Daily Summary for 2026-11-11
daily
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: It's the eleventh of November, twenty twenty-six, and this is the day's research.
Warren: 30 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 Learning Where to Look.
The day's research: Ruth: We examine how principals make decisions when they only have private information about an agent.
Warren: That asymmetry is crucial for designing better contracts and systems because understanding it helps us.
Iris: The core idea is a rule to delegate authority optimally, minimizing regret even when circumstances are unknown.
Ruth: This optimal rule balances sticking to a standard default action versus deviating from it.
Warren: How much freedom the agent gets when they choose to deviate is another key element of that balance.
Iris: Hidden commitment power shows enforcement ability can be less than we think if information about commitment is hidden.
Ruth: This suggests only measures improving the worst-case scenario are truly valuable in this information gap.
Warren: We also have work testing if linear specifications accurately describe relational data, especially in dyadic regression models.
Iris: Certain bootstrap methods consistently approximate the limiting law even when parameters are estimated or the function is curved.
Ruth: This contrasts with foundational work on identifying discount factors and payoff functions in dynamic choice models.
Warren: That research uses assumptions about per-period payoffs to find those unknown economic parameters, which is different from linear specification testing.
Iris: There is also methodological work on where to collect data efficiently using Delaunay matching to determine the best next step for an intervention.
Ruth: This suggests geometric methods can significantly reduce worst-case bias when learning about a population's response to an intervention.
Warren: The work on interacting principal-agent games is significant because it tackles incentive compatibility fragility when multiple principals design systems simultaneously.
Iris: Exact equilibrium can be recovered even when allowing small expected incentive regret, as shown in Interacting Principal-Agent Games.
Ruth: This means in complex economic settings, we do not need perfect coordination for the system to work well if regret remains small.
Warren: That is supported by analysis of a one-member team-contest specialization within that general model.
Iris: This establishes existence of an exact Bayesian-Nash principals' equilibrium under conditions where feasibility varies continuously in Interacting Principal-Agent Games.
Ruth: Delegation without Priors states a principal finds a rule minimizing regret using private information from an agent.
Warren: Specification Testing for Dyadic Regression Models develops omnibus tests to check if linear specifications accurately describe conditional means in undirected dyadic data.
Iris: Hidden Commitment Power Is Powerless A principal's ability to renege on a contract is often limited by their private information about their commitment power.
Ruth: Robust Contracts with Exploration involves the principal contracting sequentially, allowing exploration before final commitment.
Warren: Double Machine Learning for Inequality Decompositions uses double machine learning to decompose outcome gaps using a reference outcome that depends on the choice of estimator.
Iris: A Note on Qualified Majority Voting Rules characterizes every qualified majority voting rule by examining properties like anonymity, responsiveness, and q-neutrality.
Ruth: Semiparametric Identification of the Discount Factor and Payoff Function in Dynamic Discrete Choice Models identifies those parameters using restrictions from economic theory.
Warren: Learning Where to Look Delaunay Matching for Discovery and Data Collection shows using a Delaunay triangulation minimizes worst-case bias when combining observational data with small intervention samples.
Iris: Interacting Principal-Agent Games In many economic environments, multiple principals design incentive systems that strategically interact with each other.
Ruth: Combining stated and revealed preferences proposes using stated choices to control for unobserved heterogeneity to recover causal effects from actual choices.
Warren: Continuous difference-in-differences with double/debiased machine learning extends difference-in-differences to continuous treatments by proposing estimators under a double/debiased machine learning framework.
Iris: An Algebraic Approach to Continuously Updating GMM Estimation and Inference provides an algebraic solution for the continuously updating GMM estimator, allowing for efficient computation of its minimum.
Ruth: On the construction and representation of social welfare orders Satisfying Consequentialist Equity Axioms examines how to construct such orders using lexicographic preference relations.
