Improving labor market matching and screening in Addis Ababa

We conduct a two-sided RCT with two treatment arms and one control group in Addis Ababa, Ethiopia. We recruit firms posting vacancies in four occupations: Accounting, Marketing / Sales, Human Resources, and Secretarial / Receptionist Services. In parallel, we recruit jobseekers looking for jobs in the same occupations.

We then randomize both vacancies and jobseekers into one of three groups:

1. Control group: Business as usual; jobseekers and firms in this group receive no intervention.

2. Recommendation only group: Jobseekers are recommended up to five jobs via SMS and emails, with the recommended jobs consisting of vacancies in the same treatment group.

3. Recommendation and signaling group: Jobseekers are recommended up to five jobs via SMS and emails, with the recommended jobs consisting of vacancies in the same treatment group. Additionally, jobseekers receive a skill certificate with the results of a baseline skill assessments administered as part of the study.

Recommendations within the two treatment groups are based on a combination of jobseeker preferences over vacancies as well as firm preferences over applicants.

2026-03-02

2026-10-30

Jobseeker labor supply: Dummy indicating any wage work in the last 30 days. Monthly wage earnings (not conditional on wage employment).
Jobseeker match quality: Monthly wage earnings( conditional on wage employment); Estimated percentile rank of job in jobseeker preference distribution.

Vacancy-level labor demand: Number of hires; Total monthly wage bill (not conditional on having filled the vacancy)
Vacancy-level match quality: Average wage bill among hires (conditional on having filled the vacancy); Estimated percentile rank of employee in firm preference distribution.

We estimate preferences based on baseline discrete choice experiments administered to both firms and jobseekers during the baseline survey.

Firms are presented with 10 binary choices of applicants that vary along the following dimensions: education, experience, location, gender, 3 soft skills (2 decided by the firm, 1 chosen at random), hard skill (decided by the firm ).

Jobseekers are presented with 10 binary choices of jobs that vary on the following dimensions: location, job type, weekly working hours, paid leave days, training offered, additional benefits, schedule flexibility.

For both sides, we estimate the full distribution of preferences using mixed logit regressions. We then map actual characteristics of jobs or applicants into the categories and place them within the preference distribution.

All match quality measures will primarily estimated conditional on employment/hiring. To assess the effect of selection, we will also estimate unconditional treatment effects and explore the use of Attanasio bounds.

We will trace the entire application process from recommendation views, to applications, offers and acceptance.

We will estimate main treatment effects using linear regressions with treatment dummies and randomization block fixed effects for both vacancies and jobseekers, and controls chosen using LASSO. We will cluster standard errors at the firm level for the vacancy level regression.

Our main heterogeneity dimensions for the vacancy level analysis are:
1.⁠ ⁠Whether firms have a predicted probability of filling the vacancy that is above median
2.⁠ ⁠Whether firms express a desire to receive additional applicants for a given vacancy at baseline.

Our main heterogeneity dimensions for the worker level analysis are:
1.⁠ ⁠Whether workers have a predicted probability of employment that is above median
2.⁠ ⁠Whether their median rank across firms in the relevant occupation (obtained using a DCE experiment with firms) is above median

Finally, we will also run regressions whether we study whether we increased the number of jobs initiated between vacancies with below-median probability of being filled and jobseekers with below-median probability of employmen

Secondary analyses include testing for spillovers by comparing super control week observations to control observations in weeks with treatment allocation and and analysis of matches within each treatment group to assess whether the interventions created matches with recommended job seekers.

Finally, we use the data to estimate a structural model of matching in the labor market with search and matching.

Not available

We randomize using pre-defined randomization blocks, stratifying on occupation. Additionally, we randomly select three calendar weeks during the study period to serve as super control weeks with only control group allocations for both job seekers and vacancies.

We randomize both the vacancies and job seekers into treatment and control groups. The vacancy-facing randomization is clustered at the firm level (i.e. all vacancies for a firm have the same treatment status).

Yes

Ca. 1,050 firms.

1250 vacancies and 2500 job seekers.

Approximately: 40% control group and 30% in recommendation only, and 30% recommendation plus signaling.