Experimental Design Details
This study tests whether the marginal informational value of borrower-written narratives in small-loan applications varies with the gender of the officer assigned to review the file. In partnership with a commercial bank in Afghanistan, adult micro-entrepreneurs are invited to apply for a small standardized, collateral-free loan. Each application includes standard hard information and an open-ended narrative where applicants explain, in their own words, why the bank should trust their ability and willingness to repay. Each applicant can provide one narrative addressed to a male loan officer and one addressed to a female loan officer. All applications are anonymized prior to review, and loan officers cannot view borrower names, photos, or any identifying information that implies gender.
Each application is scored by multiple officers under 3 randomly assigned conditions at the officer-application level:
Version 1. Narrative Concealed: The officer sees the application (hard information), but no narrative is shown.
Version 2. Narrative Written for Female Loan Officer Visible: The officer sees the application plus the borrower’s narrative addressed to a female officer.
Version 3. Narrative Written for Male Loan Officer Visible: The officer sees the application plus the borrower’s narrative addressed to a male officer.
Officers provide standardized repayment-likelihood scores that the partner bank uses in its real lending decisions under fixed, non-negotiable terms. For each application, the difference in mean scores across versions identifies the marginal informational value of the narrative.
Following the evaluation phase, we conduct a post-review survey to assess whether exposure to borrower gender information shapes or amplifies officers’ gender-related beliefs about borrower quality or communication ability. Each loan officer receives a summary report listing the applications they reviewed along with the scores they assigned. For a randomly selected half of the officers, the report also includes the gender of each borrower; for the other half, gender remains undisclosed. Comparing responses across these groups allows us to test whether communication frictions arising from facing a loan officer of a different gender can partly explain loan officers’ gender-based stereotypes about borrowers, particularly among male officers.
AMENDMENT OF SEPTEMBER 11, 2026 — NEW VALIDATION STUDY AND LOAN-OFFICER EXPERIMENT
Loan-officer experiment. We use the original first-round borrower sample. There are 391 usable applications with at least one narrative, including 43 financed and 348 non-financed applicants; 351 applicants provided narratives for both male and female officers. Each new loan officer evaluates 150 applications: all 43 financed applications plus 107 applications randomly sampled without replacement from the non-financed pool, separately for each officer.
For applicants with both narratives, one of three versions is randomly selected with equal probability: no narrative, the neutralized narrative originally written for male officers, or the neutralized narrative originally written for female officers. For applicants with only one narrative, assignment is equally likely between no narrative and the available neutralized narrative. Each officer sees only one version of an applicant.
Officers are independently assigned, approximately equally, to one of two officer-level regimes. In Neutralized-N, all displayed narratives appear under a generic heading that does not identify the gender of the original intended addressee. In Neutralized-A, the heading truthfully identifies whether the displayed narrative was originally written for male or female loan officers. Narrative text is otherwise identical across regimes. The 150 assigned applications are randomly ordered within officer.
The primary outcome is the officer's assessed probability of repayment, from 0 to 100 in increments of 10. The 43 financed applications form an undisclosed benchmark set used for performance incentives. Officers receive a fixed payment plus a performance payment based on the Brier score of their benchmark predictions. Repayment is coded as 1 only for full repayment of AFN 10,000 by contractual maturity and 0 otherwise.
The primary source comparison uses the 351 applicants who supplied both narratives. For each applicant-gender × officer-gender combination, we compare the evaluation of the narrative originally written for the officer's gender with the narrative originally written for the opposite gender, separately in the two heading regimes. We also estimate the cross-regime difference in this source contrast. The focal cross-regime estimand is for female applicants evaluated by male officers. Primary specifications use evaluation scores in levels, applicant and officer fixed effects, inverse-exposure weights, and standard errors clustered by applicant and officer. Unweighted and log-score specifications are secondary. Narrative-versus-no-narrative comparisons and Brier-loss analyses are also secondary.
Validation survey. Before the officer experiment, we will conduct a validation survey with native Dari speakers who can read Dari and who were not involved in the earlier experiments or employed by the partner bank. Assignments are prepared for up to 100 participants. Each participant evaluates 50 passages: 46 neutralized narratives, balanced so that 23 were originally written for male officers and 23 for female officers, plus four original positive-control narratives with explicit recipient-gender cues. No participant sees both narratives from the same applicant, and passage order is randomized.
For each passage, participants make a required male/female guess about the original intended addressee and are informed that either gender is equally likely. Neutralized passages are analyzed separately from positive controls. The main validation analysis compares correct classification of neutralized passages with the 50% chance benchmark. Failure to reject a difference from 50% will be interpreted as absence of evidence that respondents classify the intended addressee better than chance, rather than as proof of exact equivalence.
The neutralized narrative corpus was produced and frozen on August 18, 2026. The validation survey is planned for September 12, 2026, and the new officer experiment is planned to begin September 19, with evaluations completed by October 3, 2026.