venture capitalists are less likely to fund entrepreneurs in countries whose citizens they trust less, and if they do, the contracts they use are different from the contracts used in countries they trust more. Giannetti and Yafeh (2012) find that culturally distant lead banks offer borrowers small loans at a high interest rate. Ahern, Daminelli, and Fracassi (2015) show that differences in level of trust between acquirer and target countries reduce M&A volume and cumulative abnormal return. Our evidence is consistent with this literature. Our dyadic analysis shows that (1) lenders bid less (more) when their counterpart is from a low (high) SC environment and (2) investments take place most often (least often) between high (low) SC regions.
The remainder of this paper proceeds as follows. Section II reviews the SC and trust literature and develops the hypotheses. Section III introduces the mechanism of online marketplace lending and institutional settings in China. Section IV describes our sample data and variables. Section V and VI present the empirical results. Section VII draws the conclusions.

A. Social Capital
The multidimensional concept of SC can be traced to Bourdieu (1985), who defines SC as advantages and opportunities accrued to people through membership in certain communities. In his seminal work, Coleman (1988) claims that three forms of SC can be taken as resources for action: (1) obligations and expectations, (2) information channels, and (3)  This article employs the broad definition of SC at society level (also termed as "civic social capital") in the spirit of Putnam (1995), who defines SC as "social organization features, such as networks, norms, and social trust, which facilitate coordination and cooperation for mutual benefit" (p. 67). Collier and Gunning (1999) argue that the economic benefits of a civic society can arise from the building of trust that lessens transaction costs, from the knowledge externalities of social networks, and from an enhanced capacity for collective action. These features, coupled with the appropriate use of sanctions in case of noncompliance, enable groups to overcome collective action problems and deal effectively with multiple social and economic issues (Bloch, Genicot, and Ray (2007)).

B. Social Capital and Trustworthiness
Societal SC can serve as a monitoring system that "rewards" honest dealings and "punishes" opportunistic behaviors (Yamagishi (1988)). In this study, SC serves as a governance institution similar to that played by the formal institution of law. Coleman (1988) argues that dense social networks make the enforcement of group cooperative behavior effective. By aggravating the cost of expropriation and breach, SC provides a mechanism for contract enforcement.
The monitoring aspect of SC can enhance its agent's trustworthiness, diminishes the cost of financial contracting, and facilitates access to external financing. For example, Hasan et al. (2017a) find that firms headquartered in U.S. counties with high SC have low spreads in bank loans and low at-issue spreads in public debt issues. Gupta, Raman, and Shang (2018) show that firms' cost of equity is negatively related to the SC environment surrounding their headquarters. Hasan et al.
(2017b) find that firms headquartered in U.S. counties with high SC pay high corporate taxes.
They interpret this result as SC, as a governance institution, constrains self-serving corporate practices that benefit shareholders at the expense of other stakeholders. Huang and Shang (2019) present evidence that firm leverage and short-term debt ratios are negatively associated with SC.
They argue that high SC alleviates agency conflicts between managers and shareholders, allowing firms to reduce the amount of debt in their capital structure and the usage of short-term debt in their debt structure. Hoi, Wu, and Zhang (2019)

A. Social Capital and Credit in China
This paper presents important evidence from the emerging market of China. In this market, laws and courts are ineffective in protecting investors (La Porta, Lopez-de-Silanes, Shleifer, Vishny (1998)), necessitating reliance on alternative institutions, such as SC. However, heterogeneities in the SC stock are substantial across Chinese regions (see Section 4.1 for details). For instance, using data from the World Values Survey, Ang et al. (2015) show that SC differences among China's 31 provinces are often greater than those among European countries.
China's financial environment is composed of a bank-dominated credit market and a relatively underdeveloped capital market (Allen, Qian, and Qian (2005)). Most credit is extended by state-owned banks to state firms or the listed sector, leaving major obstacles for private smalland medium-sized firms and individuals to secure financing (He, Xue, and Zhu (2017)). "Shadow banks," or financial firms outside the formal banking sector, primarily serve the financial needs of the vast private sector (Elliott, Kroeber, and Yu (2015)). These financial firms take various forms, such as trust companies, inter-corporate loans via financial institutions ("entrusted loans"), microfinance companies, guarantee firms, leasing companies, pawnshops, and unofficial lenders.
In the past decade, the investment and credit demand of Chinese individuals has surged along with the country's rising middle class, and technological development in finance has greatly facilitated person-to-person lending on the Internet. China has over 700 million Internet users, many of whom have developed the habit of making digital payments 5 . Data from Wangdaizhijia 6 show that the number of companies operating peer-to-peer marketplaces soared from only 10 in 2010 to 3,984 by March 2016. These firms facilitated a total of 1.745 trillion RMB (USD 268.4 billion) in loans. Although this emerging market is smaller than the country's colossal financial system, 7 by any measure of size, China is the largest peer-to-peer lending market in the world (The Economist (2017)).

