Ebook Expanding Credit Access: Using Randomized Supply Decisions to Estimate the Impacts
Expanding access to credit is a key ingredient of development strategies worldwide. The microfinance industry has grown exponentially over the past twenty years under the premise that expanding access to credit will help improve the welfare of the poor (Morduch 1999; Armendariz de Aghion and Morduch 2005). This policy push has been driven by both theoretical and empirical motivations. Theoretical models show that information asymmetries can lead to credit market failures and ensuing poverty traps (Banerjee and Newman 1993). Empirical evidence shows strong positive correlations between depth of access and poverty rates at the macro level (Levine 1997; Honohan 2004), positive impacts of access to microfinance at the micro level (Pitt and Khandker 1998), and positive impacts from expansions of bank branch networks on aggregate poverty (Burgess and Pande 2005). Policymakers, practitioners, and funders are committed to continued rapid growth.
There is less consensus on the role of consumer credit in expansion initiatives. Some microfinance institutions are moving beyond “traditional” entrepreneurial credit and offering consumer loans. But many practitioners remain skeptical about “unproductive” lending (Robinson 2001). Policy is similarly conflicted about lending at “usurious” rates. Concerns about the development of consumer credit markets are fueled by academic work highlighting behavioral biases that may induce consumers to overborrow.
There is also uncertainty about how to expand credit access. Traditional approaches to microcredit expansion creating new microfinance institutions, adding branches, using joint liability mechanisms to overcome high fixed transaction costs and poor screening, monitoring and enforcement capabilities may not be the most cost effective method to support efficient expansion. Another way to expand access to credit is simpler: liberalizing screening criteria.
We assess the impacts of liberalizing credit screening criteria by analyzing new data produced by a field experiment and follow-up survey work and data collection. The key questions are threefold. First, do credit constraints bind? Second, does relaxing any credit constraints benefit marginal borrowers? Revealed preference logic says it should: a consumer borrows only if she will benefit (weakly, in expectation). Behavioral models say not necessarily: biases in preferences and cognition may lead consumers to overborrow. The third key question is whether the lender profits from making these marginal loans.
The experiment was implemented by a consumer lender in a high rate, high risk South African installment loan market where credit constraints appear to bind. First time applicants are often rejected, even at prevailing real rates of 200% APR. Default rates average about 20% among new borrowers. A prior experiment on experienced borrowers from the same lender found far greater sensitivity to maturity than price (Karlan and Zinman forthcoming); as Attanasio, Goldberg, and Kyriazidou (forthcoming) show formally, this pattern of elasticities is further evidence of unmet demand for credit.
Measuring the causal impacts of credit expansion on borrower and lender outcomes is usually complicated by deep identification issues. Two types of endogeneity are particularly problematic: the self selection of clients into loan contracts, and targeted interventions by lenders and policymakers. These problems make it difficult to draw firm conclusions from non experimental studies without strong assumptions. A classic example concerns relatively “spunky” individuals selecting or being selected into microcredit borrowing, and thereby confounding any causal effect of access to credit with the causal effects of individual characteristics (including those that may change unobservably over time). Selection can work in the opposite direction as well; e.g., if households (lenders) tend to take (target) microcredit in anticipation of needing to smooth upcoming negative shocks. Attempts to overcome these problems using quasi-experimental, structural, and control function approaches have yielded mixed results.
We addressed the identification problem by working with a lender to engineer exogenous variation in the loan approval process. Our treatment randomly encouraged loan officers to approve some marginal applications. Specifically, the Lender added three additional steps to its normal process for new loan applicants. First, loan officers were required to label rejected applications as either egregiously uncreditworthy or marginally uncreditworthy. Second, the loan officer’s computer then instructed the loan officer to reconsider some marginal applications in real time by randomly producing a message to “approve” or “still reject.” Loan officers were instructed by management to follow the computer’s instructions in all cases. But in the third and final step, loan officers had pecuniary incentives to be risk-averse and approved the loan in only 53% of the cases when the computer instructed them to approve. Consequently our design identifies treatment on the treated effects of expanding access on a policy and strategy relevant sample: those applicants deemed by loan officers to be closest to the margin of creditworthiness. Neither the treatment (computer said “approve”) nor the control (computer said “reject”) groups were informed by the Lender that a component of the loan decision was randomized.
