Research

We find that cross-country panel estimates are too imprecise to establish the medium-term effects of personal, corporate, or value-added tax rate changes on national income. The results depend on how tax shocks are measured and how repeated tax changes are handled.

We evaluate the implications of different methodological choices for estimating the effect of tax policy on economic growth using cross-country panel data. We estimate the effect of personal, corporate, and value-added tax rate changes using alternative measures of policy shocks and employing estimation strategies both from the previous literature, and from newer methods that account for staggered treatments. Overall, we find imprecise estimates that do not credibly support claims that tax rate changes have a statistically robust medium-term impact on national income.

We develop a projection method for multivariate probability distributions using optimal transport and constrained linear regression. The method extends synthetic control to the joint distribution of several outcomes, including outcomes that are not continuous.

We develop a notion of projections between sets of probability measures using the geometric properties of the 2-Wasserstein space. It is designed for general multivariate probability measures, is computationally efficient to implement, and provides a unique solution in regular settings. The idea is to work on regular tangent cones of the Wasserstein space using generalized geodesics. Its structure and computational properties make the method applicable in a variety of settings, from causal inference to the analysis of object data. An application to estimating causal effects yields a generalization of the notion of synthetic controls to multivariate data with individual-level heterogeneity, as well as a way to estimate optimal weights jointly over all time periods.

We show that dropping incomplete matched pairs, which is equivalent to including pair fixed effects, does not generally recover the average treatment effect. Even when attrition depends only on the matching variables, the estimator reweights treatment effects toward groups more likely to remain observed. Stratifying into larger groups does not resolve the issue.

In this paper we revisit some common recommendations regarding the analysis of matched-pair and stratified experimental designs in the presence of attrition. Our main objective is to clarify a number of well-known claims about the practice of dropping pairs with an attrited unit when analyzing matched-pair designs. Contradictory advice appears in the literature about whether or not dropping pairs is beneficial or harmful, and stratifying into larger groups has been recommended as a resolution to the issue. To address these claims, we derive the estimands obtained from the difference-in-means estimator in a matched-pair design both when the observations from pairs with an attrited unit are retained and when they are dropped. We find limited evidence to support the claims that dropping pairs helps recover the average treatment effect, but we find that it may potentially help in recovering a convex weighted average of conditional average treatment effects. We report similar findings for stratified designs when studying the estimands obtained from a regression of outcomes on treatment with and without strata fixed effects.

I develop a framework for separately identifying direct effects, network-mediated effects, and their interaction when treatment itself changes social connections. The main identifying restriction is that pretreatment networks and covariates capture the common causes of network formation and outcomes.

Randomized saturation experiments measure how outcomes respond to own treatment and to the proportion assigned treatment. When treatment changes social connections, these comparisons do not reveal how much of the response operates through network changes. I propose a framework that decomposes the outcome difference between treated individuals and untreated individuals at different saturations into direct, network-mediated, treatment–network interaction, and saturation-spillover components without imposing functional-form restrictions on potential outcomes. I give conditions under which these components can be identified leveraging pretreatment networks and covariates, without estimating a network formation model. The main restriction, baseline network sufficiency, requires that network-formation factors contain no further information about the shocks governing outcomes at fixed treatment and exposure after conditioning on baseline characteristics. Under neighborhood stability and regularity conditions, I propose an estimation procedure based on debiased machine learning without sample splitting that accommodates network dependence. I propose an inference procedure that conditions on realized saturations, which is valid as cluster sizes grow while the number of clusters stays fixed. Reanalyzing an agronomy-training experiment in Rwanda, I find a negative network-mediated effect on input use for untreated farmers and a positive treatment–network interaction; these opposing effects reveal a network channel obscured by the small total effect and suggest that training changes how social connections shape the allocation of shared inputs.

We extend Romer and Romer's (2010) analysis through 2019 and find that U.S. tax cuts substantially boosted GDP in 1950–1984 but had little estimated effect in 1985–2019. Changes in the Federal Reserve's response to tax cuts are consistent with this shift.

We extend Romer and Romer's (RR, 2010) classic narrative analysis of exogenous tax shocks to cover data through 2019 and identify an exogenous tax increase in 2013 and an exogenous tax cut in 2018. We find that tax cuts have the same substantial effects on growth over the 1950-2019 period as RR found for 1950-2007. Several major events circa 1985 – the indexing of the income tax for inflation, the onset of the Great Moderation, and changes in monetary policy – could plausibly have changed the effects of tax shocks. We test this idea and find that although tax cuts substantially boosted growth in 1950–1984, they had minimal effects on GDP in 1985-2019. Changes in monetary policy before and after 1985 – specifically, in the federal funds rate's response to tax changes – are qualitatively consistent with the changing GDP response to taxes. Rough estimates of the share of the GDP response due to monetary policy span a wide range.

