One Cohort at a Time: A New Perspective on the Declining Gender Pay Gap [Preliminary version available upon request. Complete paper coming soon!]

Using data from Italy and the US, this paper provides new insights on the decline of the gender pay gap over the last few decades. We find support for a cohort-driven interpretation for the shrinking of the gender gap, according to which more recent worker cohorts (defined by their year of entry in the labor market) with smaller gender gaps have gradually replaced older cohorts with larger gender gaps. We show that the cross-cohort decline in the entry-wage gender gap among younger workers accounts for the total reduction in the overall gender gap at most until 2000, after which the labor-market exit of older cohorts becomes the prominent explanation. Next, we zoom on younger workers and ask whether the reduction in the entry-wage gender gap is driven by improvements in younger women's outcomes. Instead, we find that most of the reduction in the entry-wage gender gap is accounted for by large earning losses among younger men (relative to both older workers and younger women). Finally, we show that, in Italy, cross-cohort changes in the initial matching between workers and firms help explain a substantial portion of the decline in the overall gender gap, but this is not true of initial sector allocations.

Using Prediction Models to Design Tax Audit Policies: Incentives vs Targeting

We study how to design tax audit policies when the Tax Authority can use an inference model that predicts the true taxpayers income. We propose a decomposition of tax revenues into an “ex-ante incentives” component that depends on declared income and an “ex-post targeting” component that captures the performance of tax auditing relative to random audit. In a simple case where taxpayers are heterogeneous in income but not in their propensity to evade we show that the Authority can achieve arbitrarily good collection if the model is precise enough and taxpayers are aware of the use of model. However, the revenues from ex-post targeting of audits are negligible at the optimum: the information from the model is used to shape ex-ante incentives, while ex-post collection is close to a random audit system. Introducing heterogeneity in evasion behavior, which makes the inference problem more complicated, does not change this conclusion. This is because optimal reliance on the model compresses the declarations of taxpayers earning the same income, making the potential revenues from targeting small. We present some numerical simulations that study how the tax collection potential responds to improvements in the precision of the model to start quantifying the gains and the decreasing marginal returns of investing in better prediction models.