Daily Summary for 2026-11-08
daily
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: It's the eighth of November, twenty twenty-six, and this is the day's research.
Warren: 10 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 Debiased Conditional Independence Tests for Causal Effects.
The day's research: Ruth: We are testing conditional independence in high dimensions where selection matters most because understanding this helps us determine if treatment assignment D i Y i X 2 can be predicted from outcomes better than just looking at covariates alone.
Warren: The core idea is checking if the distributions of potential outcomes are equal under treated and control states, which boils down to checking if treatment assignment D i Y i X 2 is independent almost surely.
Iris: We have three main ways to test this conditional independence: a conditional randomization test using an inverse propensity weighted mean squared error statistic assuming known propensity scores.
Ruth: And a double-residualization approach that checks if the correlation between residuals from both treatment and outcome predictions is zero.
Warren: And a test based on an augmented inverse propensity weighted estimator built around doubly robust population moments.
Iris: These methods inherit desirable large-sample properties like Neyman orthogonality, making them computationally efficient for machine learning applications.
Ruth: When propensity scores need to be estimated from observational data, we construct specific statistics called Augmented inverse propensity weighting AIPW, double residualization DR, and double residualization with augmentation DR.
Warren: The results show that AIPW and DR tests are systematically more powerful than the conventional joint Wald test when the signal is low.
Iris: They also outperform the Wald test even when covariate-driven heterogeneity or nonzero mean effects are present.
Ruth: Furthermore, these methods remain consistent for the true conditional treatment effect even if our outcome model is misspecified.
Warren: AIPW and DR tests achieve near 100% power across various outcome distributions with as few as five hundred observations.
Iris: However, small sample performance benchmarks indicate that DR/DR and AIPW dominate when the treatment effect averages to zero.
Ruth: While the joint Wald test remains superior when E tau i X is not equal to zero.
Warren: The most significant development today concerns bounding the difference between finding undominated mixed strategies and finding undominated pure strategies in compact and continuous games.
Iris: This is addressed by Iterated Elimination of Dominated Mixed Strategies.
Ruth: This matters because it sets a tight limit on how much more work is needed when players use stochastic choices compared to purely deterministic ones, suggesting a potential large gap in the inference required for mixed strategies.
Warren: The core finding shows that if the pure-strategy elimination procedure stops at round M, the mixed-strategy procedure will either stop at M or M plus one round.
Iris: This means if you find all undominated pure strategies in M steps, you only need to check one more step for the mixed strategies.
Ruth: This relationship is further clarified by introducing grandfather dominance, which shows that iterated dominance based on mixed strategies is equivalent to iterated grandfather dominance.
Warren: Furthermore, an alternative procedure called iterated support dominance is shown to be round-for-round equivalent to both of these concepts.
Iris: The computational utility stems from this equivalence because it allows for finding m-undominated mixed strategies by searching only against m minus one-undominated pure strategies.
Ruth: This suggests a more targeted way to locate undominated mixed strategies than relying on the entire history of previously undominated sets.
Warren: The paper also establishes that if the procedure based on mixed strategies terminates in finitely many rounds, then so does the procedure based on pure strategies, implying they are equivalent in terms of termination.
Iris: This means one method stops if and only if the other does.
Ruth: Debiased Conditional Independence Tests for Causal Effects This paper proposes robust tests to measure conditional independence between outcomes and treatment assignment in high-dimensional settings.
Warren: Organizing Intelligence Agents, whether human or artificial intelligence AI, sort endogenously into competing firms, and a firm's productivity is contingent upon its agents' capabilities.
Iris: Bang-Bang Correlation Sensitivity: A Local-to-Global Dichotomy for Correlation Preference This paper establishes that correlation insensitivity in binary choice under risk is either void or global.
Ruth: Complex Menus in Discriminatory Pricing Optimal menus are generically infinite in multi-dimensional screening problems when complexity costs are present but small.
Warren: Universal Nonexistence of Monotone Equilibrium in Multiunit Auctions with Risk Aversion This paper shows that even a single strictly risk-averse bidder can rule out the existence of an admissible monotone equilibrium in multiunit auctions.
Iris: Aggregating treatment effects across multiple outcomes Researchers critically analyze common methods for aggregating related outcomes, finding a novel decision-theoretic framework for selecting interpretable weights.
Ruth: Nonparametric Subgraph Regression Estimation and inference on a single large network This paper develops kernel regression estimators to estimate outcomes defined on groups of agents within a network structure.
Warren: Bankruptcy Rules for Strategic Agents This paper identifies optimal bankruptcy rules that maximize social welfare subject to claimants truthfully reporting their claims.
