Gender discrimination in hiring in poor countries: the role of trust

Our study investigates gender discrimination in hiring within male-dominated sectors in Uganda, focusing on the roles of trustworthiness. We partner with vocational training centers to conduct an experiment with hiring managers. Hiring managers are presented with example CVs of vocationally trained workers. These CVs represent trainee candidates and are randomly assigned to be either male or female workers. This randomization allows us to test for gender-based differences in hiring preferences. Following the IRR paradigm, managers indicate their hiring preferences based on the provided CVs, knowing that we will refer them to workers with characteristics they favor. To test the effect of asymmetric information on gender differences in hiring, we randomize the provision of monitoring support to the managers via randomized audit visits. We have a pure control arm with no monitoring support, one treatment arm where monitoring support focuses on monitoring the trainees against stealing and misbehavior (monitoring trainees) and other common misbehavior practices, and one active control arm where the monitoring support focuses on workers' safety and wellbeing (monitoring firm).

Pre-Registration Update: Following initial data collection in the mechanics sector, we expand to additional male-dominated sectors due to observed heterogeneity in workplace gender diversity preferences, aiming to enhance external validity and statistical power as detailed in the attached document. Our main treatment and between-subject randomization remain unchanged, with the addition of a within-subject variation to further increase precision.

In the between-subject monitoring variation, each manager is randomly assigned to one of three groups: monitoring trainees, monitoring firm, or control. One-third of the sample receives monitoring support aimed at reducing stealing and other common misbehaviors among new workers (monitoring trainees). Another third receives monitoring support designed to prevent harassment and promote workplace safety (monitoring firm). The final third serves as a pure control group, receiving no monitoring or visits. Audit visits are performed by local team members. Both managers and workers are informed of the visits, but they are not pre-announced. [UPDATE] In the within-subject version of the intervention (that is, from the second sector onwards) managers make decisions under two of the three treatments prior to knowing their definitive treatment assignment.

2024-06-07

2025-08-31

MEET: choice of receiving the referral of a trainee like the respondent (yes/no).

- Meet: binary choice of meeting with a trainee with characteristics similar to the candidate (measure of interest in the candidate).

Pre-Registration Update: Following initial data collection in the mechanics sector, we expand to additional male-dominated sectors due to observed heterogeneity in workplace gender diversity preferences, aiming to enhance external validity and statistical power as detailed in the attached document. Our main treatment and between-subject randomization remain unchanged, with the addition of a within-subject variation to further increase precision. We note that in our first round of data collection, due to technical difficulties, we were unable to record our second main outcome, OFFER. We will therefore exclude this outcome moving forward.

- Quality: beliefs about workers worker's skills on a 0-10 scale;
- Behavior: beliefs about workers behavior (trustworthiness, honesty, skills) on a 0-10 scale;
- Effort: beliefs about workers effort (hardworking, not lazy, focused, punctual) on a 0-10 scale;
- Earnings: beliefs about workers' earnings a year from the experiment (UGX).

Each CV is randomly assigned to a combination of gender (female/male) and ability (GPA 5 out 5 / GPA 3 out of 5).

Monitoring support randomization at the manager level:

T0) Pure Control Arm: Receives no monitoring support.
T1) Monitoring Firm: Receives monitoring support in the form of audit visits focused on monitoring workers' safety and wellbeing.
T2) Monitoring Trainee: Receives monitoring support in the form of audit visits focused on monitoring trainees misbehavior (stealing, disrespect etc).

Pre-Registration Update: Following initial data collection in the mechanics sector, we expand to additional male-dominated sectors due to observed heterogeneity in workplace gender diversity preferences, aiming to enhance external validity and statistical power as detailed in the attached document. Our main treatment and between-subject randomization remain unchanged, with the addition of a within-subject variation to further increase precision.

Experimental Design: Hiring managers are presented with example CVs of vocationally trained workers. The CV content is randomized from workers' registration files and GPA data from the partner vocational training centers. Each CV has both a male and female version, and a low and high GPA version, resulting in 24 unique CVs with 96 combinations. Profiles are presented using the Incentivized Resume Rating Paradigm (Kessler et al. 2019). Gender and ability are cross-randomized at the CV level.
Each manager evaluates 24 randomly selected profiles, stratified by gender and ability. CVs are presented in random order and organized into three blocks of eight profiles, with scheduled breaks between blocks. For each profile, managers decide whether they wish to be matched with a similar worker to initiate a probationary period at the firm. Firms receive financial incentives designed to cover worker-related costs, including transportation and meals. Additionally, firms are randomly assigned to either receive or not receive monitoring support from the experimenters for supervising trainees.
Excluding those in the first sector, managers, after completing the initial evaluation round, are randomly reassigned to evaluate an additional set of 12 new CVs under one of the two alternative experimental conditions they had not previously experienced. We incentivize the exercise by matching employers with workers from various vocational training centers in Kampala. Managers' choices remain incentive-compatible also in the within-subject version of the experiment, as decisions are elicited using the strategy method before their final treatment assignment is revealed. We track whether the matching works by recording if the worker is contacted by the manager (viceversa) and if the worker ends up starting a probation period at the firm.

