Long-Run Price Elasticities of Demand for Credit: Evidence from a Countrywide Field Experiment in Mexico 1

Transcription

1 Long-Run Price Elasticities of Demand for Credit: Evidence from a Countrywide Field Experiment in Mexico 1 Dean Karlan Northwestern University Innovations for Poverty Action M.I.T. Jameel Poverty Action Lab NBER Jonathan Zinman Dartmouth College Innovations for Poverty Action M.I.T. Jameel Poverty Action Lab NBER October 2017 Abstract We use randomized interest rates, offered across 80 geographically distinct regions for 29 months by Mexico s largest microlender, to sketch the adjustment from a price change to a new equilibrium. Demand is elastic, and more so over the longer-run; e.g., the dollarsborrowed elasticity increases from -1.1 in year one to -2.9 in year three. Credit bureau data does not show evidence of crowd-out, although this and other null results are imprecisely estimated. The lender s profits increase, albeit noisily, starting in year two. But competitors do not respond by reducing rates. These findings, together with other results, suggest that informational frictions are important, and that cutting rates furthered Compartamos Banco s double bottom line of improving social welfare subject to a profitability constraint. 1 karlan@kellogg.northwestern.edu, jzinman@dartmouth.edu. Thanks to Kerry Brennan, Angela Garcia Vargas, Matt Grant, Kareem Haggag, Rachel Strohm, and Natalia Torres for excellent project management and research assistance, with special thanks to Martin Sweeney for his continued support. Thanks to Alissa Fishbane, Braulio Torres and Anna York from Innovations for Poverty Action for leadership of IPA- Mexico. Thanks to Abhijit Banerjee, Esther Duflo, Jake Kendall, Melanie Morten, David Roodman, Chris Snyder and participants in seminars at NBER Household Finance, M.I.T./Harvard and NYU for comments. Thanks to CGAP, in particular Richard Rosenberg, and the Bill and Melinda Gates Foundation for funding support. Thanks to the management and staff of Compartamos Banco for their cooperation. The authors retained complete intellectual freedom to report and interpret the results. Any opinions, errors or omissions are those of the authors. 1

2 Loan pricing is a strategic lever that drives profitability, market size, and market share. Understanding consumer and competitor responses to price changes is of course critical for firms seeking to maximize profits. But achieving such understanding is nontrivial when there are identification issues, information frictions, and uncertainty: while profit maximization may be a textbook assumption, in practice firms face some of the same challenges social scientists face. Understanding price responses is also important for policymakers, donors and investors keen to see firms maximize subject to a double-bottom line by pricing social products (like basic financial, health, and housing services) as low as possible while still maintaining profitability, thus expanding outreach. Understanding responses to price changes requires evidence on both customers and suppliers, over time horizons that allow for adjustment dynamics to any new equilibrium. We track consumer and competitor responses to randomly assigned, market-level price cuts by the largest supplier in a large social product market: microcredit in Mexico. 2 We estimate responses over 12- to 29-month horizons horizons much longer than the typical loan cycle in this market -- and explore market frictions that slow the adjustment to a new equilibrium. Credit elasticities have been difficult to identify. Indeed, the experiment in this paper came about because a leading microcredit policymaker, the World Bank-based Consultative Group to Assist the Poor ( CGAP ), re-examined its prior belief of price-inelastic demand after seeing our shorter-run estimates from South Africa (Rosenberg 2002; Karlan and Zinman 2008). CGAP introduced us to Compartamos Banco ( CB ), a for-profit, publicly-traded bank with a doublebottom line. CB was planning to cut rates but was unsure how much to cut and wanted to better understand consumer and competitor responses. The nature of any market frictions that slow adjustment to a new equilibrium is also poorly understood, in part because it is rare to have all ingredients required for convincing inference: a well-identified shock, with data on both consumer and firm responses, measured over 2 Microlending is typically defined as the provision of small-dollar loans to (aspiring) entrepreneurs, although there is some policy and practitioner debate over the definition. For reviews, see, e.g., Karlan and Morduch (2009), Banerjee (2013), and Banerjee, Karlan, and Zinman (2015). Our Table 1 summarizes related literature on price elasticities. 2

3 appropriately long horizons. Understanding frictions is critical for properly specifying the constraints facing firms and consumers in equilibrium, and/or along adjustment paths. CB worked with us to randomize the interest rate offered on its core group lending product, Crédito Mujer. We randomized at the level of 80 distinct geographic regions 3 throughout Mexico, covering 130 bank branches, thousands of borrowing groups, and tens of thousands of borrowers. Randomizing at the level of large geographic units internalizes any within-region feedback effects of price changes on learning, competitive responses in credit and product markets, and input markets. It also allows us to track competitor responses, and explore the mechanisms underlying dynamic adjustment. 4 Treatment branches implemented permanent 20 percentage point (pp) reductions in the annual interest rate (on a base of roughly 100% APR), while Control branches implemented permanent 10pp reductions. Crucially, CB left these prices in place for 29 months, permitting estimation of dynamics and long-run responses. Demand and competitor responses can be attributed to price changes per se under the assumption that no other changes are correlated with treatment assignment; e.g., we assume that screening, monitoring, and marketing did not differ across treatment and control groups. The screening and monitoring assumptions are supported by the finding that the experimentally lower interest rate does not affect bunching at CB s loan amount limits, or delinquency rates. The marketing assumption seems plausible given study protocols, the fact that both treatment and control regions experienced rate cuts (20pp vs. 10pp) relative to CB s prior rate, and that CB s senior management did not want to adulterate the experimental design. Another benefit of CB changing prices in all regions is that this reduces the likelihood that competitors viewed any price cut as temporary. We identify four types of responses to CB s price cuts. 3 Each region comprises a distinct geographic market per CB s market definition, corresponding roughly in most cases to a municipality. 4 Individual-level randomizations, implemented via e.g. direct mail as in Karlan and Zinman (2008) and Ponce et al (2017), can offer greater statistical power but tend to be relatively ill-suited for identifying within-region feedback, competitor responses, and mechanisms underlying dynamic adjustment. 3

4 First, we estimate demand elasticities and how they vary across horizons, using CB administrative data on loans and borrower characteristics. We find elasticities that are towards the strong end of previous estimates even in the shorter-run (e.g., -1.1 for pesos borrowed over the first 12 months). More critically perhaps, we find that lower prices bring in many new borrowers and that demand becomes even more elastic overer the long-run (reaching e.g., -2.9 for pesos borrowed during the final six months of the experiment, with a p-value of 0.01 on the difference between the year 1 and year 3 treatment effects). Second, we use credit bureau data to explore whether elasticities with respect to CB credit are upper or lower bounds on net demand elasticities. We find no evidence of crowd-out if anything, the point estimates suggest crowd-in. 5 The lack of crowd-out on the consumer side and isomorphically, the lack of business-stealing on the competitor side squares with several other pieces of evidence suggesting that credit constraints bind (Sections 1 and 3-H). Third, we examine effects on CB s costs and profitability, finding that the compositional shift toward new borrowers induced by lower prices is costly for CB to manage in the short-run, but that ultimately the lower price is more profitable, at least in accounting (point estimate) terms. Indeed, based on preliminary results from this study, CB instituted the lower-rate across its entire operation shortly after the experiment ended. Accounting sensibly for the net present value (NPV) of new clients not just new loans proves to be key in evaluating the impacts on profits. It also lays bare a real challenge to profit maximization: the effects of the lower rate on NPV are imprecise, even at the massive scale of our experiment. Fourth, we work with CB to collect data on competitors offerings using a standard method deployed by firms in many industries--mystery shopping. We document many close competitors, in terms of geographic and product proximity. There is little evidence that competitors had 5 These results are consistent with those from a separate study we conducted with CB on the impacts of its loans. In Nogales, an area not included in the interest rate study, we randomized the rollout of the lending program and find that CB did not crowd-out other sources of formal credit, despite substantial use of credit from close competitors (Angelucci, Karlan, and Zinman 2015). 4

