Standard bandit algorithms for online policy learning assume the reward is always observed, but in many applications—such as gauging satisfaction through voluntary review prompts—feedback is missing in ways correlated with the outcome. In this regime, I show that the classic UCB algorithm selects suboptimal policies, and introduce a nonparametric Doubly-Robust UCB algorithm that models the missingness through covariates. I show that the newly proposed algorithm achieves near-optimal worst-case regret by means of new high-probability bounds for a class of doubly-robust estimators.

Empirical work leveraging regression discontinuity designs routinely examines how treatment effects vary with pretreatment covariates, yet the most common practices are not backed up by formal results. We develop a unified, theoretically grounded framework built on a fully interacted local-linear model, with principled bandwidth selection and robust bias-corrected inference for estimating heterogeneous effects and testing differences across groups.

We introduce rdhte, a package implementing our framework for heterogeneous treatment effects in RD designs.

Using a regression discontinuity design on Italy’s largest firm-subsidy program (Law 488/92), we show that investment subsidies raised a firm's investment, employment, and survival probability. Both objective eligibility rules and local politicians’ discretion selected firms that went on to create more jobs, but discretion did so at a higher cost per job: allocating on objective rules alone would lower the cost per job by 11%, while relying on political discretion alone would raise it by 42%.

We develop prediction intervals that quantify the uncertainty of synthetic control predictions when multiple treated units and staggered adoption of treatment are involved. We provide precise non-asymptotic coverage guarantees and a careful taxonomy of the causal quantities one may target.

We introduce scpi, a package implementing our framework for uncertainty quantification in synthetic controls.

Deferring systems let a machine-learning model hand difficult cases to a human expert, yet their effect on accuracy is usually judged only by looking at the final human–AI performance. We instead propose a novel framework to evaluate the causal impacts on predictive accuracy of introducing a deferring system.

We explain how pretreatment covariates can be used in regression discontinuity designs: to improve the efficiency; to uncover heterogeneous policy effects; to change the RD parameter of interest.

The regression discontinuity estimand is local in nature. I introduce the package getaway, which implements the Angrist–Rokkanen (2015) method to extrapolate RD treatment effects away from the cutoff under an unconfoundedness assumption. I extend their framework to multi-cutoff designs, propose a data-driven algorithm to select covariates that satisfy such condition, and a toolkit for falsification of the identifying assumptions and visualization.

How long migrants choose to stay abroad is linked to their earnings in ways that need not be monotonic. Using recent documented Mexican immigrants, we document a U-shaped relationship between earnings and the intention to settle permanently in the U.S.. We then estimate a dynamic life-cycle model that rationalizes this feature of the data, and use the model to study how changes in Mexican earnings and U.S.–Mexico purchasing-power parity reshape return plans and savings.

Exploiting the allocation of subsidies in a large Italian investment-subsidy program (Law 488/92) as a regression discontinuity design, we estimate that subsidies raised the investment of marginal firms by 39% and their employment by 17% over six years. The gains are heterogeneous: smaller firms grow employment fastest, but larger and younger firms create more jobs per euro of public spending.

Can technology relax constraints on educational supply and expand local productive capacity? We exploit a quasi-natural experiment in postwar Italy to estimate the causal impact of educational television on adult skills and local economic activity. In 1960, the Italian public broadcaster and the Ministry of Education expanded televised adult education, with Non è mai troppo tardi (“It Is Never Too Late”) as a flagship program—a daily literacy course supported by a national network of collective viewing centers designed to broaden access. We digitize historical records of transmitter locations and technical characteristics to predict television reception and link these predictions to industrial- and population-census data. Our effects are identified from variation in signal diffraction caused by topography, conditional on transmitter placement. We estimate stacked long differences across the 1951, 1961, and 1971 censuses for roughly 7,700 municipalities. A one-standard-deviation increase in signal strength raised adult literacy by 0.18 percentage points during the program decade relative to the preceding decade—17% of the average decadal gain. It also increased labor-force participation by 1.0 percentage point overall and 1.4 percentage points among women, and shifted employment out of agriculture. Labor-market outcomes responded to reception in neighboring municipalities within approximately 15 kilometers, consistent with local spillovers. Estimated literacy gains were larger in initially less literate municipalities and in places with collective viewing centers. Our findings suggest that broadcast technology can expand educational provision and support structural transformation, with benefits extending across municipal boundaries.

Generative adversarial networks (GANs) are increasingly used to produce synthetic data for downstream decisions. We challenge the assumption that the decision-relevant distribution coincides with the training distribution and study regret under bounded-Lipschitz perturbations of size δ. We characterize when ordinary Wasserstein GANs yield vanishing regret and establish minimax-optimal rates. We show that no loss-independent procedure can adapt to unknown smoothness at these rates, even for smooth Lipschitz losses. Finally, when decisions depend on a known regular score, we construct a score-aware procedure that achieves minimax adaptation.