Algorithmic Profiling and the Beliefs of Caseworkers and Jobseekers: a Field Experiment in Unemployment Insurance

This RCT uses an algorithmic profiling prediction as an intervention to affect caseworker beliefs on their clients' (job seekers) chances of finding a job. The trial was conducted within unemployment insurance in Switzerland (at the public employment service agencies of the Canton of Fribourg). Based on an elaborate prediction model, an individual prediction of the expected unemployment duration of the job seeker was provided to the caseworker, in a randomized number of cases. This prediction was supposed to serve as a signal to the caseworker to revise and update her/his own expectation concerning her/his client's unemployment duration. (Earlier evidence has shown that caseworkers are on average too optimistic with respect to the their client's expected duration to finding a job.) The idea was that caseworkers may act on this corrective signal by adjusting their decisions and strategy how to treat their client (using their available instruments of counseling, job search requirements and active labor market policy measures). The final policy intent was that this revision of the treatment strategy should positively affect the job seekers' labor market outcomes.
These steps of expected reactions can be tracked by collected data on caseworkers and job seekers. Thus, this RCT provides a unique set of (linked) data tracing the caseworkers' and the job seekers' activities, behaviors and beliefs, as well as the individual labor market outcomes. A tool within the unemployment insurance register database was used to collect the mentioned data and to provide the algorithmic predictions. The collected behavioral data on job seekers can be used for two purposes: first, they provided important input for the profiling prediction model (in a calibration phase of the project); second, the variables on behaviors and personal characteristics are of analysis interest per se, to better understand job seeker's behaviors. The data is linked at the individual level to the unemployment insurance register data (covering job seeker demographics, all events in the unemployment spell and the use of caseworkers' instruments) and to social security data (covering earnings and employment status post-unemployment).