From the Fair Distribution of Predictions to the Fair Distribution of Social Goods: Evaluating the Impact of Fair Machine Learning on Long-Term Unemployment

Algorithmic fairness focuses on the distribution of predictions at the time of training, rather than the distribution of social goods that arises after deploying the algorithm in a concrete social context. However, requiring a "fair" distribution of predictions may undermine efforts at establishing a fair distribution of social goods. Our first contribution is conceptual: we argue that addressing the fundamental question that motivates algorithmic fairness requires a notion of prospective fairness that anticipates the change in the distribution of social goods after deployment. Our second contribution is theoretical: we provide conditions under which this change is identified from pre-deployment data. That requires distinguishing between, and accounting for, different kinds of performative effects. In particular, we focus on the way predictions change policy decisions and, therefore, the distribution of social goods. Throughout, we are guided by an application from public administration: the use of algorithms to (1) predict who among the recently unemployed will remain unemployed in the long term and (2) target them with labor market programs. Our final contribution is empirical: using administrative data from the Swiss public employment service, we simulate how such policies would affect gender inequalities in long-term unemployment. When risk predictions are required to be "fair", targeting decisions are less effective, undermining efforts to lower overall levels of long-term unemployment and to close the gender gap in long-term unemployment.