Iris: Iterated Elimination of Dominated Mixed Strategies This note compares iterated dominance procedures and shows that the mixed-strategy procedure can trail the pure-strategy procedure by at most one round.
Ruth: On the Theory of Multidimensional Stationary Random Processes 1941 This edition provides a translation and commentary on a multivariate extension of Wold's decomposition for stationary random processes.
Iris: Bankruptcy Rules for Strategic Agents This paper identifies optimal bankruptcy rules that maximize social welfare subject to claimants truthfully reporting their claims.
Ruth: On the Theory of Multidimensional Stationary Random Processes 1941 This edition provides a translation and commentary on a multivariate extension of Wold's decomposition for stationary random processes.
Warren: Organizing Intelligence Agents, whether human or artificial intelligence AI, sort endogenously into competing firms, and a firm's productivity is contingent upon its agents' capabilities.
Iris: Complex Menus in Discriminatory Pricing Optimal menus are generically infinite in multi-dimensional screening problems when complexity costs are present but small.
Ruth: Universal Nonexistence of Monotone Equilibrium in Multiunit Auctions with Risk Aversion This paper shows that even a single strictly risk-averse bidder can rule out the existence of an admissible monotone equilibrium in multiunit auctions.
Warren: Aggregating treatment effects across multiple outcomes Researchers critically analyze common methods for aggregating related outcomes, finding a novel decision-theoretic framework for selecting interpretable weights.
Iris: Nonparametric Subgraph Regression Estimation and inference on a single large network This paper develops kernel regression estimators to estimate outcomes defined on groups of agents within a network structure.
Ruth: Bankruptcy Rules for Strategic Agents This paper identifies optimal bankruptcy rules that maximize social welfare subject to claimants truthfully reporting their claims.
Warren: Iterated Elimination of Dominated Mixed Strategies This note compares iterated dominance procedures and shows that the mixed-strategy procedure can trail the pure-strategy procedure by at most one round.
Iris: On the Theory of Multidimensional Stationary Random Processes 1941 This edition provides a translation and commentary on a multivariate extension of Wold's decomposition for stationary random processes.
Ruth: Debiased Conditional Independence Tests for Causal Effects This paper proposes robust tests to measure conditional independence between outcomes and treatment assignment in high-dimensional settings.
Warren: Organizing Intelligence Agents, whether human or artificial intelligence AI, sort endogenously into competing firms, and a firm's productivity is contingent upon its agents' capabilities.
Iris: Bang-Bang Correlation Sensitivity: A Local-to-Global Dichotomy for Correlation Preference This paper establishes that correlation insensitivity in binary choice under risk is either void or global.
Ruth: Complex Menus in Discriminatory Pricing Optimal menus are generically infinite in multi-dimensional screening problems when complexity costs are present but small.
Warren: Universal Nonexistence of Monotone Equilibrium in Multiunit Auctions with Risk Aversion This paper shows that even a single strictly risk-averse bidder can rule out the existence of an admissible monotone equilibrium in multiunit auctions.
Iris: Aggregating treatment effects across multiple outcomes Researchers critically analyze common methods for aggregating related outcomes, finding a novel decision-theoretic framework for selecting interpretable weights.
Ruth: Nonparametric Subgraph Regression Estimation and inference on a single large network This paper develops kernel regression estimators to estimate outcomes defined on groups of agents within a network structure.
Warren: Bankruptcy Rules for Strategic Agents This paper identifies optimal bankruptcy rules that maximize social welfare subject to claimants truthfully reporting their claims.
Iris: Iterated Elimination of Dominated Mixed Strategies This note compares iterated dominance procedures and shows that the mixed-strategy procedure can trail the pure-strategy procedure by at most one round.
Ruth: On the Theory of Multidimensional Stationary Random Processes 1941 This edition provides a translation and commentary on a multivariate extension of Wold's decomposition for stationary random processes.
Warren: Debiased Conditional Independence Tests for Causal Effects This paper proposes robust tests to measure conditional independence between outcomes and treatment assignment in high-dimensional settings.
Iris: Organizing Intelligence Agents, whether human or artificial intelligence AI, sort endogenously into competing firms, and a firm's productivity is contingent upon its agents' capabilities.
Ruth: Bang-Bang Correlation Sensitivity: A Local-to-Global Dichotomy for Correlation Preference This paper establishes that correlation insensitivity in binary choice under risk is either void or global.
Warren: Complex Menus in Discriminatory Pricing Optimal menus are generically infinite in multi-dimensional screening problems when complexity costs are present but small.
