The Egyptian Test: When a Larger Loan Helps
For formal credit and SME performance, the Egyptian experiment by Gharad Bryan, Dean Karlan, and Adam Osman asks a more precise question than whether lending is beneficial in general: what happens when eligible entrepreneurs receive a relatively large enterprise loan rather than a smaller comparison loan? That focus makes the allocation problem concrete. Readers can follow the lending margin from applicant and approval process to business outcomes, then see why a small average effect does not settle which firms should receive more credit.
A real lending margin, not an abstract borrower
The study concerns approved applicants to the Tamouh program at the Alexandria Business Association, a microfinance organization in Egypt. This starting point matters. The experiment is not a survey of all Egyptian businesses, and it is not a comparison between people who chose to borrow and people who never entered a lending process. Participants voluntarily applied for the larger loan and were approved by a loan officer and the bank’s credit committee. The relevant population is therefore the group of applicants who reached this institution’s approval margin and could plausibly receive the larger enterprise loan.
That boundary changes how the result should be read. A causal estimate can be highly credible for the people and decision under study without being universal. The experiment tells us about the effect of changing loan size for eligible applicants in this program and setting. It does not automatically tell us what would happen to an unregistered business that never applied, a firm in another country, or a borrower facing a different contract and financial system. External validity begins with knowing who was actually exposed to the decision.
The institutional roles are also part of the causal story. The researchers designed the randomization and evaluation protocols and managed the data collection and analysis. The Alexandria Business Association designed and implemented the loans. In other words, the intervention was not a hypothetical amount of capital handed out for research convenience. It was a lender’s enterprise-credit product delivered through the lender’s own approval process, with the research design placed around that operation.
The evaluation combines business information with measures intended to describe the people making decisions inside those businesses. Baseline and follow-up surveys include business outcomes, cognitive measures, risk questions, financial-literacy questions, and a broad set of psychometric questions. This combination allows the researchers to study two linked but distinct relationships. The first is causal: how does the larger loan change business performance relative to the comparison condition? The second is predictive: do baseline characteristics help identify firms whose performance responds differently to that change?
The distinction is important because the outcome of a business is never just a mechanical function of the cash received. A larger loan creates room for a purchase, an expansion, or a change in operations, but the firm still has to choose how to use it. Those choices occur in a market, under a repayment obligation, and with incomplete information about demand and risk. The treatment is the loan-size change; the path from that change to profit runs through the firm’s opportunities and decisions.
The randomized comparison gives the first relationship a defensible counterfactual. Instead of asking whether firms that took a larger loan look different from firms that did not, the researchers compare outcomes under different loan-size treatment conditions among applicants inside the lending evaluation. That design helps separate the effect of the larger loan from the many pre-existing differences that would otherwise make a borrower-level comparison difficult.
It still does not reveal a directly observed causal effect for each individual firm. For any one business, we see what happened under the loan condition that business received, not what the same business would have done under the alternative condition at the same time. The experiment can estimate an average effect and investigate systematic differences in effects across predicted groups. It cannot simply read an individual treatment effect from one observed profit figure.
The sequence shows why the study can speak to the causal effect of loan size while still needing a separate method for the “who benefits?” question. The lending institution determines who reaches the margin, the research design creates the comparison, and the firm’s later decisions mediate the observed business outcome. None of those roles should be collapsed into a single label such as “borrower quality.”
The small average is the beginning of the result
The paper’s central finding is that larger loans generate small average impacts. On its own, that result could be read in several ways. It might suggest that additional capital has little productive use among these applicants. It might indicate that gains and losses offset one another. Or it might mean that the loan changes some businesses substantially while leaving others largely unaffected. The additional analysis is what distinguishes the second and third interpretations from a simple story of uniformly weak demand for credit.
Using baseline psychometric data and machine-learning methods, the authors find large differences in predicted treatment effects. The group with the highest predicted gains—called “top performers” in the paper—substantially increases profits, while profits drop for the group with the poorest predicted response. The key word is predicted. The groups are formed from information available before the outcome, and the prediction is later checked against experimental responses. They are not fixed social categories, and the labels do not mean that every member is guaranteed to win or lose from a larger loan.
This is the cancellation logic behind the small mean. Suppose one set of firms can turn more capital into productive expansion, while another set uses the same opportunity in ways that do not cover the added obligation. When their outcomes are pooled, the positive and negative movements can produce a modest net effect. The modest mean is not evidence that nothing happened. It may be evidence that opposite responses are being averaged together.
