Prediction versus discretion: human-algorithm collaboration in assignment of unemployed jobseekers

We study the use of a statistical profiling tool for assigning jobseekers to active labor market programs at the Swedish Public Employment Service (PES). The tool predicts risk of long-term unemployment and recommends support to jobseekers at highest predicted risk. Final assignment decisions, however, remain at the discretion of individual caseworkers. The assignment decision involves a fundamental trade-off: whether to prioritize the most vulnerable jobseekers — those at highest risk of long-term unemployment — or those likely to have the greatest treatment effects. The algorithm is designed to target risk, but caseworkers may pursue either objective, and may use private information not available to the algorithm.

We compare the observed caseworker allocation of jobseekers to the counterfactual allocation that would have resulted from following the algorithm. Specifically, we ask:

• How do caseworker deviations affect the alignment of the allocation with a risk-based assignment rule?
• Do these deviations increase or decrease total employment relative to the algorithmic recommendation?

To construct the counterfactual algorithmic allocation, we exploit random assignment of jobseekers to caseworkers, together with variation in caseworkers' propensity to assign jobseekers to active labor market programs, as well as variation in their propensity to deviate from the algorithmic recommendation. This allows us to identify the causal effect of caseworker discretion on both the composition of program participants and their employment outcomes. We also examine whether caseworker performance relative to the algorithm differs across groups — for example, between women and men or between individuals with and without a foreign background.

The project uses administrative data from the Swedish Public Employment Service, including information on jobseekers, profiling scores, algorithmic recommendations, caseworker decisions, and the stated reasons for any deviations from the recommendation.