We develop an empirical Bayes framework for combining a country's own taxable income elasticity estimate with evidence from other countries, accounting for structural differences in their tax systems. Applied to 27 OECD countries, it gives outside evidence more weight when a country's own estimate is imprecise.

Estimates of the elasticity of taxable income (ETI) based on data from only one country implicitly assume nothing can be learned from the experience of other countries. This assumption is wrong, but rigorously relaxing it requires a principled framework to aggregate information from different contexts. We propose a general three-step empirical Bayes framework grounded in statistical decision theory that accomplishes this by modeling cross-country heterogeneity as a function of the structural determinants of the tax system. The framework determines endogenously how much each country's estimate should be informed by external evidence, placing more weight on cross-country evidence for countries whose within-country estimates are imprecise. We apply our framework to 27 OECD countries, characterize how ETI estimates should be revised taking into account evidence from other countries, and outline how to predict behavioral responses for countries that have not adopted a given tax.

We show why training on less customer data can be more profitable for an AI platform. When premium users face higher costs of data sharing, a discounted data-sharing tier can buy useful interactions while keeping those users on restricted-use plans.

AI service platforms sell current access while deciding which customer interactions may be used for future model improvement. This creates a tension absent from standard service menus: the usage price that protects premium restricted-use plans from cannibalization may be too high to procure valuable training usage, while the price that buys data may undermine premium separation. We study nonlinear service menus in which each contract specifies an access fee, a usage price, and a data-use right. Against blanket data policies that either train on all served usage or restrict all served usage, we characterize when targeted data procurement is optimal. A low-tier data-sharing plan can buy useful interactions and, when premium users have higher training-use burdens, make the cheap plan less attractive to them. The relevant value of data is menu-adjusted: continuation-value gains net of user burdens, information rents, and any repair needed to preserve self-selection. Numerical analyses illustrate that targeted procurement can outperform train-all policies even while collecting less raw data. The analysis yields a select--price--repair--retire rule: select tiers whose data value survives these costs, price data sharing as procurement rather than a generic subsidy, repair menus that create positive self-selection cycles, and retire discounts once marginal continuation value falls below the rent burden.

I extend stacked difference-in-differences to triple differences under staggered adoption. The method lets researchers choose comparison groups separately for each treatment cohort and, under the identifying assumptions, combine cohort effects with positive weights.

Triple differences (DDD) is a workhorse quasi-experimental design in applied economics. But, under staggered adoption, its conventional three-way fixed-effects (3WFE) implementation inherits the interpretation issues now well understood in the difference-in-differences literature. I introduce stacked DDD. I extend the stacked difference-in-differences approach to the DDD setting by creating self-contained stacks, each consisting of four cells over an event window: treated and clean comparison cohorts, each with treatment-eligible and treatment-ineligible units. Appending these stacks yields a unified dataset for estimating treatment effects. I prove that, at each post-treatment event-time, a linear regression with fully saturated fixed-effects applied to the stacked dataset identifies a strictly positive, cell-size-weighted average of stack-level conditional average treatment effects, with stack weights proportional to stack-level cell sizes. Building on this characterization, I outline alternative weighting schemes that recover causal estimands with clear interpretations. Stacked DDD complements recent GMM and imputation-based frameworks by trading efficiency for regression-based transparency, pairwise (rather than global) parallel changes-in-trends, and direct control over both the comparison group for each treated unit and the aggregation weights. I provide two empirical illustrations where stacked DDD yields substantially different quantitative conclusions compared to existing procedures.

I develop a shift-share-like instrument for settings where the number of shocks varies randomly across regions and the diffusion network may be endogenous. Using patents as shocks, I find that expanding knowledge-intensive activity increases income segregation within U.S. cities.

I study whether the rise in knowledge-intensive economic activities led to increased economic disparities in U.S. cities over three decades. While existing works have sought to combine patenting activities and patenting citations to construct shiftshare instruments for knowledge-intensive economic activities, I argue this strategy does not apply cleanly, because it requires fixing the set of patent shocks a priori, whereas patenting activities arise randomly and differentially across regions. I propose a shift-share-like instrumental variable framework that leverages regional-level shocks and inter-region diffusion links as inputs in its construction. I allow shocks to arise randomly and differentially across regions, and I allow the observed diffusion network to be endogenous. Identification is achieved through assuming shocks are "as good as randomly assigned". I show that the asymptotic framework does not require observing a large set of regions, but rather large numbers of realized shocks across regions. Using the proposed instrumental variable framework, I find local expansion in knowledgeintensive activities leads to positive and nontrivial increases in income and education segregation in U.S. cities, but its effect on occupation segregation is statistically insignificant.