Experimental Design Details
We first select micro-markets that will be used in the experiment. They will be selected so that the restriction to these job x geographical areas allows to build reliable recommendation for all the treatments.
Next, we classify jobseekers registered in these markets at the French public employment service (Pôle Emploi, hereafter, the PES) according to their predicted bias w.r.t their job finding probability. We use a model trained on a rich panel about expectations and the available administrative data useful for the prediction. We then build 4 groups gathering jobseekers that are predicted to share the same type of bias: pessimistic, rational, optimistic and over-optimistic. We then design interventions that aim at improving job search, distinguishing between the pessimistic and the others.
For the pessimists (S1): the suggestions made to this group of jobseekers therefore relate to the fact that they could be more selective about the salary of the vacancies chosen, while at the same time intensifying their search efforts.
For the others (S2): the suggestions made to this group relate to the extension of searches to other sectors of activity and to the benefits of such searches on the return to employment.
The suggestions are formulated in three ways:
- T1 ("direct" direct suggestion),
- T2 ("peer" suggestion referring to the behavior of claimants with the same profile),
- T3 ("explanatory" explanation of the different biases in perceptions of the labor market leading to the adoption of a sub-optimal search strategy).
Then we sample 390,000 jobseekers from the selected micro-markets and randomly allocate them into 8 initial treatment and control groups:
- Super control: do not receive the questionnaire;
- Survey control: they receive the questionnaire but no recommendation at the end of it;
- S1 and T1
- S1 and T2
- S1 and T3
- S2 and T1
- S2 and T2
- S2 and T3
Rational, optimistic and over-optimistic claimants were treated only with S2-type recommendations, and were divided into 5 arms:
- Super control: they do not receive the questionnaire;
- Survey control: they receive the questionnaire but no recommendation at the end of it;
- S2 and T1
- S2 and T2
- S2 and T3.
For pessimistic allocated to S1, but whose reservation wage is above the suggested wage for the main job they are looking for, the S1 treatment is proposed without this precise suggestion. The following table sums up the different treatment each bias group potentially receives.
Bias group Wage level S1 S1 – without wage suggestion S2
Pessimists Lower than suggested X X
Higher than suggested X X
Others X
Treatment will then be adapted according to one answer of the survey about the subjective job-finding probability. Depending on the provided answer and more accurate model for the biases, the previous classification into pessimistic, rational, optimistic and over-optimistic is refined. At the end of the questionnaire, all groups but the super control receive an email with at least a link to the PES repository of job postings, and a reminder of the suggestions if any. This reminder is also sent two weeks later by email.
Taking into account the two steps randomization, the 390,000 job seekers will thus be randomly assign into 13 groups. The survey control group receive the survey collecting the expectations but no treatment suggestions. The control group receives no survey at all. We expect to have a 10% response rate to the survey and first stage stratification will be done to target 3000 respondents per treatment arm. We intend to estimate intention to treat (ITT) effects as well as average treatment effects (ATE) on the compliers of the survey. We also intend to estimate conditional average treatment effects (CATE), in particular with respect to previous unemployment experience, search behavior, and predicted group status.
To evaluate the impact on job search behavior and job finding outcomes we use the data made available through our partnership with the PES. This allows to characterize the job applications decisions (job, wage, type of contract, distance, required experience) as well as the results of these and the characteristics of the potential job (contract date, wage, type of contract, location). A follow-up survey one month later will provide more details on the impact on subjective outcomes such as the representation of the labor market, as well as the updated set of job search parameters. The super control group will help us controlling for potential market equilibrium effects and provide a more robust analysis.