#### Research

## Published and accepted articles

Motivated by a recent literature on the double-descent phenomenon in machine learning, we consider highly over-parameterized models in causal inference, including synthetic control with many control units. In such models, there may be so many free parameters that the model fits the training data perfectly. We first investigate high-dimensional linear regression for imputing wage data and estimating average treatment effects, where we find that models with many more covariates than sample size can outperform simple ones. We then document the performance of high-dimensional synthetic control estimators with many control units. We find that adding control units can help improve imputation performance even beyond the point where the pre-treatment fit is perfect. We provide a unified theoretical perspective on the performance of these high-dimensional models. Specifically, we show that more complex models can be interpreted as model-averaging estimators over simpler ones, which we link to an improvement in average performance. This perspective yields concrete insights into the use of synthetic control when control units are many relative to the number of pre-treatment periods.

We develop a framework for difference-in-differences designs with staggered treatment adoption and heterogeneous causal effects. We show that conventional regression-based estimators fail to provide unbiased estimates of relevant estimands absent strong restrictions on treatment-effect homogeneity. We then derive the efficient estimator addressing this challenge, which takes an intuitive “imputation” form when treatment-effect heterogeneity is unrestricted. We characterize the asymptotic behavior of the estimator, propose tools for inference, and develop tests for identifying assumptions. Our method applies with time-varying controls, in triple-difference designs, and with certain non-binary treatments. We show the practical relevance of our results in a simulation study and an application. Studying the consumption response to tax rebates in the United States, we find that the notional marginal propensity to consume is between 8 and 11 percent in the first quarter – about half as large as benchmark estimates used to calibrate macroeconomic models – and predominantly occurs in the first month after the rebate.

Since their introduction by Abadie and Gardeazabal, Synthetic Control (SC) methods have quickly become one of the leading methods for estimating causal effects in observational studies in settings with panel data. Formal discussions often motivate SC methods by the assumption that the potential outcomes were generated by a factor model. Here we study SC methods from a design-based perspective, assuming a model for the selection of the treated unit(s) and period(s). We show that the standard SC estimator is generally biased under random assignment. We propose a Modified Unbiased Synthetic Control (MUSC) estimator that guarantees unbiasedness under random assignment and derive its exact, randomization-based, finite-sample variance. We also propose an unbiased estimator for this variance. We document in settings with real data that under random assignment, SC-type estimators can have root mean-squared errors that are substantially lower than that of other common estimators. We show that such an improvement is weakly guaranteed if the treated period is similar to the other periods, for example, if the treated period was randomly selected. While our results only directly apply in settings where treatment is assigned randomly, we believe that they can complement model-based approaches even for observational studies.

Nearest-neighbor matching is a popular nonparametric tool to create balance between treatment and control groups in observational studies. As a preprocessing step before regression, matching reduces the dependence on parametric modeling assumptions. In current empirical practice, however, the matching step is often ignored in the calculation of standard errors and confidence intervals. In this article, we show that ignoring the matching step results in asymptotically valid standard errors if matching is done without replacement and the regression model is correctly specified relative to the population regression function of the outcome variable on the treatment variable and *all* the covariates used for matching. However, standard errors that ignore the matching step are not valid if matching is conducted with replacement or, more crucially, if the second step regression model is misspecified in the sense indicated above. Moreover, correct specification of the regression model is not required for consistent estimation of treatment effects with matched data. We show that two easily implementable alternatives produce approximations to the distribution of the post-matching estimator that are robust to misspecification. A simulation study and an empirical example demonstrate the empirical relevance of our results.

Encouraging vaccination is a pressing policy problem. To assess whether text-based reminders can encourage pharmacy vaccination and what kinds of messages work best, we conducted a megastudy. We randomly assigned 689,693 Walmart pharmacy patients to receive one of 22 different text reminders using a variety of different behavioral science principles to nudge flu vaccination or to a business-as-usual control condition that received no messages. We found that the reminder texts that we tested increased pharmacy vaccination rates by an average of 2.0 percentage points, or 6.8%, over a 3-mo follow-up period. The most-effective messages reminded patients that a flu shot was waiting for them and delivered reminders on multiple days. The top-performing intervention included two texts delivered 3 d apart and communicated to patients that a vaccine was “waiting for you.” Neither experts nor lay people anticipated that this would be the best-performing treatment, underscoring the value of simultaneously testing many different nudges in a highly powered megastudy.