B. Renrendai Online Marketplace
Much of our data are drawn from RRD, one of the largest peer-to-peer lending platforms in  (2012) show that on Prosper, trustworthy appearance is associated with better loan outcomes. We can safely dismiss this factor in our setting. Second, borrowers have no choice on interest rate, because RRD adopts a "posted price mechanism," which assigns interest rates and calculates monthly payments on the basis of its proprietary credit rating model 9 . This feature is useful in the institutional setting because the outcome depends directly on lenders' willingness to supply credit at the given interest rate. 10 To initiate a loan listing on RRD, users first register on renrendai.com by providing the required information, including their ID card (two-sided), bank account, and cellphone number.
For verification, borrowers must submit a photo of themselves holding their ID card (not required among lenders). In addition, they need to provide supplementary evidence of their occupation (employment contract), income (bank statement), education, marital status, home ownership, and residential address. As the most important information, residential address holds the most credibility because RRD requires a "proof of address" that includes bank statements, phone bills, and water or electricity bills. We use this variable to identify a borrower's home province.
To make loan requests (called "listing"), borrowers must supply a title, description, loan amount, and maturity. All loans are unsecured personal loans, and their maturity ranges from 1 month to 48 months. In addition, personal information about borrowers, including age, gender, education, income, marital status, house ownership, employment information, and address (city), is verified and disclosed in the platform by RRD.
8 Individual lenders on RRD can choose one of the two channels to make investments on loan listings. The "automatic bidding" (zidongbiao) channel allows lenders to lock in a sum of money at RRD's wealth management plans for algorithm-based bidding. The "manual bidding" (shoudongbiao) channel requires lenders to manually select and make investment decisions by themselves. The manual bidding channel is peer-to-peer lending in its essence, because it reflects bounded rationality of individual lenders based on the information they have, their cognitive limitations, and the finite amount of time they have to make a decision. These are the data that we use. 9 The exact credit rating model used by RRD to assign a credit rating is unknown due to its proprietary nature. However, unlike in the US where individuals' FICO scores can be obtained, in China the personal credit score system is non-existent. Each peer-topeer lending platform claims to have its own credit rating model based on available information. For example, RRD classifies borrower credit ratings into seven categories: AA, A, B, C, D, E, and HR (high risk). A minimum rating is acquired when a borrower inputs the minimum information required by RRD to open an account. If borrowers voluntarily provide more documentary proof, such as bank income statement, house-ownership certificate, then these details are verified by the website, and their credit rating will increase. Moreover, if a borrower has a good repayment history on this platform, then the borrower's credit rating will also increase. 10 Wei and Lin (2016) note that two mechanisms are popular in online peer-to-peer lending: auctions and posted prices. In auctions, the "crowd" determines the "price" (interest rate) of the transaction through an auction process. In posted prices, the platform determines the interest rate on the basis of its own "grading" of the borrower. RRD adopts the posted price mechanism.
A loan listing can be open for several days. Lenders can bid any amount in multiples of 50 RMB (USD 7.7). The majority of loans are crowdfunded by multiple lenders. A loan that reaches 100% subscription becomes binding; otherwise, the borrower receives zero funding. Once a successful loan is verified by RRD, funds are transferred from the lender(s) to the borrower, minus a platform service fee. Service fee rates vary according to borrowers' credit rating.
Subsequently, borrowers are obligated to repay the principal and interest in monthly installments.
Repayments are proportionally distributed to the lenders of a loan. If a repayment is overdue (i.e., funds in the borrower's bank account are insufficient to repay the interest), then RRD makes several attempts to collect, including sending emails and text messages, seeking the borrower's employer, and conducting on-site collections.

IV. Data and Research Design
A.

Measuring Regional Social Capital
Following the empirical literature, we construct a composite SC index of Chinese provinces.
Our SC proxies rely on provincial statistics and national surveys, which incorporate attitudinal and civicness measures of societal trust 11 . The composite SC index has four components: voluntary blood donation (Blood), NGO participation (NGO), enterprise survey (Enterprise), and citizen survey (Citizen). Each proxy is illustrated below.

Voluntary Blood Donation
Our first SC proxy, Blood, is voluntary blood donation per thousand population in a province.
Neither legal nor economic incentives are given to those who donate blood (Guiso et al. (2004)).
The act is likely driven by citizens' reciprocity and civic-mindedness. Following Ang et al. (2015), we measure this variable in milliliters of blood donated voluntarily in a province divided by its population in 2000, the only year that province-level data from the Chinese Society of Blood Transfusion were complete 12 . China's blood donation law states that blood can only be collected by the National Blood Center (NBC) and is without compensation. The NBC has operating 11 Anderson, Mellor, and Milyo (2004) categorize societal SC measures into (1) attitudinal measures, where subjects are asked if they agree that "most people can be trusted," "most people try to be fair," "most people try to be helpful," "you cannot trust strangers anymore," and "I am trustworthy;" (2) behavioral measures of "trust" suggested by Glaeser et al. (2000), including whether subjects leave their doors purposely unlocked, loan money to friends or strangers, have been a crime victim, or lie to different categories of persons; and (3) "civicness" measures, including hours spent volunteering, membership in volunteer groups, attendance in religious services, political volunteering, and voting. 12 We are grateful to Ang et al. (2015) for sharing these data with us. branches in all provinces and adopts the same medical procedures across all regions, thereby mitigating the concern that blood donation levels are affected by differences in the quality of healthcare or medical infrastructure among provinces. Panel B (column 2) of Table 1 shows a large variance among Chinese provinces, with an average blood donation of 3.433 mL/1,000 people in Shanghai and only 0.017 mL/1,000 people in Yunnan.

NGO Participation
The second SC proxy, NGO, is measured by the number of people registered in NGOs per thousand population in a province. NGOs are typically funded by charities and operated by volunteers. They aim to address poverty reduction, environment protection, and rights of that Shanghai is the province with the highest NGO participation (4.4 registered NGO members per thousand population) and that Tibet has the lowest NGO participation (only 0.03).

Enterprise Survey
Our third proxy, Enterprise, is drawn from a national survey of Chinese enterprises in 2000 (Zhang and Ke (2003)) 14 . In this survey, questionnaires were sent to over 15,000 managers of companies in every province of China. Over 5,000 usable responses were received, and respondent managers covered firms from every two-digit industry and ownership type. This variable is elicited from their answers to the question, "According to your experience, could you list the top five provinces where enterprises are most trustworthy?" Following Wu, Firth and Rui (2014), we set the SC score of a province as the logarithm of the total score given by the managers. Panel B (column 4) shows that Shanghai (22.7) leads Chinese provinces in enterprise reputation, followed by Beijing (16.6) and Guangdong (10.1). The least trusted province appears to be Hainan (0.1).

Citizen Survey
The fourth proxy, Citizen, employs data from the China General Social Survey (CGSS provinces. The Citizen variable is elicited from the response to the question "Do you trust strangers?" Responses range from 1 ("do not trust greatly") to 5 ("trust greatly"). We average the scores of respondents' choices by the provinces where they are located. Panel B (column 5) shows a considerably small variance among the scores given by the citizens of each province. Shanghai ranks second (2.40) and is surpassed by Jiangxi (2.442). The least trusting provinces appear to be Gansu (2.014) and Guizhou (2.014).

Composite Social Capital Index
Each of the four proxies could be an imperfect measure of SC. For instance, the Blood and NGO participation proxies capture outcomes more than perceptions. The Enterprise and Citizen proxies are based on survey data and capture perceptions, yet they suffer from self-esteem and ingroup bias. To account for their intrinsic biases, we construct a composite index by applying principal component analysis (PCA). Panel A in Table 1 shows the results of the PCA for our four components. This method shows that only one component has an eigenvalue larger than 1 (2.967).
All four components have positive loadings and are closely correlated with the index. Our SC_index gives roughly equal weight to Blood, NGO, and Enterprise but low weight to Citizen.
According to the SC_index (Panel B, column 1), Shanghai, Beijing, and Guangdong are the top three SC stock provinces, whereas Gansu, Guizhou, and Yunnan rank in the bottom.
[Insert Table 1 here] We first obtain information on the funding success or failure of each loan listing (FUND). For each successful loan, we obtain the loan size (AMOUNT), MATURITY (in months), SPREAD (interest rate relative to benchmarked lending rate of People's Bank of China), number of lenders (OWNERSHIP), stated loan purpose (in descriptive text), number of words used to describe a loan (WORDS), default status (DEFAULT), and BID_TIME for each fully funded loan (in minutes).

