Gender discrimination in online peer-to-peer lending : empirical evidence from China

(1)

Gender Discrimination in

Online Peer-to-Peer Lending:

Empirical Evidence from China

Abstract

©Y. WANG, UNIVERSITY OF AMSTERDAM, 2018

Despite mounting concerns about the role of gender differences in credit resource access within developing countries, empirical studies of credit-related gender discrimination are limited. As a result, this paper combines approaches by using the non-linear Blinder-Oaxaca method and a semiparametric propensity score matching to investigate gender discrimination in the largest peer-to-peer lending market in China. The results provide unambiguous evidence of taste-based discrimination against female borrowers in the high-risk subgroup and weaker evidence of statistical discrimination in favor of females in the low-risk subgroup. Further studies reveal that female applicants who receive loans pay higher interest rates despite having default rates lower than males.

Name: Yizhen Wang

Student Number: 11725915

Supervisor: dhr. dr. Rafael Perez Ribas

Word count: 12,107

MSc. Finance – Quantitative Finance

University of Amsterdam

(2)

Statement of Originality

This document is written by Student Yizhen Wang who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

I. Introduction

Since the 1980s, the gender gap in labor force participation has narrowed from 32% to 26.5%. As of 2018, women comprised 48.5% of the global labor force (Kuhn et al., 2018). However, the increasing participation rate has not transformed into equal employment opportunities or incomes between males and females. Women are more still likely to work in informal sectors earning lower incomes. Such gender gaps are especially salient in developing countries. The World Bank (2011) argues that three main factors drive gender earnings gaps: the length of time women allocate to take care of family and domestic work, the possibility of accessing financial resources (including assets and credit), and policy restraints and market failures. In this case, microfinance and crowdfunding, which are development tools that improve the functioning of financial credit markets, can help females access small loans and build a credit history of borrower performance. Although it is generally believed that females benefit more from small loans, researches that investigates whether gender discrimination exists in microfinance markets in developing countries is extremely rare. Thus, this paper applies empirical methods to determine whether gender discrimination exists in China’s online peer-to-peer (P2P) lending market.

Online P2P lending is a new e-commerce microfinance model that enables credit transactions between individual borrowers and lenders in online communities. Unlike traditional bank lending, P2P borrowers post unsecured loans and personal profiles on a platform by creating a listing that is viewed by lenders who then bid on small portions of customer loans through the platform. Since borrowers and lenders are not acquainted, they have no personal relationships. Thus, online P2P lending provides a broader context to examine gender discrimination in a real-life situation.

The central research question of this study explores the role of gender discrimination for credit access on Renrendai.com, one of the largest online P2P lending platforms in China. More specifically, this paper distinguishes whether the gender discrimination is pure statistical or taste-based. Statistical discrimination involves discrimination simulated by stereotypes based on the borrowers’ general behavior, such as average income and default rates. Meanwhile, taste-based discrimination indicates lenders who would simply prefer to lend money to either male or female borrowers. The discriminated group is usually offered higher interest rates for loans in credit markets.

To answer the above questions, I use Python to scrape the 2015 transaction data from the Renrendai website. Then, I apply the modified Blinder-Oaxaca (B-O) method for non-linear models (Yun, 2005) to examine gender gaps in the loan success rate. The B-O method can decompose the loan approval rate difference between male and female borrowers into an explanation derived from the difference in the characteristics and an ‘unexplained’ aspect resulting from unobserved factors or gender discrimination. After conducting the B-O decomposition for two subsamples, one subsample involves applications from high-risk borrowers and the other involves applications from low-risk borrowers. The results indicate

(5)

2 insignificant (less than 2%) gender discrimination in favor of male and female borrowers in high-risk subgroup and low-risk subgroup, respectively. However, the traditional B-O method suffers from shortcomings, including restriction of linear assumption and failure to recognize gender differences that support borrower distributions (Frolich, 2007; Ñopo, 2008; Pacheco et al., 2017). Furthermore, the majority of sample borrowers are male and approximately 20% are female. In order to solve the above problems and provide a fair comparison, I combine semiparametric propensity score matching (PSM) with the traditional Blinder-Oaxaca method to further examine the gender discrimination.

The PSM method builds fair control and treatment groups and let the data speak for themselves. The female and male borrowers are matched by demographic characteristics such as age, marital status, monthly income, education level and city of residence. Five matching methods are applied to examine gender gaps, including one-to-one matching, nearest neighbor matching, kernel-based matching, standard NN matching and bias-adjusted NN matching. Next, the paired data resulting from one-to-one matching is used to perform the B-O decompositions. The results suggest that taste-based discrimination (11%) against females exists if applicants have a credit grade of high-risk. For low-risk borrowers, however, little statistical discrimination is observed to favor females.

To gain a deeper understanding of gender discrimination in China’s online P2P lending market, this study further investigates whether fully-funded female or male borrowers pay approximately the same interest rates and whether they have relatively similar default rates. The result reveals that female borrowers on average pay higher interest rates for loans although they have lower default rates.

These findings are consistent with the research of Chen et al. (2017), who report that gender discrimination exists against female borrowers in China’s P2P lending market, even though females have lower default rates and pay higher interest rates. In contrast, studies in developed countries such as Germany and the United States suggest different results. For example, Barasinska (2009) examines the German P2P platform Smava.de, concluding that there is no significant gender discrimination. She also suggest that women are more willing to fund a loan with a lower credit grade and lower interest rate than men. Pope and Sydnor (2008) investigate the American P2P lending platform Prosper and report that single women have a greater chance of receiving funding, but at a higher interest rate than single men. Findings in the present paper reveal that for loan listings requested by low-risk borrowers, lenders focus more on the borrower’s investment returns instead of his or her possibility of default. As a result, lenders are more willing to lend money to females who offer higher interest rates in order to earn higher interests. In contrast, for loan listings requested by high-risk borrowers, although female characteristics are stronger than males in all aspects, lenders still lend money to males since they predict males have advantages in salary and promotion in the labor market; thus males offer greater long-term potential repay capability.

These findings have several implications for women’s empowerment and future policy intervention to eliminate discrimination. First, it is important for female borrowers to

(6)

3 maintain a high credit grade and leverage the new crowdfunding channels to receive small loans and build credit history. Second, women should improve their awareness and involvement in family financial decisions. Third, the government interventions should be done to standardize the P2P lending market and increase market competitions.

This paper contributes to the literature of gender discrimination in credit markets in three ways. First, discrimination research literature that focuses on developing countries is rare due to the lack of data. Moreover, studies based on the new online P2P lending market are also limited. Thus, the present study provides evidence of discrimination against female borrowers in the largest P2P lending market in a developing country. Second, this paper applies two empirical methods, the parametric Blinder-Oaxaca and the semiparametric PSM methods, which are widely used in gender pay gaps but are rarely used in loan approval rate gaps for P2P lending markets. Third, this research not only evidences gender discrimination but also distinguishes whether the discrimination is taste-based or statistical.

