American Psychological Association 5th Edition - Chris Snijders

1 MICRO-FINANCE DECISION MAKING: A FIELD STUDY OF PROSOCIAL LENDING Jeff Galak, Deborah Small, and Andrew T. Stephen*

Jeff Galak Assistant Professor of Marketing Tepper School of Business - Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA 15213, USA Email. [email protected], Tel. +1 (412) 268 5810, Fax. +1 (412) 268 7345

Deborah Small Wroe Alderson Term Assistant Professor of Marketing Wharton School of Business - University of Pennsylvania Email. [email protected], Tel +1 (215) 898-6494, Fax. +1 (215) 898-2534

Andrew T. Stephen Assistant Professor of Marketing INSEAD Boulevard de Constance, Fontainebleau 77305, France Email. [email protected], Tel. +33 (0)1 60 71 26 47, Fax. +33 (0)1 60 74 55 00

* Jeff Galak ([email protected]) is an Assistant Professor of Marketing at the Tepper School of Business, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, USA. Deborah Small ([email protected]) is the Wroe Alderson Term Assistant Professor of Marketing at the Wharton School of Business, University of Pennsylvania. Andrew T. Stephen ([email protected]) is an Assistant Professor of Marketing at INSEAD. The authors would like to thank Kiva for providing data and the review team for their invaluable feedback on a previous version of this manuscript. This research was partially funded by the INSEAD-Wharton Alliance Fund.

Electronic copy available at: http://ssrn.com/abstract=1634949

2 MICRO-FINANCE DECISION MAKING: A FIELD STUDY OF PROSOCIAL LENDING

Abstract

Micro-financing, or small uncollateralized loans to entrepreneurs in the developing world, has recently emerged as a leading contender to cure world poverty. Our research investigates the characteristics of borrowers that engender lending in a field setting with real world and consequential data. We observe that lenders favor individual borrowers over groups or consortia of borrowers, a pattern consistent with the identifiable victim effect. They also favor borrowers that are socially proximate to themselves. Across three dimensions of social distance (gender, occupation, and first name initial) lenders prefer to give to those who are more like themselves. Finally, we discuss policy implications of these findings.

Keywords: prosocial lending, micro-finance, micro-lending, decision making, financial decision making

Electronic copy available at: http://ssrn.com/abstract=1634949

3 MICRO-FINANCE DECISION MAKING: A FIELD STUDY OF PROSOCIAL LENDING

"For those readers who ask me what they can do to help fight poverty, one option is to sit down at your computer and become a microfinancier." Nicholas D. Kristof

Research in the area of financial decision making has primarily focused on the decision maker’s welfare, such as saving for retirement (De Bondt & Thaler, 1995). In this paper, we instead focus on prosocial financial decisions, or those with theintention to affect the welfare of others who are disconnected from the decision maker. The emerging field of microfinance is especially fertile ground for studying prosocial financial decisions. Micro-financing, or small loans made to small businesses and entrepreneurs in developing countries, has emerged as a leading effort to alleviate world poverty. These types of loans have provided $25 billion in small collateral-free loans to the poorest of the poor (Diekman, 2007), and have won favor because they appeal to economically minded individuals who believe that supporting entrepreneurial endeavors will spur growth and thus do more good than traditional charitable giving (Yunus, 1999). In spite of the upsurge in popularity of microfinance, little research has examined the decisions made by individual lenders. One reason for this may be that the decision itself is difficult to classify. Is lending to the poor a financial decision much like investing, or is it rather more like a prosocial decision much like donating to charity? We propose that this context constitutes a new hybrid decision form, which we term prosocial lending. This decision is both financial in nature because it shares many characteristics

4 with conventional financial decision making (e.g. likelihood of repayment, repayment terms, etc), but is also prosocial in nature because its stated purpose is to help others. As our results will show, psychological factors that influence charitable giving decisions also exert influence in these part-financial decisions. In other words, the financial nature of the decision does not negate the psychological influences. The research to date on charitable giving, decision making tends to emphasize factors that encourage giving (e.g., Liu and Aaker, 2008; Small & Verrochi, 2009). In much of this research, individuals are approached with a request to donate to a particular charity, in which some aspect of the request is experimentally manipulated. However, in many giving contexts, much like sequential purchase decisions (Chiang, 1991; Gupta 1988; Krishnamurthi & Raj, 1988), once a consumer has decided to give, they must also decide to whom to give among a set of potentially worthy causes or recipients. Less research involves this stage of the charitable giving decision process, despite its obvious importance. Moreover, several popular charity websites put donors in the position of allocating their donations among recipients. For example on Donorschoose.org, donors read proposals written by teachers for funds for their classes and select to which class(es) to give. 1 This is also the case with the prosocial lending context we investigate in this paper. To address this gap, we examine the decisions of lenders regarding whom to lend using data on actual micro-financing lending decisions made through Kiva.org, a popular micro-lending website where decision makers (lenders) choose among a set of loan requests from budding entrepreneurs from around the world. Despite the fact that micro-lending is increasingly popular, we know of no published research examining individual decision making in this context.

1

This is not unique to the internet, but the internet has certainly made it more common.

5 The remainder of this paper is organized as follows. First, we discuss the history and unique features of the micro-financing context. Second we discuss the current literature on charitable decision making as it pertains to our hypotheses about prosocial lending. Third, we analyze real world micro-lending decisions to test our hypotheses. Finally, we discuss the implications of our results for consumer financial decision making and policy-making.

THE MICRO-FINANCING CONTEXT Micro-finance has been heralded as the modern day solution to world poverty. However, given limited lending capital, a system of direct-to-borrower financing has been established to allow individuals to make uncollateralized loans to individual entrepreneurs and small businesses in need. Among the many micro-financing organizations that act as intermediaries between lenders in the developed world and borrowers in the developing world, one not-for-profit organization has emerged as the largest and most influential player: Kiva. As of June 30, 2010 Kiva has made a total of US$145,429,775 in loans to 370,806 entrepreneurs across 200,191 loans. The loans were sourced from 463,703 lenders from 199 countries. Amazingly, despite the fact that the loans are uncollateralized, the historic repayment rate has been 98.6%, suggesting that the loans are doing exactly what they purport to do: support budding entrepreneurs grow their businesses. Kiva accomplishes its role as an intermediary by posting solicitations for loans from entrepreneurs on the Internet that are then funded by individual lenders on the Kiva.org website. The solicitations take the form of borrower profiles and contain information regarding the composition of the borrowing entity (e.g. an individual entrepreneur or a group of entrepreneurs working as a team), and some personal information about the entrepreneur(s) including their

6 name(s), location, photo(s), and a description of the purpose and nature of the loan (e.g. loan amount, loan purpose, and loan repayment term). Kiva also provides information about the field partner that sourced the loan and that will ultimately manage the loan (e.g. name of the field partner, default rate of all loans managed by the field partner, etc). For an example of a borrower’s profile, see Appendix A. Lenders have two primary decisions to make. First, the lender chooses to whom to lend. This is accomplished by browsing through a list of borrowers who are currently seeking funding (Appendix B). Second, the lender selects an amount to lend, ranging from $25 to $5,000, in $25 increments. 2 Once the loan is fully funded (i.e., 100% of the requested amount has been raised), the borrowers are responsible for repaying the loans based on pre-determined repayment schedules (which are known to the lenders). These monies are returned to the lender who can then either withdraw the funds or re-lend them to other borrowers. Of important note is the fact that neither Kiva nor the individual lenders receive interest payments. Instead, the field partners who manage the loans keep any interest payments to cover operating expenses and, in some cases, to generate a profit. Lenders are thus donating the interest generated from their loans to the field partners.

Financial Decision Making, Pro-Social Behavior, or Both? As previously mentioned, prosocial lending could be conceived as either financial or prosocial. On the one hand, the decision to lend is financial in nature: the principal of the loan is returned to the lenders (assuming the loan does not go into default) and many investment-like metrics are provided to the decision maker (e.g. field partner rating, historic default rate, loan

2

The maximum loan size allowable has since been lowered to $500, but was $5,000 in the time of our data.