5 responded a year after CB s price cut, although this null is imprecisely estimated and may miss longer-run effects. 6 These four sets of results have several implications for pricing as a strategic lever and for the technology of profit maximization. The elasticity results suggest that pricing can be an important targeting lever for a social good, even one for which many key actors held priors of inelastic demand (given evidence on credit constraints and high marginal returns to investment for microentrepreneurs; besides CGAP, see e.g., discussion in Armendariz de Aghion and Morduch (2010)). The big differences between short- and long-run elasticities validate a canonical textbook assumption and highlight the importance of taking dynamics into account when making policy, business, or modeling decisions. For example, although the lower interest rate produced substantially greater loan volume and outreach (new clients) right away, profits did not increase until year two, and then only noisily, meaning that CB only saw benefits on both sides of its double bottom line when evaluating over a longer horizon than any previous study of credit demand elasticities (Table 1). Putting our four main results together with other pieces of evidence also sheds light on the frictions that mediate adjustment from a shock to equilibrium. Specifically, they suggest, circumstantially, that informational frictions are key on both sides of the market. On the demand side, we start by casting doubt on some candidate frictions. Contractual frictions might be important in other settings, but here loan cycles are short enough (16 weeks) to allow for much more rapid adjustment than we see over the 120-week span of the experiment. Capital adjustment costs might be important in other settings (indeed, these usually motivate the textbook assumption that demand elasticities are stronger in the long-run), but our findings on the impacts of CB s expansion into a new geographic area, not covered by the interest rate experiment, 6 That sort of slow adjustment would be interesting in its own right. Another possible explanation for a null result is that competitors respond uniformly across regions, despite CB pricing substantially lower in treatment regions. That pattern of response would also be noteworthy, as it would suggest that there are frictions discouraging competitors from responding in a more granular way to CB s differential pricing across regions. 5

6 show that marginal CB borrowers expand their microbusinesses with inventory but not with lumpy investments (Angelucci, Karlan, and Zinman 2015). It seems more likely that it takes time for borrowers to learn about the new rates (lender practices probably contribute to this, as noted below), including learning at the borrowing-group level about risk gradients with respect to price under joint liability. This fits with evidence from other markets/contexts that information frictions impede demand responses and adjustment to a new equilibrium (e.g., Jensen 2007). On the supply side, we speculate that information frictions also dampen competitor responses and create inertia at baseline prices. Prices are rarely featured in lender advertising, or are presented non-transparently (Gine and Mazer 2017). Indeed CB staff did not know, with any precision, what even close competitors were charging-- hence the need for the mystery shopping effort. Competitors received especially weak signals that CB cut rates because they did not actually lose clients (as evidenced by our results from the credit bureau data). Forecasting profitability is complex and replete with uncertainty all the more so at a new price level-- given e.g., the dynamics of customer acquisition and retention. As such our findings add to literatures on optimization frictions facing firms (e.g., Romer 2006) and help explain the substantial and growing investment in experimentation and analytics across many industries. The fact that experimentation requires investment is crucial: experimentation may prove valuable, but it is costly in various ways (Anderson and Simester 2011; Senior and Almquist 2013). Indeed, our conversations with CB and many other lenders across the globe indicate that one key cost is risk: experimenting with lower rates may shift customer reference points, making it difficult to return to higher rates if lower ones prove unprofitable. This no turning back mechanism aligns with our finding from a South Africa microloan market that borrowers are much more elastic with respect to rate increases than decreases (Karlan and Zinman 2008). These potential information frictions also help us characterize the industrial organization of the Mexican microcredit market, which has hallmarks of consumer credit markets worldwide: prevalent credit constraints and substantial price dispersion despite many competitors offering 6

7 similar products in close proximity (Zinman 2015). 7 We speculate that price dispersion persists, in part, because competitors face substantial risks and information gaps in evaluating whether it would be valuable to change prices. At the same time, prevalent credit constraints may contribute to market power that cushions the blow for lenders who have not (yet) found their optimal price, as evidenced by our finding that CB s growth in lower-rate regions did not come at the expense of competitors (and may have even produced some crowd-in). The welfare implications of CB s lower rate are clearer qualitatively than quantitatively. Textbook qualitative suggests that surplus under the demand curve increases substantially, particularly over the long-run as the demand curve rotates out and quantities increase even more at the lower price relative to the higher one. Coupled with evidence that credit constraints bind and total borrowing increases for consumers (the no crowd-out result), and that profits increase for CB, it seems reasonable to infer that welfare increases (although not as much is it would if competitors cut prices as well). Quantitatively estimating welfare changes would require accounting for many potentially important features e.g.., the shape of the demand curve, the source and magnitude of credit constraints, credit being an intermediate good, returns to that intermediate good being uncertain and/or mediated by behavioral factors, etc. that are not measured or identified in this study. 8 In the same vein, inputting our elasticities into micro-founded macro models would require substantial additional data and theory on how these parameters relate to tastes, technology, policy, market structure, etc. This strikes us as a fruitful line of inquiry for future research. Relatedly, although the Mexican microcredit market is interesting in its own right as a large market for one of the most scrutinized social goods-- the external validity of our results is uncertain. But the richness of our findings and market/institutional descriptive data allows one to identify whether and how other settings compare. Given the surprises in this paper we find much larger elasticities than documented previously, as well as evidence inconsistent with textbook assumptions that successful firms maximize profits, and that competitors will match a substantial price cut by a market leader it is 7 Proximity here, as is often the case, takes the form of geographic as well as product space: CB and its competitors often have branches on the same street. In other markets (e.g., U.S. credit cards), proximity takes the form of competitors direct-marketing close substitutes to the same customer base. 8 Zinman (2014) reviews different models of credit markets and their welfare implications. 7

8 important to understand whether and why our findings replicate or not. Our methods offer guidance for conducting similar studies in other markets: we show that one can combine a field experiment with various sources of administrative data to identify dynamics in consumer and competitor responses to a market leader s pricing change. Our unusually long time horizon generates evidence on long-run responses and adjustment frictions, and our region-level randomization adds to the small set of market-level field experiments where competitor responses (as well as consumer responses) are outcomes of interest (e.g., Busso and Galiani 2014; Andrabi, Das, and Khwaja 2017; Cunha, De Giorgi, and Jayachandran 2017) The Market Setting A. Compartamos Banco and its Target Market CB has long been the largest microlender in Mexico, 10 and it had well over one million borrowers during our study period. CB was founded in 1990 as a nonprofit organization and was converted to a commercial bank in It went public in 2007 and its parent company Gentera has a market capitalization of about $2.6 billion USD as of this writing. During our study period a super-majority of CB clients borrowed through Crédito Mujer, the group microloan product studied in this paper. Crédito Mujer nominally targets women who have a business or self-employment activity or intend to start one. Empirically, all borrowers are women, but a companion paper uses survey data from a different region to estimate that only about 52% are microentrepreneurs (Angelucci, Karlan, and Zinman 2015). Borrowers tend to lack the income and/or collateral required to qualify for loans from traditional banks and other upmarket lenders. B. CB s Cost Structure, Loan Terms and Competition Crédito Mujer loan amounts range from M$900-M$24,000 pesos (during our sample period 13 pesos, denoted M$, = $1US), with CB imposing caps that increase as groups successfully repay 9 Table 1 summarizes other studies on elasticities of demand for consumer credit and microcredit, including non-rcts and studies set in wealthier countries, and does not find any other studies using market-level randomization or a time horizon longer than 12 mo nths. 10 According to Mix Market, 8

