Daily Summary for 2026-11-04

daily

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: It's the fourth of November, twenty twenty-six, and this is the day's research.

Warren: 35 new papers came out today.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We'll take the day in one pass, then pull out the papers we're staying with.

Warren: Start us off with Policy Learning with Confidence.

The day's research: Ruth: Understanding choices helps target policies effectively, looking at network responses for deworming uptake in Kenya or classroom peer effects in Tennessee schools.

Warren: The most significant work uses discrete choice models showing how variation within a network reveals interaction effects while controlling for unobserved context, detailed in Discrete Choice with Generalized Social Interactions.

Iris: That foundational work links social structure to choice outcomes and suggests conformity to similar peers shapes decisions, which implies direct policy design implications.

Ruth: Another area involves making statistical inferences more robust by pairing recommended decisions with quality guarantees through P-certificates, as seen in Certified Decisions.

Warren: This concept lets decision-makers adopt actions safely when facing ambiguity, relevant for high-stakes calls under imperfect information.

Iris: Moving down the list, there is work on selecting policies when estimation uncertainty exists by proposing a risk-aware rule maximizing welfare for a given level of estimation risk in Policy Learning with Confidence.

Ruth: That helps allocate limited resources among social programs by ensuring delivered welfare meets a threshold with high confidence.

Warren: Then there is research into causal estimation methods using machine learning to model flexible relationships, examining how successful these are at recovering true baseline effects from nonrandomized data in Reevaluating Causal Estimation Methods with Data from a Product Release.

Iris: This addresses practical challenges of using modern, high-dimensional datasets for causal analysis.

Ruth: Finally, there is work on bias-aware inference in regularized regression models that develops confidence intervals accounting for potential bias in Bias-Aware Inference in Regularized Regression Models.

Warren: This provides more accurate measures of uncertainty when using complex statistical models under high-dimensional settings.

Iris: The most critical finding relates to how large language models reshape understanding of economic policy uncertainty via the Economic Policy Uncertainty index, as shown in Narratives to Numbers: Large Language Models and Economic Policy Uncertainty.

Ruth: When fine-tuned on human audit labels, these models identify nearly twice as many auditor-flagged articles compared to keyword-based methods and remain robust against noisy text and linguistic drift.

Warren: This means tools we use to gauge uncertainty are fundamentally changing based on how they are trained, which is a big deal for interpreting current economic signals.

Iris: This suggests reliance on certain data sources might be systematically biased by modeling choices upstream rather than just underlying economic reality.

Ruth: Furthermore, evidence-based policy choice in macroeconomics demonstrates that choosing between contemporaneous and forward guidance regimes yields lower risk across nearly all posterior draws in Bayes-optimal rule simulations, as shown in Evidence-Based Policy Choice in Macroeconomics.

Warren: This implies anticipating future paths through guidance is statistically superior to reacting to immediate shocks even when parameter uncertainty is present.

Iris: The analysis of transmission channels provides a quantitative way to break down total effects of policy shocks into specific channels within dynamic models, as seen in Transmission Channel Analysis in Dynamic Models.

Ruth: This method allows researchers to assess the strength of different economic pathways, which is crucial for understanding how a policy choice actually moves the economy.

Warren: Finally, market design principles can be applied to areas like targeted advertising platforms where merchants share data, as shown in Targeted Advertising Platforms: Pooling Audiences Despite Poaching Risk.

Iris: This model shows optimal mechanisms can balance platform revenue against merchant surplus even when facing risks of customer poaching, suggesting transferable insights into resource allocation across different markets.

Ruth: Today's papers include Discrete Choice with Generalized Social Interactions, Certified Decisions, Policy Learning with Confidence, Reevaluating Causal Estimation Methods with Data from a Product Release, Bias-Aware Inference in Regularized Regression Models.

Warren: These papers study how social interactions influence individual choices in models where people respond differently to different network members.

Iris: This paper develops a theory for statistical inference by pairing recommended decisions with guarantees of their quality called P-certificates.

Ruth: This paper introduces a rule for choosing policies that explicitly accounts for the risk associated with estimation uncertainty.

Warren: This paper analyzes how successful machine learning methods are at recovering true causal effects from nonrandomized data in high-dimensional settings.

Iris: This develops bias-aware confidence intervals for regularized regression models that account for the possible bias of the estimator.

Ruth: The Anatomy of Dependence in Multivariate Ordered Choice introduces a model to capture dependence across multiple ordered dimensions when individuals make simultaneous decisions.

Warren: Cohort-Anchored HonestDiD proposes a cohort-based sensitivity analysis method to properly interpret event study coefficients under staggered adoption.