The reminder of this thesis is structured as follows: Section II provides a background of P2P lending markets and explores the dynamics of the Renrendai platform; Section III reviews the related literature for discrimination in financial markets; Section IV describes the overview of data, variable selection, and descriptive statistics; Section V presents the empirical methodology of the parametric Blinder-Oaxaca decomposition and semiparametric PSM; section VI provides the empirical results; and lastly, Section VII discusses the implications of the results and concludes the paper.

(7)

4

II. Background of Online Peer-to-Peer Lending

Peer-to-peer lending has been used since the early 1700s, when Jonathan Swift, the Irish author of Gulliver’s Travels, lent small amounts of money to people in need without interest (Lichtenwald, 2014). During the 18th and 19th centuries, P2P lending became one of the most common ways of lending in Europe. Although with the development of banking, P2P lending got less popular in the 20th century, it has boomed in recent years because of the growth of the internet and e-finance. The first online P2P lending platform ‘Zopa’ was established in 2005 in the UK, which was also the first time the new online lending model attracted public attention (Hulme & Wright, 2006). Today it has more than 277,000 borrowers and has lent more than £2.28 billion to its customers (Torney, 2016). Since 2005, many more P2P lending platforms have emerged, such as Lending Club and Prosper in the US, Funding Circle in the UK, etc. The success of online P2P lending markets has resulted in many new companies starting up to follow the trend.

The essential idea behind online peer-to-peer lending is the following: borrowers spend less money than on a traditional bank loan, while lenders get more interest than from bank savings or an investment. Since lending and borrowing only take place online, the transaction cuts out the role of intermediaries, and there is no application fee or processing fee such as individuals have to pay during a regular bank loan. Moreover, in P2P lending market, borrowers can usually get access to their loans within one day or even within six hours, while a regular bank loan always take weeks. Lenders can also spread their assets across different loans to realize diversification of risk.

One of the most challenge problems in P2P lending markets is information asymmetry, which can lead to adverse selection and problems of moral hazard between lenders and borrowers (Jensen & Meckling, 1976). In the traditional banking loan markets, banks can use certified accounts and collaterals to guarantee a borrower’s repayment capacity. However, it is difficult to implement such banking mechanisms with respect to the online platform, because it would lead to an increase in transaction costs. Borrowers sometimes want to hide some information to increase their chances of loan approval and to get a lower interest rate. In addition, because of the default risk of borrowers, there is no guarantee that the lenders’ money will always be paid back, and the losses will be unrecoverable. As a result, it is important for lenders to get more valid information about a borrower to determine whether to lend money to him/her or not more carefully. Online P2P lending platforms help to reduce asymmetric information between borrowers and lenders by forcing borrowers to provide comprehensive information such as demographic characteristics, financial information, social information and loan-specific information. It also screens out high-risk borrowers based on their credibility to minimize risks. This lead to another problem that for borrowers with a bad credit score, it is not easy to obtain a full loan, and they might even get stuck with a high interest rate which will cost more interest in the long term.

As in much of the rest of the world, thousands of P2P lending platforms have sprung up in China as a new loan model. Today China’s P2P lending markets are the most dynamic and

(8)

5 largest in the world, driven by the demand for funds from individuals and owners of micro and small businesses, and the supply of funds from individual investors (Deer et al., 2015). The number of platforms has risen considerably after China’s first online P2P lending platform Paipaidai started operating in 2007. By the end of 2012, there were around 300 P2P lending providers in China, and this number reached 4,000 in 2015. In addition, the transaction volume topped 2.8 trillion yuan (370 billion euros) in 2016, which represented an increase of 138% over the volume of 2015 (Jin, 2017).

P2P lending from customers and businesses account for 91% of alternative financing on the Chinese mainland. According to data published by People’s Bank of China, only 25.1% of bank card holders had loans from banks in 2015, and most of them were credit card holders. The reason is that obtaining a loan from traditional bank institutions is so complex and difficult. Online P2P lending platforms make it easier for the large majority of borrowers who have a lower net asset worth to get access to funds. For instance, more than half of Paipaidai borrowers claim that they had no loan history with traditional financial institutions such as banks or credit societies (Deer et al., 2015).

However, a number of serious problems emerged with the rapid development of China’s P2P lending markets. In 2016, the Blue Book of Internet Finance reported that more than one third of P2P companies were problematic, and half of the ‘problem platforms’ were involved in fraud that benefited from loopholes in the regulations. The most famous case occurred in January of 2016, Ezubao, one of the biggest P2P lending platforms, was allegedly a Ponzi scheme. Ezubo cheated approximately 900,000 investors out of approximately 7.6 billion by creating fake investment projects to attract funds (Leng, 2016). In another fraudulent trick, some lenders divided long-run financing schemes into different short-run loan projects and used the money from new lenders to repay the original lenders.

Therefore, in August of 2016, China’s bank regulator published a set of regulations for P2P lending. It specified a maximum loan amount on each platform of only 200,000 yuan for an individual, and 1 million yuan for a company, with a total amount of 1 million yuan and 5 million yuan, respectively. P2P companies were banned from offering guarantee for lenders, taking deposits or raising money for their own use (Yu & Li, 2016). In addition, P2P lending platforms were required to work with investor fund custodians to ensure the safety of the funds. By the end of 2016, the transaction volume dropped by half and the number of China’s P2P lending platforms dropped to 2,300, with 184 of them having established relationships with fund custodians. Today, with the implementation of stricter regulations, China’s P2P lending market is finally slowing down and back on track.

For this paper, data was collected from Renrendai.com, which is one of the largest P2P platforms in China and was funded by Shishi Zhang in 2010. By the end of 2016, its total transaction volume has reached 243 billion yuan (33 billion euros), and active membership topped 3,000,000. Although Renrendai primarily operates an offline P2P lending model, it also operates directly online as a P2P lending business for smaller loans. On the Renrendai platform, borrowers can request loan amounts of 3,000 to 200,000 yuan (400-27,000 euros), on loan terms of 3 to 36 months. They are required to offer basic information such as their

(9)

6 gender, age, income, education level, occupation, house ownership, city of residence, and so on. Some ‘hard information’ in the form of certificates is also required, such as income certificate, graduation certificate and marriage certificate. The Renrendai platform classifies borrowers into different credit grade groups (AA to HR) based on their basic information and credit information. However, because of the surplus of demand, not all borrowers can get funded on the Renrendai P2P lending platform. For example, around one quarter of the borrowers obtained access to funding in 2015, while three quarters of applications failed. These statistics inspired me to research the determinants of lending success in China’s online P2P lending markets. Do any gender discriminations exist in China’s P2P lending markets? More specifically, in the United States, “anti-discriminatory laws such as the Equal Credit Opportunity Act prohibit institutional lenders from treating equally creditworthy borrowers differently based on gender, race, age, marital status and religion.” There are no similar laws to protect people’s equal opportunities to credit in China. Also, gender discrimination has a long history and is rooted in social culture. For example, male superiority was maintained for thousands of years in Chinese feudal society, and this erroneous concept has even today not been corrected completely.