9 prior loans (Section 3-F provides details). 11 Borrowers typically start with a loan of about M$4,000 and take successively larger loans over time. Loan repayments are due over 16 equal weekly installments, and are guaranteed by the group (i.e., joint liability). Aside from these personal guarantees, there is no collateral. Interest rates are marketed as monthly, add-on rates of 3.0 to 4.5%, 12 excluding value-added tax and required simultaneous savings. 13 We convert these addon rates to portfolio-weighted APRs and estimate that CB charged APRs of roughly 100% at baseline. Market research, detailed below, indicates that CB s pricing falls at the low end of the market for comparable loans. 14 Prior to this study, CB had set prices based on its cost structure and anecdotal information about competitive factors. Its most important costs are labor and of course the cost of funds. Labor costs loom large because, as detailed below, marketing and group administration (including monitoring) involves field staff. Staff receive a base salary with a performance bonus based on a proprietary model that includes factors such as revenue and repayment rates. New groups and borrowers are particularly costly because they require more bank effort (in set-up and monitoring) for lower gross payouts (small loan amounts and higher delinquency/default rates). In fact, the typical borrower is barely profitable at first but then becomes profitable with subsequent loans (Section 3-D). The cost of funds reported to us by CB averaged about 9.5% during our sample 11 Also, beginning in weeks 3 to 9 of the second loan cycle, clients in good standing can take out an additional, individual liability loan, in an amount up to 30% of their joint liability loan. 12 An add-on rate is calculated over the original loan amount and does not adjust for declining balances as an Annual Percentage Rate (APR) does. CB rates vary only based on city and a borrowing group s loan cycle, with newer groups paying the highest rates and well-performing and more-experienced groups paying lower rates. 13 Borrowers are supposed to make an upfront deposit totaling 10% of their loan amount into a personal savings account, and contribute at least 10 pesos weekly for the remainder of the loan cycle. This forced savings component at either zero or low interest is not claimed as collateral, but rather meant to instill a culture of regular deposits and generate a signal of the client s ability to generate and manage cash flow that can be used to evaluate creditworthiness for future loans. The forced savings is not necessarily held by CB; i.e., the effective APR paid by the borrower may be higher than the effective APR earned by CB. Anecdotally, neither CB staff nor group members vigorously enforce the savings requirement. Mexican law does not require advertisements or disclosures to include savings requirements or value-added tax in APR calculations. 14 See David Roodman s work for more details on microloan pricing in Mexico. His APRs are higher than ours because he includes VAT and forced savings, and compounds more frequently. 9

10 period. Given CB s access to financial markets it is effectively unconstrained in supplying credit on the margin. 15 Market research paints a picture of a microloan market that was competitive, but imperfectly so, during our study period. CB embarked on this study because it had already decided to cut rates- - based on its perception of competitive pressures-- but was undecided about how much to cut. CB field staff easily identified nearby close competitors for the mystery shopping exercise detailed in Section 3-I. Two key results emerge from this exercise. The first speaks to the competitive part of imperfectly competitive -- nearly every CB branch 16 identified at least three close competitors (with closeness defined as being both geographically proximate and offering a similar product), with many branches identifying more than three. The second key result speaks to the imperfectly part-- there is substantial price dispersion even within a set of close competitors. 17 Taken together with evidence of binding liquidity constraints for consumers, the price dispersion reinforces that how consumers respond to CB s rate cut is an open question (e.g., whether cheaper CB debt complements, or substitutes for, existing debt), as is the question of how competitors respond. C. Targeting, Marketing, Group Formation, and Screening 18 Crédito Mujer groups range in size from 10 to 50 members. When CB enters a new market, loan officers typically target self-reported female entrepreneurs and promote the Crédito Mujer product through diverse channels, including door-to-door promotion, distribution of fliers in public places, radio, and promotional events. As loan officers gain more clients in new areas, they promote less frequently and rely more on clients to recruit other members. When a group of about five women half of the minimum required group size expresses interest, a loan officer visits the partial group at one of their homes or businesses to explain loan terms and process. These initial women are responsible for finding the rest of the group members. 15 The fact that we find CB s lower rate drawing in many new borrowers (Section 3-B) is consistent with CB being unconstrained, because if CB were constrained it presumably would ration new borrowers first, since existing borrowers are far more profitable, as noted above and detailed in Section 3-D. 16 A branch is the main operating unit for field staff, as described in Section This sort of price dispersion is not so surprising, in light of similar evidence from various credit markets in the U.S. and elsewhere (Zinman 2015). 18 This sub-section also appears in Angelucci, Karlan, and Zinman (2015). 10

11 The loan officer returns for a second visit to explain loan terms in greater detail and complete loan applications for each individual. All potential members must be at least 18 years old and must present a proof of address and valid identification to qualify for a loan. Business activities (or plans to start one) are not verified; rather, CB relies on group members to screen out any uncreditworthy applicants. In equilibrium, potential members who apply are rarely screened out by their fellow members, since individuals who would not get approved are not recruited and do not to tend to seek out membership. CB reserves the right to reject any applicant put forth by the group but relies heavily on the group s endorsement. CB does pull a credit report for each individual and automatically rejects anyone with a history of fraud. Beyond that, loan officers do not use the credit bureau information to reject clients, as the group has responsibility for deciding who is allowed to join. Applicants who pass CB s screens are invited to a loan authorization meeting. Each applicant must be guaranteed by every other member of the group to get a loan. Loan amounts must also be agreed upon unanimously, subject to the upper bounds imposed by CB. Loan officers moderate the group s discussion and sometimes provide information on credit histories and assessments of individuals creditworthiness. Proceeds from authorized loans are disbursed as checks to each client. D. Group Administration, Loan Repayment, and Collection Actions Each group elects a treasurer who collects payments from each group member at each weekly meeting. The loan officer is present to facilitate and monitor proceedings but does not touch the money. If a group member does not make her weekly payment, the president (and loan officer) will typically solicit and encourage solidarity pooling to cover the payment and keep the group in good standing. All payments are placed in a plastic bag that CB provides, and the Treasurer then deposits the group s payment at either a nearby bank branch or convenience store Compartamos has partnerships with six banks (and their convenience stores) and two separate convenience stores. The banks include Banamex (Banamexi Aquí), Bancomer (Pitico), Banorte (Telecomm and Seven Eleven), HSBC, Scotiabank, and Santander. The two separate convenience stores are Oxxo and Chedraui. 11

12 Beyond the group liability, borrowers have several other incentives to repay. Members of groups with arrears are not eligible for another loan until the arrears are cured. Members of groups that remain in good standing qualify for larger subsequent loan amounts and lower interest rates. CB also reports individual repayment history for each borrower to the Mexican Official Credit Bureau. Loans that are more than 90 days in arrears after the end of the loan term are sent to collection agencies. Late payments are common, but default is rare: in our data, we find a 90-day group delinquency rate of 9.8%, but the ultimate default rate is only about 1%. 2. Study Design The research team (IPA) worked with CB to identify 80 distinct geographic areas ( regions for the purpose of the study) throughout Mexico, for the purpose of randomly assigning interest rates (Figure 1). Regions are the size of a small metropolitan area, a moderate-sized city, or a large district in the largest cities. The CB operating unit within a region is a branch (actually more like a regional or sub-regional office); the mean number of branches per region is IPA then assigned each of the 80 regions (and all branches within each region) to either low-rate or highrate. The interest rate changes applied only to Crédito Mujer. Regions were defined specifically to maximize compliance with the experimental design by providing ample geographic distance between branches offering different rates. We also verify that our main results are similar after dropping the eight branches, from five regions, in the densely populated Mexico City metropolitan area (Appendix Table 10). CB did not implement any additional operational or strategic changes that mirrored our treatment assignment, so any responses we observe are demand-driven and pure price effects, under the assumption of no other supply-side changes or constraints. There are institutional reasons to assume that no-changes assumption holds: senior management put strong incentives and monitoring in place to promote staff compliance with the design, an individual staff member interacted with customers from only one treatment arm, and the design changed prices in both treatment and control regions-- so any discrete changes in marketing, screening, or enforcement would affect both study arms. Supply constraints could bias estimates of price elasticity of demand, and perhaps differentially over time if those constraints adjust endogenously and gradually. But 12

