Papers

Researchers are often interested in the existence and extent of interference between units when conducting causal inference or designing policy. However, testing for interference presents significant econometric challenges, particularly due to complex clustering patterns and dependencies that can invalidate stan- dard methods. This paper introduces the pairwise imputation-based random- ization test (PIRT), a general and robust framework for assessing the existence and extent of interference in experimental settings. PIRT employs unconditional randomization testing and pairwise comparisons, enabling straightforward imple- mentation and ensuring finite-sample validity under minimal assumptions about network structure. The method’s practical value is demonstrated through an application to a large-scale policing experiment in Bogot´a, Colombia (Blattman et al., 2021), which evaluates the effects of hotspot policing on crime at the street- segment level. The analysis reveals that increased police patrolling in hotspots significantly displaces violent crime, but not property crime. Simulations cali- brated to this context further underscore the power and robustness of PIRT.

Generalized and Simulated Method of Moments are often used to estimate structural Economic models. Yet, it is commonly reported that optimization is challenging because the corresponding objective function is non-convex. For smooth problems, this paper shows that convexity is not required: under a global rank condition involving the Jacobian of the sample moments, certain algorithms are globally convergent. These include a gradient-descent and a Gauss-Newton algorithm with appropriate choice of tuning parameters. The results are robust to 1) non-convexity, 2) one-to-one non-linear reparameterizations, and 3) moderate misspecification. In contrast, Newton-Raphson and quasi-Newton methods can fail to converge because of non-convexity. The condition precludes non-global op tima. Numerical and empirical examples illustrate the condition, non-convexity, and convergence properties of different optimizers.

We examine the tradeoff between privacy and personalization for online content by evaluating the impact of YouTube's settlement with the Federal Trade Commission over violating the Children's Online Privacy Protection Act (COPPA). Under the settlement, YouTube removed all forms of personalization for child-directed content starting in January 2020, which included personalized ads and platform features like personalized search and recommendations. We study the resulting impact on 5,066 top American YouTube channels by comparing the child-directed content creators to their non-child-directed counterparts using a difference-in-differences design. On the supply side, we find that child-directed content creators produce 18% less content and pivot towards producing non-child-directed content. Child-directed content creators also invest less in content quality: the proportion of original content falls by 11% and manual captioning falls by 27%, while user content ratings fall by 10%. On the demand side, views of child-directed channels fall by 20%. Consistent with the platform's degraded capacity to match viewers to content, both content creation and content views become more concentrated among top child-directed YouTube channels.

We develop a model of discrimination that allows us to interpret observed differences in outcomes across groups, conditional on passing a screening test, as taste-based (employer,) statistical, or customer discrimination. We apply this framework to investigate the nature of non-white underrepresentation in the US motion picture industry. Leveraging a novel data set with racial identifiers for the cast of 7,000 motion pictures, we show that, conditional on production, non-white movies exhibit higher average revenues and a smaller variance. Our findings can be rationalized in the context of our model if non-white movies are held to higher standards for production.

We study a dynamic delegation problem in which a long-lived principal assigns experimentation tasks to short-lived agents, each active for only one period. The principal values experimentation more due to her longer horizon. We model this as a contextual bandit problem with type-dependent Poisson outcomes: one agent type has a known success probability, while the other is uncertain-either high or low-and is gradually learned. The optimal contract is a forward-reverse stopping contract (FRSC), generating cyclical experimentation, involving engagement, abandonment, and revival. Misaligned incentives lead to both overand under-experimentation, as the principal balances learning incentives against agents' short-termism.

Switchback experiments--alternating treatment and control over time--are widely used when unit-level randomization is infeasible, outcomes are aggregated, or user interference is unavoidable. In practice, experimentation must support fast product cycles, so teams often run studies for limited durations and make decisions with modest samples. At the same time, outcomes in these time-indexed settings exhibit serial dependence, seasonality, and occasional heavy-tailed shocks, and temporal interference (carryover or anticipation) can render standard asymptotics and naive randomization tests unreliable. In this paper, we develop a randomization-test framework that delivers finite-sample valid, distribution-free p-values for several null hypotheses of interest using only the known assignment mechanism, without parametric assumptions on the outcome process. For causal effects of interests, we impose two primitive conditions--non-anticipation and a finite carryover horizon m--and construct conditional randomization tests (CRTs) based on an ex ante pooling of design blocks into "sections," which yields a tractable conditional assignment law and ensures imputability of focal outcomes. We provide diagnostics for learning the carryover window and assessing non-anticipation, and we introduce studentized CRTs for a session-wise weak null that accommodates within-session seasonality with asymptotic validity. Power approximations under distributed-lag effects with AR(1) noise guide design and analysis choices, and simulations demonstrate favorable size and power relative to common alternatives. Our framework extends naturally to other time-indexed designs.