Warren: Difference-in-differences with a mediator Causal mediation analysis is advanced within the difference-in-differences framework to identify natural indirect, direct, and total effects of a treatment.
Iris: Disappointment Aversion and Expectiles recasts disappointment aversion theory using a Savage framework and connects it to econometric models of asymmetric least squares estimation.
Ruth: Tracking the economy at high frequency develops a high-frequency economic index by integrating mixed-frequency official data to monitor the economy in real time.
Warren: Risk-insurance parity introduces risk-insurance parity, which associates different classes of insurance contracts with varying notions of risk aversion.
Iris: Uniform price auction with quantity constraints studies the equilibria of uniform price auctions where many asymmetric bidders have flat demands up to their quantity constraints.
Ruth: Random Collection uses an axiomatic approach to characterize various parametric models in choice situations where a decision maker selects a subset of options probabilistically.
Warren: Organizational Lifespan as Commitment identifies when an organization's lifespan commits future decision makers to continue or terminate based on investment rules.
Iris: Equilibrium Selection in Coordination Games with Planned Actions and Scouting studies stag hunt games where players screen for cooperation by observing noisy signals about their opponents' preparations.
Ruth: Ever since Ellsberg compares subjective expected utility theory with non-EUT models of ambiguity aversion using revealed preference methodology.
Warren: Where the Gains Come From The wedge between multiple-prior and uniform-prior FOSD-rationalizability is the empirical space where non-EUT models can outperform subjective EUT.
Iris: Work that Counts, and Work that Doesn't compares worker responses to wage changes in experiments versus actual work to reconcile differences in elasticity estimates.
Ruth: The Process and Dynamics of the Nobel Memorial Prize in Economics investigates the complex dynamics governing the awarding of the Nobel Memorial Prize by analyzing network data.
Warren: Jornada de trabalho, produtividade e informalidade calculates how much productivity gain is needed for different working hour options to keep output constant.
Iris: Wage-Setting Constraints and Firm Responses to Demand Shocks investigates how institutional wage-setting constraints influence firm responses to demand shocks across different countries.
Ruth: Identification-aware Markov chain Monte Carlo in Bayesian analysis develops MCMC samplers that combine local updates with moves across observationally equivalent parameter values for better identification.
Warren: Advancing AI Capabilities and Evolving Labor Outcomes analyzes how increasing AI capabilities are associated with changes in employment status and work hours across different occupations.
Lucky paper: 2607.04885: Ruth: So we’re looking at Learning Where to Look: Delaunay Matching for Discovery and Data Collection, and this paper is trying to solve a real problem where a policymaker needs to decide whether or not to introduce an innovation based on private information from donors.
Warren: The core challenge they tackle is designing an estimator that guarantees the compensation cost stays below a finite budget, regardless of which target population the policymaker ends up choosing.
Iris: They introduce certification as a theoretical device, which helps bound the worst-case compensation cost by controlling the probability of making a mistake uniformly across all possible outcomes.
Ruth: The paper connects this certification with computational geometry by defining the Delaunay Matching Estimator using barycentric interpolation on the triangulation of donor covariates.
Warren: That means they’re using triangles formed by donors to guide where to look for new data points, which is a very geometric way to handle the uncertainty.
Iris: The geometric plan motivates a new collection rule that adds donor observations precisely where the current approximation of the effect is weakest, which they define as the center of the smallest circle containing the largest simplex.
Ruth: That sounds like they are specifically targeting areas where their current understanding is least certain, and empirically they showed this geometric plan halved costs by adding just three donor units compared to random collection.
Warren: That empirical validation on the Smartcards experiment is pretty strong, showing that the geometric approach beats random sampling significantly within twenty steps of additions.
Iris: It’s interesting how they move from a purely statistical decision theory setup to this geometric problem, which gives them analytical guarantees for the choice of weights.
Ruth: But there's a big hurdle mentioned in Remark one that the donor and target populations don't coincide, which means things like unconfoundedness and strict overlap are not holding in the traditional policy learning literature.