13 there is no evidence that CB was constrained in its ability to finance or administer new loans: it had ready access to credit facilities, and as noted above new group formation is largely delegated to borrowers. We also test and reject the hypothesis that CB loan size limits introduce a confound in Section 3-F. The motivation for randomizing at the region level, as opposed to a more granular level like branches or groups, is twofold. The first is to allow for any consumer learning and competitive response to take place at the level of a geographic market; i.e., we allow for within-market spillovers. Second, region-level assignment facilitates compliance with the randomization in a group lending setting, by ensuring that contiguous groups or contiguous branches (which would normally draw some borrowers from overlapping geographic areas) are assigned the same rate. The main drawback of region-level randomization is its lack of power for addressing certain questions; e.g., with 80 regions it would be difficult to identify a global optimum even if CB had been willing to randomize more price points along the demand curve. Table 2 summarizes various baseline (April 2007) averages for low- and high-rate regions and checks for balance on observables. Panel A presents borrower characteristics: education, age, number of children, number of dependents, and marital status. Panel B presents loan volume, both to all borrowers and to groups generally targeted by policymakers, MFIs, and donors (new borrowers and borrowers with low education; we also consider relatively poor borrowers below but lack pre-treatment data on income or wealth). Panel C presents loan characteristics: APR, loan amount, group size, and number of groups. Delinquency data is absent here because CB lacks the requisite pre-treatment data. Panel D covers market (i.e., regional) characteristics, focusing on CB's market share as measured using credit bureau data. Overall, Table 2 suggests that the randomization successfully generated similar treatment and control groups. The experiment engineered prices that were about 10 percentage points lower (in APR units) in low-rate regions. Starting May 15, 2007, CB implemented this variation by offering differential cuts from pre-treatment prices. Low-rate regions got 20 percentage point cuts from pre-treatment rates (which averaged about 100% APR, as shown in Table 2 Panel C), whereas high-rate regions 13

14 got 10 percentage point cuts. 20 CB presented these prices to (prospective) borrowers as permanent in the sense of the new normal : these were not promotional rates. CB kept these rates in place for 29 months and then implemented the low rate everywhere, along with other pricing changes. We measure price sensitivities by comparing various outcome measures in low- vs. high-rate regions for up to 29-months post-treatment (i.e, post- differential rate cuts). In some cases we can use pre-treatment data as well. The next section details our specifications and results. 3. Empirical Specifications and Results A. Specifications We estimate treatment effects of the lower interest rate on borrower demand, lender profitability, and competitor responses using OLS specifications that take advantage of our rich data and random assignment. Our main specification is: (1) Yrt = α + β1(lowrater*postt) + ρlowrater + τt + εr Y is an outcome, measured for region r in month-year t. α is the constant. For most Ys we have pre-treatment data from March and April 2007, and post-treatment data from June 2007-October 2009 inclusive. The variable of interest here is the interaction term which equals one if and only if the observation is from a low-rate region in the post-treatment period and β1 identifies price sensitivity. We also examine dynamics by decomposing LowRater*Postt into three post-treatment time periods: LowRater*Postt*Year1t for the first 12 months, LowRater*Postt*Year2t for months after the rate change, and LowRater*Postt*Year3t for months T is a vector of monthyear dummies (e.g., separate dummies for June 2007 and June 2009), and these absorb the Post 20 Compartamos advertises and administers interest rates in add-on, monthly terms, and prices each group into one of three tiers based on tenure and past performance. The randomization assigned low-rate regions to tiered pricing of 3.0%/3.5%/4.0% monthly, and high-rate regions to 3.5%/4.0%/4.5% monthly. We convert these monthly rates to balance-weighted APRs. Rates applied only to loans originated on or after May 15, 2007, but the amount of lock-in on previous loans is minimal given the loan term of 16 weeks and treatment duration of 120 weeks (our analysis horizon varies from 52 to 120 weeks). 14

15 main effect. ε is the error term; throughout the paper we cluster standard errors at the unit of randomization: the region. Our primary analysis starts by presenting the results from (1), for each six outcomes: number of loans, loan amount, revenues, costs, profits and net present value of client relationships. in Table 3. Appendix Table 1 through Appendix Table 6 report specification robustness checks for each of the six outcomes. Column 1 in each of these Appendix Tables shows results with fixed effects for both month-year and region (with the region indicators subsuming the LowRate indicator). Column 2 shows results controlling for the baseline values of the outcome variable instead of for LowRate or region indicators. Column 3 shows results for the log of the dependent variable (except in Appendix Table 5, since profits can be negative). Columns 4-7 (3-6 in Appendix Table 5) show results for winsorizing or trimming the dependent variable. The price elasticity of demand, reported at the bottom of Table 3 and other tables, is defined as the percentage change in quantity demanded divided by the percentage change in price. We calculate the former, for each specification, by dividing the coefficient of interest by the mean of Yrt across all high-rate (control-group) regions over the entire post-treatment period. We calculate the latter, again over the entire post-treatment period, by dividing the average, balance-weighted APR difference between high- and low-rate regions, and then dividing that difference by the average, balance-weighted APR in high-rate regions. As discussed above, we identify price elasticities of demand under the assumption of no impact on supply-side decisions as a result of treatment. B. Elasticities of Number of Loans Disbursed We find an increase of 200 loans disbursed by CB (i.e., taken by borrowers) per month in the low-rate regions, compared to the high rate regions (Table 3, Panel A, Column 1). The standard error is 93 and the implied elasticity is -1.4, with a 90% lower bound of and upper bound of Appendix Table 1 shows that other specifications yield similar estimates. The interest rate sensitivity increases significantly over longer horizons (Table 3, Panel B, Column 1), from about -0.8 in year one to -1.5 in year two to -2.2 in year three. The p-values on 15

16 the three pairwise differences are 0.08, 0.01, and The finding that longer-run demand is more elastic is consistent with borrower learning and/or adjustment costs. Figures 2a and 2b plot cumulative distribution functions (CDFs) of loan counts to explore effects throughout the distribution of regions as ordered by loan volume, separately for treatment and control. Figure 2a plots the average monthly loan count per region over the post-treatment period (i.e., one observation per region, so each CDF plots 40 data points). This suggests that there is a treatment effect for the larger regions (starting around the 40 th percentile) but not for smaller ones. Figure 2b takes the same approach, but counts only the final post-treatment month (and hence focuses more on the long-run). Here we see the suggestion of treatment effects starting around the 20 th percentile of loan count volume. C. Elasticities of Amount Borrowed from Compartamos Banco Table 3 Column 2 presents estimates of price sensitivities of the amount lent by (borrowed from) CB, again measured as region-month flows. This measure of demand combines the extensive and intensive margins. Given the treatment effects on the extensive margin (Table 3 Column 1) it is not straightforward to back out effects on the intensive margin alone, since the lower price may be drawing in borrowers with different elasticities and we do not have separate instruments for borrowing at all and for amount borrowed. We find an increase of about two million pesos per month in low-rate compared to high-rate regions (SE=777,000), on a base of about 6.5m. The implied point elasticity is -1.89, with a 90% lower bound of and upper bound of Appendix Table 2 shows that other specifications yield similar estimates. As with loan count, the amount borrowed is more elastic in the long-run (Table 3 Column 2 Panel B), with yearly elasticities increasing from to to The p-values on the three pairwise differences are 0.054, and Figures 3a and 3b plot cumulative distribution functions (CDFs) of amounts lent to explore effects throughout the distribution of regions, separately for treatment and control. Figure 3a plots the average lending per month per region over the post-treatment period (i.e., one observation per region, so each CDF plots 40 data points). This suggests that there is a treatment effect for the 16