7 duration, etc). All of these features might push lenders to treat the decision in a more calculative manner, which could possibly override psychological/emotional drivers (see Small, Loewenstein, and Slovic, 2007). On the other hand, the decision to lend is also pro-social in nature: lenders are donating any interest that is earned on the loan to the field partner and the stated purpose of the loan is to help those in need. Rather than force this dichotomy, we suggest that pro-social lending is a hybrid of the two. To confirm this supposition, we conducted a survey of Kiva members. Specifically, we surveyed 100 self-described Kiva members (50 male; 50 female; mean age = 43; mean number of loans made = 39; Mean number of months as a member of Kiva = 15.6) from the website kivafriends.org (an online forum devoted to discussions related to Kiva). We asked them to describe in their own words "how you decide whom to lend to." Two independent raters read through each open-ended response (mean response length = 103 words) and coded them as either containing or not containing each of two lending motivations: financial and pro-social. Of the 200 individual categorizations (two possible motivations for each of the 100 responses), the raters agreed 75% of the time (Cohen's KappaFinancial Motivation = .57; Cohen's KappaPro-Social Motivation

= .39). A third independent rater resolved all disputes. In total, 69% of the lenders

provided a lending reason that fell into one of these two categories: 23% of lenders choose whom to lend to for investing reasons, 33% choose for pro-social reasons, and 13% chose for both reasons. These results are consistent with our assertion that pro-social lending consists of elements of both financial decision making and pro-social decision making. That is, lenders explicitly refer to both motivations.

8 INSIGHTS FROM CHARITABLE DECISION MAKING There is no research on prosocial lending decisions for which we are aware, but the emerging literature on charitable decision making informs our hypotheses. In general, the causes and victims that attract generosity tend to be ones that are emotionally evocative. Emotional reactions (i.e., empathy, sympathy) to victims’ needs serve as the motivational force that overrides consumers’ otherwise self-benefitting tendencies. 3 One such emotional trigger is the identifiable victim effect: a single victim tends to evoke a stronger emotional response than multiple victims (Kogut and Ritov, 2005a; 2005b), and an identifiable victim tends to evoke a stronger emotional response than an unidentifiable victim (Small & Loewenstein, 2003). Furthermore, characteristics of donors interact with characteristics of victims and causes in important ways. In general, donors prefer to give when there exists some match or similarity between themselves and a victim or cause. Research on intergroup relations and social categorization consistently finds that people are more caring towards others in their in-group than in their out-group (e.g., Dovidio et al., 1997; Flippen, Hornstein, Siegal, & Weitzman, 1996; Levine, Cassidy, Brazier, & Reicher, 2002). Beyond group status, people are also more caring when led to believe that they are similar to the victim in values and traits (Krebs, 1975; Stotland & Dunn, 1963) and when they are led to take the perspective of the victim (Batson, Early, & Salvarani, 1997). Finally, donors target causes that help victims of the same misfortunes suffered by their friends and loved ones (Small & Simonsohn, 2008), suggesting that personal experience with misfortune facilitates misfortune-specific empathy. In all of these cases, the

3

Of course, this begs the question of whether charitable giving is ever purely unselfish. That question, however, is beyond the scope of the current paper. For a review of the literature addressing the motivations for altruistic behavior see Batson (1998).

9 purported mechanism is the reduction of social distance between victims and potential benefactors, which facilitates empathy and caring. In sum, two general mechanisms tend to affect donors’ preferences for certain victims over others. First, certain features tend to intensify emotional reactions across donors. Second, certain features correspond with donors’ unique characteristics and experiences--mitigating the social distance that typically separates victims from potential benefactors. Thus, there exists some consistency across donors in their preferences for helping certain causes and victims, and some differences based on the unique characteristics and experiences of different donors. To the extent that lenders' micro-financing decisions are, at least in part, prosocial in nature, then the factors just described should also affect them. However, if the financial considerations alter or mitigate these mechanisms, then the effects may not emerge in the data.

HYPOTHESES As we described above, several features of victims have been identified that causally influence charitable giving in experimental, largely lab-based, studies. However, as we also discussed, micro-lending is different from traditional giving. In addition, researchers and policymakers often question whether the insights gained in the lab can be extrapolated to real world situations (Levitt and List, 2007). Our primary goal is to examine the role and extent to which some of these features that predict giving in the lab also matter for real micro-financing lending decisions made by lenders on Kiva.org. Our investigation is divided into two parts, respectively addressing two research questions: (1) does borrower group size influence lending decisions? And (2) do lenders prefer to lend to borrowers that are similar to them?

10 Does Borrower Group Size Matter? First, we test the hypothesis that when lenders are deciding among potential loan recipients (i.e., borrowers), lending will decrease as group size increases, consistent with labbased evidence on charitable giving described above by Kogut and Ritov (2005). This pattern is thought to be sub-optimal because more people can benefit when resources are shared rather than concentrated on a single person. Indeed, critics of micro-financing argue that loans to individuals are often too small to influence business expansion and instead end up being spent on nonbusiness expenses (Boudreaux & Cowen, 2008). Additionally, loans to groups have had historically higher repayment rates due to peer screening (Varian, 1990) and peer monitoring (Stiglitz, 1990), making them more financially viable (Yunus, 1999). Indeed, as of the date of this writing, of all loans made via Kiva.org, only 0.55% of those made to groups of entrepreneurs defaulted, as compared to 2.07% of those made to individual entrepreneurs (χ2 (1, N = 212,998) = 67.93, p < .001). We test this hypothesis by assessing the amount of money lent and the time taken to fill a loan request as a function of the number of entrepreneurs requesting a given loan. Furthermore, we predict that, despite the presence of financially relevant information (e.g. loan risk level, loan duration, etc), the psychological influence of group size will still play a significant role. That is, controlling for the financial aspect of a lending option, the number of borrowers will continue to exert a significant influence.

Do Lenders Prefer Borrowers that are Similar to Them? Second, we examine whether lenders’ choices are influenced by perceived similarity between the borrowers and themselves. We expect that as the social distance between lenders

11 and borrowers becomes smaller, the lending likelihood increases. We test this hypothesis on three dimensions of social distance: gender, occupation, and initial of first name. The first two dimensions are more obviously important, as gender and occupation are likely salient aspects of a lender’s identity. However the third dimension, first initial of first name, may also be nontrivial. Several papers in psychology, for example, have documented a so-called “name-letter effect” in various real world preferences (Pelham et al., 2002). For example, consumers are more likely to choose a product whose brand name starts with the same first letter as their own name (Brendl et al. 2005). This effect is purportedly due to implicit egotism or a tendency to prefer things that share surface characteristics with the self. There are potentially many other dimensions of social distance that could affect lending preferences. However, we selected these three dimensions because, as described below, we have access to data about both the lenders and the borrowers on each of these three dimensions. We expect that lenders will be more likely to choose borrowers who match them on each of these three dimensions.

DATA Does Borrower Group Size Matter? Loan and Borrower Characteristics. The primary dataset used for analysis was provided by Kiva and covered 289,329 loans made to 23,024 borrowers of varying size (group size M = 1.37, SD = 1.17, min = 1, max = 20) between November 27, 2007 (the first day Kiva allowed group borrowers), and June 18, 2008. 4 Loans were made in $25 increments (M = $36.15, SD = $42.43, min = $25, max = $1,150) and the average loan request amount was $825.31 (SD =

4

Kiva originally provided partial information on 371,521 loans but we dropped some because of incomplete data (an observation was dropped if one or more variables in the regression analysis were missing for that loan).

12 $316.65, min = $75, max = $1,500). For each loan we know how much each lender provided to each entrepreneur, the total loan amount given to a borrower (loan size), the number of entrepreneurs working together as a borrower making the loan request, average borrower gender (the gender of a group of borrowers uses “average” gender, male = 0 and female = 1; historically, micro-financing has emphasized female entrepreneurs’ business activities), the loan term (in months; longer-term loans are likely less attractive to lenders), and the risk rating associated with each field partner 5 (a one to five "star" rating indicating the risk level associated with a field partner; one represents high risk and five represents low risk). The last two variables, loan term and field partner risk rating, are particularly interesting because they are factors that traditionally fall into the category of financial decision making. By comparing the effect that these variables have on the decision to whom to lend with a less financially relevant one like borrower group size, we can determine, at least in this context, the relative degree to which financial factors and non-financial, more psychological factors, play a role in lenders’ decision processes. Moreover, the last variable, field partner star rating, speaks to the creditworthiness of the borrower, an issue important in the person-to-person lending literature (Pope and Sydnor, in press; Ravina, 2010). Country characteristics. Due to the correlational nature of our data, there are several possible ways to interpret results associated with our first hypothesis. Primarily, it is possible that an unobserved variable is responsible for any relationship between the number of borrowers

5

Other factors regarding the creditworthiness (e.g. default rate and delinquency rate) of each field partner were available to us but we were unable to use them for two reasons. First, because these ratings vary over time and because Kiva did not provide us with historical values (we only have the current default and delinquency rates), we have no way of knowing what the rates were when the lenders were actually deciding whom to lend to. Second, because, as we mentioned, the default (and delinquency) rates are incredibly low, there is insufficient variance to allow for any meaningful inferences. The risk ratings, on the other hand, are relatively stable across time and provided enough variance to allow us to incorporate them into our analyses. Moreover, the risk ratings are constructed by Kiva to reflect both the objective creditworthiness metrics like default rate and the intangible ones that could only be assessed through close work with the field partner. As such, we believe that the risk rating is a superior measure of creditworthiness than any single metric that lenders are provided with.

