Providing rigorous evidence on gender discrimination in access to finance is challenging for at least two reasons. First, not just the supply of credit but also the demand for it can differ by gender. For example, women may self-select into industries that are less capital intensive or that operate at a smaller scale. Second, gender discrimination can be both direct and indirect. Even when loan officers are unlikely to reject female loan applicants straight away, they may apply conditions that make credit de facto unattainable for many women. Previous non-experimental evidence suggests that guarantor requirements may be a source of such indirect gender discrimination (Alesina et al. 2013).
In a recent study (Brock and De Haas, 2019), we test for the presence of both direct and indirect gender discrimination in small business lending in Turkey. We implemented a lab-in-the-field experiment during which loan officers evaluated loan applications in which the applicant gender had been (randomly) manipulated by us. The setting also allowed us to measure various loan officer characteristics, including their implicit gender bias and risk preferences, that normally remain unobservable.
The experiment in a nutshell
We conducted our experiment with 334 bank employees of a large commercial bank in Turkey. Each of these participants evaluated two rounds of four loan applications (so eight in total). We determined the gender of each application by assigning new names, randomising between male ones (Ahmet, Ali, Mehmet, Mustafa) and female ones (Ayse, Emine, Fatma, Zeynep). Participants had to decide whether to approve or reject each application and, in case of initial approval, whether to request a guarantor or not.
All applications occur in the data multiple times, sometimes as coming from a male applicant and sometimes from a female applicant. This allows us to obtain a within-application estimate of gender discrimination. Importantly, loan officers assessed applications that our partner bank had received in the recent past so that we know how loans subsequently performed in reality (Cole et al. 2015 follow a similar approach). We incentivised all loan decisions in line with common bank incentive schemes.
For each credit application, the participant was also asked to estimate, on a 0-100 scale, the probability that the borrower would repay. This helps us to verify that the experimental task was meaningful in the sense that loan officers could infer credit risk based on the information in the loan file. Figure 1 provides a scatterplot of the files used in the experiment. We see a tight negative correlation between expected repayment probability and the likelihood of loan rejection. This suggests that our incentive scheme worked and that participants thought the task realistic and paid attention to the information provided.
Equally important is to check whether the decision making in our lab-in-the-field correlates with what happened to the loan applications in real life. We find that this is the case. Overall, 72% of all applications that related to loans that performed well in real life were approved during the experimental sessions (green dots). This percentage is significantly lower for applications related to non-performing loans (53%, red dots) and for applications that were rejected in real life (47%, orange dots). This indicates that across the board participants correctly identified loans that performed well or badly in real life and made lending decisions in line with these subjective perceptions of loan quality.
Figure 1 Expected repayment and loan rejection rates
Notes: The horizonal axis in the within-file mean, across participants, of the subjective repayment probability. The vertical axis is the share of participants who declined the loan application. The figure is based on the first round of the experiment only.
We also implemented experimental modules to measure participants’ risk preferences and implicit gender bias. To measure the latter, participants took an implicit association test (IAT) in which they had to sort, as quickly as possible, words that appeared sequentially on an electronic tablet. In one task, “female” words had to be allocated to the category “family” and “male” words to the category “career” (the stereotypical task). In another task, “male” words had to be allocated to the category “family” and “female” words to “career” (the non-stereotypical task). We defined a participant’s implicit bias as the normalized difference in mean response times between the non-stereotypical and stereotypical task. While the scores range widely, a large majority of lending staff (87%) have a positive IAT score, indicating that they subconsciously associate business more with men than with women. This tendency is stronger among women than among men (Figure 2).
Figure 2 Participants’ implicit gender bias (IAT score)
Notes: The figure shows a local polynomial smooth with 95% confidence intervals of the variable Participant gender bias (IAT) for male (blue) and female (red) participants, respectively. The combined two-sample Kolmogorov-Smirnov test statistic is 0.181 and has ap-value of 0.01.
Our lab-in-the-field experiment with Turkish loan officers yields four main results:
- We find no evidence of direct gender discrimination. That is, one and the same loan application does not have a higher chance of being rejected when we present it with a female rather than a male applicant name. We also find no evidence of direct gender discrimination among specific sub populations of loan officers, such as male versus female loan officers or officers with more versus less lending experience.
- While unconditional approval rates are similar for male and female loan applicants, loan officers do discriminate against women in an indirect way. In particular, they are 30% more likely to require a guarantor when we present an application as coming from a female instead of a male entrepreneur.
- For this indirect discrimination, we find a consistent and intuitive pattern of statistically significant heterogeneous treatment effects. When we present the application as coming from a woman instead of a man, loan officers are more likely to ask for a guarantor when they are less experienced; younger; and/or display more implicit gender bias during our IAT. Importantly, there is no difference between male and female participants in how they treat female applicants. This implies that earlier evidence based on observational studies (e.g. Beck et al. 2013), and which suggested that male and female loan officers treat female loan applicants differently, may partly reflect deeper personal characteristics that are usually unobservable (such as implicit gender bias) rather than loan officers’ gender per se.
- Lastly, we find that discrimination is concentrated among loans that performed well in real life, making it potentially costly to the bank. The bars in the middle and at the right of Figure 3 show that for lower-quality applications it does not matter whether we present them as coming from male or female entrepreneurs. In contrast, the first two bars show a large and statistically significant gender difference. This is confirmed by a parametric analysis: also when we control for loan officer covariates as well as file and city fixed effects, are women 12.4 percentage points more likely to be asked for a guarantor in case of high-quality loans.
Figure 3 Guarantor requirements, by loan quality and applicant sex
Notes: The figure shows the percentage of loan applications that were approved during the experiment and for which participants requested a guarantor. Separate bars are shown for approved loans that were repaid in real life (left), approved loans that were defaulted on in real life (middle), and loan applications that were rejected in real life (right). In each case, separate bars indicate applications that were shown to participants as coming from a female (red) or male (blue) entrepreneur. The whiskers indicate one binomial standard deviation. The sample is restricted to the first round of the experiment.
Our results are mostly in line with models of implicit gender discrimination. As such, policies to mitigate the impact of loan officers’ implicit biases may be called for. This could include simply making sure that loan officers have sufficient time to evaluate loan applications. Banks could also set bank-wide or branch-wide goals for lending to women without a guarantor. Management could then hold those that deviate from this norm accountable through comply-or-explain procedures. Interventions like these may be more effective than explicit diversity training programs, which can make gender differences more salient and even generate a backlash (Bohnet 2016). Measuring the effectiveness of different types of interventions to contain the negative impact of implicit gender bias among loan officers provides a fruitful area for future experimental research.
Alesina, A F, F Lotti and P E Mistrulli (2013), “Do Women Pay More for Credit? Evidence from Italy”, Journal of the European Economic Association 11(s1): 45-66.
Beck, T, P Behr and A Guettler (2013), “Gender and Banking: Are Women Better Loan Officers?”, Review of Finance 17(4): 1279-1321.
Bohnet, I (2016), What Works. Gender Equality by Design, Harvard University Press.
Brock, J M and R De Haas (2019), “Gender discrimination in small business lending. Evidence from a lab-in-the-field experiment in Turkey”, EBRD Working Paper No. 232, European Bank for Reconstruction and Development.
Cole, S, M Kanz and L Klapper (2015), “Incentivizing Calculated Risk-taking: Evidence from an Experiment with Commercial Bank Loan Officers”, Journal of Finance 70(2): 537-575.