The result also changes the meaning of “scale.” A lender considering a larger loan is not deciding only whether the average borrower benefits. The lender is choosing which applicants receive additional capacity and which applicants bear a larger repayment obligation. If borrowers respond differently, moving a larger loan toward a high-benefit firm could raise aggregate income more than assigning it without regard to conditional response. Assigning it to a low-benefit firm could produce little gain or greater exposure to repayment difficulty. The authors therefore argue that credit-allocation decisions matter for aggregate income and that existing practice leads to substantial misallocation.
That implication should not be inflated into a tested policy rule. The experiment does not show that a lender can immediately replace its approval process with a score, nor does it establish an optimal allocation for every market. It shows why the allocation question is economically meaningful: when treatment effects differ, the identity of the recipient can change the return generated by a fixed pool of lending capacity. The finding is about the relationship between loan size, firm response, and allocation—not a universal ranking of entrepreneurs.
How opportunity can become risk
The authors propose that some entrepreneurs are overoptimistic and squander the opportunity created by a larger loan by taking on too much risk. This is a mechanism hypothesis, not a diagnosis that the randomized treatment proves for every low-benefit firm. The causal evidence identifies what happened when loan size changed. The proposed mechanism explains why the same increase in capacity might produce different business outcomes.
The chain is straightforward enough to understand without making it simplistic. A larger loan can make an attractive opportunity financeable. If the owner recognizes demand, controls costs, and chooses an expansion that the business can support, the extra capital may raise revenue and profit. But the same loan can also make a risky plan feasible, encourage commitments that are difficult to reverse, or increase the scale of a mistake. Once repayment is required, the downside is not limited to an opportunity that failed to materialize; it can include a larger financial burden than the business can comfortably service.
The appendix’s survey instruments reflect this concern. They include questions about willingness to take risks, hypothetical risky investment, financial literacy, cognitive performance, and many aspects of decision-making and personality. These measures are useful for asking whether baseline information predicts different responses. They do not establish that one answer causes a particular profit outcome. A survey response may stand in for several underlying conditions, and a model may combine many weak signals in a way that is difficult to interpret.
The study’s ethical discussion makes the stakes visible without turning them into a morality tale. Bank management had been hesitant about providing larger loans. Participants expected to benefit, and both borrowers and bank staff viewed the intervention as an opportunity to test whether larger loans would help clients and the bank. The researchers designed the study to limit harm: applicants volunteered, and loan officers and the credit committee approved them. Even with those safeguards, the results indicated that some participants did not benefit, with decreases in profits and related outcomes. The lesson is not that larger loans are inherently dangerous. It is that a loan can transfer both opportunity and exposure, and a responsible evaluation has to look for both.
This also explains why a predictive signal should not be treated as a psychological verdict. Calling an entrepreneur “overoptimistic” can sound like an explanation while hiding the actual uncertainty. Perhaps the model is detecting planning behavior, risk tolerance, business conditions, or a combination that changes with the market. Perhaps a characteristic predicts response in this sample because it correlates with an unmeasured constraint. The source supports the proposed opportunity-risk mismatch; it does not justify reducing it to a single personality label.
From access to allocation
The Egyptian case preserves the value of expanding credit while making the allocation problem unavoidable. The broad evidence suggests that formal loans can improve SME employment, sales, and profits on average. This experiment shows why a lender still needs to ask what happens inside that average when the treatment is a relatively large enterprise loan. If the average is small because gains and losses offset, then expanding the pool without improving allocation may fail to capture the available benefit.
The relevant counterfactual is not simply “more loans versus fewer loans.” It can be “the same lending capacity directed in different ways.” A lender might ask whether a larger loan should go to an applicant for whom finance appears to be the binding constraint, or to an applicant whose expected response is less favorable. A policymaker might ask whether a public program should preserve its targeting rather than broaden eligibility without evidence that the additional recipients face the same constraint. These are applications of the study’s logic, not interventions tested by it.
The distinction also protects against a false choice between human judgment and machine prediction. The appendix’s loan-officer analyses do not provide a definitive verdict that officers cannot identify beneficial borrowers, and the paper does not establish that a machine-learning ranking is ready for deployment. The stronger lesson is that ordinary allocation practice may fail to capture some treatment-effect variation, while a predictive method can reveal a signal worth testing. Any operational use would still need prospective validation, error analysis, attention to fairness, and monitoring of repayment and market effects.
The case ends where the method begins. We know that the larger-loan treatment has a small average impact and that predicted groups respond differently. We also know that the lowest predicted group’s label is not a causal explanation, and that the study’s applicant pool limits how far the result can travel. The next question is therefore technical but practical: how did the researchers keep the prediction from being built on the same outcomes used to test it? Understanding that sequence is necessary before treating the observed contrast as evidence about allocation rather than as an attractive story told after the fact.