Experimental Design
Loan officers are asked to evaluate 45 hypothetical loan applications, in three sets. Each set is a treatment arm. The order of the treatment arms is fixed, but each hypothetical application is randomly selected into one of each arm for each participant.
The applications are grouped into two loan profiles, which vary by amount requested and repayment rate. The first loan profile is a Ush. 1 million, with 6 months repayment; the second loan profile Ush. 5 million, with 12 months repayment. Notably, I exclude Tier 1 banks from my sample as they generally do not lend as low as Ush. 1 million.
Within each loan profile, the applications vary by characteristics; in each application, the applicants' age and residence are defined statically, while the applicants' occupation, income, portrait, body mass are cross-randomised. In particular, to avoid suspicion, portraits are randomized without replacement across the full set of applications each respondents receives.
Using this procedure, I build 45 loan applications and randomly assign each to a treatment arm. In the first treatment arm, respondents only see the picture, age and loan profile requested. In the second treatment arm, respondents only see the picture, age and loan profile requested but can choose to learn occupation and income of the applicant. In the third treatment arm, participants see all the information (including the picture) at once.
Then, the remaining applications are shown in random order. For each application, the loan officer answers the following questions:
[Main outcomes]
(1) A Yes/No choice of meeting the applicant to discuss and verify information; (2) a 1-6 Likert scale assessment of the probability of referring the person to the next step of the application process (3) a 1-6 Likert scale assessment of the applicants' probability of receiving the requested loan.
[Secondary outcomes]
(4) A 1-6 Likert scale assessment of the probability that the applicant makes productive use of the money.
(5) A 1-6 Likert scale assessment of the probability that the applicant will repay/collateral will be collected, conditional on the investment not being successful.
For loan profiles where the loan officer has full discretionally on the application decision, question (2) and (3) are substituted by question (6) a 1-6 Likert scale assessment of the probability of granting the requested loan.
The set of branches is obtained by randomly selecting 70 branches and headquarters in Kampala. For each branch, 1 to 3 loan officers are interviewed. The procedure is as follows: The enumerator asks who are the people that directly in touch with with (potential) borrowing customers. If there are more than 3, the enumerator randomly selects 3 of them to be interviewed. If there are less than 3, everyone is interviewed.
When answering the questions, the loan officers are informed that the applicants are hypothetical. Honest answers are incentivised by informing the loan officers that their answers will be used to provide referrals to prospective clients whose characteristics match with their preferences. The matching is performed using a machine learning algorithm, following Kessler et al. (2019).
The design allows to investigate whether (1) body mass affects perceived creditworthiness and referral probability, (2) whether reducing asymmetric information on income changes the relevance of body mass in determining the decision, (3) at which point of the decision making process does discrimination bites.
Relevant heterogeneity analysis include: 1) gender; 2) body mass of the loan officer; 3) financial literacy/experience; 4) perception of high body mass. If heterogeneity in the share of applicant invited is detected (either at the loan officers' level or at the institution level), an interesting heterogeneity analysis compares "lemon dropping" types to "cherry picking" ones. The design builds on Kessler et al. (NBER WP 2019) and Bartoz et al. (2016).
At the end of the experiment, loan officers answer a set of questions on financial literacy; perception of high body mass;