Iris: Universal Nonexistence of Monotone Equilibrium in Multiunit Auctions with Risk Aversion This paper shows that even a single strictly risk-averse bidder can rule out the existence of an admissible monotone equilibrium in multiunit auctions.
Ruth: Aggregating treatment effects across multiple outcomes Researchers critically analyze common methods for aggregating related outcomes, finding a novel decision-theoretic framework for selecting interpretable weights.
Warren: Nonparametric Subgraph Regression Estimation and inference on a single large network This paper develops kernel regression estimators to estimate outcomes defined on groups of agents within a network structure.
Iris: Bankruptcy Rules for Strategic Agents This paper identifies optimal bankruptcy rules that maximize social welfare subject to claimants truthfully reporting their claims.
Ruth: Iterated Elimination of Dominated Mixed Strategies This note compares iterated dominance procedures and shows that the mixed-strategy procedure can trail the pure-strategy procedure by at most one round.
Warren: On the Theory of Multidimensional Stationary Random Processes 1941 This edition provides a translation and commentary on a multivariate extension of Wold's decomposition for stationary random processes.
Lucky paper: 2611.05608: Ruth: Let's look at Debiased Conditional Independence Tests for Causal Effects now. The authors are Susan Athey, Maria Nareklishvili, and Jann Spiess from Stanford and Glasgow respectively.
Warren: The paper focuses on robust tests for causal effects in potentially high-dimensional outcomes where selection into treatment might occur.
Iris: They specifically design these tests to measure conditional independence between outcomes and treatment assignment by checking if the treatment can be predicted from outcomes better than just looking at covariates alone.
Ruth: So the core null hypothesis is that the distributions of potential outcomes are equal under treated and control states, which means testing if the treatment assignment D i Y i X two is independent almost surely.
Warren: That's equivalent to checking no predictability: P(D i equals one Y i conditional on X) is equal to P(D i equals one X).
Iris: They introduce three primary procedures for this test: first, a conditional randomization test based on an inverse propensity weighted mean squared error statistic assuming propensity scores are known.
Ruth: And then you have a double-residualization approach that involves partial out covariate effects from both treatment and outcome-based treatment predictions and checks whether the correlation between those residuals is zero.
Warren: The third method they propose is a test based on an augmented inverse propensity weighted AIPW estimator built around doubly robust, Neyman-orthogonal population moments necessary for conditional independence.
Iris: These proposed measures are designed to test for the joint distributional causal effect of treatment on outcomes under unconfoundedness, which is Assumption one in the paper.
Ruth: They inherit desirable large-sample properties like Neyman orthogonality and double robustness under the null hypothesis, which makes them computationally efficient for machine learning applications.
Warren: When propensity scores need to be estimated from observational data, they construct three specific statistics: Augmented inverse propensity weighting AIPW, double residualization DR, and double residualization with augmentation DR.
Iris: These statistics are based on efficient Neyman-orthogonal doubly robust scores that measure the predictability of treatment D i from outcomes Y i conditional on covariates X i.
Ruth: The paper shows that AIPW and DR tests are systematically more powerful than the conventional joint Wald test in low-signal regimes.
Warren: They also outperform the joint Wald test even when covariate-driven heterogeneity or nonzero mean effects are present in the data.
Iris: Crucially, these methods remain consistent for the true conditional treatment effect even under misspecification of the outcome model.
Ruth: The AIPW requires known propensity scores, but DR and DR accommodate observational settings where propensity scores must be estimated.
Warren: The analysis details convergence properties: Part one shows that theta hat AIPW, theta f0, theta hat DR, theta f0 under Assumption four.
Iris: Part two establishes the consistency of the variance estimators showing that the estimated variances converge in probability to the asymptotic variances derived from Theorem five.
Ruth: This convergence depends on terms involving nuisance function errors that vanish under Assumption four.
Warren: Part three proves Central Limit Theorems for the test statistics T k n, where for AIPW and DR, sqrt n theta hat k minus theta k f N zero, sigma squared k plus delta squared.
Iris: Under the null hypothesis, theta f0 equals zero almost surely, which leads to T k n N zero one for all three tests.
Ruth: That means every test statistic converges to a standard normal distribution under the null hypothesis.
Warren: Part four confirms asymptotic size alpha under the null hypothesis.
Iris: Under the alternative, if the true outcome model is well-specified, consistency is achieved when theta f0 is greater than zero and theta f0 is greater than zero.
Ruth: If the true model is misspecified, consistency requires those parameters to remain positive as well.
Warren: Performance comparisons show that AIPW, DR, and DR tests achieve consistently high power across various outcome distributions like Normal and Mixtures.
Iris: They reach near one hundred percent power even with as few as five hundred observations, whereas the Wald test remains below twenty percent in complex designs like Mixtures and Heavy Tails.