Warren: That’s a big caveat because it means their guarantee needs to hold conditional on the specific collection design they chose, not just on an average collection over all possible designs.
Iris: The paper then shows how they can use certified decisions, where an estimator is defined by a boundary based on the error rate alpha and a scalar gamma of the estimator, which essentially defines a region where making a mistake is bounded.
Ruth: And Lemma one confirms that if you choose an estimator that satisfies certain conditions related to this certification boundary, the worst-case compensation cost is bounded by the maximum between gamma of the estimator and alpha taubar.
Warren: So basically, for the PM's objective to be met, they just need to commit to an estimator where that bound is less than or equal to their budget B0.
Iris: That links back nicely to Definition two where the goal is setting an ex-ante estimator such that the supremum of the loss over all possible outcomes and estimators stays below that budget.
Ruth: This whole paper contributes by highlighting how studying data-driven decisions from a geometric perspective offers concrete, affordable guarantees for cost-efficient collection plans in policy choice problems.
Warren: It moves beyond just saying random sampling is bad by giving them a specific, geometrically motivated rule for where to look next when collecting new donor observations.
Iris: It really shows the value of statistical decision theory when you apply these geometric results to design actual data collection strategies.
Lucky paper: 2507.13552: Ruth: So we're looking at Combining stated and revealed preferences, which is paper number two of this show. The core thing here is using stated choices to identify unobserved heterogeneity to correct for endogeneity when actual and hypothetical choices are observed together.
Warren: That’s a big deal because it addresses that problem of endogeneity in demand function estimation, especially when we have unmatched data sets.
Iris: The identification framework assumes an agent makes a binary decision based on an endogenous attribute X, and the structural choice function m(x, eta) is what we want to recover.
Ruth: You mentioned the structural choice function m(x, eta), how does that relate to what’s being reported in stated preferences?
Warren: Stated preferences are modeled by assuming the stated choice function m-tilde is driven by the exact same vector of unobserved characteristics, eta, as the structural one.
Iris: Assumption two guarantees that a unique eta can be recovered from the set of all stated preference reports P equals m-tilde (x, eta).
Ruth: If we can recover that vector eta from the reports P, what exactly does it let us do with the actual choices?
Warren: It lets us identify m(x, p), which is the structural function based on actual choices and stated preferences.
Iris: The fundamental idea is that stated preferences provide information on individual heterogeneity, allowing us to correct for its influence in demand function estimation using actual choices.
Ruth: You mentioned bounds on causal effects derived from this combination of data sets. What are those bounds actually telling us in practice?
Warren: Corollary one gives bounds on the average structural function µ(x) based on the distributions of actual choice outcomes and stated preference reports.
Iris: The inference procedure uses a one-sided confidence bound for the upper bound UBx using a bootstrap procedure proposed by Fang and Santos from two thousand nineteen.
Ruth: That sounds computationally intensive. Can you tell us more about how that bootstrap works in practice?
Warren: It yields a confidence region defined by the quantile of b phi prime n, which is related to the distribution of the estimated parameter difference between the true value and our estimate.
Iris: The simulation results show coverage rates close to one for the true value of µ(zero), meaning we expect most of our confidence regions to actually capture that true value.
Ruth: That’s encouraging, but what about when you only have scalar or bivariate stated probability reports? Does it still work well then?
Warren: The paper notes that this inference procedure is best suited for environments with scalar or at most bivariate stated probability reports.
Iris: They are also looking into using methods from the large factor model and group fixed effects literature in ongoing research to identify a finite number of latent types directly from elicited preferences.
Ruth: So, beyond just correcting for endogeneity, what is the broader implication for researchers who only look at actual choices?
Warren: The authors suggest these bounds are useful even when you aren't combining stated and revealed preferences in exactly the way they did in this work.