13 requesting a loan and lenders’ propensity to fund it. To deal with this concern, we have collected additional data on a number of variables that could plausibly be responsible for any effect we might observe. By controlling for these variables, we hope to strengthen our conclusions. Our first set of control variables deal with the borrower’s home country. Specifically, for each loan request we observe the borrower(s) country (which lenders observe). For each country we then identify the 2010 GDP and infant mortality rate (World Factbook, 2010), which we treat as proxies of national wealth, and the five Hofstede Cultural Dimensions (Power-Distance, Individualism, Masculinity, Uncertainty Avoidance, and Long-Term Orientation; Hofstede 2001 6). Subjective loan characteristics. In the same vein as the country specific characteristics, we sought to measure each loan on a number of subjective dimensions that could plausibly differ across borrowers of different sizes 7. However, given the size of this task, conventional manual human coding using research assistants was impossible. Instead we turned to the Amazon Mechanical Turk (AMT) service. This service provides a forum where real people (as opposed to computerized systems) complete small “human intelligence tasks” in exchange for small amounts of compensation. For this project we created a “task” that required each AMT user to read the entirety of a loan description (exactly as it was seen by the lenders in our data set) and rate the loan on five dimensions. Specifically, for each loan request an AMT answered the following questions on five point scales: “To what extent do you think this loan will help fight poverty in the region?” (1 = Not at All, 5 = Very Much), “How complex is the task that the borrower wishes to accomplish?” (1 = Very Simple, 5 = Very Complex), “How well off is the

6

We have incomplete information on these five dimensions and did not include cases from countries for which data were missing 7 We thank an anonymous reviewer for suggesting this additional set of controls.

14 borrower(s) requesting this loan?” (1 = Very Bad Off, 5 = Very Well Off), “How likely is this borrower(s) to pay back this loan in full?” (1 = Very Unlikely, 5 = Very Likely), and “How big of a project is this borrower(s) undertaking?” (1 = Very Small Project, 5 = Very Large Project). To increase the reliability of the rating task, each borrower request was coded by three independent AMT users. Each AMT user was paid $0.05 per task (which is a typical payment rate in this platform). A total of 895 unique AMT users completed all 121,272 tasks (three per borrower request), resulting in, on average, 135.5 tasks per AMT user (SD = 244.0; min = 1; max =2,525). We subsequently computed a simple average for each loan request for each of the five measures. By controlling for these loan characteristics in our analyses, we attempt further to rule out the possibility that any differences between borrower groups of different sizes is caused by some unobserved variable.

Do Lenders Prefer Borrowers that are Similar to Them? The second hypothesis regarding the effect of social distance on lenders’ choices required additional data collection for each lender and borrower featured in the Kiva-supplied dataset. To test this hypothesis we needed data on lenders’ and borrowers’ characteristics on three social distance dimensions (gender, occupation, name). With such data, we could examine whether the prevalence of lender-borrower pairs with matching characteristics was higher than what would be expected if lenders randomly chose borrowers to fund. Borrower Data. Kiva provided near-complete data for 40,424 borrowers spanning the time period March 30, 2005 to April 11, 2008. In addition to the variables described above, we also know the gender for 40,238 borrowers (99.5%), occupation (as indicated by the stated loan

15 purpose, e.g., if it was to build a farm then the occupation was agriculture) for all borrowers, and name data for 40,264 borrowers (99.6%). Lender data. Kiva did not provide structured data for the lenders. However, lenders have the opportunity to provide such data on their profile pages on Kiva.org, and we used a data scraping program to collect these data from the public website. The program iterated over all borrower profile pages on the website and extracted the usernames of all lenders who contributed to that particular borrower. Some lenders chose to remain anonymous and so their usernames were unavailable. Through this data scraping program we identified 163,746 unique lenders who made a total of 694,552 individual loans, resulting in, on average 4.24 loans per lender (SD = 16.71, min = 1; max = 2,997). Once the usernames were identified, a second data scraping program automatically visited each lender’s public Kiva.org profile page and extracted the following lender-supplied information (when it was available): a photograph, a short description of their job or occupation, and their name. Of the 163,736 unique lenders we identified, 67,718 provided a photograph (41.4%), 118,835 provided job/occupation information (72.6%), and all provided a name. Much of the lender data was in an unstructured form that could not be directly analyzed. For example, we did not have exact gender data but in many cases gender could be inferred from a lender’s photograph. Much like with the coding of loan characteristics described above, we relied on manual human coding of the automatically collected lender data to fill these gaps. Accordingly, we again used the Amazon Mechanical Turk service. For this project we created a “task” that required each AMT user to identify the following for each lender: (1) whether or not the name provided by the lender was real (e.g. “Bob”) or not (“rainbow123”); (2) if the name was a real name, if it was that of a single person (e.g. “Bob”) or a group (e.g. “Bob and Sally”);

16 (3) if the name was real and of a single person, what that person’s first initial was (e.g. “B” for Bob); (4) if a picture was provided, was the person male, female, multiple people, or unidentifiable (many photos were of random items such as pets instead of the lenders themselves); and (5) to categorize the self-described occupation into one of the 15 occupations used by Kiva to categorize the borrowers (see Table 4 for a list of the occupation categories). To increase the reliability of the coded data, each lender was coded by two independent AMT users. Each AMT user was paid $0.01 per task. A total of 2,803 unique AMT users completed all 327,472 tasks (two per lender), resulting in, on average, 116.83 tasks per AMT user (SD = 357.74; min = 1; max = 5,847). We found coder agreement on 103,876 lenders’ first initials (Cohen's kappa = .94), 32,735 genders (Cohen's kappa = .84), and 26,075 occupations (Cohen's kappa = .36). Because many lenders lent more than once, the coding resulted in first initial data for 355,805 loans (55.9%), gender data for 152,522 loans (24.0%), and occupation data for 93,437 loans (14.7%). Country and subjective loan characteristics. We have the same data regarding country and subjective loan characteristics as described above for testing the first hypothesis for this one as well.

Data Limitations The data supplied by Kiva to test the first hypothesis included actual loan data, but for privacy reasons only provided information on the borrowers of these loans, not the lenders. Put simply, for the first hypothesis we can examine how borrower characteristics (in particular, the size of a group) affects loans, but not how specific lender characteristics also play a role. For the second hypothesis we can examine the similarities between lenders and their chosen borrowers

17 utilizing data, on both borrowers’ and lenders’ characteristics. For this analysis, we lack data on the dollar values of the loans, which for privacy reasons, could not be supplied by Kiva. In other words, we can assess whether certain borrowers are more likely to loan to certain lenders (as a function of social distance), but cannot assess whether social distance is related to the dollar amounts associated with each loan.