Ruth: Small sample performance benchmarks indicate that while DR/DR and AIPW dominate when the treatment effect averages to zero, KCI, GCM, and the joint Wald test dominate when E tau i X is not equal to zero.
Warren: The main development they highlight is bounding the difference between finding undominated mixed strategies and finding undominated pure strategies in compact and continuous games.
Iris: This is addressed by Iterated Elimination of Dominated Mixed Strategies, which sets a tight limit on how much more work is needed when players use stochastic choices compared to purely deterministic ones.
Ruth: The core finding shows that if the pure-strategy elimination procedure stops at round M, the mixed-strategy procedure will either stop at M or M plus one round.
Warren: This implies that if you find all undominated pure strategies in M steps, you only need to check one more step for the mixed strategies.
Iris: This relationship is further clarified by introducing grandfather dominance, which shows that iterated dominance based on mixed strategies is equivalent to iterated grandfather dominance.
Ruth: Furthermore, an alternative procedure called iterated support dominance is shown to be round-for-round equivalent to both of these concepts.
Warren: The computational utility stems from this equivalence because it allows for finding m-undominated mixed strategies by searching only against m minus one-undominated pure strategies instead of the much larger set of all m minus one undominated mixed strategies.
Iris: This suggests a more targeted way to locate undominated mixed strategies than relying on the entire history of previously undominated sets.
Ruth: The paper also establishes that if the procedure based on mixed strategies terminates in finitely many rounds, then so does the procedure based on pure strategies, implying they are equivalent in terms of termination.
Warren: This means one method stops if and only if the other does.
Lucky paper: 2611.07128: Ruth: So we're looking at Complex Menus in Discriminatory Pricing and the paper suggests optimal menus are generically infinite when complexity costs are small.
Warren: That means even if a seller has a very simple contract, like a two-part tariff, they might need an infinite menu because of how selection works.
Iris: It's interesting that this complexity comes from screening considerations alone and not from non-linearities in production costs or preferences.
Ruth: The paper specifically says that for generic type distributions, revenue maximizing menus cannot be achieved by simple screening contracts like two-part tariffs.
Warren: And it goes further, implying that the necessity of infinite menus means optimal mechanisms must divide the market into infinitely many segments.
Iris: Then we have this section about complexity costs, which are defined as costs that decrease menu size but grow without bound as those costs decrease.
Ruth: The results show that optimal menus are generically infinite if those complexity costs remain small.
Warren: That leads to an implication that simple pricing schemes we see in practice might actually be driven by relatively large complexity costs instead of the absence of them.
Iris: The paper explores four reasons why adding new options might be costly or difficult, including setup costs and buyer navigation costs.
Ruth: Specifically for generic type distributions, the optimal menu size grows without bound as those frictions vanish because that's what technological progress does.
Warren: Technological progress in information technology is expected to lower those complexity costs, which in turn increases the extent to which sellers price discriminate through vertical differentiation.
Iris: Cognitive costs can also increase the optimal menu size because a seller has to leave every buyer who inspects the menu enough expected surplus to cover that increasing cognitive cost.
Ruth: Theorem one proves that for every number k, the set of type distributions for which every optimal mechanism has a menu size greater than k is strongly generic.
Warren: That means for generic type distributions, every optimal mechanism has an infinite menu size.
Iris: Corollary five shows that for generic type distributions, the optimal menu size along any sequence of allocation spaces tends to infinity as the mesh of those spaces goes to zero.
Ruth: This result really hammers home how robust this finding is across different mathematical settings, which is a big part of what makes this paper so interesting.
Warren: It shows that even if we change the way we measure things slightly, like changing the allocation space, the menu size keeps going up without limit.
Iris: The paper also discusses friction effects on menu size, looking at setup costs and technological constraints as reasons why adding new options might be costly or difficult.
Ruth: It seems like this research connects theoretical screening models directly to real-world pricing decisions in a very concrete way.
Warren: I wonder if this finding about infinite menus really changes how we think about how firms set prices when they have many product choices available.
Lucky paper: 2611.06644: Ruth: We're looking at the Universal Nonexistence of Monotone Equilibrium in Multiunit Auctions with Risk Aversion from Wei He, Yeneng Sun, and Hanping Xu.
Warren: This paper shows that even one strictly risk-averse bidder can rule out a monotone pure-strategy equilibrium in both uniform and discriminatory auctions.
Iris: Specifically, under uniform pricing for any number of bidders n at least two and supply m at least three there are independent type distributions where neither the highest-losing-bid nor the lowest-winning-bid pricing rules admit an admissible monotone equilibrium.