Iris: It suggests a new way to use stated choices to control for unobserved heterogeneity, which is crucial when trying to recover causal effects from actual choices.
Ruth: How does this relate to what Meng worries about at scale? Does this method become too slow or complex when you try to deploy it in a production system?
Warren: From an engineering standpoint, the inference procedure involves a bootstrap and quantile calculations, which means its performance depends on the sample size n.
Iris: The confidence bands are expected to narrow both with the sample size and somewhat with the tuning parameter x n, which is something to watch for when scaling up.
Ruth: Let's see what Lalam says about how this connects to real-world AI modeling.
Lalam: Based on the literature I have access to, this method provides a principled way to inject information from stated preferences into structural models without needing strong structural assumptions on those reports, which is very useful for complex preference elicitation tasks.
Warren: It seems like a solid tool for when you are trying to bridge the gap between what people say they want and what they actually do.
Iris: It's more than just a correction; it’s about finding bounds on causal effects in situations where perfect knowledge is impossible.
Lucky paper: 2408.10509: Ruth: We're looking at Continuous difference-in-differences with double/debiased machine learning today.
Warren: That paper by Lucas Zheng Zhang extends difference-in-differences to settings with continuous treatments.
Iris: It uses a conditional parallel trends assumption and establishes asymptotic normality within the double/debiased machine learning framework.
Ruth: So, what does that mean for the actual estimation of those average treatment effects on the treated, or ATT?
Warren: The paper defines it as AT T d equals E of Y t d minus Y t zero D equals d, where d is any given continuous treatment intensity.
Iris: To estimate that ATT, you first have to estimate infinite-dimensional nuisance parameters, like the conditional density of the continuous treatment variable.
Ruth: That sounds like a big hurdle because estimating those densities can introduce substantial bias into the result.
Warren: The authors tackle that by proposing estimators for causal parameters under the double/debiased machine learning framework and showing their asymptotic normality.
Iris: They use an approximate causal parameter, AT Th, built with a kernel function to avoid directly applying standard double/debiased machine learning methods.
Ruth: What are the asymptotic properties they establish for these estimators?
Warren: They show that AT T d minus AT T d sigma N d over the square root of N converges to zero times a normal distribution with zero mean and one variance.
Iris: The asymptotic variance has two parts, both depending inversely on the kernel bandwidth h; one comes from the kernels in the orthogonal score psi h, and the other from linear expansion with respect to kernel density estimator.
Ruth: That dependence on bandwidth sounds like a practical challenge for us when we try to deploy these models.
Warren: They construct a cross-fitted variance estimator and provide uniform confidence bands using a multiplier bootstrap procedure.
Iris: The effectiveness of the approach is shown by revisiting the one thousand nine hundred eighty-three Medicare Prospective Payment System reform, reframing it as difference-in-differences with continuous treatment.
Ruth: And what did that empirical application find about the reform?
Warren: The results suggested significant heterogeneity in the impact of that reform across hospitals based on how many Medicare inpatients they had.
Iris: Monte Carlo simulations demonstrated that these estimators perform well even though there is a highly non-linear relationship between the continuous treatment and the high-dimensional covariates.
Ruth: And what did those simulations show regarding their distribution?
Warren: The simulation estimates appeared to follow a normal distribution in each case, which is consistent with their asymptotic theory.
Iris: The coverage of the estimated confidence intervals was close to ninety-five percent across those cases.
Ruth: So, for someone just listening on the show, what does this mean for policy analysis?
Warren: It gives researchers new tools specifically for examining the impacts of continuous treatment variables in causal studies.
Iris: The paper concludes that their estimators perform well even despite that highly non-linear relationship between the continuous treatment and those high-dimensional covariates.
Lucky paper: 2604.24049: Ruth: Alright, let's talk about Difference-in-differences with a mediator and see how this paper handles the identification challenges. This is lucky paper number four of five for us today, so we'll keep that in mind as we go through it.