Iris: Best Feasible Conditional Critical Values for a More Powerful Subvector Anderson-Rubin Test proposes an improved test for subvector inference in linear instrumental variables models that uses data-dependent critical values.

Ruth: Evidence-Based Policy Choice in Macroeconomics derives decision rules for macroeconomic policy choice by explicitly accounting for estimation risk under different policy regimes.

Warren: Narratives to Numbers: Large Language Models and Economic Policy Uncertainty evaluates how using machine-generated data affects the interpretation of economic policy uncertainty indices.

Iris: Transmission Channel Analysis in Dynamic Models proposes a framework to decompose impulse response functions into effects flowing through specific transmission channels in dynamic models.

Ruth: Targeted Advertising Platforms: Pooling Audiences Despite Poaching Risk proposes a model for designing targeted advertising platforms that incentivize data sharing among merchants despite poaching risks.

Warren: Matching-LLM: Historically Anchored Synthetic Prediction of Public Opinion and Elections introduces a framework to improve large language model predictions for public opinion and elections by incorporating historical human data.

Iris: Aggregative Efficiency of Bayesian Learning in Networks quantifies the information loss that occurs when agents in a social network learn about an unknown state from private signals.

Ruth: A LATE-Targeting Experimental Design for Multi-Arm RCTs studies a two-stage experimental design to identify local average treatment effects under imperfect compliance in multi-arm randomized controlled trials.

Warren: Bias-robust causal inference for panel data develops a bias-robust framework for causal inference in panel data models that explicitly accounts for counterfactual approximation error.

Iris: Endogenous Identity in a Social Network uses social identity theory and network analysis to show how interactions across socioeconomic classes drive social mobility.

Ruth: Ranking Experiments under Sequential Sampling compares two orders of statistical experiments when observations are inexpensive and acquired sequentially until a decision is made.

Warren: Execution and Evaluation: Task Content, Labor Demand, and the Evaluator Pipeline distinguishes between the execution of tasks by machines and the evaluation of those tasks by humans in AI research.

Iris: Urban Reconstruction and Population Redistribution: Evidence from Tokyo after the Great Kanto Earthquake examines how post-earthquake population distribution in Tokyo was shaped primarily by market-based price adjustments.

Ruth: Carbon Regulation and Competition in the European Airline Industry estimates how increased carbon pricing impacts airline competition, consumer surplus, and airline profits across different market segments.

Warren: Short-Run Multi-Outcome Effects of Nightlife Regulation in San Juan evaluates the short-run economic effects of a late-night alcohol sales ordinance using a multi-outcome synthetic control.

Iris: European Defence Readiness assesses Europe's defence readiness by examining mobilisation readiness and whole-of-society resilience under deteriorating security conditions.

Lucky paper: 2502.10653: Ruth: We’re looking at Policy Learning with Confidence today, and it introduces a rule for policy selection when there’s estimation uncertainty that explicitly accounts for estimation risk to guarantee welfare exceeds a threshold with high confidence.

Warren: It contrasts this with the intuitive plug-in rule, which is empirical welfare maximization, or EWM—simply replacing the population objective with its sample analog.

Iris: The authors point out that estimation error could result in suboptimal policy choice because a policy with high estimated welfare might actually have low precision.

Ruth: The proposed rule belongs to the class of risk-aware rules on the efficient decision frontier, which are characterized as policies offering maximal estimated welfare for a given level of estimation risk.

Warren: So these rules select a policy from that efficient decision frontier by trading off higher sample estimates against smaller sample variance.

Iris: The PoLeCe rule is determined by balancing this tradeoff using the tangent line to the frontier whose slope is set by the critical value of a one-sided upper confidence band for welfare following an approach developed for intersection bound inference.

Ruth: This PoLeCe rule delivers a reporting guarantee ensuring with high confidence that actual welfare delivered exceeds a lower threshold while also having the favorable regret properties of all risk-aware rules.

Warren: Specifically, Proposition three shows that with probability at least one minus alpha minus r sub n, the true welfare of the PoLeCe rule is bounded below as LV1 minus alpha times pi.

Iris: This bound provides a natural performance guarantee since the true welfare of policy PoLeCe is greater than or equal to LV1 minus alpha times pi with probability at least one minus alpha minus r sub n.

Ruth: The authors apply this approach to allocating a limited budget among social programs using estimates of their marginal value of public funds and associated standard errors.

Warren: In empirical examples, such as those involving programs targeting individuals older than twenty-five the PoLeCe rule allocates funds differently than EWM by spreading out funds over an efficient portfolio similar to Markowitz portfolio allocation.

Iris: Simulations calibrated to three empirical applications show that when there is heterogeneity in the estimation risk for policies and sigma pi zero, which is the lower bound on the risk of the best policies, is small, PoLeCe outperforms EWM in terms of tail regret.