Ruth: That means even when you have three units or more, you can find scenarios where those standard pricing rules fail to support a monotone outcome.
Warren: And under discriminatory pricing, Theorem two proves universal nonexistence for any number of bidders n at least two and some integer m at least three.
Iris: The paper notes that this result holds even without admissibility or no-overbidding restrictions, which is quite broad.
Ruth: So the findings apply to any profile of strictly increasing, concave utilities as long as at least one bidder is strictly risk-averse.
Warren: That's interesting because it suggests that when bidders are risk-averse, they have to allow for bidding behavior that isn't monotone in marginal values.
Iris: The paper also looked at the two-unit case, where the two-unit case with CARA utilities still admits a monotone no-overbidding pure-strategy equilibrium for any number of bidders and type distributions.
Ruth: But when you introduce a third unit, the utility cost associated with paying more for the first unit stays unchanged while the additional surplus from that third unit yields less utility, which potentially reverses bidding incentives.
Warren: That sounds like a structural change in how the surplus is distributed across those units, making monotonicity break down.
Iris: They also examined all-pay auctions and finite bid grids, demonstrating nonexistence for continuous-bid all-pay auctions on specific sets of bids.
Ruth: Proposition nine shows that there are type distributions and bid grids where the all-pay auction has no pure equilibrium whose third bid for bidder one is coordinatewise nondecreasing on a full measure set.
Warren: It’s important to remember that the analysis for finite bid grids confirms this nonexistence under uniform pricing and discriminatory pricing when the integer r is sufficiently large.
Iris: Lemma eight further shows that if certain grid conditions are met, the continuous-bid auction must also have a monotone pure-strategy equilibrium, which contrasts with those nonexistence results.
Ruth: So they're showing this conflict between continuous and discrete settings depending on the specific grid conditions used in the analysis of Universal Nonexistence of Monotone Equilibrium in Multiunit Auctions with Risk Aversion.
Lucky paper: 2611.07240: Ruth: Alright team, let's get into paper number four of four which is about Nonparametric Subgraph Regression by Olivia Ding from Yale University.
Warren: This paper tackles the challenge of estimation and inference on a single large network because there's dependence among observations.
Iris: It studies nonparametric regression for outcomes defined on groups of p agents, like the indicator that three agents in a network form a triangle.
Ruth: The main application is subgraphs of a network, but the methods apply to any outcome defined on groups of agents.
Warren: The setup involves observing a vector of covariates Xi for N agents and an ordered p-tuple Yip for every ordered p-tuple ip equals i1,..., ip.
Iris: The object they are interested in is the regression function m of w equals E of Yip Xip equals w, where Xip stacks the covariates of the p agents.
Ruth: Outcomes on disjoint tuples are independent, but those on tuples that share an agent are generally dependent through the characteristics of the agents they share, which is called polyadic dependence.
Warren: That dependence complicates estimation and inference in two ways: first, N factorial over N minus p factorial observed outcomes contain far less information than an independent sample of the same size.
Iris: And second, standard tools for nonparametric regression with dependent data do not apply.
Ruth: The authors study the subgraph analogue of the Nadaraya-Watson kernel regression estimator.
Warren: They establish a variance bound of order N h dX to N to the power of minus one for the kernel numerator.
Iris: Furthermore, they show that the kernel numerator converges uniformly in probability at a rate of ln N over N h dX to N raised to the power of minus one over two.
Ruth: That convergence region expands at a polynomial rate.
Warren: They also establish pointwise asymptotic normality of the estimator at the rate of N h dX to N raised to the power of minus one over two.
Iris: The results deliver two lessons extending Graham et al from dyads to groups of any size.
Ruth: First, the relevant sample size is the number of agents N, not the number of tuples which is order N p.
Warren: Second, the convergence rate depends on dX which is the dimension of a single agent’s covariates rather than on p dX.
Iris: This means the optimal uniform rate would be ln N over N raised to beta divided by two beta plus p dX, but they got a rate that is that of a dX dimensional problem.
Ruth: The paper contributes to three literatures: first, the theory of nonparametric estimation including optimal rates for weakly dependent observations.
Warren: Second, studies of estimation and inference with dyadic and exchangeable array data.
Iris: And third, uses subgraph frequencies to learn about network formation.
Ruth: This paper establishes a variance bound uniform convergence rates in probability and pointwise asymptotic normality for the subgraph analogue of the Nadaraya-Watson kernel regression estimator.
Warren: The leading application is the conditional probability that a group of agents forms a given subgraph, such as a triangle.
Iris: Two questions remain open regarding the minimax lower bound for p greater than or equal to three and consistent variance estimation.
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