We then estimate the average impacts of expanding credit access by comparing outcomes across applicants assigned to the treatment and control groups. Our outcome data comes from the Lender’s records on repayment and profitability, from credit bureau reports over two years after the start of the experiment, and from household surveys conducted by an independent firm at the home or workplace of the marginal applicants six to twelve months after the start of the experiment. The survey measures borrowing activity, loan uses, and a range of proxies for household well-being.
Our results corroborate the presence of binding liquidity constraints. Control applicants did not simply obtain credit elsewhere; conversely, treated applicants borrowed more overall in the 6-12 months following the experiment, and changed their lender type composition.
Measuring the ultimate impacts of consumer credit on borrowers presents several challenges. There is no natural summary statistic for household utility; hence we follow evaluations of social policy interventions and measure treatment effects on a range of variables that capture economic behavior and subjective well-being (Kling, Liebman and Katz 2007). But treatment effect channels may vary across households; e.g., some households may smooth consumption by making critical purchases, others may use loan proceeds to maintain employment in the face of adverse shocks to transportation or family health, others may make investments as more traditionally defined (in self-employment, housing, schooling, or health), while others may benefit in less-tangible ways (becoming more hopeful about future prospects, or acquiring more bargaining power in the household). Consequently we use summary index tests that aggregate across outcomes to address the problem of multiple inference (Anderson 2007; Kling, Liebman and Katz 2007).
We find that expanded access to credit significantly improved average outcomes. Over the 6 to 12 month horizon, applicants in the treatment group were significantly more likely to retain their job over the study period, and treatment group incomes were significantly higher. Treated households were also less likely to experience hunger, and had more positive outlooks on their prospects and position. We do find a significant and negative impact on other aspects of mental health (depression and stress). But the average treatment effect across all of our economic and subjective outcomes is significant and positive. Over 15 to 27 month horizons, we find a positive impact on having a credit score, and no impact on the score itself. The effects on credit scores cast doubt on the hypothesis that positive treatment effects will turn negative over longer horizons due to debt traps or other delayed realizations of the cost of borrowing.
Perhaps most critically, the confidence intervals for treatment effects on our summary impacts (the overall index of survey outcomes, and credit scores) rule out substantial negative effects. This is important because the default policy regime for consumer credit is restricted access based on the presumption of negative effects on the margin.
The Lender agreed to implement this experiment because its senior management believed that branch staff applied inefficiently strict underwriting criteria. Our estimates of loan returns suggest that this prior was well-founded. The evidence suggests that the marginal loans were profitable in an absolute sense, although substantially less profitable than inframarginal loans. Exactly how profitable depends on several assumptions about marginal costs and risk-weighting.
In all our results suggest a role for welfare improving interventions in consumer credit markets but come with other important caveats. We only directly measure the ultimate borrower outcomes of interest at 6 to 12 month horizons, and some costs and/or benefits may only materialize over longer horizons. The external validity of treatment effects in the South African cash loan market is unknown.
Despite these limitations, our results and methodology offer some novel insights into the motivation, design, and evaluation of credit market interventions. We demonstrate that randomized controlled trials can be used to help identify the severity of liquidity constraints, and to evaluate efforts to expand credit access. Most practically, our results suggest that liberalizing screening criteria can benefit both borrowers and lenders, and our methodology demonstrates how lenders can hone in on their sustainability/outreach frontier by taking controlled risks using randomized experimentation.
The paper proceeds as follows. Section II provides background information the applicants, the Lender, and the cash loan market. Section III details the design and implementation of our experiment and data collection methods and empirical strategy. Section IV presents estimates of treatment effects on borrowing and credit access. Section V presents estimates of treatment effects on component and summary index measures of ultimate outcomes of interest. It also presents our estimates of effects on credit scores 15-27 months after treatment, and details our estimates of Lender profits on marginal and inframarginal loans. Section VI concludes with a discussion of external validity and other questions for future research.
Download
PDF Ebook Expanding Credit Access: Using Randomized Supply Decisions to Estimate the Impacts
Posted in :
