The Effects of Transactions Costs and Social Distance: Evidence from a Field Experiment

Perhaps, even more indicative of the high profile of microfinance, no less a cultural touchstone than The Simpsons devoted a recent episode (“Loan-A-Lisa,” original airdate October 3, 2010) to the idea and featured Yunus as a guest star. In the episode, Lisa Simpson browses a Kiva-like website – and chooses to loan to a borrower in her own town, illustrating the potential influence of social distance on funding decisions.

See http://www.kiva.org/about/facts.

MFIs can be NGOs or for-profit organizations. The total amount originated by a loan officer is low relative to the industry standards, even in developing countries. The reason is that loans are small and that loan officers that operate in villages serve only a few borrowers. According to Kiva, the broad majority of the interest payments reflect operating expenses from serving borrowers rather than financial expenses of borrowing the money to lend to micro-entrepreneurs. Rigbi (Forthcoming) explores the effects of interest rates in Prosper.com, a similar online person-to-person loan marketplace operating in the US only.

See http://fellowsblog.kiva.org/kivablog/2010/02/10/update-on-recent-change-in-default-protection.html.

In addition to translating foreign language loan requests, Kiva’s volunteers edit requests that were originally written in English. The experimental design described above applies also to English loan requests. We find that editing English requests only slightly increases their readability, as measured by the indices described below, and that it has a small effect on the pace at which these requests are funded. Thus, we bundle edited and non-edited English requests throughout the analysis. Full results are available on request.

Due to space limitations, only the main body of the request is reproduced. The full page, including the amount requested and information about the MFI, can be viewed by following the links.

Data on languages are taken from the CIA World Factbook – http://www.cia.gov/library/publications/the-world-factbook/fields/2098.html Canada is the only country in our data that has two of the languages of interest as official languages. We define Canadian lenders as either English or French speakers, depending on their province of residence. 81.8% of Quebecois named French as their primary language (Stat Canada 2006). We therefore define Quebecois lenders as French speakers, while other Canadian lenders are defined as English speakers.

We obtain three common readability indices: Kincaid, Flesch, and Fog. The readability indices are based on the average number of words in a sentence, the average number of syllables in a sentence, and the percent of complex words defined as words with three or more syllables. For more information on the Fathom PERL package and the readability indices, see http://search.cpan.org/dist/Lingua-EN-Fathom/lib/Lingua/EN/Fathom.pm.

Based on Census ZIP code level data, we find that, on average, American Kiva lenders are at the 87th percentile of the ZIP code average income distribution and at the 92nd percentile of the share of Master’s degree and above holders ZIP code distribution.

We attempted to approximate this concept by creating observations corresponding to a loan granted by a lender on a given request as well as any loan request that was active when that loan was granted. For instance, if a lender gives into one particular request while there are six total active requests, six observations are created with the one the lender gave into being designated as a positive response. The essential assumption is that there is no instance in which lenders browse requests but decline to make any loan; this assumption, while limiting and somewhat arbitrary, reduces the possible combinations by several orders of magnitude. Regardless, the sample still results in over 160 million observations, each representing a potential lender-request combination. The percentage of potential lender-request combinations resulting in a positive response is a mere 0.10%. The results we found are similar in spirit to those in Section 6.1 and appear to provide support for the transactions costs of translation mechanism. However, the effects are necessarily extremely small, on the scale of 0.05% point changes in probability. Given the low baseline response, the need to use a set of linear probability models rather than a choice model, and the very small effects, we exclude the results of this exercise from the paper; they are available on request.

One possible concern is that since all requests are funded and lenders are participating for philanthropic reasons, they may be fairly indifferent about which requests they fund. It seems reasonable to believe, though, that lenders are not pure altruists and that they feel more a warm glow from funding requests that appeal to them.

A natural alternative to the estimated specification would be to pool requests written in a foreign language (Spanish/French) together. The results are similar to the results obtained from specification 1. However, for the results of this specification to be easily interpretable, two very strong assumptions are required. First, lenders from Spanish (French)-speaking countries face the same transactions costs for loans posted in Spanish and in French. Second, Spanish (French)-speaking country residents do not feel social distance from French (Spanish) requests. Therefore, we feel that separating the languages apart is the best approach.

The value that the dependent variable takes across equations is not independent, since the share funded by speakers of one language crowds out funding by the speakers of other languages. One way to account for this is by estimating equations simultaneously. We find that the estimates are nearly identical if equations are estimated simultaneously; the coefficients generally differ in the fourth significant digit. Note also that we exclude the share funded by speakers of languages other than English, Spanish, and French, so those shares may not sum to 1. Including an equation for the share of funds from those other countries in the jointly estimated model does not substantially alter our results; additionally, there is no obvious pattern in the coefficients of interest in that equation. Full results are available on request.

Another potentially relevant variable is the identity of the MFI originating the loan. For example, lenders might favor requests posted by one MFI due to better practices it employs in screening borrowers or due to the time the MFI has been partnering with Kiva and its portfolio performance. Few MFIs write their loan requests in more than a single language, making it difficult to identify language effects separately from MFI effects. We discuss this issue further in Section 6.2.1.

Another explanation is that requests that were translated into English might be more or less understandable than requests that were originally written in English. To test the hypothesis that readability affects lending behavior, we calculate several readability indices, designed to measure comprehension difficulty, for loan requests that were posted in English. We discuss results based on the Kincaid index, though similar results are obtained if we use the Flesch or the Fog indices instead. The value of a passage’s Kincaid index should be interpreted as the mean number of schooling years required to understand the passage. Requests that were originally written in English require 8.41 years of schooling in order to be understood; the corresponding years of schooling for translated Spanish and French request are 8.81 and 9.07, respectively. These differences are statistically significant. In addition, translated requests have lower variance. To test whether differences in loan requests’ readability are associated with changes in lenders’ giving behavior, we included a normalized readability index value in our specifications. We find no effect of a request’s readability and our main effects are qualitatively unchanged. We cannot, however, reject the possibility that more readable requests are easier to follow, resulting in lower transactions costs and more funding, but that requests written in a more sophisticated language (reflected in a higher readability index) makes them seem to be of a higher quality, also attracting more funders. If both mechanisms are operative and similar in magnitude, we would find no effect. Given the short length of these passages, the observable information on the MFI’s quality, and that lenders have opted to participate in Kiva’s mission of providing funds to those living in poverty in lesser-developed countries, the latter mechanism seems unlikely. It, therefore, seems plausible that readability does not have a large effect on funding.

See http://blog.kiva.org/2009/08/11/some-loans-may-look-a-little-different.html.

The sum of shares is lower than 1 due to missing country of residence for nearly 5% of the lenders (3874 out of 77,592).

The length of the text of a loan request might be a relevant dependent variable if there are cognitive costs associated with reading and processing requests or if the length of a request is used by lenders as a quality signal. We experimented with specifications that include the length of the loan request. We used either the number of characters or number of words as the length variables, and included each of them as a linear variable and as a second-order polynomial. We find that the estimated treatment effects are unchanged and that the length of the loan request is uncorrelated with the funding share.

Jenq, Pan, and Theseira (2011) find that the loan request characteristics we focus on, as well as other traits reflected in the borrower’s pictures such as the skin color and attractiveness of the borrowers have significant effect on the time it takes for a request to get funded.

We also estimate our model for Africa and Latin America separately. Requests from African MFIs are written in either French or English, while those from Latin America are written in either Spanish or English. The treatment effects in these models are similar to those in Table 3.

There are seven MFIs that write their loan requests in multiple languages. These MFIs are responsible for about 18% of the loan requests in our sample.