B. Variables of Interest and Controls
For each unsuccessful loan, we obtain the proportion of campaign proceeds out of the total amount (FRACTION).
For each borrower, we obtain their unique ID, age, gender, resident province, marital status, income range, education, work experience, home ownership status, and borrowing history on RRD.
We also obtain their credit rating assigned by RRD (in seven categories, i.e., AA, A, B, C, D, E, and HR). For provincial variables, other than the four SC proxies, we include GDP per capita (PGDP) to measure their economic environment and the number of law offices per ten thousand residents (LAW_OFFICE) to capture the legal environment. LOAN is the ratio of total bank loans to provincial GDP, which we use to measure the financial development of a province (Rajan and Zingales (1998)). In our regressions, the institutional variables of a province in year t-1 are matched with loans originating in year t.

C. Summary Statistics
Our sample is composed of 247,115 loan listings on RRD from 2011 to 2015. Panel A of Table   2 reports that approximately 24.9% of loan listings are fully funded. Of the 61,577 fully funded loans, the mean of loan size varies significantly from 3,000 RMB (USD 437) to 3 million RMB (USD 461,538). On average, the loan rate is 2.13 times the benchmark lending rate, with significant variation of 0.76-5.38 times the benchmark lending rate. Relative to the stability of China's benchmark lending rate, these large pricing differences reflect, at least in part, the differences in borrower risks. The mean (median) loan maturity is 18.79 (19) months. We construct an additional variable LONGTERM, which is a dummy variable that equals one if the loan maturity is over 12 months and zero otherwise. The variable shows that 80% of borrowers request a long-term loan.
Ownership also varies considerably across loans. The average loan has 35.5 lenders in the range of 1-1,370 lenders. The average bid time for each fully funded loan is 69 minutes. Finally, approximately 5.4% of completed loans incur default.
Panel B reports the summary statistics of borrower characteristics. Most borrowers are young, male, married, do not have a bachelor's degree, and have low credit scores. The median income level of borrowers is less than 10,000 RMB (USD 1,538) per month. Only 44% of borrowers own a house, and 15.8% of borrowers report having a home mortgage loan. Panel C reports the summary statistics of (borrower) provincial-level variables. It shows a large variation in economic, legal, and financial development across Chinese provinces. We do not include province-level or borrower-level fixed effects in most regressions because our SC_index is time invariant for all borrowers in the same province 15 .

D. Research Design
Our study is motivated by the prevailing role of SC on its home trustee's trustworthiness and its home trustor's generalized trust. Regional SC has major effects on trust-intensive contracts, that is, debt, stock, and venture capital, particularly in cases of severe information asymmetry and limited information exchange. These problems are highlighted in online marketplace lending where lenders are unsophisticated investors.
First, we postulate that a region's SC stock positively affects its home trustee (borrower)'s trustworthiness. If so, then SC should be associated with better both ex ante outcomes of finance, such as funding success and loan size. We also hypothesize that the marginal effects of SC on ex ante outcomes are stronger when borrowers are less educated, borrowing for the first time, and have lower credit score. Apart from "perceived" trustworthiness, we test whether SC affects "actual" trustworthiness by investigating the ex post defaults in fully funded loans.
Next, we examine whether SC affects its home trustor (lender)'s generalized trust. If lenders in high SC environment are inherently more trusting on others, then, controlling for loan and borrower properties, lenders from high SC regions are more likely to bid, and when they do, they bid larger amounts and larger fraction of loan requests. We also expect that the impact of SC on lender's trust propensity is larger on inexperienced lenders, who are more likely to engage in (SCinduced) coarse thinking. Finally, to gain insights into the consequence of high generalized trust, we examine their investment success through ex post default rates.
We employ a battery of robustness tests to tackle the potential endogeneity. Regional SC is clearly not randomly assigned, nor is it a choice. Accordingly, we treat SC of one's home province as historically and econometrically predetermined. We can safely dismiss the possibility of reverse causality because each micro-economic transaction is clearly too small to influence SC among regions. Hence, the main identification challenge is not reverse causality, but whether our SC index is correlated with other (omitted) factors that simultaneously affect the debt crowdfunding outcomes 16 . Section 5.3 discusses our various empirical strategies.

V. Empirical Results on Social Capital and Trustworthiness
We start by testing how borrowers' SC (B_SC_INDEX) affects their debt crowdfunding outcomes. We infer borrowers' trustworthiness from their funding success (dummy and fraction), loan terms, and default rates. We also consider how the effects of (borrowers') SC vary across heterogeneous borrower characteristics, such as education, credit history, and credit score. Table 3 reports the results of borrowers' SC on loan funding success, the number of lenders for a given loan (OWNERSHIP), and loan size (AMOUNT). SUCCESS is a dummy variable that equals 1 if a borrower's loan is fully funded. FRACTION is the proportion of proceeds relative to the loan amount. Columns 1 and 2 use probit models for funding success, and we report the marginal effects for each variable. [Insert Table 3 here]

B. Heterogeneity Tests
If our proposition is true that lenders use borrowers' home SC as impression of trust, then theories of adverse selection (Akerlof (1970)) predict that the marginal benefit of SC would be large for low-quality borrowers. To test this proposition, we partition the sample on the basis of quality indicators, such as borrower's education level, credit history on RRD, and credit grade. Table 4 shows the results. 17 Lenders are attracted to this market due to its promised high return, where the pre-determined interest rates are several times (not simply basis point) higher than the rate that potential lenders could earn in banks. In the beginning, loan lists with high interest rates will mechanically accumulate many more bids than those with low interest. However, high interest rate also signals high moral hazard of borrowers (Stiglitz and Weiss (1981)). If lenders are rational and they can perceive borrowers' quality from offered interest rate, then borrower's lists with high interest are less likely to be fully funded.
Prior works show that an individual's human capital is closely correlated with education (Lusardi and Mitchell (2008), Behrman et al. (2012)) and that borrowers with low human capital tend to have high financial constraints. In Panel A, a borrower is classified highly (low) educated if his or her highest qualification is a bachelor's degree or above (post-tertiary or below).
Consistent with adverse selection, SC has little impact on the funding success of highly educated borrowers (columns 2 and 4) but has large and significant impact on undereducated borrowers (columns 1 and 3). As for loan ownership, SC has a negative and statistically significant effect on the number of lenders in both subsamples, and the difference between the undereducated and highly educated groups is statistically insignificant. Finally, the positive effect of SC on loan amount is large and statistically significant among undereducated borrowers, whereas it is insignificantly negative among highly educated borrowers. The difference between them is statistically significant at the 5% level.
Panel B re-runs the regressions by partitioning the sample into repeat and non-repeat borrowers.
A borrower is a repeat borrower if he or she appears more than twice on RRD; otherwise, the borrower is a non-repeat borrower. The effects of borrower SC on funding success, fraction, loan ownership, and amount are highly significant in the subsample of non-repeat borrowers but insignificant among repeat borrowers. The differences between the two groups in terms of funding success and loan ownership are statistically significant.
Panel C re-runs the regressions by partitioning the sample into low-versus high-credit-score borrowers. A borrower is classified as high (low) grade if his or her credit score is below 5 (above or equal to 5). The effects of borrower SC on funding success, fraction, and amount are significant in low-grade borrowers but not among high-grade borrowers. The only exception is loan ownership, that is, the impact of borrower SC on the reduction of the number of bidders is significant among high-grade borrowers but not among low-grade borrowers. The difference between the two groups is not statistically significant.
Taken together, the cross-sectional evidence validates our proposition that lenders use borrowers' home SC as impression of trust. Consistent with adverse selection, our proposition benefits borrowers with low education, little credit history, and low credit grade.