RESULTS Does Borrower Group Size Matter? We examine the impact of borrower group size in two ways. First, we utilize the magnitude of loans made as the dependent measure, and second, we utilize the time it took to fill a loan as the dependent measure. Together, these shed light on the relative attractiveness of different borrowers. To assess whether group size affected the sizes of loans made, we estimated a random effects regression where we regressed the amount of money that a lender loaned to a borrower on the three sets of characteristics described above in the Data section: loan and borrower characteristics (including our focal independent variable, borrower group size), borrower country characteristics, and subjective loan characteristics. We controlled for the borrower/loan industry (e.g., agriculture, retail) with a set of industry dummy variables, and for the borrower (since data were at the individual loan level) with a borrower random effect (which helps control for heterogeneity across borrowers and unobserved borrower-related characteristics that could also affect loan size). Importantly, our dependent variable was not the raw loan size. Because borrowers’ requested loan amounts decrease over time as partial loans are made, each successive potential

18 lender for a given borrower has less of an opportunity to lend larger amounts (even if they wanted to). Therefore, the amount loaned to a borrower at any time will depend on the available, incomplete requested loan size. We computed Pij as the proportion of remaining requested funding for borrower j that lender i gives. For example, if a borrower requests $1,000, the first lender can lend any amount up to $1,000. If she lends $25, then the next lender can lend any amount up to $975, and so on. This “shrinking pie” characteristic was accounted for by using Pij as a dependent variable representing the proportion of what could possibly be loaned to a given borrower at the time immediately before each loan was made. For example, if a borrower started with a request for $1,000 and loan 1 was $25, this measure was 25/1000 = .025. If loan 2 was $50, the measure was 50/(1000-25) = .0513. Thus, we regressed Pij (i.e., the proportion of the “incomplete” requested loan amount that was given) on the sets of variables listed above. [INSERT FIGURE 1 ABOUT HERE] We plot the average proportional loan size (Pij) for each borrower group size observed in the data in Figure 1. As this figure clearly illustrates, there is a nonlinear decrease in loan size as the group size increases. Because of this nonlinear distribution of the dependent variable (conditional on group size, the main predictor of interest) we estimated a generalized linear random effects model. Specifically, given the nature of the dependent variable and the apparent distribution conditional on group size in Figure 1, we fit an exponential regression model to the data. 8 For robustness, we also fit familiar first-order (linear) and second-order (quadratic) polynomial regressions (these alternative models gave similar results; see Table 1). [INSERT TABLE 1 ABOUT HERE]

8

In an exponential regression model (which is a generalized linear model) the dependent variable, Y is assumed to be exponentially distributed: Y ~ exponential(λ), where λ > 0 is the scale parameter. Then, log(λ) = Xβ where X is the matrix of regressors/covariates, and β is the vector of parameters to be estimated.

19 Standardized parameter estimates are reported in Table 1. The main model is the full exponential model (column 4). We also report a reduced exponential model with only economic factors as predictors of lending (column 3), and the linear and quadratic regression models for robustness (columns 1 and 2, respectively). The best-fitting model is model 4, the full exponential model. Consistent with our first hypothesis, there was a significant negative effect of group size on loan size (b = -.28, t = -98.79, p < .001). As the size of the borrower group increased, the proportional loan sizes decreased. This effect appeared to be robust to model specification and was also found in models 1 and 2 (linear and quadratic regressions). Further, many of the other variables also affected loan size (including a number of economic and country characteristics). However, group size had the strongest effect on proportional loan size: a test of the null hypothesis that the absolute value of this effect (.28) was equal to the absolute value of the nextlargest effect (.14) was easily rejected (p < .001). Note that it is possible that this result is an artifact of how our dependent variable has been constructed. For example, if larger loans are requested by groups than individuals then it could happen that proportionally less is given to groups than individuals whereas in absolute terms more is given to groups. We checked for this by re-running this analysis using the raw loan size as the dependent variable. 9 The same negative effect of group size on loan amount is found, indicating that this result is not merely an artifact of the construction of our dependent variable. As further evidence of the robustness of this group size effect we estimated another model using a completely different dependent variable: the time (in minutes) it took for a loan to

9

In this analysis we also controlled for the amount of the loan that had already been filled (which is embedded in the proportional loan size dependent variable in the previous analysis). The results are reported in Table WA1 of the Web Appendix.

20 be “filled” once it was announced. A loan was filled once the total amount requested by a borrower had been given by lenders. For the time-to-filled outcome we estimated a Cox proportional hazards model, regressing time-to-fill (minutes) on the same sets of variables used in the previous analysis and again focusing on the effect of borrower group size. Results are reported in Table 2. Consistent with the previous finding, we found that as borrower group size increased, the probability of an unfilled loan being filled decreased (i.e., the hazard of loan fulfillment decreased as group size increased; b = -.17, χ2 = 74.10, p < .001). A number of the other factors (e.g., economic characteristics of the loan/borrower) also affected time-to-fill, and borrower group size was one of the strongest effects. 10 [INSERT TABLE 2 ABOUT HERE] Interestingly, these results hold even when controlling for the subjective evaluations of how likely the given loans are to help the cause of curing poverty. For instance, one could argue that the reason that lenders are more likely to lend to groups of borrowers over an individual borrower is that these types of loans will be more likely to spread benefits to more individuals in need and thus help fight poverty faster. Indeed, this is exactly what the AMT users reported during the subjective loan characteristics coding task. Specifically, we observe a significant positive correlation between the number of borrowers requesting a loan (group size) and the subjective sense of how likely the loan is to solve poverty (r = .106, p < .001). Put differently, people think that loans to larger groups should have a bigger social impact. However, when controlling for this variable (and other subjective loan characteristics), we find that lenders nonetheless prefer smaller groups.

10

In this case the absolute effect size for group size was smaller than two other effects (both country characteristics): uncertainty avoidance (.21) and long-term orientation (.20). Group size was not significantly different from the largest of these two effects (p = .47).

21

Do Lenders Prefer Entrepreneurs that are Similar to Them? Our second hypothesis predicts that lenders prefer to fund borrowers who are more similar to them; i.e., less socially distant. Given the available data, we tested this hypothesis for three dimensions of social distance: gender, occupation, and first name initial. Since we did not have loan size data linked to borrowers and lenders (as discussed above) we took a different approach to testing this hypothesis with the available data. Our approach was based on comparing empirical probabilities of matches on each social distance dimension to base rates computed based on an assumed lending process whereby lenders randomly select borrowers. For example, for female lenders we compared the empirical probability of a female borrower being matched with a female lender to the base rate, which was the probability of randomly selecting a female borrower from all available borrowers (i.e., 1/Nfemale_borrowers). Since lenders’ borrower choices are not completely driven by social distance it was necessary to compute the empirical probabilities after controlling for other observed variables that could conceivably affect this choice. We conducted this analysis in three different ways to ensure that the effects were robust to variations in statistical methodology. We now describe the details of these analyses. Note that only individuals were studied since we cannot objectively identify matches between groups. Analysis counting overall social distance matches. First, for each lender-borrower pair (i, j) and for each social distance dimension (k = 1, 2, 3) we computed the variable matchijk where matchijk = 1 for lender i and borrower j if they were the same on dimension k, and 0 otherwise. For example, if lender i and borrower j were both female (e.g., lender Joanne and borrower Maureen) then matchij.gender = 1 but matchij.name = 0; however if lender i’s first name was Maureen (female) and borrower j’s first name was Mark (male) then matchij.gender = 0 but matchij.name = 1.

22 As a first test of our similarity hypothesis we compared the observed number of matches for each lender across dimensions to the number of matches that would be expected by chance given each lender’s gender, occupation, and first name initial and the distributions of the levels of these dimensions among borrowers. This involved four steps: 1. For each loan (i.e., i, j pair) we summed the matchijk variables over k. This gave us, for each loan, the number of similarity dimensions that matched (between 0 and 3; M = .628, SD = .596). 2. We computed the base rates for each lender i for each dimension k: baseratei,k. Each is the probability of a match on dimension k if lender i with a given characteristic selected a borrower randomly. For example, for gender baseratei.gender depends on lender i’s gender: if the lender is a male then baseratei.gender = 1/Nmale_borrowers and if the lender is a female then baseratei.gender = 1/Nfemale_borrowers. 3. For each loan we computed the expected number of matches if selecting borrowers randomly equal to the sum of three Bernoulli (0/1) random variables, with their respective probabilities being baseratei,k across the three similarity dimensions. 11 The mean expected chance matches (between 0 and 3) was .552 (SD = .582). 4. Finally, we tested whether the observed mean number of matches (.628) was greater than the expected number of chance matches (.552), controlling for lender (since many lenders made multiple loans), with a within-subject analysis. Indeed this was the case (F(1, 358076) = 3601.78, p < .001). Analysis counting social distance matches within each lender category. As a second test we extended this analysis to within each level of each dimension (i.e., male lenders, female

11

A Bernoulli random variable has a value of 1 (match) or 0 (no match), and has probability p = baseratei,k.

