This is a randomized control trial where the treatment variable is the gender of the individual that submits the loan request to the assigned loan officer. In particular, this is an RCT implemented in the form of a corresponding study where loan officers randomly receive loan requests from either male or female borrowers according to randomization.
A secondary treatment is an information intervention where one-paragraph email about the importance and the prevalence of gender discrimination is randomly sent from the governmental authority to half of the loan officers considered for this study. Here the treatment variable is whether the loan officer receive the message or not.

2018-07-12

2018-09-27

We analyze response, approval, and rejection rates across gender (external margin), as well as amount, interest rate, and term of approved loans (internal margin). Please see the pre-analysis plan for details

Response rate: proportion of loan requests responded by bank executives (from total of loan requests sent by loan applicants).
Approval rate: proportion of loan requests approved by bank executives (from total of loan requests sent by loan applicants).
Rejection rate: proportion of loan requests rejected by bank executives (from total of loan requests sent by loan applicants).
Amount: amount of approved loan
Interest rate: interest rate of approved loan
Term: loan term of approved loan
Please see the pre-analysis plan for details

Please see the pre-analysis plan for details

Please see the pre-analysis plan for details

By using a self-selected sample of incentivized loan applicants ("actors"), we match 404 male and female profiles on demographics and credit history information to build a balanced sample of actors across gender. The recruitment of actors was made based on randomly selected emails from DEMRE data, which contain detailed information about emails and test scores of people that have taken the PSU test from 2000 onward (PSU = Prueba de Seleccion Universitaria, a national test on math, verbal, etc. taken every year by any student who applies to the university system). Then, our sample is composed by relatively young professionals with short credit history. Information about incomes and employment status is obtained through wage settlements and social security contributions reports attached by the testers. Finally, data on credit history of each actor is obtained through SBIF administrative records (SBIF = Superintendencia de Bancos e Instituciones Financieras), the national authority for banking regulation. Overall, our prettreatment information allows us to fairly match the set of variables typically used by the banks to evaluate consumer loan applications.

Applicants are asked to send 4 loan requests through email to 4 different and randomly assigned bank executives who are in charge of selling loans to clients. Applicants receive around US$20 for sending the 4 loan applications, for a total sample of 1,616 loan applications distributed across more than 600 bank executives. Loan requests are text-standardized, but randomly vary across amount and loan term so we can test heterogeneous effects across the two dimensions. For each loan application, the actor receives the exact text and email address of the bank executive, so they copy and paste the text with the loan request and send it to the email address from their personal accounts. Responses from the executive are tracked by the actor who resend the information to the research team.

Bank executives are randomly selected from a nationally representative survey of bank executives administered through SBIF. The survey includes information on gender, education, and years of experience, as well as a series of gender-preference elicitation pretests.

Finally, we further conduct a salience experiments where an email about the importance and the prevalence of gender discrimination is randomly sent from the governmental authority to half of the loan officers included in our study, so that we can test the effectiveness of information treatments on reducing gender discrimination in the consumer credit market. The emails were sent at the beginning of the experiment to all the executives at the same time, the half of which randomly contains treatment information.

Randomization was done in office by a computer

There are several randomizations:
1. We randomly assign loan amount to each loan request. There are 9 types. Randomization unit is the loan application, stratified by applicant.
2. We randomly assign loan term to each loan request. There are 9 types. Randomization unit is the loan application, stratified by applicant.
3. We randomly assign 4 loan requests to each tester.
4. We randomly assign loan requests to loan officers. We stratify by region-bank where the loan officer works as well as for loan officer's gender.
5. We randomly assign salience treatment to loan officers. All receive email from SBIF, but half of them additionally receive the information treatment.

Yes

The treatment status of each loan request is given by the tester's gender. The outcome of the loan application (responded, approved, rejected) is endogenous to the decision making process made by the bank executive. Each bank executive will never evaluate more than one application coming from the same actor. Thus, from the eyes of the econometrician, each application is independent from each other. However, these may well be correlated within region-bank, so we cluster the standard errors by region-bank.

Total number of actors: 404 Number of males in sample: 233 Number of females in sample: 171 Number of applications per actor: 4 Total number of loan applications: 404*4 = 1,616

Number of males in sample: 233
Number of females in sample: 171
Number of applications per actor: 4
Number of female applications in sample (treatment): 233*4 = 932
Number of male applications in sample (control): 171*4 = 684
Total number of loan applications: 1,616

Information Treatment:
Number of loan requests assigned to treated loan officers (treatment): 787
Number of loan requests assigned to control loan officers (control): 795

For a sample of 684 loan applications made by men and 932 made by women (n1/n2 ratio=0.73), our experimental design is able to identify across-gender differences on the order of 1.1% in any of the external margin outcomes (i.e., response, approval, and/or rejection rate), at the 95% level of statistical significance and with a statistical power of 80%. Likewise, assuming a standard deviation of 1, our experimental design is able to identify across-gender differences on the order of 0.14 standard deviations in any of the external margin outcomes, at the 95% level of statistical significance and with a statistical power of 80%. As detailed in the pre-analysis plan, our hypothesis is that there is discrimination against females (not against males), thus we will report one-sided test of the null hypothesis that discrimination is against female borrowers.

Pre-Analysis Plan