Data cleaning and analysis points:

- We drop observations where the respondent refused to answer the question (-88) .
- We exclude from our analysis observations for which there is no variation in the primary outcomes across a block of 8 CVs.
- We exclude from our analysis observations for which the respondent did not pass the understanding checks with respect to their treatment assignment. We drop observations where respondents assigned to T1 do not select "safety" as the purpose of the visits, and respondents assigned to T2 do not select "no stealing and misbehavior" as the purpose of the visits.
- We cross-randomize ability and gender in the CVs to be able to test for the interaction between gender and ability on gender differences in hiring (statistical discrimination test).
- We estimate gender differences in hiring from a regression including CV fixed effects and manager fixed effects. We cluster standard errors at the manager level.
- [updated see below] We plan to run the following heterogeneity analysis by managers' characteristics at baseline: 1) optimal gender composition in firm (median); 2) whether they normally hire from vocational training centers and 3) hiring manager's gender.

[UPDATE]:
- We measure preferences for gender diversity in the workplace using the following question: ``What is the best gender composition of workers in your firm?" (0 out 10 --- 10 out of 10 women).
- We estimate heterogeneity in gender differences in hiring by managerial preferences for gender diversity in the workplace.
- Our primary specification evaluates the effect of monitoring treatments on gender differences in hiring using between-subject variation (first 24 CVs), with regressions that include CV fixed effects and standard errors clustered at the manager level. This specification focuses specifically on managers with positive preferences for gender diversity, for whom we can meaningfully measure a decrease in the preference for hiring women versus men. This analysis leverages data from all the sectors.
- We supplement this approach with an additional analysis using data from the subsequent data collection (first 24 CVs plus the additional 12 CVs), leveraging within-subject variation to increase statistical power. This analysis includes both CV- and manager-level fixed effects but excludes the mechanics sector, for which we did not plan within-subject variation.
- We estimate heterogeneity in the effects of the "Monitoring Trainee" treatment by examining interactions with managerial preferences regarding optimal gender composition within the firm, employing non-parametric specification.

Randomization done in office by a computer.

Profile gender: CV-ability level randomization; Monitoring support: manager level randomization in between-subject; cv-block level randomization for monitoring in within-subject (first 24 CVs + 12 additional CVs).

Yes

We plan to collect data from 300 hiring managers in a sector. Should we be unable to reach our target due to insufficient firms within the Kampala metropolitan area, we will expand to additional sectors. [UPDATE]: The first sector was mechanics. Although we reached 300 firms, we identified power and external validity issues as detailed in the attached document. Consequently, we have updated our plan to expand to welding, carpentry, and gas stations. The final number of clusters will depend on the number of firms we can reach in each sector. Based on previous literature, 300 is considered an upper bound to the total number of firms within each sector.

The total number of observations equals at most the total number of respondents, times the number of CVs. In the between-subject randomization: 24 CV per manager. Within-subject randomization: 36 CV per sector in the within-subject specification.

Gender treatment: 50% male CV and 50% female CV.
Monitoring treatment: 300 managers (12 to 24 CVs in each arm) per sector.

[Pre-Registration Update]: Following initial data collection in the mechanics sector, we expand to additional male-dominated sectors due to observed heterogeneity in workplace gender diversity preferences, aiming to enhance external validity and statistical power as detailed in the attached document. Our main treatment and between-subject randomization remain unchanged, with the addition of a within-subject variation to further increase precision. The expansion of our data collection to new sectors increases our sample size by up to 300 observations (managers) per sector. In addition, the within-subject variation adds 12 observations (CVs) per manager.

Pre-Registration Update: After our initial data collection, we updated the preregistration to include three additional male-dominated sectors beyond the initial sector (mechanics). This expansion aims to increase statistical power and enhance external validity, following the discovery of significant and unexpected heterogeneity in gender bias within our initial sector of interest. This heterogeneity substantially reduced our power to detect the effect of our "Monitoring Trainee" treatment. Our treatment and randomization procedures remain unchanged. We now report updated power calculations. Gender treatment effect: at the initial sample size (24 CVs x 300 managers), we are highly powered to detect the effect of gender on hiring choices (Meet:0-1). Monitoring Trainee treatment effect: At our initial sample size (N=300), statistical power to detect the effect of monitoring on gender bias was only 33%. Increasing the sample size to our target of 1,200 nearly doubles our power, raising it to 56.6%. While we would prefer an even larger sample, we are limited by the number of available businesses within each sector. Given the disproportionate influence of the minority of managers who explicitly lack gender diversity preferences and for which we would not expect treatment effects, a more cost-effective approach to enhancing power is to focus exclusively on respondents who express at least some preference for workplace diversity. Increasing our sample size to 300 firms across 4 sectors, while also focusing on managers with any preference for diversity greater than zero when testing for the effect of Monitoring Trainee, our power raises to 82.8%. At the current sample size, we are powered to detect an effect of Monitoring Trainee leveraging within-subject variation.