Learning Across Genders: Evidence from the Microcredit Market in Ghana

We work with a large microcredit company in Ghana. The intervention consists of providing clients with information about the average amounts requested by similar men, similar women, or both when they call to request a new loan. We will analyze how providing clients with this information affects their final loan request, the amount of loan they receive, and their loan repayment rates.

2023-11-20

2024-06-01

- Final loan amount requested

- Final loan amount requested:
This is the final amount that the call center agent records. In all the arms with information provision, this is the amount after information provision. In the control groups, it is the final amount requested, absent the information treatment. We will look at the absolute final amount, the difference between the final amount and the initial amount requested (absolute and percentage), and whether a revision was made at all (a dummy variable).

- Amount of money received
- Default rates

- Amount of money received:
The amount the borrower actually received, after the application was processed by the company. We look at both the amount and the difference (absolute and percentage) from the borrower's final requested amount, as well as whether the amount was reduced at all by the loan officers (a dummy variable).

- Default rates:
A measure of repayment performance (how much money the borrower still owes after the loan has matured). We will look at both the absolute amount, the percentage of the loan still owed, and whether any amount is owed at all (a dummy variable).

We will conduct an RCT with potential clients of a large microfinance institution in Ghana. Each client will be randomly assigned to one of three treatment arms (or to a control group that receives no information) and will be provided with information about the average amounts requested by other similar borrowers (men, women, or both men and women).

We will intervene at the first stage of the loan application process, where potential clients apply for loans via a telephone call with a call center agent of the microcredit company. After collecting basic information (including the initial loan amount requested), in the randomized conditions the call center agent elicits prior beliefs about the typical amounts requested by other men and women with similar characteristics. The client is then provided with information about what other similar borrowers request on average. Exactly what priors are elicited and what information is provided to the client depends on the treatment group into which the client was randomized, as follows:

1. Control group: both priors elicited, no additional information provided

2. Own-gender-info-only treatment: same-gender prior elicited, information about average amount requested by other clients of the same gender

3. Other-gender-info-only treatment: opposite gender prior elicited, information about average amount requested by other clients of opposite gender

4. Double-info treatment: priors elicited for both genders, information about average amount requested by both other male and female clients, order of elicitation and provision of information is randomized.
The average loan amounts provided in the information treatments come from either an XGBoost algorithm or a regression model run on the company's historical data. Both models perform similarly on average, but there are differences in predictions at the individual level. Within each treatment group, we provide either the lower or higher amount with a 50% probability. In the double-info treatment, this implies 4 different combinations, each chosen with a 25% probability.

This design will allow us to test whether and how people respond to information about their own gender, the opposite gender, or both genders, and to pinpoint the mechanisms at play, including whether there are any psychological biases or inconsistencies.

We will conduct the main analysis pooling both genders (and coding treatments as same gender vs. opposite gender) as well as separately for men and women.

We will also conduct the following heterogeneity analyses:

- Heterogeneity by gender, priors and initial request amount.
- Heterogeneity by old vs. new borrowers, as we hypothesize that the largest effect would be for first-time borrowers, since they have the least experience and information about market conditions.

Not available

Randomization is done by computer in Stata

Individual level (client)

No

We expect to have at least 1000 observations (clients). The treatment is not clustered.

We expect to have at least 1000 observations (clients). However, since we don't know how many calls will be received during the experimental period, we are uncertain about the total number of observations.

The total sample will be split into four groups: with a probability of 0.1, an individual will be assigned to the control group, and with a probability of 0.3, an individual will be assigned to each of the three treatment conditions. These probabilities will be stratified by gender and conducted separately for each call center worker.

Importantly, we will have both a randomized and a non-randomized control group. We indend to complement our control group with a "non-randomized control" group consisting of observations recorded at the microcredit company during, before and after our intervention. We intend to test whether this "non-randomized control" is comparable to our experimental control group. The "non-randomized control" may differ from the randomized control group due to the selection of call center agents and the absence of prior elicitation in the "non-randomized control". We expect the non-randomized control group to be at least one order of magnitude larger, perhaps two, than our randomized control group, leading to greater precision.

The high vs. low treatment will be stratified by treatment group. In T1 (own-gender-info-only) and T2 (other-gender-info-only), high and low treatment will be assigned with 50% probability each. In T3 (double info-treatment), we will have 4 combinations (high-low, low-high, low-low, high-high), each with a 25% probability of occurring.

We aim to have a control sample size of at least 100 and a treatment sample size of at least 300 for each treatment arm.