Lucky paper: 2012.14823: Ruth: So we're looking at Bias-Aware Inference in Regularized Regression Models, and the core idea here is developing estimators based on regularized propensity score regression to solve the bias variance tradeoff.

Warren: That two step process, running a penalized propensity score regression of the control variables on w i using a penalty function Pen(times) and then using those residuals as an instrument in a univariate regression of the outcome Y i on the estimated control coefficients w tilde i, is what exactly solves that tradeoff.

Iris: The key result here is that these estimators yield bias-aware confidence intervals, and those CIs are constructed using a critical value that explicitly incorporates the worst-case bias of the estimator.

Ruth: That inherent feature means these CIs are valid in finite samples under an idealized homoskedastic Gaussian error assumption.

Warren: But what about when things get messy? The paper extends this framework to more realistic scenarios involving heteroskedasticity and an unknown error distribution.

Iris: In that setting, the bias-aware CIs are formed by utilizing heteroskedasticity-robust variance estimators, and the authors provide rigorous conditions for their asymptotic validity in high-dimensional asymptotics where the number of regressors k significantly exceeds the number of observations n.

Ruth: That high dimensional convergence rate analysis is what allows them to derive novel optimal rates of convergence for these bias-aware CIs, which is important when dimensionality grows faster than the sample size.

Warren: However, there's a limitation they flag: this approach requires an explicit specification of a penalty bound C, and it's fundamentally impossible to automate the choice of that bound using purely data-driven methods.

Iris: That is because obtaining a sufficiently informative upper bound for C from the data itself is infeasible, but they do provide a constructive method to obtain a one-sided lower CI for C, which acts as an essential specification check.

Ruth: So, even with that explicit bound requirement, the paper demonstrates that both the proposed estimators and their corresponding bias-aware CIs are near-optimal in terms of mean squared error and CI length for finite samples under the idealized setting.

Warren: Under those specific conditions of homoskedastic Gaussian errors, they show near optimality in mean squared error and CI length for finite samples.

Iris: When we look at the application, how does this methodology change how we think about uncertainty when using complex statistical models in practice?

Lucky paper: 2205.05779: Ruth: We are looking at The Anatomy of Dependence in Multivariate Ordered Choice, and this paper is introducing a new class of General Rectangular Structure Models that separates behavioral interdependence from correlation arising from latent unobservables.

Warren: So the core idea is distinguishing between how decision rules link different dimensions versus how those dimensions are just correlated because of some hidden factor.

Iris: That separation means researchers can test empirically whether observed dependence reflects a structural link in decision rules or if it's just a correlation among unobserved factors.

Ruth: The authors formalize this by allowing the decision thresholds governing choices in one dimension to be a function of the outcome indices across all other dimensions, which introduces that second layer of dependence.

Warren: That richer specification lets them model cross-partial effects and even variable sign changes in partial effects, which is something standard lattice models don't allow.

Iris: And they introduce coherency as a formal concept to ensure those rectangular regions defined by the thresholds actually partition the latent space, which is rigorously characterized by Proposition one.

Ruth: To estimate all these complex parameters, they outline a hierarchical identification procedure starting with corner regions, specifically R j one R j D, where initial thresholds alpha are identified using conditions a and c.

Warren: Then it gets more complicated because the process moves sequentially through stages like Stage two focusing on rectangular regions near the borders using finite thresholds from Stage one.

Iris: Stages three and four involve allowing discrete processes in specific dimensions to take interior categories while others remain at boundary values, which is a key part of their refinement.

Ruth: The identification relies on analyzing the positive set of a derived expression utilizing supporting hyperplanes defined by condition b, where their contact points lie within the observed index range.

Warren: Those supporting hyperplanes are crucial because they are linearly independent and uniquely determine the remaining unknown thresholds through that analysis.

Iris: The authors conclude that this General Rectangular Structure provides a more powerful econometric toolkit than simpler lattice models because it captures nuanced patterns of behavioral interdependence absent in those restricted frameworks.

Lucky paper: 2509.01829: Ruth: We're looking at Cohort-Anchored HonestDiD today, which is number four of five in this show.

Warren: This paper by Ziyi Liu, Kirill Borusyak, and Yiqing Xu introduces a cohort-anchored framework for robust inference in event studies with staggered treatment adoption.

Ruth: The core idea here is shifting the focus from aggregated event-study coefficients to cohort-period level coefficients.

Iris: The authors introduce block bias as the parallel trends violation for a specific cohort relative to its fixed initial control group, and they show that any overall stacked bias delta vector can be decomposed invertibly into these interpretable block biases delta.

Ruth: That linear mapping with W allows us to visualize pre-treatment coefficients for estimators like the imputation estimator, which is pretty transparent.

