Whom Would You Rather Work With? Experiment 3

The experiment will offer to successful alumni of VT institutions in Uganda the opportunity of referring someone to a 6-week subsidized internship at their company. We will show to each of them a pair of profiles that match their respective trainings. The pair of profiles will always have different gender. Respondents will also have the chance to refer a person they know. We will cross-randomize the sample in four ways.

First:

(1a) Group 1 (High quality male and High quality female): respondents are shown a randomly generated male profile with 11 months of work experience and a randomly generated female profile with 11 months of work experience;
(1b) Group 2 (High quality male and Low quality female): respondents are shown a randomly generated male profile with 11 months of work experience and a randomly generated female profile with 6 months of work experience;
(1c) Group 3 (Low quality male and High quality female): respondents are shown a randomly generated male profile with 6 months of work experience and a randomly generated female profile with 11 months of work experience;
(1d) Group 4 (Low quality male and Low quality female): respondents are shown a randomly generated male profile with 6 months of work experience and a randomly generated female profile with 6 months of work experience.

We will tell respondents that profiles are truthful and willing to accept the internship if offered it. This element of deception differs this experiment from its predecessors (AEARCTR-0011599 and AEARCTR-0008036) and serves the purpose of avoiding having respondents guess which parts of the profile are hypothetical or not. This different component sets us between the IRR literature and traditional audit studies.

Second:

(2a) Public referral (same as AEARCTR-0011599 and AEARCTR-0008036): the business owner will know who referred the candidate and the respondent will be warned that his/her referral will be PUBLIC;
(2b) Private referral (same as AEARCTR-0011599):We will NOT inform the employer that the employee referred the candidate, and we will never mention his/her name. However, we will tell the employer that the referral came from a worker in the company;
(2c) Referred by BRAC: the EMPLOYER will be told that the matched candidate was selected by BRAC, which is truthful, with no mention to any worker in the company;
(2d) "Private" referral (same as AEARCTR-0008036): We tell the respondent that the business owner will NOT know who referred the candidate and that his/her referral will be PRIVATE. However, we have anecdotal evidence that the interpretation of this treatment arm was ambiguous during AEARCTR-0008036, for which we changed it in AEARCTR-0011599. We are reintroducing it now (keeping its ambiguity) to gauge how respondents interpreted it in AEARCTR-0008036. After outcomes are realized (i.e., respondents rate profiles and choose the candidate), we clarify the ambiguity and respondents have to choose between either (2b) Private referral (as in AEARCTR-0011599) and (2c) BRAC Referral.

Third, we also cross-randomize a vignette experiment to understand the mechanisms of referrals:

(3a) Vignette group 1: we present two scenarios, one in which a man applies to a job by walking in and another in which a man applies to a job through a referral.
(3b) Vignette group 2: we present two scenarios, one in which a woman applies to a job by walking in and another in which a woman applies to a job through a referral.
(3c) Vignette group 3: we present two scenarios, one in which a woman applies to a job by walking in and another in which a man applies to a job through a referral.
(3d) Vignette group 4: we present two scenarios, one in which a man applies to a job by walking in and another in which a woman applies to a job through a referral.

In the vignette experiment, the order of the scenarios shown is also randomized.

Fourth, we cross-randomize a question to prime respondents to incorporate employers' preferences:

(4a) Priming: After the first profile is shown, respondents are asked how much they would like to work with the profile and, right after, how much they think their employers would like to have the profile as their employee. Then, we elicit the remaining outcomes.
(4b) No priming: Only after all other outcomes are realized and respondents make their referring choice among the two profiles, we elicit their second-order beliefs about how much they think their employers would like to have the profile as their employee.

With this, we want to understand to what extent second-order beliefs on employers' preferences can affect referral decisions.

Randomization 1 is stratified by gender of the respondent, dummy =1 if respondent was trained in a female dominated sector, dummy = 1 if respondent was wage-employed the last time we interviewed, and dummy = 1 if respondent was found in the previous wave. Randomization 2 is stratified by the exact same variables plus treatment assignment in Randomization 1. Randomization 3 and Randomization 4 are both stratified by assignment in Randomization 1 and assignment in Randomization 2.

2023-11-01

2024-05-31

We want to assess to what extent gender biases exist in referrals performed by our respondents. Specifically, we are interested in measuring the existence and relevance of two potential biases: against females and against the non-stereotypical candidates, which we define as candidates whose gender does not match the gender that prevails in the sector. We identify the primary outcomes of interest:

Outcome 1: Probability of picking a female profile AND Difference in probability of picking the female profile and male profile (respondent-level)
First, we look at the probability that the respondent picks the female candidate for the 6-week internship. Similarly, we look at the gap in probabilities of choosing the profile of one gender or the other.

Outcome 1b: Difference in probability of picking the female profile and male profile in each of the main treatment arms (1a, 1b, 1c, 1d).