Warren: The paper from Fred Hutchinson Cancer Center and UC San Diego looks at identifying natural indirect, direct, and total effects when you have a mediator. It sets up the framework by defining the total effect as E of Y1(one) minus Y1(zero) when G equals one.
Iris: That framework is key because it defines the natural indirect effect as switching treatment associated with the mediator to get tauIE, and switching treatment associated with the outcome to get tauDE.
Ruth: So we have three distinct effects: total effect, natural indirect effect, and natural direct effect. How does the paper handle those identification assumptions in a real observational setting?
Warren: The identification relies on Assumption two requiring that conditional on baseline covariates and the effect on the mediator, the time trend in untreated potential outcomes is identical across groups. That avoids potential collider bias.
Iris: And Assumption three sequential ignorability, which means no unmeasured confounding between treatment assignment and the mediator, and no unmeasured confounding between the post-treatment mediator M(one) and the change in potential outcomes. Those are pretty stringent conditions for observational studies.
Ruth: It sounds like they're using a difference-in-differences structure to manage this identification problem, which is interesting since we usually need randomized trials for that kind of clean setup.
Warren: Exactly, they adopt a difference-in-differences framework and impose those assumptions to identify the effects in the treated group G equals one. They then derive efficient influence functions using semiparametric theory to estimate these estimands.
Iris: The paper notes that these estimators are multiply robust and locally nonparametrically efficient, provided either two of the propensity score, mediator distribution, and outcome regression models are correctly specified. That's a lot of assumptions packed into one estimation method.
Ruth: Okay, let's look at the application to the Job Corps Study data they use. What did they find regarding those specific effects?
Warren: Applying this methodology to data from the Job Corps Study shows that job training significantly increases both short-term and long-term earnings after controlling for the indirect effect through how many weeks a person was employed.
Iris: The estimated total effect is zero point zero six zero nine with a standard error of zero point zero six two three, and the ninety-five percent confidence interval goes from negative zero point zero six one two to positive zero point one eight three, which tells us something about the overall impact.
Ruth: That's a relatively small total effect size, but what about the components? What did they find for the natural indirect and direct effects?
Iris: The natural indirect effect is estimated at negative zero point zero four four seven with a standard error of zero point zero four three six, which suggests that joining the training program slightly reduces earnings by reducing work time.
Warren: But the natural direct effect is estimated at positive zero point one zero five five with a standard error of zero point zero four four five, indicating that joining the training program leads to higher earnings if you hadn't reduced your work time.
Ruth: So the direct effect is bigger than the indirect one in this case, which seems counterintuitive if we think about how mediation usually works. What about those controlled effects?
Iris: They also looked at the controlled direct effect, or CDE, which they define as E of Y1(one) minus Y1(zero) when G equals one and the mediator is fixed. The CDE curve shows a positive direct effect of training across a broad range of employment levels for both continuous and ordinal mediators.
Warren: That pattern suggests that in the longer run, the Job Corps program enhances earnings capacity conditional on how much work you do.
Ruth: So, to wrap up this segment, what's the main conclusion they draw about this framework? What's the big picture for someone just listening?
Iris: The paper concludes that while the natural indirect effect is insignificant, the natural direct effect is significant. And importantly, it suggests that for both continuous and ordinal mediators, that controlled direct effect is significantly positive across a broad range of employment levels.
Lucky paper: 2507.07450: Ruth: We're looking at Tracking the economy at high frequency, Freddy García-Albán and Juan Jarrín, who are building this new High-Frequency Economic Index for Ecuador using a Bayesian dynamic factor model on mixed data.
Warren: It's really interesting because they are dealing with very sparse and asynchronous official data there.
Iris: The paper describes constructing a novel dataset by integrating variables at weekly, monthly, and quarterly frequencies to create this high-frequency indicator for Ecuador.
Ruth: They introduce a pseudo-week structure where each month is divided into four equal periods to harmonize the timing of observations across those different frequencies.