(10)

7

III. Discrimination in Financial Markets Across the Literature

There is a long history of behavioral and experimental economic studies that have examined discrimination across various markets. Theories of discrimination generally fall into two categories: taste-based discrimination and statistical discrimination. The taste-based discrimination model first proposed by Becker (1957) indicates that individuals who have a discrimination taste would prefer to work with one group over another. In this case, the discriminated group must offer more favorable terms, such as higher interest rates, in order to receive loans in financial markets. Phelps (1972) and Arrow (1973) introduce an alternative statistical discrimination model which considers discrimination simulated by stereotypes based on the minority group’s general behavior. In this model, when information is asymmetric, decision makers use stereotypes to make rational choices to maximize economic benefits. However, it is difficult to check for discrimination and even harder to determine the type of discrimination prevalent in financial markets.

A. Borrowing Discrimination in Traditional Financial Markets

Discrimination against female and minority borrowers has been empirically identified in traditional financial markets such as mortgage lending, small business lending and microfinance lending markets. Mortgage lending studies report that minorities have higher denial rates as mortgage applicants than whites (Munnell et al., 1996; Munnell et al., 1996; Hunter et al., 1996). However, no research confirms a significant gender difference for loan approval in mortgage markets. Stefani and Vacca (2013) propose that women-owned small and medium-sized companies in Germany are less likely to receive loans from banks while being more likely to face stricter loan conditions than companies owned by men. In fact, Agier and Szafarz (2013) conduct a study on a Brazilian microfinance institution, revealing a “glass ceiling” effect. More specifically, when the scale of the borrower’s project increases, the gender gap with reference to loan size also increases. Alesina et al. (2013) indicate that female business owners pay higher interest rates than men for bank credits, even if they have better credit history. Alesina et al. also find that loan costs decrease in longer relationships between banks and individual borrowers, although females benefit less than males. Hu et al. (2011) use the semiparametric PSM approach to study gender and racial interest rate differences within small business lending, concluding that companies owned by blacks and Hispanics pay higher interest rates on average than white-owned companies. However, there is no evidence suggesting interest rate gender gaps between white women and white men.

B. Borrowing Discrimination in Peer-to-peer Lending Markets

Peer-to-peer lending, as a form of internet lending, is different from traditional bank lending or microfinancing lending. It is one kind of crowd funding that one loan applicant always funds by many lenders Borrowers and lenders are not acquainted; thus, they have no personal

(11)

8 relationships. Meanwhile, their behavior may deviate from bankers and applicants who have personal contacts.

Although P2P lending is a relatively new field of study, it has already produced a number of scientific contributions. For example, Prosper published the entire platform’s comprehensive data in 2007 and enabled a wave of empirical research. After that publication, researchers from fields such as economics and social science have been investigating the new online P2P lending markets. The research direction examining determinants of funding success, final interest rate and default risk, has mainly focused on three aspects: borrowers’ demographic characteristics, borrowers’ financial characteristics and borrower’s social capital.

Regarding a borrower’s demographics, certain characteristics such as gender, age and race may cause a borrower to be rejected for a loan because of discrimination. In terms of gender discrimination, Pope and Sydnor (2008) reveal that single women have a 0.4% greater chance of receiving funding, but at a 0.4% higher interest rate with a 2% greater estimated return than single men. More recently, Chen et al. (2016) study the largest P2P lending platform in China and find that female borrowers are more likely to receive such loans than males of equal conditions. For borrowers who receive the loans, default rates for female borrowers are lower than males while on average females pay higher interest rates. Chen et al. conclude that statistical discrimination and taste-based discrimination exist in China’s P2P lending market. Similarly, Barasinska (2009) examines the effect of a lender’s gender on the loan risk variables, indicating that men are more risk averse than women while lending money to borrowers. Also, women are more willing to fund a loan with a lower credit grade and lower interest rate than men. Barasinska interprets that as suggesting that female lenders have more compassion and are more easily driven by altruistic motives.

A borrower’s race and age are determinants which cannot be ignored in P2P lending markets. Pope and Sydnor (2008) find that African Americans have a 25%–32% smaller chance of being fully funded than white Americans with similar credit grades. In the same year, Herzenstein et al. confirm that finding, reporting that African Americans are less likely to obtain full funding than others. However, Ravina (2007) contends that gender, race and age do not affect the probability of receiving funding. Yet, racial discrimination exists in the final interest rates paid by African Americas, which are generally 1.45% higher than the rates paid by white Americans. In addition, borrowers younger than 35 years of age are 0.9% more likely to receive funding than the medium group of borrowers between 35–60 years of age, while borrowers older than 60 years of age are 2.3% less likely to be funded than the same medium group (Pope & Sydnor, 2008).

C. Contributions

The present paper contributes to the literature of gender discrimination in financial and credit markets in three ways. First, because P2P lending is a relatively new credit method, literature about discrimination in developing countries is extremely rare. Nevertheless, this paper

(12)

9 reveals that gender discrimination exists in China’s P2P lending market, which is the largest and most dynamic P2P market in the world. Compared to studies in developed countries focusing on such platforms as Germany’s Smava (Barasinska, 2009) and the United States’ Prosper (Pope and Sydnor, 2008), this thesis provides a different pattern of gender discrimination.

Second, all published studies related to discrimination in P2P lending markets use simple regression models as their research methods, such as linear regressions, Logit regressions and Probit regressions. However, some papers use the Blinder-Oaxaca method or PSM, or a combination of these two methods, to investigate gender discrimination in labor markets and other financial markets. For instance, Pacgeco et al. (2017) apply PSM to the Blinder-Oaxaca decomposition to assess the gender wage gap in New Zealand. Hu et al. (2011) use the Blinder-Oaxaca and the PSM approaches to examine the gender and racial interest rate gap in the American small business lending market. This thesis applies the aforementioned advanced methods to discrimination research in P2P lending markets.

Third, most studies find that discrimination widely exists in financial markets, although few have examined whether the discrimination is purely statistical or taste-based. Nonetheless, both statistical and taste-based discrimination appears to exist in China’s P2P lending market given the Blinder-Oaxaca decomposition results.

(13)

10

IV. Data

A. Data Overview

In China, peer-to-peer lending platforms such as Renrendai do not publish information regarding transactions between their borrowers and lenders. Also, there are no official agencies which collect and publish statistics. In order to collect data about borrower’s characteristics and loan decision variables, I can only write programs to scrape the Renrendai website. Each webpage is a structured document written using HTML, and each web page corresponds to one unique uniform resource locator (URL), i.e. https://www.renrendai.com/loan/2184225. All relevant information is stored at that web address, such as the interest rate, the borrower’s marital status, credit grade, loan status and so on. It is ideal to gather all this information through web scraping and preserve the structure of each loan at the same time by using Python. I collected 23,827 listings after randomly scraping data for the whole year of 2015 from the Renrendai lending platform. Some of the listings were not complete, so I deleted these incomplete listings and kept 18,955 listings in my sample.