23 lenders, each lender occupation type, and each lender first name initial). For each group (e.g., male lenders) we compared the proportion with matches on that category (e.g., gender) to what would be expected by chance using the appropriate baseratei,k from above. The same withinsubject analysis was run for each lender group, and we computed the effect size for each comparison (partial η2). Thirty-six valid comparisons were run (three occupations and four first initial letters were excluded because there were no lenders in the data with those characteristics). A significant positive (at the p < .01 level) effect in the hypothesized direction was found in 17 cases (47%). Non-significant effects were found in 11 cases (31%). Significant negative effects in the opposite direction were found in eight cases (22%)—all first name initials. Meta-analyzing across the 17 significant positive effects, the mean effect sizes for gender, occupation, and first name initial were, respectively, .011, .021, and .003 (SDs = .009, .029, .003). Although this meta-analysis sample size is too small to permit statistical contrasts, the similarity effect is clearly much smaller for first name initial (by an order of magnitude). Analysis comparing matches to base rates controlling for other factors. Finally, as a third test of this hypothesis, as well as to show robustness of this finding to an alternative statistical approach and a series of control variables, we estimated models that allowed us to compute empirical match probabilities controlling for other factors that could possibility affect lenders’ choices of borrowers. In the previous two analyses we counted within-lender matches without any consideration for other observable factors that could affect the occurrence of these matches. We used the same factors as in the previous group size analysis: borrower/loan characteristics, country characteristics, and subjective loan characteristics. For this analysis we estimated three logistic regressions, one for each social distance dimension. In each model we also controlled for matches on the other two social distance dimensions (e.g., in the model for gender matching we

24 controlled for occupation and first name initial matches), since matches across dimensions may not be independent. Finding similarity effects using this approach whereby we control for other factors that could potentially affect match probabilities will provide further evidence of the robustness of this effect. In particular, it will help confirm that the results reported in the more straightforward similarity analyses above are due to lenders’ preferences for borrowers who are similar to them on some observable personal characteristics. To more clearly illustrate our approach, consider the model for gender matching. Here, matchij.gender was the binary “choice” variable. This was logistically regressed on the following covariates: matchij.occupation, matchij.name, the sets of loan and borrower characteristics mentioned above, and, importantly, a dummy variable for the lender’s gender. 12 The parameter estimates are not of primary interest but are reported for completeness in the Web Appendix in Tables WA2WA4). These estimates were used to compute the empirical probabilities of gender matches at the means of all the covariates for each category of the lender variable (in this example, lender gender). This was done by computing the predicted match probability using the logistic regression estimates under each lender category (e.g., male, female) at the means of the covariates. The dummy variable effects for lender category allow for lender category-specific intercepts. These probabilities are the empirical likelihoods of matches for each lender category. We then compared these estimated empirical match probabilities for each category of each lender social distance dimension to its corresponding base rate. If the empirical probability was significantly larger than its base rate (based on a χ2 test for comparing two proportions) we concluded that there was evidence supporting our hypothesis that the smaller the social distance

12

For occupation and first name initial models where there are more than two categories, M – 1 dummy variables are used to represent the lender’s characteristic, where M is the number of categories.

25 between a lender and a borrower, the more likely a lender is to give funds to them. We report summaries of these tests in Tables 3, 4, and 5 for each social distance dimension respectively, and an overall summary of these results across social distance dimensions in Table 6. [INSERT TABLES 3-6 ABOUT HERE] We found generally consistent evidence in support of our hypothesis across the three social distance dimensions, particularly for gender and occupation. The results were also very similar to those found in the previous analysis. For both male and female lenders, we found strong support for a gender match preference. Of the 12 represented lender occupation categories there were ten that strongly supported our hypothesis of an occupation match preference, one that only directionally supported our hypothesis, and one that went against our hypothesis (albeit not as strong as the collective evidence in support of our hypothesis). In Tables 3-5 we report odds ratios for comparing the empirical to base rate probabilities. An odds ratio significantly greater than 1 indicates support for our hypothesis. Based on Table 4, setting the odds ratios to 1 if the test was not significant, the mean odds ratio across the 12 occupations was 1.59, and the median odds ratio was 1.69. Finally, of the 22 represented letters of the alphabet there were seven for which we found strong support of our hypothesis predicting a first letter of name match preference, and two that supported our hypothesis only directionally. We did not find supporting evidence for the remaining letters. 13 Overall, across all categories of the three social distance dimensions we tested, the majority of cases provided statistically significant support for our hypothesis. Based on odds ratios in Tables 3-5 (an odds ratio significantly greater than 1 supports our hypothesis), the mean odds ratio across all tests after setting non-significant tests’ odds ratios to 1 was 1.19 (std. dev. =

13

Based on Table 5, setting the odds ratios to 1 if the test was not significant, the mean odds ratio across the 26 letters of the alphabet was .95, and the median odds ratio was 1.

26 .59; median = 1.05). This analysis suggests that lenders, in general, displayed preference for lending to borrowers with whom they shared observable personal or professional characteristics.

GENERAL DISCUSSION With the advent of micro-financing, and specifically, the ability for individuals to lend to the impoverished of the world, a richer understanding of how people make choices in this domain is paramount. This paper is a first step in understanding the psychology that underlies such prosocial lending decisions. It is also among the first large-scale field studies to examine psychological factors involved in prosocial lending. We demonstrate that, in this consequential setting, two psychological mechanisms help determine to whom lenders lend: number of borrowers and social distance. With regard to the former, we observe that lending is less likely as borrower group size increases, consistent with previous lab-based experiments on the identifiable victim effect (Kogut & Ritov, 2005a, 2005b). With regard to the latter, we observe that two dimensions of social distance influence lending decisions quite strongly, gender and occupation, and one dimension influences lending decisions to a lesser extent, fist name initial. Specifically, for each of these three, lenders prefer borrowers who are similar to them, consistent with labbased research in in-group and similarity effects in helping behavior (Dovidio et al., 1997; Flippen, Hornstein, Siegal, & Weitzman, 1996; Levine, Cassidy, Brazier, & Reicher, 2002). Importantly, this was the case even when controlling for financially relevant information such as loan duration and field partner risk ratings, suggesting that despite the financially oriented nature of micro-finance, people are nonetheless affected by group size and social distance.

27 Limitations There are at least four limitations of the results derived from the preceding analyses. First, because all conclusions are correlational in nature, we cannot conclusively rule out either omitted variable explanations nor reverse correlational accounts. It may be the case that groups of entrepreneurs fundamentally differ from individual entrepreneurs in ways other than just their size. Though we do control for as many such possibilities as the data permit, this alternative account remains a possibility. Nonetheless, the consistency between this field-based correlational finding and previous lab-based experimental research strengthens our confidence in this conclusion. Moreover, because we are unable to collect process measures to understand the underlying psychology behind lenders’ decisions, we cannot, for example, unpack the nature of the relationship between group size and propensity to lend. As can be seen in Figure 1, it appears that lenders are most likely to lend to individual borrowers and to groups of two borrowers. Though we argue that identifiability explains the high propensity to lend to individual borrowers, we can only speculate as the reason for the high propensity to lend to groups of two borrowers. It is possible, for instance, that groups of two borrowers represent a married couple, which is more unitized than two individuals (see Burson, Smith, and Faro, 2010). It is also possible that some lenders lend as units (e.g., married couples) and they actually identify more with other couples (e.g. two borrowers requesting loans together) than with individuals 14. This proposition would be interesting to explore further as it would be a natural extension of the existing identifiability literature. Second, during the time period of our data, all loans were fully funded, preventing us from drawing stronger conclusions about the characteristics of borrowers that lead to loan

14

We again thank an anonymous reviewer for this insight.