Bootstrapping
Hypothesis testing using a large sample like ours can possibly yield a Type I error. To check robustness, we implement a bootstrapping method. Specifically, we draw a subsample that includes half as many observations as the whole sample and repeat our regression analysis for this subsample. We then replicate this procedure 1,000 times and obtain the bootstrap statistics. Panel A (columns 1-4) of Table 5 shows that our results are qualitatively unchanged; that is, (borrowers') SC positively correlates with funding success and loan amounts and negatively correlates with the number of lenders.

Selection Bias
Much of our empirical analysis uses data from fully funded loans, which account for 24.9% of all loan listings. To account for potential bias, we employ the Heckman two-step treatment effects procedure. In the first equation, we estimate the probability that a loan will be fully funded; here, the dependent variable is a dummy of funding success. This equation uses the same specification as in column 1 of Table 3. In the second equation, we use the inverse Mills ratio to correct the selection bias for the performance equations. These equations use the same specifications as (6) and (8) of Table 3. Panel A (columns 5 and 6) of Table 5 presents the results of the Heckman selection model. The effect of SC on loan ownership and amount remains significant.

Familiarity Bias
We are also concerned about familiarity bias in our result. The top group of high SC provinces consists of key provinces, that is, Shanghai and Beijing. By contrast, the bottom group consists of unpopular provinces, that is, Gansu, Qinghai, and Ningxia, where most people in populated coastal areas may never meet someone from these places. To mitigate familiarity bias, we re-run our model specifications that exclude the top group and the bottom group, that is, Shanghai, Beijing, Gansu, Qinghai, and Ningxia. Panel A (columns 7-10) of Table 5 presents the results. The coefficients of SC carry the same signs and remain statistically significant.

Social Capital or Economic Development?
Another concern is that our SC index appears to correlate with economic development of the provinces. Our results will be spurious if, for whatever reason, borrowers in economically prosperous regions are more trustworthy. Although we have controlled for economic development (per capita GDP) and other institutional variables in all specifications, ruling out the confounding impact of economic development is not sufficient. To address this concern, we employ two methods.
First, we re-estimate our basic specifications, splitting the sample between provinces with low economic development (per capita GDP below the median) and provinces with high economic development (per capita GDP above the median). Panel B of Table 5 shows that our results are not driven by either subsample. The negative relationship between SC and loan ownership seems strong among the low economic development regions. This result indicates that the number of lenders for a given loan (OWNERSHIP) responds more to SC in less developed regions.
Next, we perform a difference-in-differences test by investigating how a negative shock to SC (unrelated to economic development) affects peer-to-peer lending. The shock we exploit is the Guo Meimei incident 18 . In June 2011, a woman nicknamed "Guo Meimei baby," who claimed herself the general manager of the Chinese Red Cross, showed off her wealth on a blog. This incident provides an ideal laboratory for the following reasons. First, it generated a severe trust crisis for the Red Cross Society of China (RCSC), causing donations to suffer 19 . Second, it was an explicit, temporary shock to trust. A police investigation in 2012 showed that Guo Meimei's wealth was actually not from the RCSC, and the RCSC gradually restored its reputation in the following months. Third, the incident isolates the effects of SC from local economic conditions, as the incident was unrelated to local economic conditions. If SC has a real impact on lending, then the shock could temporarily change investors' beliefs about the risk of their assets being stolen, causing them to withdraw or reduce their investments.
Thus, we design a difference-in-differences test surrounding the Guo Meimei incident. The incident date is set as month 0, and we focus on six months prior to and six months after the incident 20 . POST is a dummy variable for the months following the Guo Meimei incident (i.e., [1,6]). We divide our sample into two groups according to the borrower SC index. B_SC_INDEX_H is a dummy variable that equals 1 if the SC index in a region is above the median, and zero otherwise. The coefficients on POST*B_SC_INDEX_H allow us to estimate the differences in the changes in lending activities between regions with different SC. Panel C of Table 5  [Insert Table 5 here]