Figure 1. Funded and Non-funded Listings across Time by Funded Status in 2015

Figure 1 shows the number of funded and unfunded listings posted on the platform each month in 2015. The number of total applications dropped suddenly by half after January 2015, reaching 1,379 listings per month in February, rising slightly to more than 2,000 by September, and decreasing to only 640 listings per month by December 2015. However, the number of applications that received full funding is much smaller. As we can see from the solid line, the number of funded listings decreased slightly in February and then increased steadily, until they reached a peak of 861 listings per month in August. After that, they dropped dramatically in September and October to 248 listings per month. Between

(14)

11 November and December, the number of funded listings rose and remained static at around 640 per month.

When it comes to non-funded listings, indicated by a dash line, the number dropped suddenly in February, from 2,060 to 927 listings per month, and then fluctuated for the following seven months. The broken lines show a strange trend in October, when the number of funded listings and non-funded listings are approximately same. Moreover, the number of non-funded listings went down to zero in November and December, which means the probability of getting full funding was 100%. The reason is that Renrendai upgraded its platform and improved the examination and verification systems for loan application release at the end of 2015. After that, the number of loan applications posted on the platform decreased significantly, with the number of lenders far outweighing the number of borrowers. Because of this, the probability of loan funding success stabilized at almost 100%. The large number of unfunded loans motivated my interest in analyzing how this peer-to-peer lending market determines which applications to fund.

B. Determinants of Loan Funding Success

Table A1 in the Appendix provides a list of independent variables included in this paper. An earlier study by Herzenstein et al. (2008) has included two classes of factors that may influence the outcome of loan funding: the borrower’s attributes, such as gender, race, credit grade and debt-to-income ratio, etc.; and the loan decision variables, such as duration, interest rate offered, and size of loan amount. Similarly, two analogous factors are classified in the present work: (1) the borrower’s characteristics, such as gender, marital status, credit grade and city of residence in China, etc., and (2) the loan decision variables, including interest rate, duration and size of loan amount. For variable selection, the correlation test is used and some variables are omitted, such as those which are high correlated with each other (including credit grade and credit score), and the binary logistic regression model is estimated using the forward stepwise maximum likelihood method (In Lee & Koval, 1997). Loan purpose and month effects are also controlled for since although the dummy variables of loan purpose and month are individually insignificant, they are jointly significant after the F-test.

Table 1 presents descriptive statistics for loan applications during 2015. The columns illustrate the means or percentages for all variables offered for the full sample as well as respectively for females and males. The last column also indicates whether the female and male subgroups are significantly different, corresponding to each variable. The 2015 sample provides 18,955 loan listings which are unfairly split between genders. For example, only 3,758 (19.8%) of the loans were requested by women. Table 1 reveals that female borrowers have a higher average probability of being funded than male borrowers (54.2% versus 35.3%, respectively).

1. Borrower Characteristics

Three types of borrower characteristics are utilized in this empirical research: demographic characteristics, financial solvency and loan purpose.

(15)

12

Table 1. Descriptive Statistics

Variable Full Sample Female Male Significant

Difference

Loan Success Rate 0.391 0.542 0.353 0.189***

Demographic Characteristics Age 33.818 34.841 33.565 1.276*** Income 12,649 13,342 12,477 865*** Eastern 0.413 0.450 0.404 0.046*** Married 0.577 0.619 0.567 0.052*** Bachelor 0.242 0.247 0.241 0.007 Credit Situation HR 0.610 0.454 0.648 -0.194*** Certificate_num 6.247 6.857 6.096 0.761*** House Loan 0.249 0.334 0.228 0.107*** Car Loan 0.065 0.063 0.065 -0.002

Loan Decision Variables

Interest Rate 12.026 11.997 12.033 -0.035** Duration 21.818 24.943 21.045 3.898*** Loan Amount 62,403 70,031 60,517 9,514*** Loan Purpose Personal Consumption 0.192 0.242 0.180 0.062*** Capital Turnover 0.453 0.457 0.453 0.005 Home Repairs 0.130 0.102 0.137 -0.035*** Car Purchases 0.040 0.031 0.042 -0.011*** Investment/ Entrepreneurship 0.101 0.093 0.103 -0.010* Wedding 0.014 0.006 0.015 -0.009*** Home Purchases 0.020 0.018 0.020 -0.003 Education Expenses 0.008 0.007 0.008 -0.001 Medical Expenses 0.003 0.003 0.003 0.000 Unclear/ Other 0.039 0.040 0.038 0.002 Month January 0.143 0.130 0.147 -0.017*** February 0.072 0.067 0.074 -0.006 March 0.084 0.077 0.085 -0.008 April 0.095 0.093 0.096 -0.003 May 0.101 0.088 0.104 -0.016*** June 0.101 0.097 0.102 -0.004 July 0.113 0.117 0.112 0.005 August 0.106 0.105 0.106 -0.001 September 0.087 0.088 0.087 0.001 October 0.029 0.031 0.029 0.002 November 0.034 0.049 0.030 0.019*** December 0.034 0.058 0.028 0.030*** Sample Size 18,955 3,758 15,197

(16)

13

Demographic characteristics: As previously discussed, borrower demographic

characteristics are important factors to determine whether the loan application will be approved or rejected. In this case, gender, age, marital status, education level, income and city of residence are included in the demographic characteristics. In addition, gender gaps are expected to exist. The average age of all applicants is 34, and male borrowers in the sample aregenerally one year younger than females. Moreover, the average income of borrowers in the full sample of listings is 12.6k yuan per month (1.6k euros), and 41.3% of borrowers live in the eastern cities of China. More than half (57.7%) of borrowers are married, while 42.3% are single, divorced or widowed. A minority of 24.2% appear to have a bachelor’s degree or higher, while 75.8% do not hold such degrees. When comparing the results of the demographic characteristic variables between female and male borrowers, females on average have significantly higher incomes (111 euros) and are more likely to be married (5.2%) and live in the eastern part of China (4.6%).

Financial solvency: In P2P lending markets, the borrower financial characteristics related

to his or her financial solvency are also important indicators for loan success rates. Such characteristics typically include debt-to-income ratios, credit grades, house ownership, bank card utilization and open credit lines. Iyer et al. (2009) suggest that a borrower’s credit grade has an effect on funding approval. Lenders can distinguish borrowers in the most creditworthy credit grade category (AA) and the highest risk credit grade category (HR) through the credit grades defined by borrowers’ current delinquencies, credit inquiries or debt-to-income ratios. Furthermore, creditworthy credit grade borrowers are more likely to pay a lower interest rate. Klafft (2008) explains that a borrower’s credit grade is the most important determinant of interest rate, while debt-to-income ratio has a smaller but still significant effect. Klafft also points out that weak credit grade borrowers who experience difficulties to receive funding in the traditional banking system are unlikely to be funded through P2P lending. Freeman and Jin (2008) demonstrate that the average success rate of obtaining full funding increased by 1.6% between 2006 and 2008 owing to the improved financial information that P2P lending platforms provides to lenders.

Several borrower characteristics are included in the present model: number of total delinquencies, number of credit lines, credit grade, credit score, number of failed loan applications, number of successful loans, number of certifications, home ownership, home loan, car ownership and car loan. These variables indicate a borrower’s financial burden, creditworthiness and prior ability to obtain a mortgage. However, some of these financial solvency variables provide a high correlation after performing a correlation test, which may reveal a multi-collinearity problem (i.e. in respect of credit grade and credit score). Therefore, to represent a borrower’s financial solvency I only include credit grade (high risk dummy), number of certifications, house loan and car loan.