17 larger regions (starting around the 40 th percentile) but not below. Figure 3b takes the same approach, but counts only the final post-treatment month (and hence focuses more on the longrun). Here we see treatment effects starting around the 10 th percentile of peso volume. D. Elasticities of Revenue, Costs, Profits, and Net Present Value of Client Relationships A key question for equilibrium is whether the lower interest rate was profitable, or at least sustainable, for CB. Table 3 Columns 3-6 present estimates of the lower rate s effects on revenues, costs, profits, and the net present value of the client relationship. The treatment effect coefficients on revenue (Column 3) are positive but statistically indistinguishable from zero. This is true for the main specifications as well as all robustness specifications in Appendix Table 3. Note that there are offsetting effects here: the lower interest rate induces substantially more lending (Columns 1 and 2) but of course generates less gross revenue ceteris paribus. From an accounting perspective (ignoring confidence intervals, as accounting is wont to do), the additional gross revenue from the marginal loans is the 1.96m treatment effect on amount borrowed per month, multiplied by the periodic interest rate in the treated regions (80/12/100=6.7%), for a total of about 131,000 pesos. The foregone gross revenue is the difference between treatment and control periodic rates (10/12/100=0.8%), multiplied by the inframarginal loan amount. We estimate the latter using the average monthly amount borrowed in control group regions during the experimental period (9.1m pesos, as shown near the bottom of Table 3 Column 2). Estimated foregone region-month revenue is thus 9.1m*0.008= about 76,000 pesos. Comparing additional to foregone revenue, back-of-the-envelope accounting estimates a 131,000-76,000=55,000 peso increase, which is reassuringly similar to the estimate of 73,000 obtained from simply plugging in CB s reported revenues into equation 1 (Table 3 Column 3 Panel A). Treatment effect coefficients on costs--which include charge-offs are comparable in magnitude to those on revenues but more precisely estimated and hence statistically distinguishable from zero. Appendix Table 4 shows similar results for the other specifications. Appendix Table 7 suggests that about 40% of the cost increase is due to personnel. Appendix Table 8 shows that personnel costs fall in the average number of members per group and increase with the number of groups. Appendix Table 9 shows some evidence that the lower rate increases the 17

18 number of groups and group size (Column 1-4), but that loan officers do not handle significantly more groups or clients. In all, we find little evidence of economies of scale given operations and firm structure at the time of the experiment. Innovations with respect to lending technology, such as mobile devices with customized software for credit officers (that are now being implemented by many lenders, including CB after our study ended), may help lower marginal costs and capture economies of scale in the future. Column 5 examines treatment effects on short-run profits, defined simply as (Revenues-Costs), and confirms what one would expect from eyeballing Columns 3 and 4: the point estimates suggest modest negative effects on profits initially, with the effect perhaps turning positive in Year 3. These estimates are imprecise, with each of the t-statistics well below 1. Appendix Table 5 shows similar results for the other specifications. Column 6 reports treatment effects on the NPV of clients. NPV moves beyond realized profits to expected profits and is the key profitability metric from CB s perspective because it more fully accounts for the lower interest rate s changes on portfolio composition. Specifically, it is important to account for the dynamic that new clients are less profitable than more seasoned clients. Thus as the new clients drawn in by the lower interest rate age -- into lower default rates, large loan sizes, and lower marginal costs-- they may become profitable over horizons even longer than the 29- month analysis period for realized profits. We estimate NPV by first estimating the average profitability of loan cycle (first loan, second loan, etc.) and treatment status combination. Next we age each client through a life-cycle of loan cycles based on the data for likelihood of repeat borrowing (conditional on treatment status). The probability of repeat borrowing starts at 59.7% and 60.8% for the first to second loan for control and treatment groups, respectively, and then increases almost perfectly monotonically for the first nine loans until it reaches 67.8% and 67.6%, respectively. We then aggregate these estimates to the region-month level: (2) NPV of portfoliort = Σclients Σ[t,T] (Still client probabilitycrt * (loan revenuecr costcr loss to defaultcr)) / discount ratet T-t where r and t index region and month as before, T represents the number of future periods (period = month, from t=1 to T=224, with 224 being the most number of months anyone continuously 18

19 borrows in the data), and c represents tenure of the client. Our NPV measure probably produces conservative estimates of the treatment effect on profits because it does not account for cross-sells or referrals (both of which should increase monotonically in the number of clients and hence be higher in the treatment group). The coefficients on the NPV treatment effect are positive but statistically indistinguishable from zero. This is true for the main specifications as well as nearly all robustness specifications in Appendix Table 6. Figure 4 explores the dynamics by plotting the NPV estimates month-bymonth, separately for treatment versus control. The control group (=higher rate) reaps higher NPVs initially but gets surpassed by the treatment group (=lower rate) around month 9, with the gap widening as time proceeds. Indeed, after seeing preliminary results over the 29-month horizon, which ends in October 2009, CB implemented the lower interest rate in the control group regions in CB made this change roughly in concert with implementing more granular risk-based pricing across their entire operation, which unfortunately renders moot using post-experiment data to examine any convergence, since a lot changed in their pricing after our experiment ended. So why had not CB cut interest rates earlier? After all, given weakly increasing profits, elastic demand, and a double bottom line of helping clients to generate social and economic value, as well as the possibility of increased cross-sells, the lower interest rate has turned out to be beneficial for CB ex-post. We speculate that costly experimentation may discourage some lenders from deriving their demand curves. In addition to direct costs, there may be a risk to cutting rates if demand curves develop kinks at current market rates that make subsequent increases quite costly. 21 We discuss some implications of costly experimentation for market equilibria and future research in the conclusion. E. Do Lower Rates Improve Outreach? Table 4 explores whether the lower interest rate increased take-up by groups that are often the focus of outreach intended to expand access to microcredit. Columns 1 and 2 re-estimate the 21 Karlan and Zinman (2008) finds extremely elastic demand to rate increases but not decreases in South Africa, an asymmetry the lender there had anticipated. 19

20 pooled and yearly version of equation (1) with the number of loans to new clients as the outcome. 22 Columns 5 and 6 do the same for loans to clients with low education. Columns 7 and 8 repeat the exercise for loans to clients with relatively low income. 23 The estimates are imprecise, but the overall pattern is consistent with the lower rate bringing in substantial numbers from targeted groups: each of the 12 point estimates is positive, and for each of the groups we see the pattern of demand becoming more elastic over longer horizons, suggesting that lower rates do improve outreach over the long-run. F. Mechanisms: Why is demand much more elastic over longer horizons? Why does it take substantial time to reach steady-state price sensitivity? Leading possibilities include capital adjustment costs, contractual frictions, and information frictions. We start by casting doubt on the importance of the first two mechanisms in our setting. Regarding capital adjustment, although Angelucci, Karlan, and Zinman (2015) do find evidence that marginal CB borrowers in Nogales (a region not included in the interest rates experiment because CB was not operating there yet) expand their microbusinesses over a two-year period, their evidence suggests that expansion takes the form of additional inventory rather than fixed investments. Regarding contractual frictions, recall that CB loan cycles are short enough (16 weeks) to allow for much more rapid adjustment than we actually see over the 120-week span of the experiment. Information frictions seem more likely to be important. Lack of transparency in pricing and advertising practices in the microcredit market may well slow the spread of new information about prices (Gine and Mazer 2017). Relatedly, it seems reasonable to assume substantial search costs given evidence from other credit markets (Zinman 2015) and the substantial price dispersion we document in this market (sub-section H below). Joint liability could further dampen the price response initially if groups underestimate any advantageous selection produced by the lower rate 22 We define a loan to a new borrower as one disbursed to someone who had not borrowed from Compartamos in any previous month. Columns 3 and 4 use loans to existing clients as an outcome: these are clients who had borrowed before and are either renewing a current loan or coming back for a new one. 23 We measure educational attainment for each borrower using CB application data, and define loweducation as not having completed high school. We estimate income for each borrower using application data that CB collected starting in June 2007, defining low-income as having a below-median poverty likelihood in our sample per the formula in Schreiner (2010). Given the lack of pre-treatment data on poverty likelihood we estimate versions of equation (1) that drop the interaction term and identify the treatment effect with the LowRate indicator. 20