Instrumental Variable Approach
This section employs an instrumental variable approach to tackle the potential omitted variable bias. A valid instrument should induce changes in our key explanatory variable (regional SC) but should have no independent effect on the dependent variable (debt crowdfunding outcomes), other than through its impact on regional SC.
We employ two instrumental variables. The first instrument traces back a province's agricultural specialization of growing rice versus wheat. Subsistence style theory argues that some forms of subsistence require more functional interdependence than other forms, and ecology narrows the types of subsistence that are possible. Talhelm et al. (2014) find that Chinese regions with a history of farming rice have a more cooperative norm than those with a history of growing wheat. This is because paddy rice requires irrigation and high labor demand, causing farmers in rice-growing regions to form cooperative labor exchanges. By contrast, wheat does not need to be irrigated, and wheat farmers can rely on rainfall, which does not require coordination with their neighbors. On the one hand, societies that need to cooperate intensively develop more interdependent culture and accumulate higher SC stock over time 21 . On the other hand, a region's environmental suitability for rice, which relies on soil, climate, and topographic factors, should not have a direct impact on today's urban consumer credit, except for its impact on the formation of regional cooperative culture and SC 22 . Specifically, we calculate the logarithm of the "rice suitability" index of Chinese provinces (RICE_SUIT). The index is a z-score of the environmental suitability of each province for growing wetland rice according to the United Nations Food and Agriculture Organization's Global Agro-ecological Zones database.
Our second instrument exploits the ethnic diversity in Chinese provinces. The ethnic diversity in China provides an exogenous driving force behind the regional variation in SC 23 . Prior crosscountry studies show that diversity of ethnic groups in a country increases communication costs, social fragmentation, and probability of civil conflict (Easterly and Levine (1997) group in a province, which should be positively correlated with regional SC stock. However, conceiving a direct impact of regional ethnic diversity on outcomes of nation-wide debt crowdfunding is difficult, other than through its impact on the local SC stock.
Using the full sample, Table 6 reports the results from the instrument variables RICE_SUIT and ETHNIC. Columns 1 and 2 run the regression of the probit and linear models for funding success, and column 3-5 run the linear regression models for fraction, loan ownership, and amount, respectively. We control for loan and borrower variables, regional variables, and year fixed effects, but omit their coefficients for brevity. Consistent with expectation, the first-stage results in Panel B show that RICE_SUIT and ETHNIC variables are positive and significantly correlated with the SC index. The second-stage results in Panel A validate our baseline results that regional SC is an important determinant of funding success, loan ownership, and loan amount. In addition, we conduct an over-identification test because the number of instrument variables is greater than the number of endogenous variables. The reported p-value of Hansen-J statistics is larger than 0.05, we conclude that the over-identification restriction is valid.
A complete specialization on rice requires highly seasonal labor demand, which often cannot be procured locally and expose farmers to high risk against production failure or decreasing prices (Klasen, Priebe, and Rudolf (2013), Di Falco and Chavas (2008)).
In the long run, regions with high rice specialization might develop deeper agricultural credit and insurance market. This effect, however, is indirect, and pertinent to agricultural finance as opposed to urban consumer credit. 23 Ethnic diversity, which requires long duration of uninterrupted human settlements (Ahlerup and Olsson (2007)) is typically treated as an exogenous explanatory factor in economics. For a good review of this literature, see Alesina and La Ferrara (2005).
A valid instrument should satisfy the relevance condition and exclusion restriction condition.
The p-value for the F statistics for the joint significance of the instrument variables is 0.000, which is sufficient to alleviate the relevance concern. We implement tests from Stock and Yogo (2005) for weak instruments. Panel B gives critical values for 2SLS at the 10% level 24 . The reported minimum eigenvalue statistic greatly exceeds the critical value of 19.93 and is large enough to reject the null hypothesis of weak instruments. Another concern is that the instruments ( both instrument variables in the benchmark regressions. For simplicity, we include but do not report borrowers' personal characteristics, regional economic variables, and financial variables.
The SC index yields consistent results, but all the estimated coefficients of the instrumental variables are statistically insignificant.
[Insert Table 6 here] 4. City-level Evidence 24 We also use the LIML estimator at 10% level and obtain similar results.
Our SC measure at province-level may be too coarse 25 . China is a large country, and each of its provinces is comparable to a European country by population 26 . People would certainly not consider that all Italians, French, and British are alike in trustworthiness. Fortunately, one of our SC proxies, Citizen, employs data from the CGSS for 125 cities in 28 provinces. The number of cities in CGSS varies from 1 city in Hainan Province to 7 cities in Guangdong Province 27 . Citylevel analysis provides more variation and testing power than provincial-level analysis. Columns 1-4 of Table 7 present the results, which are qualitatively unchanged. City-level citizen proxy, Citizen_city, positively correlates with funding success and loan amounts, and negatively correlates with the number of lenders.
To further exploit the SC variations within a common region, we turn to a smaller sample of 11 cities in three neighboring provinces of Yunan (3), Guizhou (2), and Sichuan (6)  5. Time-series Evidence 25 We thank an anonymous referee for raising this issue. 26 For example, eight Chinese provinces have population that is comparable to Italy (60.5 million), France (65 million), and UK (68 million). These provinces are: Guangdong (104 million), Shandong (100 million), Henan (94 million), Sichuan (81 million), Jiangsu (79 million), Hebei (72 million), Hunan (66 million), and Anhui (60 million). 27 One limitation of city-level analysis is that borrowers are located in 200 cities of China, but the data of city-level SC are available only for 87 cities. 28 The analysis based on one single province will produce inaccurate estimation due to a small sample of cities. 29 Southwest China, in a narrow sense, covers only three provinces of Sichuan, Guizhou, and Yunnan. In a broad sense, it also covers Chongqing municipality and Tibet autonomous region. 30 The region covering the three provinces was historically jointly governed by the state of Shu Han during the Eastern Han Dynasty (220-280). In the 13th century, the Mongolian army conquered the Southern Song Dynasty and created Sichuan, Yunnan, and Guizhou administrative region (Fei (2017)). Since then, the region was administratively governed by Mongols in the Yuan Dynasty One criticism on our results is that our proxy for SC is cross-sectional in nature. Indeed, the common proposition is that a society's SC, which accumulates over a long time, is highly persistent (Putnam, Leonardi, and Nanetti (1994)). However, certain shocks may cause societal SC to change quickly (Algan and Cuhuc (2014), Guriev and Melnikov (2016)), which will bias our result.
Fortunately, one of our provincial-level measures of SC, NGO, is time-varying. We thus include borrower-level fixed effects to control for time-invariant unobservable heterogeneity. Columns 9-12 of Table 7 show the results. It shows that NGO participation in year t−1, NGO_t−1, is significantly positively related with funding success and loan size, and it is negatively related with the number of lenders (ownership).

D. Social Capital and Default Rates
In this section, we use the ex post measure of default rates to test whether borrowers from high SC regions are indeed trustworthy. To test this proposition, we run probit models in which the dependent variable is DEFAULT, which takes the value 1 if borrowers do not make a repayment on time and the value of 0 otherwise.
Column 1 in Table 8 shows a negative relation between SC and default. The coefficients of marginal effects are statistically and economically significant. A one-standard-deviation increase in borrowers' SC index leads to a decline in default rate of approximately 0.4 percentage points or 8% of the sample mean. In an extreme case, a loan made to a borrower in Gansu (SC index of −1.887) has a probability of default that is approximately 1.7 percentage points higher than that for a loan made to a borrower in Shanghai (SC index of 5.768); this value is approximately onethird of the sample mean. Column 2, which is based on OLS regression, presents a similar result.
As shown in Columns 3-4, we separately investigate the impact of SC on default rates for lowand high-educated borrowers. Among high-educated borrowers, SC does not significantly predict default. By contrast, in the sample of low-educated borrowers, SC is negatively related to default.
This result suggests that SC constrains the opportunistic behavior of low-educated borrowers more than it does on high-educated borrowers. As shown in Columns 5, 6, and 7, we employ the Heckman selection model, bootstrapping method, and instrument variable regression 33 in our probit model of default. We find a significantly negative relationship between SC and default rates, which validates our baseline results. 33 Column 7 of Table 9 reports the results from the instrument variables RICE_SUIT and ETHNIC.