Renrendai uses seven credit grades to represent borrower credit levels and default risks: AA, A, B, C, D, E and HR. The majority (61%) of total listings constitute borrowers who belong to Renrendai’s high-risk (HR) credit grade. Individuals with an HR credit grade usually require compensation for risk by offering higher interest rates. However, the HR credit grade only constitutes 3.4% of funded loan listings, suggesting that credit grade is an

(17)

14 important factor and that borrowers with poor credit grades have few chances to receive complete funding (see Table A2 in the Appendix). Based on the results presented in Table 1, 64.8% of male borrowers have credit grades of HR, which is 19.4% higher than females (45.4%). This discrepancy may explain why female borrowers have an 18.9% higher probability to receive funding than males.

On average, six certificates are included with an application: an income certificate, marriage certificate, work certificate and education certificate, etc. These certificates help confirm the authenticity of borrower information. The other indicators of credit are whether the borrower maintains mortgages and auto loans, which comprise 24.9% and 6.5% in the total listings, respectively. Table 1 reveals that, on average, female borrowers have 0.8 more certificates and 10.7% higher probabilities to have mortgages than males.

Loan purpose: Renrendai provides 10 loan purposes, classified by the borrower’s

self-claimed descriptions: personal consumption, capital turnover, home repairs, car purchases, investment and entrepreneurship, wedding, home purchases, education expenses, medical expenses and unclear reasons. Table 1 illustrates that approximately 45.3% of listings claim the loan reason as “capital turnover.” Another popular purpose (19.2% of all listings) is personal consumption (for example, “buy a piano for my child” or “buy home electronics”), which results in a much higher probability of full funding (37.6%). Home repairs and investment or entrepreneurship are also common reasons, representing 13% and 10.1%, respectively. Other reasons account for smaller percentages, such as money for car purchases (4%), home purchases (2%), weddings (1.4%), education expenses (0.8%) and medical expenses (0.3%). When comparing the gender differences, female borrowers are 6.2% more likely to apply for loans due to personal consumption, while male borrowers are more likely to do so because of home repairs, car purchases or a wedding (3.5%, 1.1% and 0.9%, respectively).

2. Loan Decision Variables

In P2P lending markets, borrowers must list the loan amount, the loan term or duration and the offered interest rate they are willing to pay. These elements constitute the loan decision variables. In Table 1, the average interest rate, loan duration and requested loan amount of total listings are 12%, 21.8 months and 62.4k yuan (8k euros), respectively. In particular, the interest rates offered by female applicants are on average lower than those offered by males. The small difference of -0.035 supports the view mentioned above, where male borrowers with relatively lower credit grades compensate for risk by offering higher interest rates. In terms of loan duration and amount, females have longer loan duration and larger loan amounts than males (4 months and 9.5k yuan, respectively). Furthermore, 12 dummy variables are used to control for monthly effects. The trend of funded and unfunded listings over 2015 is discussed regarding Figure 1, so it is not repeated here.

In addition to a borrower demographic and financial characteristics and loan decision variables, social capital such as friends and groups are also important indicators. Herrero-Lopez (2009) reports that a borrower’s social network not only has a positive effect on the

(18)

15 loan success rate but can also help borrowers to get more reasonable interest rate. The relationship between the borrower and his/her family, friends or business partners can benefit him/her with indirect trust, to attract lenders within their second- or even third-degree social network to bid. In addition, loans with friend endorsements have a lesser default risk and obviously larger rates of return (Freeman & Jin, 2008). Greiner and Wang (2009) find that borrowers belonging to a group attract significantly lower interest rates than borrowers outside of a group. They also show that members in a group have a lesser default risk, so they are more likely to be fully funded than borrowers with no group affiliation. However, the data for members who create social networks by developing friendships with other members or joining a group after meeting the membership criteria on Renrendai are not available. As a result, social capital variables are not included in the present model.

(19)

16

V. Empirical Methodology

A. Blinder-Oaxaca Decomposition

Given the characteristics detailed in section IV, the next step investigates how large the gender differences are in lending success rates, which variables explain the gender gap and what percentage of the gap is left unexplained. However, it is hard to examine discriminations in markets, and even more difficult to assess the types of discriminations. This paper uses two empirical methodologies to test statistical and taste-based discriminations in China’s P2P lending markets.

The first of the two econometric methodologies is the traditional Blinder-Oaxaca decomposition approach (Oaxaca, 1973; Blinder, 1973). This approach decomposes the difference between two groups into an explained part (characteristics effects), derived from the difference in characteristics, and an unexplained part (coefficient effects), resulting from the inability to involve unmeasurable factors. The unexplained effect is considered a measure for discrimination.

The Blinder-Oaxaca decomposition approach is largely based on linear regression models. In contrast to the continuous variables of linear regressions, the dependent variable in the present study is equal to 1 if the loan is fully funded; otherwise, it equals 0. Thus, the estimation of non-linear regression models applied due to ordinary least squares (OLS) may produces inconsistent parameters and misleading decomposition results. However, several studies apply Blinder-Oaxaca decompositions for non-linear (i.e. logit or probit) models. For example, Fairlie (1999) proposes a new Blinder-Oaxaca decomposition technique based on binary logistic regression models to analyze racial differences in self-employment rates. Yun (2004) develops a probit Blinder-Oaxaca decomposition to examine differences in the first moment when non-linear models were used for estimation.

For standard Blinder-Oaxaca decompositions based on linear regression models, two equations are estimated, one for male borrowers and the other for female borrowers:

𝑌_𝑖𝑀₌_𝛽𝑀_𝑋

𝑖𝑀 + 𝜀𝑖𝑀 (1)

𝑌_𝑖𝐹₌_𝛽𝐹_𝑋

𝑖𝐹 + 𝜀𝑖𝐹 (2)

where superscripts M and F denote male and female applicants, Xi implies vectors of independent variables of the ith loan application, as provided in Table A1.

The lending outcome gender gap is calculated and decomposed as follow:

𝑌̅𝑀_{− 𝑌̅}𝐹₌_𝛽𝑀_𝑋̅𝑀_{− 𝛽}𝐹_𝑋̅𝐹₍₃₎

(20)

17

where 𝑋̅𝑗_{stands for the vector of average values of characteristics as listed in Table A1 and}

𝛽𝑗_{represents coefficient vectors estimated in the lending outcome equations. On the}

right-hand side of equation (4), the first term 𝛽𝑀_(𝑋̅𝑀_{− 𝑋̅}𝐹_{) represents the explained component}

and the second term (𝛽𝑀_{− 𝛽}𝐹₎_𝑋̅𝐹_{represents the unexplained component.}

In terms of Blinder-Oaxaca decompositions based on non-linear regression models, Yun (2004) supposed that a dependent variable is a linear equation of a combination of independent variables, such that:

Y = F (X β) (5)

where F implies a mapping of X (X β) to Y. The decomposition between M and F is represented as:

𝑌̅𝑀 _{− 𝑌̅}𝐹 _{= [ 𝐹(𝑋}̅̅̅̅̅̅̅̅̅̅̅̅̅ - 𝐹(𝑋𝑀_𝛽𝑀₎ ̅̅̅̅̅̅̅̅̅̅̅̅ ] + [ 𝐹(𝑋𝐹_𝛽𝑀₎ ̅̅̅̅̅̅̅̅̅̅̅̅ - 𝐹(𝑋𝐹_𝛽𝑀₎ ̅̅̅̅̅̅̅̅̅̅̅̅ ] (6) 𝐹_𝛽𝐹₎

Wherein for the right-hand side of the function, the first component [ 𝐹(𝑋̅̅̅̅̅̅̅̅̅̅̅̅̅ - 𝐹(𝑋𝑀_𝛽𝑀₎ ̅̅̅̅̅̅̅̅̅̅̅̅ ] 𝐹_𝛽𝑀₎

and the second component [ 𝐹(𝑋̅̅̅̅̅̅̅̅̅̅̅̅ - 𝐹(𝑋𝐹𝛽𝑀) ̅̅̅̅̅̅̅̅̅̅̅̅ ] represents total characteristics and 𝐹𝛽𝐹) coefficient effects, respectively. Furthermore, a detailed composition function which presents the contribution of each variable follows:

𝑌̅𝑀_{− 𝑌̅}𝐹 _{= ∑} _𝑊 𝑖𝛥𝑋 𝑇 𝑖=1 [ 𝐹(𝑋̅̅̅̅̅̅̅̅̅̅̅̅̅ - 𝐹(𝑋𝑀𝛽𝑀) ̅̅̅̅̅̅̅̅̅̅̅̅ ] 𝐹𝛽𝑀) + ∑𝑇_𝑖=1𝑊_𝑖𝛥𝛽[ 𝐹(𝑋̅̅̅̅̅̅̅̅̅̅̅̅ - 𝐹(𝑋𝐹𝛽𝑀) ̅̅̅̅̅̅̅̅̅̅̅̅ ] (7) 𝐹𝛽𝐹) where 𝑊_𝑖𝛥𝑋₌(𝑋̅𝑖𝑀−𝑋̅𝑖𝐹) 𝛽𝑖𝑀 (𝑋̅𝑀_−𝑋̅𝐹_{) 𝛽}𝑀 , 𝑊𝑖𝛥𝛽 = 𝑋̅_𝑖𝐹 (𝛽_𝑖𝑀−𝛽_𝑖𝐹) 𝑋̅𝐹_(𝛽𝑀_−𝛽𝐹₎ , and ∑𝑖=1𝑇 𝑊𝑖𝛥𝑋 = ∑𝑇𝑖=1𝑊𝑖𝛥𝛽 = 1.

As equation (7) reveals, detailed decomposition is easily obtained if the coefficient estimates are offered.

B. Propensity Score Matching

The modern way to assess the gender gap in lending outcome applies the semiparametric PSM method. Frolich (2007) and Ñopo (2008) argued that the parametric Blinder-Oaxaca decomposition method has shortcomings, such as restricting linear assumption and failing to recognize gender differences in support of borrower characteristic distributions. By applying the PSM method, the control and treatment groups could be constructed based on observed characteristics that did not interfere with the data as opposed to imposing a relationship function between the dependent and independent variables. Furthermore, the semiparametric PSM estimator is more efficient in dealing with an asymmetrically distributed sample (i.e. approximately 20% of applicants are female in our sample) since it uses sample data efficiently.

(21)

18 The first process is to perform binary logistic models for male and female borrowers in both the HR and non-HR subsamples. I assume that zi represents the likelihood of being a

female borrower, a larger value of zi indicates a higher probability of being female. As

opposed to continuous variables in linear regressions, the dependent variable ‘Female’ is equal to 1 if the borrow is female and equal to 0 otherwise. To convert this formula into the probability of being female (a number between 0 and 1) in the binary logistic regression, let: Pi = Probability (Y = 1│X) (8)

where Pi means the probability of being female.

Moreover, n independent variables are linearly related to zi as shown in the following

equation:

zi = β0 + β1xi1 + β2xi2 + … + βnxin + εi (9)

where xi implies the independent variable of listing i, and n means the number of independent

variables.

Then a logistic transformation is used to link the independent variable to the independent variables (Tranmer & Elliot, 2008):

Logit (Pi) = ln ( 𝑃𝑖 1−𝑃𝑖) = zi 𝑃𝑖 1−𝑃𝑖 =e zi_{= Exp (z} i) (10)

Now, we can get the probability function of being female, or propensity score:

Pi =

𝑒𝑧𝑖 1+𝑒𝑧𝑖 =

1

1+𝑒−𝑧𝑖 (11)

After obtaining the propensity score for each observation, all female borrowers were resampled with replacement and each female observation was matched to one or several synthetic males by applying one-to-one matching, nearest neighbor matching (N=2), kernel-based matching (EPANECHNIKOV), standard NN matching (N=3), and bias-adjusted NN matching (N=3) techniques. With one-to-one matching of HR subgroup as an example, the matching algorithm is as follows:

 Step 1: Select one female observation from the HR subsample (with replacement).  Step 2: Select one male observation to match the female who has the same (or the

nearest) propensity score.

 Step 3: Assign the lending outcome of the male observations to the female observation to check her potential application result.

(22)

19

VI. Empirical Results

A. Funding Probability - Blinder-Oaxaca Decomposition

Table 2 reports results for the Blinder-Oaxaca decomposition approach with binary logistic estimates for male and female borrowers.1_{More specifically, the table separately details the}

results for the high-risk (HR) and low-risk (non-HR) borrowers since lenders consider borrower credit grade as the most important indicator of repayment capability and default risk. Therefore, borrowers with a higher credit grade are more likely to receive full funding, in which the HR credit grade constitutes 3.4% of funded listings. The twofold Blinder-Oaxaca approach decomposed the gender gap of lending success rates into an explained component (characteristics effects) that is attributed to gender differences in characteristics and a residual unexplained component (coefficient effects). The unexplained part may derive from unobserved predictors in characteristics between females and males or gender discrimination in the lending market.

In Table 2, the final three rows present gender differences (males minus females) and the explained and unexplained components of HR and non-HR borrowers. In this case, male borrowers with credit grades of HR have a higher probability (0.692%) of being fully funded in China’s P2P lending markets. The aggregate characteristics effect, or explained component, and coefficient effect (also explained component) suggest that characteristic differences account for 21.5% (=0.149/0.692) of the total differences, while 78.5% (=0.543/0.692) are attributed to unobserved predictors, seen as a measure for gender discrimination. In contrast, female borrowers with non-HR credit grades (AA, A, B, C, D or E) are 1.773% more likely to receive full funding. Additionally, the unexplained aspect of the gender gap is larger than the unexplained, at 20.5% and 79.5% respectively.