We investigate the optimal design of experimental studies that have pre-treatment outcome data available. The average treatment effect is estimated as the difference between the weighted average outcomes of the treated and control units. A number of commonly used approaches fit this formulation, including the difference-in-means estimator and a variety of synthetic-control techniques. We propose several methods for choosing the set of treated units in conjunction with the weights. Observing the NP-hardness of the problem, we introduce a mixed-integer programming formulation which selects both the treatment and control sets and unit weightings. We prove that these proposed approaches lead to qualitatively different experimental units being selected for treatment. We use simulations based on publicly available data from the US Bureau of Labor Statistics that show improvements in terms of mean squared error and statistical power when compared to simple and commonly used alternatives such as randomized trials.

Policy-makers are increasingly turning to behavioural science for insights about how to improve citizens’ decisions and outcomes. Typically, different scientists test different intervention ideas in different samples using different outcomes over different time intervals. The lack of comparability of such individual investigations limits their potential to inform policy. Here, to address this limitation and accelerate the pace of discovery, we introduce the megastudy—a massive field experiment in which the effects of many different interventions are compared in the same population on the same objectively measured outcome for the same duration. In a megastudy targeting physical exercise among 61,293 members of an American fitness chain, 30 scientists from 15 different US universities worked in small independent teams to design a total of 54 different four-week digital programmes (or interventions) encouraging exercise. We show that 45% of these interventions significantly increased weekly gym visits by 9% to 27%; the top-performing intervention offered microrewards for returning to the gym after a missed workout. Only 8% of interventions induced behaviour change that was significant and measurable after the four-week intervention. Conditioning on the 45% of interventions that increased exercise during the intervention, we detected carry-over effects that were proportionally similar to those measured in previous research. Forecasts by impartial judges failed to predict which interventions would be most effective, underscoring the value of testing many ideas at once and, therefore, the potential for megastudies to improve the evidentiary value of behavioural science.

Concerns about the dissemination of spurious results have led to calls for pre-analysis plans (PAPs) to avoid ex-post “p-hacking.” But often the conceptual hypotheses being tested do not imply the level of specificity required for a PAP. In this paper we suggest a framework for PAPs that capitalize on the availability of causal machine-learning (ML) techniques, in which researchers combine specific aspects of the analysis with ML for the flexible estimation of unspecific remainders. A “cheap-lunch” result shows that the inclusion of ML produces limited worst-case costs in power, while offering a substantial upside from systematic specification searches.

The ability to distinguish between people in setting the price of credit is often constrained by legal rules that aim to prevent discrimination. These legal requirements have developed focusing on human decision-making contexts, and so their effectiveness is challenged as pricing increasingly relies on intelligent algorithms that extract information from big data. In this Essay, we bring together existing legal requirements with the structure of machine-learning decision-making in order to identify tensions between old law and new methods and lay the ground for legal solutions. We argue that, while automated pricing rules provide increased transparency, their complexity also limits the application of existing law. Using a simulation exercise based on real-world mortgage data to illustrate our arguments, we note that restricting the characteristics that the algorithms allowed to use can have a limited effect on disparity and can in fact increase pricing gaps. Furthermore, we argue that there are limits to interpreting the pricing rules set by machine learning that hinders the application of existing discrimination laws. We end by discussing a framework for testing discrimination that evaluates algorithmic pricing rules in a controlled environment. Unlike the human decision-making context, this framework allows for ex ante testing of price rules, facilitating comparisons between lenders.

Machines are increasingly doing “intelligent” things. Face recognition algorithms use a large dataset of photos labeled as having a face or not to estimate a function that predicts the presence y of a face from pixels x. This similarity to econometrics raises questions: How do these new empirical tools fit with what we know? As empirical economists, how can we use them? We present a way of thinking about machine learning that gives it its own place in the econometric toolbox. Machine learning not only provides new tools, it solves a different problem. Specifically, machine learning revolves around the problem of prediction, while many economic applications revolve around parameter estimation. So applying machine learning to economics requires finding relevant tasks. Machine learning algorithms are now technically easy to use: you can download convenient packages in R or Python. This also raises the risk that the algorithms are applied naively or their output is misinterpreted. We hope to make them conceptually easier to use by providing a crisper understanding of how these algorithms work, where they excel, and where they can stumble—and thus where they can be most usefully applied.

## Working papers presented at peer-reviewed conferences

We show how to optimally regulate prediction algorithms in a world where an agent uses complex ‘black-box’ prediction functions to make decisions such as lending, medical testing, or hiring, and where a principal is limited in how much she can learn about the agent’s black-box model. We show that limiting agents to prediction functions that are simple enough to be fully transparent is inefficient as long as the misalignment is limited and first-best prediction functions are sufficiently complex. Algorithmic audits can improve welfare, but the gains depend on the design of the audit tools. Tools that focus on minimizing overall information loss, the focus of many explainer tools, will generally be inefficient since they focus on explaining the average behavior of the prediction function. Targeted tools that focus on the source of incentive misalignment, e.g., excess false positives or racial disparities, can provide second-best solutions. We provide empirical support for our theoretical findings using an application in consumer lending, where we document that complex models regulated based on context-specific explanation tools outperform simple, fully transparent models. This gain from complex models represents a Pareto improvement across our empirical applications that are preferred both by the lender and from the perspective of the financial regulator.