Outcome 2: Probability of picking a profile of a non-stereotypical gender (respondent-level)
Second, we consider the probability that the respondent picks the candidate whose gender does not match the dominance of the sector.

Outcome 3: Probability of being picked (profile-level)
Additionally, in line with other audit studies that analyze the success of candidates using callbacks (Bertrand and Mullainathan, 2004; Booth and Leight, 2010; Becker et al., 2019), we intend to analyze the performance of the fake candidates individually by looking at their probabilities of being selected by the respondent. For each candidate, the probability of being selected is one if the respondent refers him/her or zero otherwise. The dependent variables shall be the gender of the profile, the quality of the profile, and interaction term.

Outcome 4: Probability of picking a network member (respondent-level)

Outcome 5 (if selected in the lottery): respondent agreed to work with the candidate (matched or network), employer agreed to take the candidate (matched or network), candidate agreed to take up the internship.

We will carry out heterogeneity analysis on the following variables: gender of respondent, gender dominance of the sector of training, sector of training, respondent IS wage-employed when interviewed, gender of employer, size of the business, share of male workers in the business (and dummy for above and below median), share of male customers in the business (and a dummy for above and below median), gender attitudes of the respondent, perceived gender attitudes in the training area, perceived gender attitudes of customers and employer, number of referrals made in the last 12 months (above and below median), and share referrals made to men in the last 12 months.

Gender attitudes of the respondent is an index of 8 components (2 questions on gender attitudes in the household, 6 questions on gender attitudes in the labor market). Gender attitudes in the sector is an index of 6 components (6 questions similar to those asked to the respondent). Gender attitudes of the employer is an index of 3 components (3 questions similar to those asked to the respondent -- but we ask what would the employer say) and a question from 0-10 on "How open is your employer to hiring workers from the non-stereotypical gender?". Gender attitudes of customers is a question from 0-10 on "How open are your customers to being served/handled by workers of the non-stereotypical gender?".

- Perceived likability of candidates (profile-level)
- Second-order belief about employers’ perceived likability of candidates (profile-level)
- Perceived probability of retention of candidates (profile-level)
- Perceived accuracy of perceived probability of retention of candidates (profile-level)
- Perceived fit of candidates to the tasks of the business (how good candidates would be in performing those tasks) (profile-level)
- Second-order belief about coworkers’ perceived likability of candidates (profile-level)
- Number of months of work experience that the respondent recalls, at the end of the interview, that the candidate shown had (profile-level)
- (As a robustness check): these same outcomes but looking only at the first anonymous profile shown.
- Vignettes: how happy the employer will be with the candidate, how much the coworkers will like to work with the candidate, how likely the candidate is to have all qualifications for the job, and how likely the candidate is to stay in the job for one year.

This study is an extension of previously registered trials (AEARCTR-0008036 and AEARCTR-0011599), which were, on their hand, was a spin-off of the Meet Your Future Project (MYF), an RCT designed to analyze a career-coaching program for Ugandan vocational students aimed at reducing labor market access inequities for socioeconomically disadvantaged students. Both the main and the spin-off studies are conducted in collaboration with the NGO BRAC, who implemented all data collections and interventions, and five reputable vocational training institutes (VTIs) in Central and Eastern Uganda. The experimental evidence from this study comes from a cohort of 1,000 successful alumni from vocational training institutes which the research team has been following for more than 4 years as part of the MYF Project. We leveraged a follow-up survey of the MYF project and added an extra module to explore how gender preferences in segregated occupations interplay with a referral-based hiring system.

We showed to respondents pairs of randomly generated profiles based on real-life candidates and asked them to refer one of the candidates for a 6-week subsidized internship program in their companies. Selected candidates will be called for the program based on a lottery system (approximately 200 will be selected). Respondents also had the chance to name up to two network members and select one of them instead of the anonymous candidates. To ensure all respondents had quality-seeking behaviors, we offered to all of them a monetary reward if the candidate is retained. Non-wage-employed respondents answered the survey only hypothetically, as their choices will not be part of the lottery and they will be offered no money.

Not available

Randomization done in office by a computer on a statistical software (STATA).

Individual-level randomization.

No

1000 VTI alumni.

1000 VTI alumni.

Main randomization:

(1a) Group 1 (High quality male and High quality female): 250
(1b) Group 2 (High quality male and Low quality female): 250
(1c) Group 3 (Low quality male and High quality female): 250
(1d) Group 4 (Low quality male and Low quality female): 250

Second:

(2a) Public referral (same as AEARCTR-0011599 and AEARCTR-0008036): 250
(2b) Private referral (same as AEARCTR-0011599): 250
(2c) Referred by BRAC: 250
(2d) "Private" referral (same as AEARCTR-0008036): 250

Third:

(3a) Vignette group 1: 250
(3b) Vignette group 2: 250
(3c) Vignette group 3: 250
(3d) Vignette group 4: 250

Fourth:

(4a) Priming: 500
(4b) No priming: 500

N/A