Who Benefits—and What the Score Cannot Explain

The Egyptian experiment makes the allocation question visible, and its prediction method shows that the question can be tested. The next issue for formal credit and SME performance is what information sits behind the ranking—and whether that information can be used responsibly. This chapter separates baseline signals from causal explanations, compares them with bounded loan-officer judgments, and sets out what must be validated before a predictive score influences who receives a larger loan.

What the baseline signals are—and are not

The online appendix examines a wide set of information collected before the loan-size treatment. It includes business characteristics, cognitive performance, risk attitudes, financial literacy, and psychometric questions about planning, decision-making, motivation, flexibility, and responses to uncertainty. The researchers use a lasso analysis to summarize a seven-variable prediction of individual treatment effects and a random-forest analysis to report feature importance. They also compare baseline characteristics of groups at opposite ends of the estimated-impact distribution.

The purpose of these analyses is not to produce a personality test. It is to ask whether observable information helps predict how a firm responds to a larger enterprise loan. That is a narrower and more defensible claim. A lender may learn that certain combinations of planning behavior, risk-related responses, cognitive measures, or business characteristics are useful for ranking expected treatment response in this sample. The lender has not thereby learned that one personality type is inherently suited to borrowing, or that a survey answer determines a firm’s future profit.

The distinction becomes clearer when reading feature importance. A high importance score means that a variable helped the model distinguish among estimated treatment effects. It does not, by itself, show whether a higher answer raises or lowers the predicted response. It does not say that the variable is the cause of the difference. Nor does it show what would happen if a lender changed the variable or tried to select for it. A model can use a variable as a proxy for several conditions that are not separately observed.

The group comparisons have a related limitation. The appendix reports many statistically different psychometric characteristics between top and bottom estimated-impact groups. Those differences can make the groups easier to describe, but they do not turn the descriptions into mechanisms. A group that scores differently on planning or risk questions may also differ in business opportunities, constraints, networks, or other characteristics. The prediction tells us that a combination of information is associated with different responses to the larger loan; it does not tell us which component is doing the causal work.

This is why the proposed overoptimism-and-risk explanation needs careful language. The authors argue that some entrepreneurs may squander the opportunity created by a larger loan by taking on too much risk.