When machine-learning algorithms are used in high-stakes decisions, we want to ensure that their deployment leads to fair and equitable outcomes. This concern has motivated a fast-growing literature that focuses on diagnosing and addressing disparities in machine predictions. However, many machine predictions are deployed to assist in decisions where a human decision-maker retains the ultimate decision authority. In this article, we therefore consider in a formal model and in a lab experiment how properties of machine predictions affect the resulting human decisions. In our formal model of statistical decision-making, we show that the inclusion of a biased human decision-maker can revert common relationships between the structure of the algorithm and the qualities of resulting decisions. Specifically, we document that excluding information about protected groups from the prediction may fail to reduce, and may even increase, ultimate disparities. In the lab experiment, we demonstrate how predictions informed by gender-specific information can reduce average gender disparities in decisions. While our concrete theoretical results rely on specific assumptions about the data, algorithm, and decision-maker, and the experiment focuses on a particular prediction task, our findings show more broadly that any study of critical properties of complex decision systems, such as the fairness of machine-assisted human decisions, should go beyond focusing on the underlying algorithmic predictions in isolation.

Researchers often run resource-intensive randomized controlled trials (RCTs) to estimate the causal effects of interventions on outcomes of interest. Yet these outcomes are often noisy, and estimated overall effects can be small or imprecise. Nevertheless, we may still be able to produce reliable evidence of the efficacy of an intervention by finding subgroups with significant effects. In this paper, we propose a machine-learning method that is specifically optimized for finding such subgroups in noisy data. Unlike available methods for personalized treatment assignment, our tool is fundamentally designed to take significance testing into account: it produces a subgroup that is chosen to maximize the probability of obtaining a statistically significant positive treatment effect. We provide a computationally efficient implementation using decision trees and demonstrate its gain over selecting subgroups based on positive (estimated) treatment effects. Compared to standard tree-based regression and classification tools, this approach tends to yield higher power in detecting subgroups affected by the treatment.

## Current working papers

In many settings, interventions may be more effective for some individuals than others, so that targeting interventions may be beneficial. We analyze the value of targeting in the context of a large-scale field experiment with over 53,000 college students, where the goal was to use “nudges” to encourage students to renew their financial-aid applications before a non-binding deadline. We begin with baseline approaches to targeting. First, we target based on a causal forest that estimates heterogeneous treatment effects and then assigns students to treatment according to those estimated to have the highest treatment effects. Next, we evaluate two alternative targeting policies, one targeting students with low predicted probability of renewing financial aid in the absence of the treatment, the other targeting those with high probability. The predicted baseline outcome is not the ideal criterion for targeting, nor is it a priori clear whether to prioritize low, high, or intermediate predicted probability. Nonetheless, targeting on low baseline outcomes is common in practice, for example because the relationship between individual characteristics and treatment effects is often difficult or impossible to estimate with historical data. We propose hybrid approaches that incorporate the strengths of both predictive approaches (accurate estimation) and causal approaches (correct criterion); we show that targeting intermediate baseline outcomes is most effective, while targeting based on low baseline outcomes is detrimental. In one year of the experiment, nudging all students improved early filing by an average of 6.4 percentage points over a baseline average of 37% filing, and we estimate that targeting half of the students using our preferred policy attains around 75% of this benefit.

Econometric analysis typically focuses on the statistical properties of fixed estimators and ignores researcher choices. In this article, I approach the analysis of experimental data as a mechanism-design problem that acknowledges that researchers choose between estimators, sometimes based on the data and often according to their own preferences. Specifically, I focus on covariate adjustments, which can increase the precision of a treatment-effect estimate, but open the door to bias when researchers engage in specification searches. First, I establish that unbiasedness is a requirement on the estimation of the average treatment effect that aligns researchers’ preferences with the minimization of the mean-squared error relative to the truth, and that fixing the bias can yield an optimal restriction in a minimax sense. Second, I provide a constructive characterization of treatment-effect estimators with fixed bias as sample-splitting procedures. Third, I show how these results imply flexible pre-analysis plans that include beneficial specification searches.

📎 PDF download · 📎 PDF download of longer 2018 version with additional results

What is the purpose of pre-analysis plans, and how should they be designed? We propose a principal-agent model where a decision-maker relies on selective but truthful reports by an analyst. The analyst has data access, and non-aligned objectives. In this model, the implementation of statistical decision rules (tests, estimators) requires an incentive-compatible mechanism. We first characterize which decision rules can be implemented. We then characterize optimal statistical decision rules subject to implementability. We show that implementation requires pre-analysis plans. Focussing specifically on hypothesis tests, we show that optimal rejection rules pre-register a valid test for the case when all data is reported, and make worst-case assumptions about unreported data. Optimal tests can be found as a solution to a linear-programming problem.