Table 2 further reports detailed decompositions, such as coefficients and standard errors of each variable, and the gap shares of individual variables contribute to the total explained and unexplained gaps. For instance, the second column illustrates the binary logistic decomposition for male borrowers with HR credit grades. The results demonstrate that 7 of 11 variables involved in the model significantly affect the lending outcome. In general, given the decompositions of four subsamples, the age and income coefficients are both positively significant at a 99% confidence level, which implies that older borrowers with higher incomes are more likely to be fully funded due to lenders estimating that older, higher-income borrowers are more capable of loan repayment. However, age differences only explain -0.091% of the gap in the HR subgroup and 0.080% in non-HR subgroup (0.017% and 0.043% for income differences, respectively). Such limited explanation power results from the weak relationship between age, income and lending success rate.

1_{I did all calculations in Stata by applying the following commands: oaxaca (Jann, 2008) and nldecompose} (Sinning et al., 2008).

(23)

20

Table 2. Blinder-Oaxaca Decomposition Results

Dependent Variable: Loan Success Rate

HR Non-HR Estimate for Males Estimate for Females Characteristics Effect (%) Coefficient Effect (%) Estimate for Males Estimate for Females Characteristics Effect (%) Coefficient Effect (%) Age 0.062*** 0.063** -0.091 0.596 0.055*** 0.035 0.080 2.991 (0.013) (0.030) (0.016) (0.029) Ln(Income) 0.416*** 0.562* 0.017 4.471 0.657*** 0.534** 0.043 9.098 (0.099) (0.300) 0.131 (0.255) Eastern -0.049 -0.377 -0.015 0.206 -0.070 0.739 -0.000 0.122 (0.192) (0.562) (0.253) (0.483) Married 0.013 0.545 -0.019 0.112 0.029 0.148 0.020 0.247 (0.172) (0.479) (0.227) (0.445) Bachelor 0.144 -0.231 -0.005 0.138 -0.045 -0.087 -0.038 0.006 (0.175) (0.557) (0.207) (0.448) Ln(Loan Amount) -1.034*** -1.420*** 0.327 -1.930 -1.957*** -1.300*** 1.100 -56.039 (0.094) (0.297) (0.179) (0.369) Interest Rate 0.476** -0.836 -0.024 25.260 -2.670*** -2.554*** 0.437 -44.662 (0.217) (0.716) (0.157) (0.324) Duration -0.075*** 0.090 0.113 -4.331 0.367*** 0.346*** -2.366 14.699 (0.029) (0.094) (0.020) (0.037) Certificate_num 0.579*** 0.696*** -0.119 2.565 -0.105*** -0.021 -0.000 -8.082 (0.030) (0.095) (0.031) (0.073) House Loan -0.141 0.377 0.000 -0.161 0.163 -1.040** -0.006 0.636 (0.221) (0.593) (0.225) (0.480) Car Loan 0.500* -1.397 0.006 0.055 -0.553* 0.187 -0.003 -0.062 (0.273) (2.006) (0.294) (0.749) Loan Purpose X X -0.036 0.713 X X -0.016 -0.122

Month Fixed Effects X X -0.005 -0.236 X X 0.290 0.764

Constant -7.813*** 6.733 41.036*** 35.140*** (2.387) (7.819) (2.771) (5.540) Observations 9,852 1,707 5,345 2,051 Blinder-Oaxaca Difference 0.692% -1.773% Explained Component 𝛽𝑀₍_𝑋̅𝑀_-_𝑋̅𝐹₎ _0.149% _-0.364% Unexplained Component (𝛽𝑀_-_𝛽𝐹₎_𝑋̅𝐹 _0.543% _-1.409%

(24)

21 On the other hand, the disparity between male and female loan decision variables explains a substantial percentage of the gender gap. In general, a significant negative relationship exists between the requested loan amount and the funding probability, likely because larger loan amounts imply greater payment uncertainty. As predicted, a longer duration loan with a lower offered interest rate may have a higher probability of success. One explanation is that higher interest rates and shorter durations imply higher default risks since the borrower must pay a higher loan interest over a limited period.

With respect to the unexplained component, the main driver of coefficient effects stems from interest rate (25.26%) and loan amount (-56.04%) in HR and non-HR groups, respectively. Variables such as age and number of certificates are also important coefficient effects.

Overall, the unexplained component as an integral part of the gender gap suggests that there are unobserved predictors of characteristics and gender preferences as well as discrimination. However, the results do not confirm the gender discrimination due to small gender differences, 0.692% in the HR subgroup and -1.773% in the non-HR subgroup.

B. Funding Probability - Propensity Score Matching Decomposition

An additional method to evaluate the gender gap in lending outcomes is the semi-parametric approach, PSM. In the sample, the majority of applicants are males and approximately 20% are females. To propose a fair comparison of these two applicant groups, female and male borrowers, matched control (male) and treatment (female) groups were constructed by applying the PSM approach, as discussed in section IV. The matching began with a binary logistic regression of the female dummy variable for age, income in log form, marriage status, education level and city of residence. The results in panel A of Table 3 indicate that majority of demographic characteristics are significant in predicting the female dummy.

Based on the propensity scores calculated in the logit model, the matched control and treatment groups were selected using five matching estimators: one-to-one matching, nearest neighbor matching (using N=2 male borrowers), kernel-based matching (EPANECHNIKOV), standard NN matching (N=3) and bias-adjusted NN matching (N=3).2 Panel B of Table 3 illustrates the loan approval probabilities of male and female borrowers in addition to the results of gender differences for the three PSM estimators. The results fully affirm that gender gaps are significant for loan approval rates on the Renrendai platform. In the matched HR subsamples, male borrower loan success rates are between 0.3%–7.8% higher than male borrowers across the three PSM estimators. In contrast, female loan success rates are between 1.7%–2.6% higher than males in the matched low-risk subsample, or non-HR.

To further test result stability, after one-to-one matching the new data set was utilized to perform the binary logistic Blinder-Oaxaca decomposition, as done previously in Table 2.

2_{I did all calculations in Stata by applying the following commands: psmatch2 (Nichols, 2008) and nnmatch (Abadie et al.,}

(25)

22

Table 3. Propensity Score Matching Using Logit Regressions

Panel A. Regression Results

Dependent Variable: Female

Independent Variable HR Non-HR

Age 0.007** -0.004** (0.002) (0.002) Ln(Income) -0.035** -0.016 (0.018) (0.018) Married 0.096** -0.097** (0.031) (0.035) Bachelor 0.062* -0.081** (0.034) (0.035) Eastern -0.094** 0.006 (0.032) (0.032) Constant -1.001*** -0.195 (0.158) (0.173) Observations 11,559 7,396 Pseudo R-Squared 0.0042 0.0024

*significant at 10% level; **significant at 5% level; ***significant at 1% level

Panel B. Gender Difference in Lending Success under Propensity Score Matching

HR Non-HR

Propensity Score Matching Estimator

Male Female Difference Male Female Difference

One-to-one Matching 0.094 0.016 0.078*** 0.955 0.981 -0.026*** (0.008) (0.006) Nearest Neighbors (N=2) 0.023 0.016 0.007* 0.962 0.980 -0.019*** (0.004) (0.005) EPANECHNIKOV 0.023 0.016 0.007* 0.963 0.980 -0.018*** (0.004) (0.005) NN Match, N=3 0.006* -0.017*** (0.004) (0.005) NN Match, N=3 (bias-adj.) 0.006* -0.017*** (0.003) (0.004)

The method ensures that each female applicant is paired with a male applicant of the same demographic characteristics. The new HR pooled sample includes 1,684 observations with female borrowers and 1,684 matched observations with male borrowers, while the low-risk pooled sample involves 2,015 females and 2,015 males.

