Discriminatory Lending: Evidence from Bankers in the Lab
How to cite
Haas, Ralph De; Brock, Michelle. 2022. "Discriminatory Lending: Evidence from Bankers in the Lab." AEA RCT Registry. AEARCTR-0009025. https://doi.org/10.1257/rct.9025-1.1
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
We implement a lab-in-the-field experiment with 334 Turkish loan officers to document gender discrimination in small business lending and to unpack the mechanisms at play. Each officer reviews multiple real-life loan applications in which we randomize the applicant's gender. While unconditional approval rates are the same for male and female applicants, loan officers are 26 percent more likely to require a guarantor when we present the same application as coming from a female instead of a male entrepreneur. A causal forest algorithm to estimate heterogeneous treatment effects reveals that this discrimination is strongly concentrated among young, inexperienced, and gender-biased loan officers. Discrimination mainly affects female loan applicants in male-dominated industries, indicating how financial frictions can perpetuate entrepreneurial gender segregation across sectors.
Intervention
Participants evaluated four applicant forms (henceforth "loan applications'') in the main part of the experiment. We randomly presented these applications as coming from a woman or a man. Participants had to decide whether to approve or reject each application and, in case of initial approval, whether to request a guarantor or not. For each application, participants also had to provide a subjective repayment probability between 0 and 100. We did not constrain the time participants had to evaluate the applications and there was no feedback to participants about their decisions during the session. The sessions were framed as a general training exercise and no gender-related issues were mentioned.
Experimental design
We conducted our experiment in cooperation with a large commercial bank in Turkey. Over a two-month period, 22 experimental sessions were held with a total of 334 bank employees across eight cities (Adana, Ankara, Antalya, Bursa, Gaziantep, Istanbul, Izmir, and Trabzon). We also conducted a pilot session with 32 loan officers in Istanbul but do not use these pilot data. The bank operates a regional office in each of these cities and participants were randomly selected from all bank employees involved in small business lending (which makes up two-thirds of the bank's loan portfolio.
We incentivized decisions in line with common bank incentive schemes. Participants earned ten points (equivalent to ten Turkish lira) for each completed review (quantity) and an additional five points when they correctly approved a loan that performed well in real life (quality). Five points were deducted when they incorrectly accepted a loan that was defaulted on in real life. When participants approved a file that had been declined in real life, we gave them a 50/50 chance that the file was counted as performing, thus yielding the extra five points. We did not penalize incorrect rejections in order to mimic the incentive scheme at the bank, and the bank cannot realistically know when a rejection is incorrect.
Participants evaluated four applicant forms (henceforth "loan applications'') in the main part of the experiment. We randomly presented these applications as coming from a woman or a man. Participants had to decide whether to approve or reject each application and, in case of initial approval, whether to request a guarantor or not. For each application, participants also had to provide a subjective repayment probability between 0 and 100. We did not constrain the time participants had to evaluate the applications and there was no feedback to participants about their decisions during the session. The sessions were framed as a general training exercise and no gender-related issues were mentioned.
The task closely mimicked the daily choices that participants make in real life at work. Specifically, we presented all loan applications electronically and in the standard application format that bank staff normally process on their computers. The loan applications contained all the information that was required for determining creditworthiness of an applicant and that was available at the time the application was processed. The loan applications did not include information about whether in real life a guarantor was requested.
We use 100 applications, selected from an initial sample of 250. These 250 applications were a stratified random sample of all first-time applications by existing SMEs (that is, no start-ups) that the bank received in the three to six years before the experiment. Using this earlier period allows us to track what happened to each application in real life. The strata were region, gender, firm size, and whether the application was performing, non-performing or declined in real life. By using applications from applicants who had never before borrowed from our partner bank, we minimize the influence of soft information generated over time. All applications were gender neutral except for the randomly assigned name.
Design details
Cited by
No citation data available for this registration yet.