Instrumental variables (IV) regression is widely used to estimate causal treatment effects in settings where receipt of treatment is not fully random, but there exists an instrument that generates exogenous variation in treatment exposure. While IV can recover consistent treatment effect estimates, they are often noisy. Building upon earlier work in biostatistics (Joffe and Brensinger, 2003) and relating to an evolving literature in econometrics (including Abadie et al., 2019; Huntington-Klein, 2020; Borusyak and Hull, 2020), we study how to improve the efficiency of IV estimates by exploiting the predictable variation in the strength of the instrument. In the case where both the treatment and instrument are binary and the instrument is independent of baseline covariates, we study weighting each observation according to its estimated compliance (that is, its conditional probability of being affected by the instrument), which we motivate from a (constrained) solution of the first-stage prediction problem implicit to IV. The resulting estimator can leverage machine learning to estimate compliance as a function of baseline covariates. We derive the large-sample properties of this weighted IV estimator in the potential outcomes and local average treatment effect (LATE) frameworks, and provide tools for inference that remain valid even when the weights are estimated nonparametrically. With both theoretical results and a simulation study, we demonstrate that compliance weighting meaningfully reduces the variance of IV estimates when first-stage heterogeneity is present, and that this improvement often outweighs any difference between the compliance-weighted and unweighted IV estimands. These results suggest that in a variety of applied settings, the precision of IV estimates can be substantially improved by incorporating compliance estimation.

In this paper we present tools for applied researchers that re-purpose off-the-shelf methods from the computer-science field of machine learning to create a “discovery engine” for data from randomized controlled trials (RCTs). The applied problem we seek to solve is that economists invest vast resources into carrying out RCTs, including the collection of a rich set of candidate outcome measures. But given concerns about inference in the presence of multiple testing, economists usually wind up exploring just a small subset of the hypotheses that the available data could be used to test. This prevents us from extracting as much information as possible from each RCT, which in turn impairs our ability to develop new theories or strengthen the design of policy interventions. Our proposed solution combines the basic intuition of reverse regression, where the dependent variable of interest now becomes treatment assignment itself, with methods from machine learning that use the data themselves to flexibly identify whether there is any function of the outcomes that predicts (or has signal about) treatment group status. This leads to correctly-sized tests with appropriate p-values, which also have the important virtue of being easy to implement in practice. One open challenge that remains with our work is how to meaningfully interpret the signal that these methods find.

We consider learning personalized assignments to one of many treatment arms from a randomized controlled trial. Standard methods that estimate heterogeneous treatment effects separately for each arm may perform poorly in this case due to excess variance. We instead propose methods that pool information across treatment arms: First, we consider a regularized forest-based assignment algorithm based on greedy recursive partitioning that shrinks effect estimates across arms. Second, we augment our algorithm by a clustering scheme that combines treatment arms with consistently similar outcomes. In a simulation study, we compare the performance of these approaches to predicting arm-wise outcomes separately, and document gains of directly optimizing the treatment assignment with regularization and clustering. In a theoretical model, we illustrate how a high number of treatment arms makes finding the best arm hard, while we can achieve sizable utility gains from personalization by regularized optimization.

Shrinkage estimation usually reduces variance at the cost of bias. But when we care only about some parameters of a model, I show that we can reduce variance without incurring bias if we have additional information about the distribution of covariates. In a linear regression model with homoscedastic Normal noise, I consider shrinkage estimation of the nuisance parameters associated with control variables. For at least three control variables and exogenous treatment, I establish that the standard least-squares estimator is dominated with respect to squared-error loss in the treatment effect even among unbiased estimators and even when the target parameter is low-dimensional. I construct the dominating estimator by a variant of James–Stein shrinkage in a high-dimensional Normal-means problem. It can be interpreted as an invariant generalized Bayes estimator with an uninformative (improper) Jeffreys prior in the target parameter.

The two-stage least-squares (2SLS) estimator is known to be biased when its first-stage fit is poor. I show that better first-stage prediction can alleviate this bias. In a two-stage linear regression model with Normal noise, I consider shrinkage in the estimation of the first-stage instrumental variable coefficients. For at least four instrumental variables and a single endogenous regressor, I establish that the standard 2SLS estimator is dominated with respect to bias. The dominating IV estimator applies James–Stein type shrinkage in a first-stage high-dimensional Normal-means problem followed by a control-function approach in the second stage. It preserves invariances of the structural instrumental variable equations.