Warren: So, if we have this decomposition, how does it help researchers compare pre-treatment and post-treatment effects?

Iris: It lets you do a direct comparison between pre-treatment and post-treatment effects because you can see them separated by these block biases.

Ruth: And to keep things robust, they impose two key restrictions on delta: the relative magnitudes restriction and the second differences restriction.

Warren: The relative magnitudes restriction bounds the post-treatment changes in a cohort’s block bias by a benchmark derived from pre-treatment trends.

Iris: And that simplifies down to considering the single largest square defined by those bounds, which helps set identify the true treatment effect under those constraints.

Ruth: So, empirically, they found this approach yields better-centered confidence sets than traditional aggregated approaches when pre-trends vary across cohorts.

Warren: That superiority is supported by simulations and an empirical reanalysis of minimum-wage effects on teen employment, which is interesting because that's a real setting.

Iris: The framework is most suitable in settings with multiple cohorts, provided there's enough within-cohort precision to estimate those cohort-period coefficients reasonably well.

Ruth: But we have to keep an eye on the limitations: this framework doesn't accommodate missing data or estimators that use pre-treatment covariates for unit weighting, like inverse probability weighting or doubly robust estimators.

Warren: Those methods can change the fundamental bias-decomposition identity linking overall and block biases, which is a big caveat when you try to apply it to more complex real-world datasets.

Iris: So, the main contribution is both the framework itself and that underlying mechanism, which offers a powerful tool for interpreting pre-treatment coefficients in staggered adoption settings.

Lucky paper: 2601.17843: Ruth: We're moving into paper number five of five today, and this is Best Feasible Conditional Critical Values for a More Powerful Subvector Anderson-Rubin Test by Jesse Hoekstra and Frank Windmeijer.

Warren: So, this paper deals with subvector inference in linear instrumental variables models when you allow for weak instruments under homoskedasticity.

Iris: The authors propose a conditional subvector Anderson and Rubin test that uses data-dependent critical values to adapt to the strength of the parameters not under test.

Ruth: They show this test has correct size and strictly higher power than the test that uses standard asymptotic chi-square critical values.

Warren: The paper investigates a data-dependent critical value function for this subvector Anderson and Rubin test by conditioning on the second-smallest eigenvalue of a data-dependent matrix.

Iris: They find that this approach maintains correct size and strictly higher power than the Guggenberger, Kleibergen, and Mavroeidis test when the number of parameters not under test is larger than one.

Ruth: The authors state that this procedure is termed "feasible" when restricting conditioning to a single component of the spectral summary of identification strength, which allows for feasible inference via a known highly tractable approximate conditional density that relies solely on standard scalar-input functions.

Warren: The subvector Anderson and Rubin test statistic is defined as the minimum eigenvalue of a data-dependent matrix, and the Guggenberger, Kleibergen, and Mavroeidis critical value function conditions on the largest eigenvalue of this non-central Wishart-distributed matrix.

Iris: They investigate using data-dependent critical values that condition not on the largest but on the second-smallest eigenvalue because that eigenvalue is considered the appropriate indicator of weak identification.

Ruth: For p greater than two, if the p minus two largest eigenvalues of the p times p non-central Wishart-distributed matrix are all large and increasing to infinity, then the joint distribution of those two smallest eigenvalues is the same as that of a two by two Wishart matrix with a change in degrees of freedom.

Warren: They show that when those p minus two largest eigenvalues decrease to less than infinity, the rejection probability decreases, which means the test has size less than or equal to its nominal size under the null in all circumstances.

Iris: For p greater than two, as the critical value when conditioning on the second-smallest eigenvalue is smaller than that conditioning on the largest eigenvalue, this Anderson and Rubin subvector testing procedure has strictly higher power than the Guggenberger, Kleibergen, and Mavroeidis test when mW is greater than one.

Ruth: The authors conclude that this phi cp minus one test is the best feasible test within the class of tests that use a conditional critical value function, conditioning on a single eigenvalue.

Warren: They emphasize that by isolating the second-smallest eigenvalue, they condition on the component most informative about identification strength.

Iris: The test is shown to control size and has power strictly higher than the Guggenberger, Kleibergen, and Mavroeidis test when mW is greater than one.

Ruth: This approach is also robust to an approximate Kronecker product structure of conditional heteroskedasticity, maintaining its power advantage in that setting as well.

Warren: The empirical analysis confirms the practical importance of the proposed test, showing it not only strengthens results reported in Guggenberger, Kleibergen, and Mavroeidis but also affects some of its conclusions.

Iris: The final result is that for mW greater than one, phi cp minus one has strictly higher power than the Guggenberger, Kleibergen and Mavroeidis test phi c1 when mW is greater than one.

More episodes

← Home