The decomposition results after PSM are provided in Table 4. After using more reliable samples, the gender differences in lending outcome are 11 and 1.5 times those before matching in the HR and non-HR subsamples, respectively. More specifically, male borrowers with credit grades of HR have a 7.84% higher probability of being fully funded than females in China’s P2P lending markets. The total gender difference (7.838%) is composed of characteristics effects (-3.09%) and coefficient effects (10.92%). The negative explained component reveals that female characteristics are 3.09% better than males, in total.

(26)

23

Table 4. Propensity Score Matching Decomposition Results

Dependent Variable: Loan Success Rate

HR Non-HR Estimate for Males Estimate for Females Characteristics Effect (%) Coefficient Effect (%) Estimate for Males Estimate for Females Characteristics Effect (%) Coefficient Effect (%) Age -0.019 0.087** -0.015 -10.403 0.035 0.038 -0.050 9.427 (0.022) (0.032) (0.023) (0.031) Ln(Income) -0.052 0.542* -0.002 -4.572 0.815*** 0.613** -0.046 15.722 (0.165) (0.302) (0.227) (0.269) Eastern 0.440 -0.334 0.000 0.457 -0.118 0.584 0.011 0.419 (0.325) (0.566) (0.389) (0.494) Married -0.189 0.578 -0.003 -0.267 -0.296 -0.005 0.005 -0.043 (0.279) (0.481) (0.346) (0.459) Bachelor -0.425 -0.196 0.000 -0.597 0.341 0.011 0.012 0.290 (0.318) (0.560) (0.324) (0.460) Ln(Loan Amount) -1.153*** -1.521*** 0.838 -24.795 -2.008*** -1.370*** 1.992 -68.357 (0.166) (0.312) (0.259) (0.381) Interest Rate 0.774** -0.684 -0.100 55.190 -2.917*** -2.613*** 2.096 -71.736 (0.347) (0.725) (0.241) (0.330) Duration -0.134** 0.073 0.480 -11.972 0.360*** 0.359*** -5.595 17.411 (0.047) (0.095) (0.030) (0.040) Certificate_num 0.851*** 0.701*** -3.840 30.645 -0.098** -0.070 -0.014 -6.333 (0.064) (0.096) (0.046) (0.072) House Loan 0.410 0.351 -0.107 0.180 0.013 -0.885* 0.019 0.466 (0.393) (0.596) (0.338) (0.483) Car Loan 2.090*** -1.492 -0.098 0.468 0.114 0.469 -0.003 0.024 (0.497) (2.120) (0.525) (0.760) Loan Purpose X X 0.053 1.097 X X -0.006 0.349

Month Fixed Effects X X -0.291 -5.314 X X 4.571 -1.883

Constant -1.025 5.32 44.007*** 35.624*** (3.735) (7.959) (4.215) (5.560) Observations 1,684 1,684 2,015 2,015 Blinder-Oaxaca Difference 7.838% -2.581% Explained Component -3.085% -1.122% Unexplained Component 10.923% -1.459%

(27)

24 Nonetheless, females are less likely to be funded due to the large coefficient effects (10.92%), often interpreted as taste-based discrimination against females in lending markets. On the other hand, low-risk male borrowers have 2.58% lower probability to receive full funding than females. Compared to the results provided in Table 2, the explained component increased from 20.5% to 43.5% (=1.122/2.581) after matching. The residual unexplained component (-1.459% of -2.58%) suggests that gender discrimination in low-risk subsample is negligible.

In the detailed decompositions, demographical characteristics have insignificant effects on lending outcomes and rarely explained the gender gap. This lack of explanatory power resulted from male and female applicants having similar demographic characteristics after one-to-one matching. In addition, the matching magnified the impact of loan decision and financial solvency variables on the gender gap. As with the results of traditional Oaxaca decomposition, smaller loans of longer duration with lower offered interest rates may have higher probabilities of success. A previously mentioned explanation argues that larger loans with shorter duration and higher interest rates imply higher default risks. In fact, lenders prefer to lend money to borrowers who can repay money within the borrowing time. For instance, given the descriptive statistics table (Table A3 in the Appendix), the average loan amount of male applicants is 3.6k yuan lower than females. Furthermore, the decomposed characteristics effect indicates that the negative gender difference in loan amount could explain 0.84% of the gender difference in lending outcomes as a result of the negative effect of loan amount on lending outcome.

In terms of the unexplained component in the HR subsample, the main driver of coefficient effects stems from interest rate and number of certificates. More specifically, the unexplained coefficient effects total of 10.92% is composed by the following elements: 55.19% interest rate, 30.65% number of certificates, 24.80% loan amount, 11.97% duration and -10.40% age.

In general, the Blinder-Oaxaca decompositions performed after PSM reveals clear taste-based gender discrimination against females if the borrower has an HR credit grade and insignificant statistical discrimination against males if the borrower is with a non-HR credit grade.

C. Interest Rate and Default Rate

To further understand why gender discriminations exist in funding probability in China’s P2P lending market, this paper investigates whether there are any gender differences in offered interest rates and default rates conditional on a loan is fully funded.

In terms of offered interest rates, the loan status was replaced with the interest rate as the new dependent variable and use the same combined method as in part B. First, use a propensity score matching method based on borrowers’ age, income in log form, marriage status, education level and residence city to ensure that each female applicant is paired with a male applicant. Then use the new paired data to do the linear Blinder-Oaxaca decomposition.

(28)

25

Table 5. Estimates of Gender Gap in Interest Rate.

Dependent Variable: Interest Rate Blinder-Oaxaca Decomposition

Difference -19.844%

Explained Component -21.802%

Unexplained Component 1.958%

Observations 4,008

Table 6. Estimates of Gender Gap in Default Rate.

Female Male Observations

Bad Debt 35 138 173

Closed 2,003 5,232 7,235

Observations 2,038 5,370 7,408

Due to the small sample size of HR-funded subsample, it does not show the results for the high-risk and low-risk borrowers separately. Table 5 shows the general result that female borrowers have to pay a 19.84% higher interest rate on their funded loans than males. The total gender difference (-19.84%) can be decompose into a -21.80% explained component and a 1.96% unexplained component.

In the aspect of default rates, Table 6 presents that the majority of borrowers who get fully funded repaid the money on time while only 2.34 percent of the borrowers defaulted on their loans. 138 out of 173 bad debts are from male borrowers while only 35 bad debts are from female borrowers. This implies that female borrowers in general have lower default probabilities.

In conclusion, in order to get funded in China’s P2P lending market, female borrowers have to offer higher interest rate despite their lower default probability. This result is inconsistent with Emekter et al.’s study which suggests that higher interest rates charged on the borrowers with higher default probabilities in American P2P lending market. It proves that there is costly taste-based discrimination against females in China’s P2P lending market.

Gender discrimination in online peer-to-peer lending : empirical evidence from China