Warren: That sounds like a necessary step to avoid these irregular aggregation artifacts when you're trying to stitch data together.
Iris: The core methodology uses a Bayesian dynamic factor model cast in state-space form, which helps accommodate that mixed-frequency data and time-varying volatility.
Ruth: The model assumes there is a single latent weekly cyclical factor, ft, that captures most of the comovement between all the variables in the system.
Warren: How they are estimating it—using a Gibbs sampler within a Bayesian framework—that's where I want to see some rigor.
Iris: They use that Gibbs sampler to jointly infer latent variables, time-varying volatilities, and missing observations through the Kalman filter and simulation smoother.
Ruth: The resulting High-Frequency Economic Index is interpreted as a pseudo-weekly economic index scaled to match the mean and standard deviation of the year over year growth rate of real GDP.
Warren: So they are trying to get this index to mirror major phases of Ecuador’s business cycle, like the two thousand eight two thousand nine financial crisis and COVID nineteen.
Iris: The paper shows that this index accurately tracks those major phases, including domestic episodes like the two thousand fifteen two thousand sixteen oil-related downturn.
Ruth: They compared their High-Frequency Economic Index with the business cycle estimates published by the Banco Central del Ecuador based on real GDP, and they got a simple correlation coefficient of approximately zero.
Warren: A zero correlation suggests the index is capturing something different than just what GDP is telling us directly.
Iris: The authors suggest that incorporating stochastic volatility in the factor yields more interpretable and economically plausible estimates, even if it isn't strictly favored by model selection criteria.
Ruth: They decompose the latent factor time series into a time varying mean mu t and a noise term epsilon t to feed into a second stage Markov-switching model for estimating recession probabilities.
Warren: That regime switching model allows them to distinguish between heterogeneous recessions and expansions, which is more granular than just one overall state.
Iris: The estimated recession probabilities provide a consistent signal about the likelihood that the economy was in a recession at any given time under this framework.
Ruth: The paper documents four well defined episodes of economic contraction using their index, which contrasts with the six contraction periods identified by GDP alone.
Warren: That difference in counting cycles is quite significant for short term monitoring applications, right?
Iris: The High-Frequency Economic Index captures nuances like the unusual resilience of deposits during the pandemic that traditional indicators might obscure.
Ruth: When we look at the data details, they recovered adequate employment series by using linear interpolation guided by administrative records from the Ecuadorian Social Security Institute.
Warren: And they also reconstructed weekly series for private sector credit and deposits through automated extraction from the Boletín Monetario Semanal publications.
Iris: They constructed trade variables from transaction level customs data, specifically classifying export data into oil and non-oil categories following the official methodology of the BCE.
Ruth: The state vector of the model is defined by a combination of factor components and idiosyncratic errors across quarterly, monthly, and weekly observations.
Warren: That structure must be quite complex to handle that level of temporal resolution while maintaining a manageable estimation process.
Iris: The matrix H in the state space form has components like Hq, Hm, and Hw which are defined by specific formulas involving the factor loadings and innovation standard deviations.
Ruth: Model comparison using the Deviance Information Criterion or DIC reveals that models with stochastic volatility in the idiosyncratic errors tend to produce smoother cyclical movements.
Warren: So while structural complexity might give you a good statistical fit, smoothness in transitions seems to be what they prioritize for economic plausibility here.
Iris: The model incorporating stochastic volatility in the factor and heterogeneous dynamics strikes a balance between flexibility and interpretability because it yields more plausible estimates.
Ruth: Ultimately, the paper develops this High-Frequency Economic Index using a Bayesian dynamic factor model estimated on mixed-frequency data to provide real time monitoring in Ecuador.
Warren: It's very practical for short term monitoring, especially since the framework can be extended to other emerging markets as more high frequency administrative data becomes available.
Iris: The key contribution is this new index that captures nuances beyond what is embedded solely in GDP alone.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19