E. Lender Fixed Effect
Thus far, the results using each loan as a unit of observation show the collective wisdom that borrowers from high SC regions are more trustworthy. To see whether the same is true among individual lenders, we use each lender's bid as a unit of observation. In debt crowdfunding, a borrower can obtain funding from multiple lenders. Each lender also bids on different borrowers.
Our 61,577 fully funded loans are composed of 2,172,520 bids made by 114,119 unique ID lenders.
Although RRD assigns each lender a unique user ID, it does not require lenders to provide personal information required from borrowers. Thus, we control for lenders' fixed effects to examine how borrowers' SC affects lenders' bids. The regression model is: (1) _ , ( , ) = 0 + 1 _ _ + 2 , + + + , where _ , ( , ) represents the bid amount (DEFAULT) of lender i in borrower j in time t. _ _ is the home SC of borrower j, and , represents the loan and borrowers' characteristics and regional variables. , represent the lender fixed effects and time fixed effects, respectively. , is the standard error.
The results presented in Table 9 confirm our baseline finding. That is, Columns 1-3 show that individual lenders make larger investments to borrowers from higher SC regions. A one-standarddeviation increase in a borrower's home SC increases a lender's bid size by 86.1 RMB (USD 13), an increase of almost a fifth in the median amount of a lender's investment. The effects are significant at the 1% confidence level. We also construct a variable BID_RATIO, which is the fraction of lender i's bid relative to the loan amount requested by borrower j ( _ ).
Columns 4-6 show that higher SC is associated with a larger BID_RATIO. Finally, the results in Columns 7-9 confirm that a borrower's SC significantly reduces default probabilities.

A. Dyadic Analysis
Although the results suggest that lenders use information about potential borrowers' SC when making lending decisions, little is known about the influence of SC on lenders' generalized trust.
One unique advantage of our study is that we can extract crucial details in a subsample of lending in which borrower and lender information is available. In other words, we can identify a specific lending relationship (i.e., who is borrowing and from whom) and examine how lenders' and borrower's SC affects lending decisions. We proceed in two steps. First, we identify lender characteristics from borrowers' information set and construct lender-borrower pairs. We then study how SC affects lender bids and the consequences of loans in terms of observable outcomes (i.e., the probability of default).

B. The Lender-borrower
We match a lender's user ID with a borrower's ID, yielding a nontrivial group of borrowers who bid on the same platform. The data show 1,743 unique lenders (bidding-borrowers) with investments in 21,727 loans, accounting for more than one-third of total fully funded loans. We first identify the factors that affect the likelihood of borrowers becoming bidders in the platform.
The dependent variable equals 1 if borrowers bid in the RRD online lending market, and zero otherwise. The main variable of interest is the level of SC. The control variables are (1) listing and loan characteristics, (2) borrower characteristics, and (3) provincial environment. Table 10 shows the results of probit and logistic regressions, and we report the marginal effects for each variable. Clearly, high SC increases the chance of a borrower bidding in the lending market. This result suggests that compared with borrowers who never bid in the market, individuals in regions with high SC are more likely to extend loans. In addition, bidding borrowers are more likely to be male, married, younger, and highly educated. Bidding borrowers are also likely to have more working experience and are more likely to own properties. However, in contrast to borrowers who never bid in the market, bidding borrowers tend to have lower credit ratings and personal income. Bidding borrowers are also more likely to come from less developed regions and regions with higher ratio of total bank loan to GDP. We interpret this surprising result as follows: lower credit rating and income borrowers from less developed regions typically have less access to finance and investment opportunities. Once they are familiar with the platform and become aware of investment opportunities available to them, they are more likely to become bidders. By contrast, in regions with high ratio of private debt to GDP, borrowers have easy access to credit and can afford to be profligate. As a result, they are more likely to bid in pursuit of higher return.

C. Lenders' Social Capital and Investment
Next, we focus on bids made by these borrowers, whose information is available. As a result, we have borrower and lenders characteristics, which yield 49,759 lender-borrower bid pairs in 21,727 fully funded loan projects.
Panel A of We first examine how lenders' SC affects their bid behaviors. We use this sample to estimate how lenders' SC affects their bid amount, while controlling for borrower fixed effects. Our regression specifications mimic that in Column (3) of Table 9, except that all control variables are on the side of lenders. Our working hypothesis is that lenders from higher SC regions have higher level of generalized trust, which positively predicts investment. Panel A of Table 12 reports results that are consistent with our hypothesis. Columns 1-3 show that lenders' SC index is positively related to bid amount 36 . These results suggest that for the same borrower, lenders from regions with higher SC are more trusting; in turn, they bid larger. This finding is consistent with that of 34 Borrowers need their repayment ability assessed by the platform to be allowed to borrow. However, lenders have no eligibility requirement to be a bidder. 35 One natural concern is that borrower-lenders can differ from non-borrower lenders in systematic ways. Assuming that is true, then we should find systematic differences in loan properties between our average and paired-loan sample. However, as Panel B shows, no differences in loan terms are statistically significant, as reported in Table 2. 36 The results remain qualitatively unchanged when we use the ratio of bid amount over loan size. Bottazzi et al. (2016), who find a positive relationship between generalized trust and investment in the context of venture capital.
The results so far reflect how lenders' SC affects the magnitude of loans conditional on observing lenders' SC. The potential problem is that pairs with lending relationships are only observable when borrowers bid in the same platform. For example, the results in Table 10 show that borrowers in regions with high SC are more likely to bid. However, important differences in SC may exist in lender-borrower pairs for which we do not observe lenders' SC.
We address this issue by implementing a Heckman selection model, which considers the selection bias arising from considering only lender-borrower pairs with observable information.
Columns 4-6 report the results of the second-stage Heckman estimation. The first stage of Heckman is a probit model, mimicking Next, we ask whether and how the differences in regional SC between borrower and lender affect lending transactions. To isolate the effects of SC and eliminate alternative explanations, we control for the distance between lender and borrower provinces, as well as other observable differences between lenders and borrowers that affect investments.
(2) _ , = 0 + 1 _ _ + 2 _ , + 3 + + + Of interest is the finding that the coefficients of lndistance are statistically significant in columns 1-2, suggesting that lenders tend to bid more for distant borrowers. Prior work shows that investors tend to trust counterparties that are close to home more than they do those in remote regions (Coval and Moskowitz (1999), Grinblatt and Keloharju (2001), Petersen and Rajan (2002), and Chan, Covrig, and Ng (2005)). To disentangle the effect of home bias from that of SC on investment, we employ the following strategy. First, we exclude investments in which lenders and borrowers come from the same province. Second, we include an indicator variable that equals 1 if 37 We also examine how borrowers' and lenders' social capital affects bid ratio and obtain similar results. For brevity, we do not report the results. 38 We do not control for lender and borrowers' fixed effects when we include B_SC_INDEX and L_SC_index separately in our regression, as the fixed effects are captured by their corresponding social capital. 39 This evidence is consistent with the findings of Giannetti and Yafeh (2012), who find that culturally distant lead banks offer borrowers small loans at a high interest rate.
the two provinces share the same border (BORDER); it equals zero otherwise. We then repeat the similar regressions in column 2 of Panel B and include Border and the interaction term between BORDER and the SC index. The results in column 3 show that both coefficients on BORDER and B_SC_INDEX are statistically insignificant. In sum, we find no evidence that home bias eliminates the effect of SC on investment.
We also implement a Heckman selection model to address selection bias. Columns 4-6 report the results of the second-stage Heckman estimation. The first stage is a probit model, which is the same as Panel A. The second stage is an OLS regression that includes control variables capturing the difference in the other explanatory variables between lender and borrower, and the inverse Mills ratio. The results show that the OLS regression is robust to this correction for selection. In sum, lenders bid less (more) for borrowers from lower (higher) SC provinces.