28 fulfillment. This was partly due to the fact that demand for lending (i.e., number of lenders) far outweighted supply of loans (i.e., number of borrowers). Clearly that is not the case in all lending situations. In fact, on June 30, 2010, there were 1,338 unfunded loans available on Kiva.org. To the extent that our results generalize to situations where there are more borrowers than lenders, the characteristics we identify that lead to differential preferences for certain borrowers may make a large difference in determining which borrowers receive any funding at all. Third, we are further limited in understanding the mechanisms driving the preference for socially proximate lenders. Although past research theorizes and finds increased empathy and caring toward similar others, in the context of prosocial lending it could be that other psychological ramifications of interpersonal similarity (e.g., trust), influence decisions. Moreover, in the first name initial effect, we cannot identify whether lenders choose borrowers with similar names because of implicit egotism or because first initials are correlated with other similarity dimensions, such as culture (i.e., certain first initials are more common in certain culture). Again, to the extent that we control as best we can for these cultural factors by including a number of relevant covariates in our analysis, we feel confident in our conclusions. Fourth, there has been a recent debate regarding the veracity of the name-letter effect (Pelham et al., 2002). Specifically, Simonsohn (2010) found that many of the original nameletter effects, those showing that people prefer spouses, places to live, and occupations with names similar to their own, were actually spurious and, instead were a function of other variables such as cohort, geography, and ethnicity. To the extent that the same criticisms apply to our name-letter analysis, our results may too be spurious. However, in our analysis we controlled for gender, occupation, and a number of other non-demographic related variables, thus alleviating some of the criticisms of Simonsohn. Moreover, where geography is concerned, 86.7% of all

29 loans originated in the United States or Canada, during our data window, and no loans were solicited by Americans or Canadians. This suggests that the likelihood of geography explaining the name-letter findings is minimal. Most importantly, however, even if the name-letter results we report are indeed the result of some unobserved factor that leads to a heightened sense of similarity between lenders and borrowers, our primary hypothesis still holds. In other words, if the name-letter finding is a result of a different similarity dimension, it is still the case that lenders are more likely to lend to borrowers who are similar to themselves.

Optimality of Lending Decisions One question that these results raise is whether or not lenders are acting in an optimal manner? Is a disproportionately higher preference for lending to individuals and to those who are similar to oneself reflective of optimal lender decision making? With regards to individual versus groups of borrowers, research in economics seems to suggest that lenders are, in fact, acting sub-optimally. For example, loans to individuals are often too small to effect substantial change (Boudreaux & Cowen, 2008), and so when lenders choose to lend to individual entrepreneurs, they may be helping to elevate the individual out of poverty, but they may not be helping to cure the problem of poverty, the stated goal of microfinancing. Moreover, micro-financing works best when peer screening (Varian, 1990) and peer monitoring (Stiglitz, 1990) are in place. These are systems whereby groups of entrepreneurs team up and work together on single projects to self-govern and police themselves (e.g., when one entrepreneur in a group fails to repay a their part of the loan, the remaining entrepreneurs are held accountable). Because of these peer monitoring mechanisms, the social pressure placed upon entrepreneurs to work hard in order to repay loans on time is quite strong. In fact, it is this

30 social pressure that has been herelded as the reason for the success of micro-loans and their increddibly high repayment rates (Yunus, 1999). If these peer monitoring mechanisms fall to the wayside (because of a general preference among lenders for indidvidual borrowers), then the potential successes of micro-financing may be undermined. On the other hand, recent research from experimental economics suggest that lending to groups may lead to business ventures that have higher risk profiles (Gin, Karlan, Jakiela, & Morduch, in press). It is unclear if these higher risk ventures are less desirable for the cause of curing poverty than their low risk counterparts, but to the extent that they are, funding individual entrepreneurs may actually be optimal. Of course, given that lender decisions may be sub-optimal, the obvious next question is how should the system of prosocial lending be changed to better serve it's goal of curing poverty. One option may be to create an allocation system that allows the lender to hand over the decision of to whom to lend to the lending organization (e.g. Kiva). The lender would then simply choose his or her loan amount and allow the lending organization to pick the borrower(s) who are most in need of funding based on a pre-determined allocation policy (or algorithm). This would potentially benefit the micro-lending system by eliminating any biases that the lender may have since the allocation decision would be made systematically and without the lenders input. On the other hand, one of the reasons for Kiva’s success is the level of personal involvement lenders have in choosing borrowers. Kiva asserts, and we concur, that lenders create a personal (even if only superficial) connection with borrowers and are thus more willing to loan in the first place. To the extent that taking the allocation decision away from the lender would reduce overall lending, the benefit of optimally systematic allocation of funds may need to be weighed against a potential decrease in the overall funding rate.

31 Conclusion Finally, our findings give confidence that the psychological effects on prosocial behavior, demonstrated in lab-based studies play an important role in a natural and noisy environment with real money and livelihoods at stake. More importantly, they persist in a lending context when controlling for financial attributes. Given the vast amounts of monies that are lent in this manner, understanding the underlying psychology of such lending decisions is an important first step in maintaining and improving the benefits that micro-financing has already had on the world.

32 REFERENCES Batson, C. Daniel (1998). Is there genuine altruism? In A. Nordgren & C. G. Westrin (Eds.), Altruism, society, health care: Acta Universitatis Upsaliensis. Studies in bioethics and research ethics 3 (pp. 21-36). Uppsala: Uppsala University Press. Batson, C. Daniel, Sannon Early, and Giovanni Salvarani (1997), "Perspective Taking: Imagining How Another Feels Versus Imagining How You Would Feel," Personality and Social Psychology Bulletin, 23(7), 761-758. Boudreaux, Karol and Tyler Cowen (2008), "The Micromagic of Microcredit," The Wilson Quarterly, 32(1), 27-31. Brendl, Miguel, Amitava Chattopadhyay, Brett W. Pelham and Mauricio Carvallo (2005), “Name Letter Branding: Valence Transfers when Product Specific Needs are Active,” Journal of Consumer Research, 32(3), 405-415. Burson, Katherine, Robert Smith, and David Faro (2010). Flocks, Herds, and Families: The Influence of Victim-Unitization on Charitable Giving. Working Paper. Chiang, Jeongwen (1991), "A simultaneous approach to the whether, what and how much to buy questions," Marketing Science, 10(4), 297–315. De Bondt, Werner. F.M. and Richard H. Thaler (1995), "Financial Decision-Making in Markets and Firms: A Behavioral Perspective." In R. Jarrow (ed), Handbooks in Operations Research and Management Science (9 ed., pp. 385-410). Elsevier. Diekman, R. (2007). Microfinancing: An Emerging Investment Opportunity. Uniting social investment and financial returns. Deutsche Bank: Frankfurt, Germany.

33 Dovidio, John F., Kerry Kawakami, Craig Johnson, Brenda Johnson, and Adaiah Howard, (1997), "On the Nature of Prejudice: Automatic and Controlled Processes," Journal of Experimental Social Psychology, 33, 510-540. Flippen, Annette. R, Harvey A. Hornstein, William E. Siegal, and Eben A. Weitzman (1996), "A Comparison of Similarity and Interdependence as Triggers for In-Group Formation," Personality and Social Psychology Bulletin, 22(9), 882-893. Gin, Xavier, Dean Karlan, Pamela Jakiela, and Jonathan Morduch (in Press), "Microfinance Games", American Economic Journal: Applied Economics. Gupta, Sunil (1988), "Impact of sales promotions on when, what, and how much to buy", Journal or Marketing Research, 25 (4), 342-355. Hofstede, Gurte (2001), Culture's Consequences, Sage Publication Inc, Thousand Oaks, California. Kogut, Tehila and Ilana Ritov (2005a), "The “identified victim” effect: an identified group or just an individual?" Journal of Behavioral Decision Making, 18, 157-167. Kogut, Tehila and Ilana Ritov (2005b), "The singularity effect of identified victims in separate and joint evaluation," Organizational Behavior and Human Decision Processes, 97, 106-116. Krebs, Dennis (1975), "Empathy and altruism," Journal of Personality and Social Psychology, 32 (6), 1134-1146. Krishnamurthi, Lakshman and S. P. Raj (1988), "A Model of Brand Choice and Purchase Quantity Price Sensitivities," Marketing Science, 7(1), 1-20. Levitt, Steven D. and John A. List (2007), " What Do Laboratory Experiments Measuring Social Preferences Reveal About the Real World?," Journal of Economic Perspectives, 21 (2), 153174.