21 and/or simply are somewhat risk-averse with respect to guaranteeing new loans; i.e., there may be learning about the price-risk gradient by groups. We also consider two alternative explanations/confounds. One is that we not identifying adjustment per se but rather responses to a macro shock that changes price sensitivity. Appendix Table 11 tests this by adding controls for macroeconomic variables that we can map to our regional level at high-enough frequencies to be useful (Columns 2 and 5), and their interactions with treatment status (Columns 3 and 6), to our main specifications for the number of loans disbursed (Column 1) and dollars disbursed (Column 4). The macro controls do not change our inferences, as evidenced by the p-values from tests comparing treatment effects across specifications shown near the bottom of each panel. 24 Another hypothesis is that CB loan size limits dampen demand elasticities, differentially for treatment vs. control and/or over time. Appendix Table 12 tests this using our main specification, at the loan- instead of region-level, with the dependent variable an indicator for whether the loan amount is the maximum allowed. Since that maximum varies with loan cycle, we estimate separately by cycle (Columns 2-5) as well as for all cycles pooled together (Column 1, where that regression includes indicators for each loan cycle). Panel A tests whether treatment-region loans are differentially likely to be at the max. Four of the five coefficients are indeed positive, but none have t-stats greater than one. The supply constraint is most likely to be bind for a group/borrower s first loan (45%, vs. 14% and 9% for the second and third loans), and there we have the one negative point estimate. Focusing on Column 1, overall we find little evidence of differential supply constraints by treatment status, although the confidence intervals include effect sizes of about 10% of the mean. Panel B tests whether the treatment effect changes over time, and finds no evidence that it does: our t-tests of equality across years produce only one marginal rejection out of 15, which is about what one would expect to find by chance. We again caution that we lack the power to precisely estimate any null effect. But we do not find cause for concern that supply constraints contribute to elasticities increasing over time. 24 Appendix Figure 1 provides some related evidence by plotting month-by-month estimates of the treatment effect, on each of the six outcomes. These monthly point estimates are very imprecisely estimated, but inspecting them for any sharp breaks that would be consistent with a differential response to a macro shock across treatment vs. control regions reveals no evidence of discontinuities. 21

22 G. Does Delinquency Fall with Price? We now turn to an analysis of whether the lower interest rate led to lower average delinquency (Table 5, Column 1 and 4). We have already considered this question implicitly in our costing and NPV exercises (Table 3), but looking directly at delinquency has value for identifying frictions due to asymmetric information and/or liquidity constraints (Karlan and Zinman 2009). We also examine whether interest rates have delinquency effects that vary with measures of the prevalence of new borrowers in a region-month (Table 5, Columns 2, 3, 5, and 6), under the hypothesis that any asymmetric information problem may be relatively severe for borrowers who have not contracted with CB before. Our outcome of interest in Panel A is the proportion of groups, in a region-month, that are behind on their repayments. Our main empirical model and sample is the same as for demand estimation, with two exceptions. First, CB did not track delinquency systematically prior to the experiment, so we lack pre-treatment data, drop the interaction term from equation (1), and identify treatment effects with the LowRate indicator. Second, we count only groups that could possibly be late in the denominator; i.e., we exclude groups that are too new to be delinquent. For the serious delinquency measure this excludes 163 region-months that are comprised entirely of groups with loans disbursed in the previous four months. Table 5 Columns 1-3 examine moderate delinquency any lateness and show little evidence of price response. The point estimates are all negative, however (consistent with lower rates mitigating asymmetric information), and the confidence intervals do not rule out economically meaningful effect sizes relative to the base rates at the region-month level, which are 0.14 for all groups, 0.15 for groups with >75% new members, and 0.19 for new groups. Columns 4-6 show a similar pattern of results for severe delinquency (later than 90 days). Again the point estimates are all negative, with confidence intervals that contain large effects relative to the base rates of 0.10 to We do not find evidence of time-varying effects (results not tabulated to conserve space). Panel B repeats the analysis with the borrowing group as the unit of observation, and an indicator for delinquency as the outcome, and finds similar results. H. Rate Cuts in Equilibrium: Evidence on Net Elasticities from Credit Bureaus 22

23 Credit bureau data enables us to examine another important aspect of general equilibrium responses to interest rate changes: net elasticities that account for any crowd-out or crowd-in of credit from other lenders. We have data from two bureaus: the Mexican Official Credit Bureau and the Circulo Bureau. Both bureaus allow us to focus on loans most comparable CB s core product. 25 The Official Bureau has more comprehensive coverage (compare the means in Columns 1 and 2 in Table 6), but has the disadvantage of including CB loans in the region-level data we were able to obtain. 26 The Circulo Bureau has the advantages of including Azteca, a main competitor of CB, and of excluding CB loans. The bureau data (Table 6) has three main differences from the CB data used to estimate our main results on elasticities of demand for CB debt. First, we were only able to obtain two snapshots from the bureaus: April 2007 (one month before the start of the pricing experiment), and December In all we have data from 79 regions at both points in time, for a total of 158 region-month observations. As such, Table 6 reports estimates from a two-period version of equation (1): we simply regress a measure of region-level borrowing on an indicator for the treatment period observation (Postt), an indicator for treated branches (LowRater), and their interaction, with Postt*LowRater identifying the treatment effect. Second, the Official Bureau data lacks loan amounts and the number of borrowers, so our main measure of aggregate demand is the number of loans. Third, the credit bureau captures stocks, not flows. Given these differences, Table 6 Column 3 also reports a comparable estimate for the stock of CB loans measured from CB s data, using the same months and specification as we use for the credit bureau data in Columns 1 and 2. We find the expected positive treatment effect on the stock of CB loans of about 800 (SE=427), with an implied elasticity of We define comparable loans in the Mexican Credit Bureau as loans from a: Bank, Bank loan PFAE, Credit line from bank, Non-bank loan for (PFAE), Financial non-bank loan for PFAE, Personal finance company, Medium market commerice, Medium market NF service, or Credit Union. In the Circulo bureau, comparable loans are those which institutions report as personal loans. 26 The credit bureaus would only provide us with data at the level of the municipality*lending institution type, so we cannot identify individual lenders or borrowers. To aggregate to regions we exclude the 3% of municipalities that straddle our regional boundaries, and the 16.5% of municipalities with <20 Compartamos clients. The latter exclusion improves power by focusing on the parts of regions where Compartamos actually has a presence (regions are large enough that Compartamos does not necessarily operate throughout the entire region). 23