D. Lenders' Social Capital and Investment Performance
After exploring the direction and magnitude of loans, we focus on understanding how lenders' SC affects investment performance. We first examine if lender SC predicts default while controlling for borrower fixed effects. Our regression specifications mimic Panel A of it equals 0 otherwise. Probit models have difficulty dealing with lender fixed effects. Thus, we run linear probability models controlling for characteristics from the lender side and borrower fixed effects 40 .
Our working hypothesis is that lenders from regions with higher SC are more trusting on others. On the one hand, lender SC could positively predict investment; on the other hand, it may induce investment in high-risk projects, leading to high default rates. We report the results in Panel A of Table 13. Columns 1-3 show a marginally higher default probability for bids made by lenders from regions with high SC than those made by lenders from regions with low SC.
We also implement a Heckprobit model to address selection bias (probit model with sample selection). Columns 4-6 report the results of the second-stage Heckprobit estimation. The first stage is a probit model, which is the same as Panel A of Table 10. The results from the second stage are consistent, suggesting that a lender's SC is positively related to the probability of loan 40 We also implement a Heckprobit model (probit model with sample selection) to check the robustness of our results.
default. The coefficient of athrho is statistically significant, indicating the importance of addressing selection bias.
Next, we focus on lender-borrower pairs and explore how the difference of regional SC between borrowers and lenders affects the probability of default. We adopt the same estimation method as equation (2), except that the dependent variable is default. Columns 1-3 of Panel B report the results for the probability of default in a given loan with different specifications. We first consider a simple regression of B_SC_index and L_SC_INDEX on default. Column 1 reports that the coefficients of borrowers' SC (B_SC_INDEX) are negative and statistically significant at the 1% level. By contrast, the coefficients of lenders' SC (L_SC_INDEX) are positively related to default rates. Consistent with our previous findings, borrowers from regions with high SC are more trustworthy, whereas lenders from regions with high SC have higher generalized trust, which leads to higher default rates.
Of particular interest is the positive coefficient of D_SC_index in column 2 after controlling for borrower and lender fixed effects. It indicates that the higher default rates incurred by lenders from high SC regions are likely explained by their investment to borrowers from lower SC regions, the latter are more likely to default. In column 3, we repeat the regressions in column 2 of Panel B and include the dummy variable BORDER and the interaction term between BORDER and the SC index. Both coefficients on BORDER and B_SC_INDEX are statistically insignificant. Columns 4-6 of Panel B show the same specifications on default rates by correcting the selection bias. We find similar results for SC's effect on the probability of loan default. In addition, the results in column 6 show that being in neighboring provinces reduces the probability of loan default, but interaction term BORDER*B_SC_INDEX is statistically insignificant. In sum, we find no evidence that investors' home bias eliminates the effect of SC on default.

E. Does Bad Experience Affect Generalized Trust?
The results above suggest that lenders from high SC provinces are more likely to bid, and when they do, they bid more, but incur more defaults. It indicates that SC affects its trustors' generalized trust. However, we expect that trustors' propensity to trust others would be affected by past experience, especially when a trustor had bad experience in trusting others.
To test this hypothesis, we partition lenders into those who had experienced default (experienced), and those who had not (inexperienced). We re-estimate the lender-borrower pair regressions, mimicking Table 12 and Table 13. The results are reported in Table 14 41 . In each group, we first examine if lender SC affects their bid behaviors and predicts default while controlling for borrower fixed effects. We then turn to examine the effect of the differences in regional SC while controlling for both lender and borrower fixed effects. Columns 1-6 show the results of inexperienced lenders, and columns 7-12 report that of experienced. Consistent with expectation, the impact of lender SC on default is positive and significant on inexperienced lenders, but not significant on experienced lenders. Moreover, experienced lenders reduce their investment amount to borrowers from lower SC environment. Taken together, our evidence suggests that lenders learn from their experience on the platform, and the instantaneous impact of SC on generalized trust concentrates on inexperienced trustors.

F. Social Capital and Regional Capital Flows
To observe clearly how regional SC affects cross-border investment flows, we prepare a twoby-two matrix. First, we classify Chinese provinces into high and low SC regions on the basis of the sample medium in our SC index. Next, we divide lenders and borrowers into those from high and low SC regions. We then calculate the (1) number of bids, (2) mean/medium size of bids, and (3) total amount of investment for each pair. Table 15 reports the findings. Approximately 63% of total investments (28,148 bids with a total size of 32.3 million RMB (USD 4.97 million)) flow from high SC regions to high SC regions. The mean and medians of bid size are 1,150 and 300 RMB, respectively. By contrast, only approximately 4.2% of total investments (2,708 bids with a total size of 2.14 million RMB (USD 0.33 million)) flow from low SC regions to low SC regions. In addition, approximately 21% of investment flows from lenders in high SC regions to borrowers in low SC regions, and 11.7% of investment flows from lenders in low SC regions to borrowers in high SC regions. The difference between the medium of each two groups (high-low) is statistically significant at the 1% level.
The results in Table 15 suggest how cross-regional investment flows are affected by the aggregate level of trust among trading partners. The aggregate level of trust is strongest when counterparties come from high SC regions and is weakest when counterparties are from low SC 41 We obtain similar results when we exclude lenders who bid only one time.
regions. If the trust level is too low, then trade opportunities are unlikely to be realized. This evidence is consistent with the findings of Guiso et al. (2009) who show that trade and investment flows are larger between countries and exhibit higher mutual trust.

