Reducing racial gaps in referrals and hiring: two experiments with the Colombian Public Employment Services

This project has two interventions.

First, after collecting information about the implicit bias of firms’ human resource personnel we will provide feedback on the result of their own implicit bias. To estimate the causal effect of this information we will allocate firms randomly into two groups. Group 1 will receive an email with the IAT feedback a few weeks after taking the IAT. Group 2 will receive the same feedback approximately 6 months after taking the IAT.

The second intervention uses information about firms’ behavior to change job counselors' referral of Afro-descendants. In the interview with job counselors, we will elicit their beliefs about the probability that a CV from an afro-descendant will be interviewed and hired after being referred. We then can contrast this belief with the same observed probability at the PES and to the information from the interview to firms’ HR. If there is a mismatch, the belief is lower than the actual probability from the SPE, we will send this information to a random group of job counselors expecting them to update their belief and increase the referral rates of Afro-descendants. If the job counselors' beliefs are above the actual rate observed in the PES data, we could use the declared likelihood by firms in our interview. We expect this likelihood to be an upper bound of the actual probability of hiring due in part to social desirability bias by the firms. A possible complication in this case is that job counselors consider in their calculations the fact that firms may overestimate their willingness to hire afro-descendants and our information campaign may not be believable. We will discuss this possibility with CCFs and ASOCAJAS to tailor the intervention. In the case that counselors' beliefs are accurate, we will move to an intervention that uses information about the efforts by governments and institutions in the region to reduce the racial employment gap. In that instance, we will contrast their beliefs against the possible change in the social norm that could be driven by policies or private efforts.

Regardless of the information we provide, once again we divide job counselors into two groups. Group 1 will receive this information a few weeks after we finish data collection. Group 2 will receive the information approximately 4 months later.

2024-08-01

2024-09-01

The main outcome of the first intervention is the probability of hiring an afro-descendant job seeker. For the second intervention, the focus is the probability of referring an afro-descendant job seeker to a job post.

This information comes from the data sets of all job openings and all job seekers who interact through the SPE. We will use one year of information including 6 months before data collection, and 4 months after the interventions. For each job post, we observe its characteristics and the job seekers that are referred or apply to the post using the SPE. We also observe the CVs that are hired for each job post. For each CV, we observe demographic characteristics including belonging to an ethnic minority, which includes afro-descendants.

The first part of this project is a description of different indicators of bias against afro-descendant job seekers. We will measure de facto (differential probability of referring or hiring an afro-descendant compared with the probability of a non-afro-descendant), explicit (using direct questions and a list experiment), and implicit (using an implicit association test) biases of firms’ HR personnel and job counselors. In addition, we will document the correlations among these indicators. This analysis will use information from the interviews with firms and job counselors, in combination with SEP applications before the interventions.

These estimations use pre-intervention data.
We aim to use data on de facto, implicit, and explicit bias to evaluate possible heterogeneity in the impacts of our intervention.
We will also use ML tools to analyze the heterogeneity of our results with respect to firm, vacant, job center, job counselor, and job seeker characteristics.

We will randomize by intervention.

Intervention 1 - information to firms' personnel about their own IAT results.
Randomization is at the firm level in two groups. We will stratify randomization by firm size, region, and sector.
Treated: They will receive the IAT results a few weeks after completing baseline data collection.
Control: They will receive the IAT results about four (4) months after completing baseline data collection.


Intervention 2 - Information to job counselors about firms' beliefs/attitudes about afro-descendant job seekers.
Randomization is at the job center level in two groups. We will stratify by region and job center size.
Treated: They will receive information about firms a few weeks after completing baseline data collection.
Control: They will receive information about firms about four (4) months after completing baseline data collection.

Not available

By computer

The first intervention will be randomized at the firm level. The second intervention will be randomized the employment center level.

Yes

First intervention: 1500 firms
Second intervention: 350 job counselors in 80 job centers

First intervention: Personnel for at least 1500 firms Second intervention: 350 job counselors in 80 job centers

Intervention 1: We expect 750 firms by arm.
Intervention 2: We expect 40 job centers (125 job counselors) by arm.

We will check that job centers in Intervention 2 are balanced within the Intervention 1 group to examine heterogeneity in the impact of Intervention 1 with respect to the information given in Intervention 2.

First intervention: Our goal is to reach 1,500 firms who posted a job opening before our intervention and will be likely to post at least one job offer after our intervention. According to our calculations, the minimum detectable effect from this analysis will be 0.12 standard deviations. Second intervention: With 350 counselors working in 80 job centers we aim to estimate effects above 0.28 standard deviations.