24 The main inference we can make from the credit bureau data is a lack of strong evidence for crowd-out. Focusing on the Circulo data (Column 1), which excludes CB loans, we see a positive treatment effect, with a p-value of 0.16, suggesting that crowd-in is more likely than crowd-out. Crowd-in can result if there are both liquidity constraints and non-convexities in investment (Banerjee and Newman 1993), and Angelucci, Karlan, and Zinman (2015) also finds some evidence that CB lending increases generate crowd-in. Column 2 (from the Official Bureau) also shows a large, positive point estimate that is much larger than the direct effect on CB borrowing (compare to Column 3), but this point estimate is imprecisely estimated as well (p-value = 0.2). Several other data points buttress the interpretation that the lack of crowd-out on the consumer side (isomorphically, the lack of business-stealing on the lender side) is due to binding credit constraints. We do not find differential results in study regions where staff identified more versus fewer nearby competitors (Table 6 Columns 4-6), as one would expect if a pure market power/segmentation story were driving the results. Nor did CB staff have any difficulty identifying multiple nearby competitors for the mystery shopping exercise (sub-section I below), again suggesting that the market is competitive. Borrower-level survey data, from the Nogales region, shows substantial overlap in borrowing from CB and close competitors, and also that when CB expanded credit supply in Nogales, the additional credit taken by borrowers did not crowd-out other formal sources of credit (Angelucci, Karlan, and Zinman 2015). 27 Note that we cannot simply measure overlap within the interest rate sample, for two reasons: 1) the Nogales region is not in our interest rate sample because CB entered there subsequent to the interest rate experiment; 2) the credit bureau data for our interest rate sample is aggregated to the municipality level. It may also be the case that CB s lower rate, which ends up being toward the low end of the market (Table 7), draws new borrowers into the market who had reservation prices at or slightly below preexperiment prices. I. Rate Cuts in Equilibrium: Evidence on Competitors (Non-)Responses Our last bit of evidence on general equilibrium effects comes from estimates of whether and how competitors responded to CB s lower interest rate (Table 8). The data come from a standard 27 The AEJ sample is comprised of people in the Nogales region who are eligible for Compartamos credit and reached for an endline survey approximately two years after Compartamos entered the region. 24

25 mystery shopping market research exercise. We worked with CB s senior management to lay out some simple protocols for the bank s branch staff: 1) collect data on interest rates and maturities for at least the top three competitors in each region; 2) collect data by having someone pose as a prospective client seeking a 4,000 peso loan (a typical loan size for a new microcredit client); 3) collect data at two or more points in time, first in May 2007 (just prior to the start of the study), and then one at least one year after the start of the experiment (although some branch managers reported back as early as April 2008); 4) collect data on the same competitors in the preand post-periods, and also on new competitors in the post-period as merited by any changes in the competitive environment. These mystery shopping trials produced 616 observations on competitor loan offers, with 201 from the pre-treatment period and 415 from the post-treatment period (Table 7). 28 Observations are generated by 107 branches in 80 regions. 76 of these branches and 72 of these regions are present both pre-treatment and post-treatment. 29 The 616 offer-level observations come from 106 unique competitors, with the top five competitors comprising 48% of the observations. The number of competitors reported per branch*reporting date (N=183) ranges from 1 to 10, with a mean and median of 3, and a standard deviation of 1.4. Seventy-one percent of the observations are for the requested loan amount of 4,000 pesos, and 48% of the observations are for the standard CB loan term (maturity) of 16 weeks (94% are for 52 weeks or less). 30 Our main objective here is to estimate whether competitors change their prices, and so we calculate the Annual Percentage Rate (APR) from the periodic rate supplied by CB branches in the raw data. 31 Starting from our sample of 616 observations, 3 lack information needed to calculate the APR, and 5 observations reported non-positive interest rates (suggesting data entry/capture errors), leaving us with 608 usable observations. The mean APR is 154% and the 28 Many branches reported on multiple dates and/or multiple loan offers (different loan amounts and maturities) from the same competitor on the same date, most likely in response to reminders sent by senior management that were targeting a handful of straggling branches that had not yet reported any mystery shopping data during the post-treatment period. 29 Results are unchanged if we limit the simple to the 76 branches, or the 72 regions, present in both the baseline and endline mystery shopping data. 30 Controlling flexibly for loan amount and term does not change the results. 31 The raw data also includes other loan terms that allow us to verify that the periodic rate is in fact the correct rate to use in calculating the APR. 25

26 median is 116% (both somewhat higher than CB s APRs %), with substantial dispersion; e.g., a standard deviation of 150 percentage points. (See also Table 7, which shows dispersion across competitors in Column 2, and to a lesser extent within them in Columns 3 and 4). Table 8 reports estimates of competitor responses using the two-period version of equation (1). 32 In Table 8 Columns 1-5 the dependent variable is a function of the APR. Columns 1-3 use each of the 608 CB branch-competitor offer-reporting date observations as the unit of analysis. Column 1 shows that our point estimate of the competitor response in level terms (recall that Post*LowRate estimates the differential effect in treatment vs. control areas) is actually positive, and large: 36pp (se = 25.7pp). We can rule out a matching price cut (10pp) with 92% confidence. But Columns 2 and 3 caution that our estimate of the level effect may be unduly influenced by APR outliers: our point estimates on log(apr) and median(apr) are essentially zero. These null effects are imprecisely estimated, and do not reject the hypothesis of a matching price cut (p-values of 0.29 and 0.20). Columns 4 and 5 use the minimum APR observed (level or log) at the branch*reporting date level, and again find imprecisely estimated zeros for the treatment effect. These results suggest that the welfare gains (to borrowers) of CB s price cut are somewhat muted by the lack of similar cuts from competitors. Columns 6 and 7 use the competitor count (level or log) at the branch*reporting date level, and suggest weakly negative effects. This would be consistent with CB s price cut driving out highercost competitors. All told, the results in Table 8 suggest that competitors responded by exiting if at all, 33 although our estimates are imprecise and hence do not rule out economically meaningful responses. We note two additional caveats re: our estimates of competitor non-response. First, we have only a single snapshot, during Year 2 of the experiment, and consequently might be missing longer-run effects. That sort of slow adjustment would be noteworthy as well, particularly in light of the fact that typical loan maturities are much shorter than a year; in an information-rich environment, it 32 Clustering standard errors at the competitor or the competitor-region level, instead of the region level, produces somewhat smaller standard errors. 33 We also estimate treatment effects on loan maturity and loan amount and do not find evidence of effects on either, although again these nulls are imprecisely estimated. 26

27 would not take competitors long to ascertain the best response to CB s price cut. Second, it may be the case that competitors respond uniformly in all regions. Such uniformity also would be noteworthy, as it would suggest frictions-- informational, technological, etc.-- that discourage competitors from responding in a more granular way to CB s differential pricing across regions. What could explain non-response by competitors? A variety of factors suggest that lenders (like consumers) face informational frictions that hinder rapid and precise responses to competitor behavior. Contracting data is not readily available (hence the motivation for the mystery shopping trials), so competitors may not have even recognized CB s (differential) price cuts. Competitors received especially weak signals because they did not actually lose clients (Sections 3-H and 3-J). Price-setting is fraught by the difficulty of forecasting profitability, as indicated by our work in Section 3-D and the fact that CB chose to engage in costly experimentation to help pin down its preferred price. A closely related problem is that experimentation is more costly than it appears at first glance: our conversations with CB and other lenders around the world indicate a belief that past prices serve as reference points for customers, and hence that experimenting with lower rates is quite risky because it endogenously makes turning back to higher rates unprofitable. This inference accords with our finding from South Africa that borrowers are much more elastic with respect to rate increases than decreases (Karlan and Zinman 2008). J. Why no business stealing? Putting together the competitor and consumer results, we ask why borrowers do not substitute away from competitors if CB offers lower prices and its competitors do not? One possibility is product differentiation, although we view this as unlikely given the evidence of substantial price dispersion even amongst suppliers offering very similar products (Table 7). Another is search and/or switch costs, which have been linked to price dispersion in other major consumer credit markets (Stango and Zinman 2016; Alexandrov and Koulayev 2017). Yet another, and complementary, explanation is liquidity constraints: if borrowers had excess demand at pretreatment market rates, they might borrow more on the margin, from CB, without reducing their inframarginal borrowing. IV. Conclusion 27

