Digital Credit and Borrowers’ Welfare in Ethiopia: A Pilot Random Control Trial Study

2024-03-20

2024-04-15

The primary outcome variable is profitability of the digitl credit.

Employment and household consumption.

Sample draws from digital credit applicants who are eligible for credit and who are in the waiting list will be randomly assigned into treatment and control groups. The first treatment group will receive digital credit; the second treatment group will receive digital credit plus financial literacy training; while the control group will not receive any treatment for at least four months. The randomization unit will be at individual MSE level. The main potential identification challenge will be compliance problem from both the treatment and control groups. Analysis of administrative data from our partner bank show that, on average, 9% of the customers rejected the digital credit after being offered in the last 16 months of COOP’s digital service provision period, where the percentage of customers who rejected the credits remained fairly the same across time. Similarly, the control group whose application will be rejected may apply for credit somewhere else. These non-compliant customers may systematically differ with the customers who comply, causing potential bias. In this case, we propose three approaches. The first approach is to estimate intent-to-treat. The second approach is to consider the time period that the customers actually receive the credit as treatment indicator, particularly if the non-compliance rate will be very high. The third approach is to use propensity score matching technique and evaluate the impact only on customers who comply, for that we increased the sample size by about 20% as discussed below.

Not available

Randomization will be done in office by a computer. Eligible SMEs will be divided in to three groups by generating a random number (Bernoulli distribution) corresponding to the list of SMEs.

The randomization unit will be individual SMEs.

No

1,980 small and medium sized formal business enterprises

1,980 small and medium sized formal business enterprises

The sample size will be equally divided in to two treatment arms and a control group, each treatment arm having 660 sample size. That is, 660 SMEs will receive digital credit, 660 will receive digital credit and training about financial literacy and the remaining 660 SMEs will not receive digital credit and finincial literacy training. We will stratify the sample by gender: 31% female and 65% male, implying a random sample of 614 female and 1366 males. The gender distribution of the sample is based on information from previous digital credit applicants administrative data.

We calculate the minimum detectable effect (MDE) to be 0.16 standard deviations with significance level of 0.05 and power of 0.8. Given that we have multiple treatment arms, we use Bonferroni correction and divide the level of significance by the number of experimental arms. Thus, with our new level of significance (0.0166), we can detect the proposed effect size if we have 1,650 sample size. To account the potential compliance problem and to have a sufficient sample size, we will increase the sample size by 10%, implying 1,815 sample size. Moreover, there could be another potential compliance problem from the control group who may apply for digital credit in other digital credit providers. We do not have data on this issue, nevertheless, will increase the sample size by another 10% to have sufficient sample size without the compliance group, implying a total sample size of 1,980. The sample size will be equally divided into treatment and control groups. In addition, evidence from administrative data shows that gender is an important factor for decision making in digital credit application and individual’s life outcomes. Hence, we stratify our sample using gender. Administrative data from COOP shows that 31% of digital credit borrowers were women. Using the gender composition of our target population, (31% female and 65% male), we propose a stratified randomization, implying a random sample of 614 female and 1366 males, totaling a sample size of 1,980 number of beneficiaries. Given the magnitude of treatment effects in the latest studies, we believe that this study is well powered for economically meaningful comparisons against control.