34 Levine, Mark, Clare Cassidy, Gemma Brazier, Stephen Reicher (2002), "Self-categorisation and bystander non-intervention: two experimental studies," Journal of Applied Social Psychology, 7, 1452-1463. Liu, Wendy and Jennifer Aaker (2008), "The Happiness of Giving: The Time-Ask Effect," Journal of Consumer Research, 35(3), 542-557. Pelham, Brett. W., Mauricio Carvallo, and John T. Jones (2005), "Implicit Egotism," Current Directions in Psychological Science, 14(2), 106-110. Pope, Devin G., and Justin R Sydnor (in press), “What’s in a Picture? Evidence of Discrimination from Prosper.com,” Journal of Human Resources. Ravina, Enrichetta (2010), “Love & Loans: The Effect of Beauty and Personal Characteristics in Credit Markets”, working paper, Columbia University, New York City, NY. Simonsohn, Uri. (2010), "Spurious? Name Similarity Effects (Implicit Egotism) In Marriage, Job and Moving Decisions," Unpublished Manuscript. Philadelphia, PA. Small, Deborah. A and George Loewenstein (2003), "Helping a Victim or Helping the Victim : Altruism and Identifiability," The journal of risk and uncertainty, 26(1), 5-16. Small, Deborah. A. and Uri Simonsohn (2008), "Friends of Victims: Personal Experience and Prosocial Behavior, "Journal of Consumer Research, 46(6), 777-787. Small, Deborah. A., George Loewenstein and Paul Slovic (2007), "Sympathy and callousness: The impact of deliberative thought on donations to identifiable and statistical victims," Organizational Behavior and Human Decision Processes, 102(2), 143-153. Small, Deborah.A. and Nicole M. Verrochi (2009), "The Face of Need: Facial Emotion Expression on Charity Advertisements," Journal of Marketing Research, 46(6), 777-787.

35 Stiglitz, Joseph. E. (1990), "Peer Monitoring and Credit Markets," World Bank Economics Review, 4(3), 351-366. Stotland, Ezra and Robert E. Dunn (1963), "Empathy, self-esteem, and birth order," The Journal of Abnormal and Social Psychology, 66(6), 532-540. Varian, Hal. R. (1990), "Monitoring Agents With Other Agents," Journal of Institutional and Theoretical Economics, 146, 153-174. World Factbook (2010) Central Intelligence Agency, Washington, DC. Yunus, Muhammad (1999). Banker to the Poor. New York, NY: Public Affairs.

36 FIGURE 1 AVERAGE LOAN SIZE AT DIFFERENT LEVELS OF BORROWER GROUP SIZE

Note: error bars represent one standard error above/below the mean.

37 TABLE 1 FACTORS AFFECTING LOAN SIZE Dependent Variable: Percentage of Remaining Requested Money that is Loaned to Borrowers (Pij) (1) Linear Regression

(2) Quadratic Regression

(3) Exponential Regression

(4) Exponential Regression

Loan and Borrower Characteristics Loan term (months)

-1.96* (-31.30)

-2.10* (-33.61)

-.02* (-8.30)

-.09* (-21.39)

Field partner rating

.48* (6.15)

.32* (4.15)

.06* (18.76)

.03* (5.58)

.71* (12.74)

.96* (17.20)

-3.80* (-88.87)

Borrower gender (0 = male, 1 = female; average gender if group) Borrower group size (number of people) Group size2

---

---

.04* (10.90)

-6.71* (-85.65)

---

-.28* (-98.79)

.68* (44.30)

---

---

Borrower Country Characteristics log(GDP)

2.45* (29.80)

2.21* (26.91)

.03* (9.08)

.10* (18.91)

Death rate (deaths per 1,000 live births)

-.10 (-1.68)

-.03 (-.51)

-.02* (-9.21)

.00 (.18)

Power distance

.97* (16.19)

.56* (9.35)

---

.01* (3.08)

Individualism

-1.11* (-19.43)

-1.02* (-17.80)

---

-.06* (-14.93)

Masculinity

-2.34* (-26.83)

-2.19* (-25.24)

---

-.07* (-12.79)

Uncertainty avoidance

-3.40* (-34.23)

-3.50* (-35.31)

---

-.14* (-21.33)

Long-term orientation

-2.54* (-28.41)

-2.62* (-29.35)

---

-.10* (-16.42)

38

Subjective Loan Characteristics Ability of this loan to fight poverty in region

.02 (.32)

.05 (1.01)

---

.00 (-.08)

Complexity of project

-.22* (-4.81)

-.19* (-4.15)

---

-.01* (-3.60)

How well off is the borrower

-.12* (-2.54)

-.13* (-2.68)

---

-.01* (-2.56)

Likely to pay back loan

-.19* (-4.06)

-.16* (-3.39)

---

-.01* (-3.29)

Size of project

-.48* (-9.87)

-.41* (-8.53)

---

-.03* (-8.36)

Intercept

14.16* (28.61)

13.87* (28.12)

2.69* (122.24)

2.63* (81.32)

Borrower random effect variance

553.10* (380.34)

549.38* (380.34)

2.00* (471.00)

2.36* (373.83)

Significant* borrower industry dummy effects?

Yes

Yes

Yes

Yes

Number of observations

289,329

289,329

289,329

289,329

-2 Log-likelihood

2,648,389

2,646,440

1,567,095

1,070,151

Akaike Information Criterion (AIC)

2,648,455

2,646,508

1,567,137

1,070,217

Bayesian Information Criterion (BIC)

2,648,804

2,646,868

1,567,359

1,070,566

Pseudo-R2

.21

.22

Standardized parameter estimates are reported. Values in parentheses are t-statistics. * p < .01.

.06

.22

39 TABLE 2 FACTORS AFFECTING TIME TO FILL LOAN Time-to-Filled (minutes) Loan and Borrower Characteristics Loan term (months) Field partner rating Borrower gender (0 = male, 1 = female; average gender if group) Borrower group size (number of people)

-.08* (12.09) -.05 (4.65) .09* (22.00) -.17* (74.10)

Borrower Country Characteristics log(GDP) Death rate (deaths per 1,000 live births) Power distance Individualism Masculinity Uncertainty avoidance Long-term orientation

-.06 (2.84) .15* (44.35) .04 (2.07) -.14* (50.43) .13* (10.83) .21* (13.76) .20* (17.72)

Subjective Loan Characteristics Ability of this loan to fight poverty in region Complexity of project How well off is the borrower Likely to pay back loan Size of project Significant* borrower industry dummy effects? Number of observations (borrowers) -2 Log-likelihood Akaike Information Criterion (AIC) Bayesian Information Criterion (BIC)

.03 (2.86) -.01 (.49) -.01 (.09) .05* (10.49) -.02 (1.49) Yes 4,590 67,562 67,622 67,815

Standardized parameter estimates are reported. Values in parentheses are χ2 statistics. * p < .001.

40 TABLE 3 COMPARISON OF EMPIRICAL PROBABILITIES TO BASE RATES FOR GENDER SIMILARITY

*

Lender’s Gender

Base Rate

Empirical Probability Estimated with Logit Model

Male

.252

Female

.748

p < .001.

χ2

Supports Hypothesis?

(df = 1)

Odds Ratio (Empirical/Base)

.437

5707.58*

1.74

Yes

.829

277.76*

1.10

Yes

41 TABLE 4 COMPARISON OF EMPIRICAL PROBABILITIES TO BASE RATES FOR OCCUPATION SIMILARITY

Lender’s Occupation

Base Rate

Empirical Probability Estimated with Logit Model

χ2 (df = 1)

Odds Ratio (Empirical/Base)

Supports Hypothesis?

Agriculture

.168

.284

281.50*

1.69

Yes

Arts

.030

.051

209.43*

1.71

Yes

Clothing

.107

.292

79.40*

2.74

Yes

Construction

.019

.034

66.45*

1.78

Yes

Education

.002

.001

14.05*

.70

No

Food

.274

.390

126.32*

1.42

Yes

Health

.018

.032

301.32*

1.77

Yes

Housing

.024

.040

39.31*

1.68

Yes

Manufacturing

.020

.024

2.15

1.20

Directional

Retail

.198

.241

43.87*

1.22

Yes

Services

.103

.117

87.85*

1.14

Yes

Transportation

.028

.062

1073.87*

2.18

Yes

The following lender occupations were excluded from this analysis because there were no observations where there was a match: entertainment, personal use, wholesale. Statistical tests therefore could not be computed for these occupation categories. * p < .001.