28 We study the long-run (up to 29-month) effects of a 10 percentage point interest rate reduction by the largest microlender in Mexico, using a field experiment implemented at the level of 80 distinct geographic regions. Demand for CB loans is quite elastic, with much more elasticity over longer horizons. For example, the average elasticity of amount borrowed over the 29 months is about -1.9, with a Year 1 elasticity of about -1.1, and a Year 3 (months 25-29) elasticity of about There is no strong evidence of crowd-out in credit bureau data, although these and other null results are imprecisely estimated. The lower rate is profitable for CB in accounting (point estimate) terms, and the bank has maintained the lower rate post-experiment, confirming that the lower rate has been sustainable for CB. We find no evidence that competitors responded by cutting rates. These findings suggest several avenues for future research. First, understanding why demand is more elastic over longer-runs e.g., the relative importance of learning vs. adjustment costs is important for modeling and policy analysis. A closely related line of inquiry is whether and why elasticities differ for group versus individual lending products. Another closely related line of inquiry would unpack relationships between elasticities of demand for credit and for saving. Estimates of the elasticities of demand for savings are as low as zero (Hall 1988; Karlan and Zinman 2016), and most are strictly below the elasticities we find here. There remains much to learn about microlending production functions as well. Our results suggest an absence of commonly-assumed economies of scale, and it would be useful to know what prevents CB (and, presumably, other microlenders) from capturing scale economies. Is it something inherent to the group lending model? Labor market frictions? A longer transition path to scale economies than our study window-- perhaps it takes time to re-optimize fixed costs to handle the increased loan volume? Obtaining more precise estimates of aggregate credit demand and net elasticities is also important. Our finding of no crowd-out, even though competitors do not price-match, raises possibilities that switch costs have big effects on market outcomes, and/or that liquidity constraints work in surprising ways: the standard story is that liquidity constraints make agents price-inelastic, but perhaps this is true only of responses to price increases. The possibility of asymmetric responses to price increases and decreases raises another possibility that fits with our results: multiple equilibria borne of costly experimentation along the profit-maximizing frontier. Holding profits constant, a lower interest rate would seem to be at least 28

29 weakly better for a lender like CB, since the lower rate delivers social benefits from the largely beneficial average impacts of increased access to microcredit (Angelucci, Karlan, and Zinman 2015), in addition to public relations benefits and increased opportunities for cross-sells. In this sense, a natural follow-up question is why firms (including CB) do not charge lower rates to begin with. One possibility worth exploring is that experimenting with lower rates is risky: a lower rate may reset customer expectations of a fair/market rate, and create a kink in the demand curve. Karlan and Zinman (2008) find evidence along these lines in South Africa. If this dynamic holds, then cutting rates may reduce or eliminate the option to increase rates in the future (e.g., if it turns out that the lower rate was not as profitable as the initial rate). In this case policymakers might consider interventions to spur learning about pricing. 29

30 References Alan, Sule, and Gyongyi Loranth Subprime Consumer Credit Demand: Evidence from a Lender s Pricing Experiment. Review of Financial Studies 26 (9): Alessie, R., S. Hochguertel, and G. Weber Consumer Credit: Evidence from Italian Micro Data. Journal of the European Economic Association 3 (1): Alexandrov, Alexei, and Sergei Koulayev No Shopping in the U.S. Mortgage Market: Direct and Strategic Effects of Providing Information. SSRN Scholarly Paper. Rochester, NY. Anderson, Eric T., and Duncan Simester A Step-by-Step Guide to Smart Business Experiments. Harvard Business Review March. Andrabi, Tahir, Jishnu Das, and Asim Khwaja Report Cards: The Impact of Providing School and Child Test Scores on Educational Markets. American Economic Review 107 (6): Angelucci, Manuela, Dean Karlan, and Jonathan Zinman Microcredit Impacts: Evidence from a Randomized Microcredit Program Placement Experiment by Compartamos Banco. American Economic Journal: Applied Economics 7 (1): Armendariz de Aghion, Beatriz, and Jonathan Morduch The Economics of Microfinance. 2nd ed. Cambridge, MA: MIT Press. Attanasio, Orazio, Pinelope Goldberg, and Ekaterini Kyriazidou Credit Constraints In The Market For Consumer Durables: Evidence From Micro Data On Car Loans. International Economic Review, International Economic Review, 49 (2): Banerjee, Abhijit Microcredit Under the Microscope: What Have We Learnt in the Last Two Decades, What Do We Need to Know? Annual Review of Economics 5: Banerjee, Abhijit, Dean Karlan, and Jonathan Zinman Six Randomized Evaluations of Microcredit: Introduction and Further Steps. American Economic Journal: Applied Economics 7 (1): Banerjee, Abhijit, and Andrew Newman Occupational Choice and the Process of Development. Journal of Political Economy 101: Bengtsson, Niklas, and Jan Pettersson The Outreach and Sustainability of Microfnance: Is There a Tradeoff? Working Paper. Bhutta, Neil, and Benjamin J. Keys Interest Rates and Equity Extraction during the Housing Boom. American Economic Review 106 (7): Busso, Matias, and Sebastian Galiani The Causal Effect of Competition on Prices and Quality: Evidence from a Field Experiment. Cambridge, MA. Cunha, Jesse M, Giacomo De Giorgi, and Seema Jayachandran The Price Effects of Cash Versus In-Kind Transfers. DeFusco, Anthony A., and Andrew Paciorek The Interest Rate Elasticity of Mortgage Demand: Evidence from Bunching at the Conforming Loan Limit. American Economic Journal: Economic Policy 9 (1): doi: /pol Dehejia, Rajeev, Heather Montgomery, and Jonathan Morduch Do Interest Rates Matter? Credit Demand in the Dhaka Slums. Journal of Development Economics 97 (2): Gine, Xavier, and Rafael Keenan Mazer Financial (Dis-)Information: Evidence from a Multi- Country Audit Study. 30

31 Gross, David B, and Nicholas S Souleles Do Liquidity Constraints and Interest Rates Matter for Consumer Behavior? Evidence from Credit Card Data. The Quarterly Journal of Economics 117 (1): Hall, Robert E Intertemporal Substitution in Consumption. Journal of Political Economy 96: Jensen, Robert The Digital Provide: Information (Technology), Market Performance, and Welfare in the South Indian Fisheries Sector. The Quarterly Journal of Economics 122 (3): Karlan, Dean, and Jonathan Morduch Access to Finance. In Handbook of Development Economics, edited by Dani Rodrick and M. R. Rosenzweig. Vol. 5. Elsevier. Karlan, Dean, and Jonathan Zinman Credit Elasticities in Less-Developed Economies: Implications for Microfinance. American Economic Review 98 (3): Observing Unobservables: Identifying Information Asymmetries with a Consumer Credit Field Experiment. Econometrica 77 (6): Price and Control Elasticities of Demand for Savings. Ponce, Alejandro, Enrique Seira, and Guillermo Zamarripa Borrowing on the Wrong Credit Card? Evidence from Mexico. The American Economic Review 107 (4): Romer, David Do Firms Maximize? Evidence from Professional Football. Journal of Political Economy 114 (2): Rosenberg, Richard Microcredit Interest Rates. Consultative Group to Assist the Poor Occasional Paper (1). Schreiner, Mark Seven Extremely Simple Poverty Scorecards. Enterprise Development and Microfinance 21 (2): Senior, John, and Eric Almquist Your A/B Testing Isn t Working Nearly as Well as You Think. WIRED. Stango, Victor, and Jonathan Zinman Borrowing High vs. Borrowing Higher: Price Dispersion and Shopping Behavior in the U.S. Credit Card Market. Review of Financial Studies 29 (4): Zinman, Jonathan Consumer Credit: Too Much or Too Little (or Just Right)? Journal of Legal Studies 43 (S2 Special Issue on Benefit-Cost Analysis of Financial Regulation): S Household Debt: Facts, Puzzles, Theories, and Policies. Annual Review of Economics 7 (1):

32 Figure 1. Randomized Pricing by Study Region high rate low rate Bhutta and Keys (forthcoming AER) Ponce et al (2016 wp) DeFusco and Paciorek (forth AEJ: Pol)* 32

33 Control Group= Higher-rate regions Treatment Group= Lower-rate regions 33

34 Control Group= Higher-rate regions Treatment Group= Lower-rate regions 34