VII. Conclusion
This paper presents the first empirical evidence on the impact of regional SC in a noninstitutional lending setting. Using highly granular data from a Chinese peer-to-peer lending website, we show regional SC affects lending decision and outcome through its impact on borrowers' trustworthiness and lenders' generalized trust. Ceteris paribus, borrowers from high SC regions have high funding success, large loan size, concentrated loan ownership, and low default rates. The effect is particularly strong among low-quality borrowers, and robust to endogeneity concerns. By contrast, lenders from higher SC regions make larger investments but have lower success. Regional heterogeneities in SC also affect investment flows. Cross-regional transactions are most (least) easily to be realized when counterparties are from high (low) SC regions. Our results suggest that Fintech users use non-standard soft information such as regional SC to facilitate their decision making, and SC is an important antecedent of cross-border transactions. Year dummies are also included. Panel A reports the results for the social capital index. Columns 1-2 use probit models. Columns 3-8 use OLS regressions. Panel B reports the results for the four proxies of social capital. Borrowers' personal characteristics and regional economic and financial variables are included but not reported. Robust standard errors are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. For variable definitions and details of their construction, see Appendix I.   Panel A reruns regressions by partitioning the sample into undereducated versus highly educated borrowers. A borrower is classified as highly educated if his or her highest qualification is a bachelor's degree or above (posttertiary or below). Panel B reruns regressions by partitioning the sample into repeat borrowers, or those who appear more than twice on the RRD platform (Rep), and non-repeat borrowers (Non-rep). Panel C reruns regressions by partitioning the sample into low-versus high-grade borrowers. A borrower is classified as high (low) grade if his or her credit score is below 5 (above or equal to 5). Borrowers' characteristics and regional variables are included. Dif represents the difference of the coefficient of B_SC_INDEX between two groups. Year dummies are also included. Robust standard errors are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. For variable definitions and details of their construction see, Appendix I.

Table 5 Robustness Tests
Panel A reports the robustness tests on the impact of social capital on trustworthiness. Columns 1-4 implement a bootstrapping method, which draws a subsample with half as many observations as the whole sample, and repeat our regression analysis for this subsample. Columns 5-6 employ the Heckman two-step treatment effect procedure to correct the selection bias. Columns 7-10 report the estimates that exclude Shanghai, Beijing, Gansu, Qinghai, and Ningxia. Panel B reruns regressions by partitioning the sample into low-versus high-economic development regions. Dif represents the difference of the coefficient of B_SC_INDEX between two groups. Panel C reports the difference-in-differences results using the Guo Meimei incident as shock to social capital. POST is a dummy variable for the months following the Guo Meimei incident (i.e., [1,6]). B_SC_INDEX_H is dummy variable that equals 1 if the borrower social capital index in a region is above the median, and zero otherwise. Borrowers' characteristics, regional variables, and year dummies are included. Robust standard errors clustered at province level are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. For variable definitions and details of their construction, see Appendix I.

Table 7 City-level Analysis and Time Series Evidence
This table presents the results from the regressions of the success indicator, fraction, loan size, and ownership onto citizen survey at the city level (citizen_city) and NGO Participation in year t−1 (NGO_t−1), as well as a set of control variables. Columns 1-4 present the results of city level analysis for all cities. Columns 5-8 present the results of city-level analysis for a smaller sample of 11 cities in three neighboring provinces of Yunan, Guizhou, and Sichuan. Columns 9-12 present the results of NGO participation by controlling for borrower fixed effects. Borrowers' characteristics, regional variables, and year dummies are included. Robust standard errors clustered at province level are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. For variable definitions and details of their construction, see Appendix I.

Table 8 (Borrower) Social Capital and Default Rates
This table presents the regression results of default rates for a given loan onto B_SC_INDEX, as well as different sets of control variables. Columns 1 and 2 implement probit and OLS regressions, respectively. Columns 3˗4 rerun the regression by using subsamples of undereducated versus highly educated borrowers. Columns 5, 6, and 7 employ the Heckman selection model, bootstrapping method, and instrument variable regressions, respectively. The first stage results are not reported here for brevity. Loan, borrowers' personal characteristics, and regional economic and financial variables are included, but they are also not reported. Year fixed effects are included. Robust standard errors clustered at the province level are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. For variable definitions and details of their construction, see Appendix I.

Table 11 Lender-borrower Pairs
Panel A reports the summary statistics for lenders and borrowers. We conduct t-value tests for the mean difference and Wilcoxon signed-ranks tests for the median difference. Panel B reports the summary statistics of lenders' bids. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. For variable definitions and details of their construction, see Appendix I.

Table 13 (Lender) Social Capital and Default in Lender-borrower Pairs
Panel A estimates the impact of a lender's social capital on the probability of loan default controlling for borrower fixed effects. Lenders' personal characteristics, regional economic and financial variables are included but are not reported. Panel B presents OLS and Heckprobit regressions for the impact of borrower's and lender's social capital on the probability of loan default. The difference in the other explanatory variables between lender and borrower are included but are not reported. OLS regressions include observations in which lenders' information is available. Heckman regressions include all borrowers in regressions. Year dummies are also included. Robust standard errors are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.

Table 14 Inexperienced versus Experienced Investors
This table re-estimates lender-borrower pair regressions by partitioning lenders into those who had experienced default (experienced) and those who had not (inexperienced). Columns 1, 4, 7, and 10 control only for borrower fixed effects, whereas the rest of the columns control for borrower and lender fixed effects, Borrowers' characteristics and regional variables are included. Year dummies are also included. Robust standard errors are reported in parentheses. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively. For variable definitions and details of their construction see Appendix I.

Table 15 Social Capital and Regional Capital Flows
This two-by-two matrix shows how investment flows from lenders in high/low social capital regions to borrowers in high/low social capital regions. A province is classified as a high (low) social capital region if it is above/below the sample medium. N is the number of bids, Mean/medium is the mean/medium size of the bid, and Total is the total amount of investment. We conduct t-value tests for the mean difference and Wilcoxon signed-rank tests for the median difference. ***, **, and * denote statistical significance at the 1%, 5%, and 10% levels, respectively.