42 TABLE 5 COMPARISON OF EMPIRICAL PROBABILITIES TO BASE RATES FOR FIRST NAME INITIAL SIMILARITY Empirical Supports Lender’s Probability Hypothesis? Odds Ratio χ2 First Base Rate Estimated (Empirical/Base) (df = 1) Initial with Logit Model A Directional .107 .108 .02 1.00 * B No .031 .019 39.00 .60 * C Yes .034 .042 51.12 1.25 * Yes D .029 .034 10.45 1.17 * E No .034 .026 7.46 .76 * F No .035 .001 47.21 .03 * G No .030 .026 8.95 .84 * H No .027 .006 54.91 .23 * I No .022 .007 14.38 .33 J Directional .051 .051 .21 .99 * K No .052 .042 25.93 .82 * L Yes .035 .052 114.70 1.50 ** M Yes .131 .136 5.12 1.04 N No .058 .053 2.16 .92 O No .014 .009 3.21 .67 P Directional .029 .031 1.35 1.07 * R No .052 .047 12.81 .89 ** S Yes .069 .073 5.63 1.06 * T No .033 .026 31.85 .78 * V No .024 .008 37.78 .35 * Yes Y .014 .031 17.88 2.19 * Z Yes .013 .026 372.37 1.95 The following first letter initials were excluded from this analysis because there were no observations where there was a match: Q, U, W, X. Statistical tests therefore could not be computed for these occupation categories. * p < .001, ** p < .05.

43 TABLE 6 SUMMARY OF SOCIAL DISTANCE MATCHING RESULTS

Social Distance Measure

Number of Lender Categories in Model

Gender

2

Occupation

12

First name initial

22

Totals

36

Support Hypothesis Significant in Hypothesized Direction

Directional Support Only

2 (100%) 10 (83.3%) 7 (31.8%) 19 (52.8%)

0 (0%) 1 (8.3%) 2 (9.1%) 3 (8.3%)

Against Hypothesis Directional Significant in Support in Opposite Opposite Direction Direction 0 0 (0%) (0%) 1 0 (8.3%) (0%) 10 3 (45.5%) (13.6%) 11 3 (30.6%) (8.3%)

44 APPENDIX A SAMPLE BORROWER LOAN REQUEST PAGE ON KIVA.ORG

45 APPENDIX B SAMPLE LIST OF BORROWERS REQUESTING LOANS ON KIVA.ORG

46 WEB APPENDIX TABLE WA1 FACTORS AFFECTING LOAN SIZE: DEPENDENT VARIABLE = RAW LOAN SIZE Dependent Variable: Raw Loan Size = Amount of Money that Borrower j Loans to Lender i Exponential Regression Loan and Borrower Characteristics Loan term (months)

.022* (3.86)

Field partner rating

.013 (1.70)

Amount of loan already filled by prior borrowers

.001* (12.60)

Borrower gender (0 = male, 1 = female; average gender if group)

.005 (.91)

Borrower group size (number of people)

-.010* (-2.46)

Borrower Country Characteristics log(GDP)

-.024* (-2.60)

Death rate (deaths per 1,000 live births)

-.005 (-1.00)

Power distance

.020* (-3.38)

Individualism

-0.010 (-1.82)

Masculinity

.020* (2.08)


.038* (3.25)


.035* (3.39)

47

Subjective Loan Characteristics Ability of this loan to fight poverty in region Complexity of project

-.003 (-.76) .000 (.01)


-.01* (-2.07)


-.007 (-1.54)

Size of project Intercept

.010* (2.27) 3.562* (93.22)

Borrower random effect variance

2.117* (262.40)

Significant* borrower industry dummy effects?

Yes

Number of observations

289,329

-2 Log-likelihood

494,401

Akaike Information Criterion (AIC)

494,467

Bayesian Information Criterion (BIC)

494,816

Pseudo-R2

.33

Standardized parameter estimates are reported. Values in parentheses are t-statistics. * p < .05.

48 TABLE WA2 LOGIT MODEL ESTIMATES FOR SOCIAL DISTANCE MATCHING ANALYSIS: GENDER

Standardized Covariates

Intercept Lender gender = male Lender gender = female matchij.occupation matchij.name

Standardized Estimate (χ2-statistic) .33* (2925.17) -.91* (36709.02) Baseline .09* (389.46) -.03* (39.82)

Economic Factors Loan term (months)

.08* (166.53)

Log(GDP)

-.11* (173.61)

Death rate (deaths per 1,000 live births) Field partner rating

-.04* (36.34) -.01* (6.59)

National Culture Power distance

.02* (12.17)

Individualism

.09* (291.46)

Masculinity

.04* (21.24)

49


-.20* (312.42)


-.32* (931.24)

Borrower Factors Borrower group size (number of people)

-.23* (846.37)

Perceptions of Loan Ability of this loan to fight poverty in region Complexity of project How well off is the borrower Likely to pay back loan

-.01* (8.85) .02* (14.42) .01 (2.54) .03* (23.08)

-.00 (.10) Number of observations = 131,302, -2LL = 293,281, AIC = 293,321, BIC = 293,529. * p < .001. Size of project

50 TABLE WA3 LOGIT MODEL ESTIMATES FOR SOCIAL DISTANCE MATCHING ANALYSIS: OCCUPATION


Intercept Lender occupation = agriculture Lender occupation = arts Lender occupation = clothing Lender occupation = construction

Standardized Estimate (χ2-statistic) -2.72* (11984.25) 1.80* (1781.82) -.19* (21.63) 1.84* (202.40) -.63* (82.12)

Lender occupation = education

-4.13* (2166.58)

Lender occupation = food

2.28* (2285.19)

Lender occupation = health

-.70* (335.21)

Lender occupation = housing

-.46* (31.50)

Lender occupation = manufacturing

-.97* (68.81)

Lender occupation = retail Lender occupation = services Lender occupation = transportation matchij.gender

1.58* (1459.74) .71* (716.05) Baseline .18* (300.67)

51

matchij.name

-.06* (28.99)

Economic Factors Loan term (months)

.12* (83.81)

Log(GDP)

-.14* (48.87)

Death rate (deaths per 1,000 live births) Field partner rating

-.19* (161.27) -.06* (32.16)


-.21* (172.02)

Individualism

.10* (38.58)

Masculinity

.14* (51.66)


-.03 (1.47)


.03 (1.70)


.09* (38.58)

Perceptions of Loan Ability of this loan to fight poverty in region Complexity of project

-.07* (47.45) .13* (125.09)


.05* (22.56)


-.08* (48.53)

52

Size of project

.02 (3.92)

Number of observations = 233,523, -2LL = 72,553, AIC = 72,613, BIC = 72,824. The following lender occupations were excluded from the model because there were no observations where there was a match: entertainment, personal use, wholesale. * p < .001.

53 TABLE WA4 LOGIT MODEL ESTIMATES FOR SOCIAL DISTANCE MATCHING ANALYSIS: FIRST NAME INITIAL


Intercept

Standardized Estimate (χ2-statistic) -3.64* (10644.07)

Lender first name initial = A

1.53* (1220.42)

Lender first name initial = B

-.33* (14.26)

Lender first name initial = C

.52* (130.01)

Lender first name initial = D

.29* (25.83)

Lender first name initial = E

.08 (1.07)

Lender first name initial = F

-3.11* (42.23)

Lender first name initial = G

-.04 (.28)

Lender first name initial = H

-1.42* (53.44)

Lender first name initial = I

-1.27* (20.87)

Lender first name initial = J

.71* (288.04)

Lender first name initial = K

.52* (101.85)

Lender first name initial = L

.74* (215.60)

Lender first name initial = M

1.79* (2200.18)

54

Lender first name initial = N

.76* (144.18)

Lender first name initial = O

-1.01* (21.61)

Lender first name initial = P

.19* (8.52)

Lender first name initial = R

.62* (183.19)

Lender first name initial = S

1.10* (728.85)

Lender first name initial = T

-.01 (.05)

Lender first name initial = V

-1.12* (44.59)

Lender first name initial = Y

.19 (1.12)

Lender first name initial = Z

Baseline

matchij.gender

-.07* (65.68)

matchij.occupation

-.06* (43.60)

Economic Factors Loan term (months) Log(GDP) Death rate (deaths per 1,000 live births) Field partner rating

-.01 (.54) .21* (159.80) -.02 (1.34) -.07* (55.38)


-.13* (96.53)

Individualism

-.03* (6.75)

55

Masculinity

.01 (.53)


-.21* (103.93)


-.27* (181.12)


.11* (57.68)

Perceptions of Loan Ability of this loan to fight poverty in region Complexity of project

.06* (37.78) .01 (2.09)


.00 (.14)


-.01 (.68)

.04* (17.64) Number of observations = 240,578, -2LL = 103,051, AIC = 103,131, BIC = 103,547. The following lender first name initials were excluded from the model because there were no observations where there was a match: Q, U, W, X. * p < .001. Size of project