Maximum score and non-smooth M-estimators with multiway dependence
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "Maximum score and non-smooth M-estimators with multiway dependence".
Warren: Here's the gist of the paper. The paper investigates the asymptotic properties of maximum score estimators in settings characterized by multiway dependence and non-smooth M-estimators, drawing upon a general theory for such estimators.
Ruth: Under i.i.d.
Warren: sampling, the maximum score estimator converges at a cube-root rate to a non-Gaussian limit.
Ruth: However, under multiway dependence and low-level conditions, it exhibits asymptotic normality at a parametric rate in the smallest number of clusters and achieves rates comparable to regular smooth parametric estimators when cluster-level variance is non-degenerate.
Title and authors: Ruth: The paper is titled "Maximum score and non-smooth M-estimators with multiway dependence." This is a pretty descriptive title for what they’re tackling, focusing on the type of estimator and the kind of data structure they are dealing with.
Warren: It tells you right away that this isn't just about basic regression anymore; it’s about estimators that aren't perfectly smooth, which is a major hurdle in applied econometrics. And the multiway dependence part suggests they are looking at situations where observations aren't totally independent across all dimensions.
Iris: I think what’s important is that they are connecting this to a general theory for non-smooth M-estimators, which gives them the tools to analyze these estimators systematically, rather than just guessing how they perform.
Ruth: Right. And then they explore how this general theory applies when you have multiway dependence and low-level conditions. It seems like the main goal is to establish some kind of stability or convergence rate under those more challenging data structures.
Warren: So, instead of just saying, "this estimator works here," they are trying to define the boundaries—what conditions are necessary for these estimators to behave in a predictable way when things get complicated.
Iris: That’s the econometrician’s angle. She wants to know what variation is actually doing the work here. Is it some specific structure in the dependence that helps, or does it just make everything more complicated?
Ruth: And they seem to be finding that under certain conditions, like when cluster-level variance is non-degenerate, these flexible estimators can converge at rates comparable to regular smooth parametric estimators.
The paper's summary: Warren: So, the summary of the paper seems to boil down to this: under i.i.d. sampling, the maximum score estimator converges at a cube-root rate to a non-Gaussian limit, which is what we see in simpler models like Kim and Pollard's work on logit maximum likelihood convergence at rate n to something related to Chernoff limits.
Ruth: That cube-root rate under i.i.d. sampling is the baseline they start with, right? But when you introduce multiway dependence and low-level conditions, things get interesting because the estimator can exhibit asymptotic normality at a parametric rate in terms of the smallest number of clusters involved.
Iris: That sounds like they are finding a way to tame that cube-root rate under these specific dependence structures, moving towards something more familiar, like asymptotic normality. And it depends on how you define those clusters and what kind of conditions you have on the data.
Warren: And when the cluster-level variance is non-degenerate—meaning that variance isn't zero or degenerate in some way—they find that the flexible semiparametric model doesn't lose its rate relative to regular smooth parametric models. The estimation errors are of the same order, even if it means the flexibility only affects how fast it converges asymptotically.
Ruth: That’s a key point: under multiway dependence and those non-degenerate variances, the flexibility of this estimator doesn't hurt its convergence speed compared to something simpler, like probit. It just changes the asymptotic variance part.
Iris: So they are showing that you can still get good rates even with this extra flexibility, as long as the underlying structure of the data allows for it to be well-behaved asymptotically.
The paper's improvements: Warren: Now, looking at what they suggest for improvements, they introduce a modified bootstrap procedure that is first-order valid given a consistent estimate of the Hessian. This is important because it gives them a way to do some kind of inference—a bootstrap—that actually works reliably under these non-smooth conditions.
Ruth: That makes sense from an econometric perspective. If the standard methods aren't perfectly valid because of the non-smoothness, you need a better way to estimate the uncertainty around your estimates, and this modified bootstrap provides that consistency when you have a good Hessian estimate.
Iris: And they show that for two-way clustering specifically, even with this new procedure, a cube-root term can still linger. But they then introduce high-level conditions where symmetric bootstrap intervals maintain uniform control over size for certain scalar parameters.
Warren: So they’re giving us a way to handle the lingering cube-root term by controlling the size of the intervals uniformly across different parameter settings, which is crucial when you're building confidence regions.
Ruth: And then they tackle that for K=two where there are two dimensions, and they find that a second term competes with the Gaussian one <ref:2604.10232#pg2>. They show how this competition can be managed through their modified bootstrap, which inflates that cube-root term by a factor of twenty-one/three <ref:2604.10232#pg3>.
Iris: Inflating it rather than mis-centering it is important because it means the inequality used in Anderson's work, Lemma one makes the intervals conservative instead of invalid <ref:2604.10232#pg1>. That’s a nice piece of inference theory they are leveraging.
Conclusion: Ruth: To wrap up, the paper "Maximum score and non-smooth M-estimators with multiway dependence" shows that under i.i.d. sampling, we get a cube-root rate to a non-Gaussian limit, but under multiway dependence and certain variance conditions, we can achieve asymptotic normality at a parametric rate in terms of the smallest number of clusters.
Warren: And when the cluster variance is non-degenerate, these flexible models keep their convergence speed on par with regular smooth parametric models, just with an adjusted asymptotic variance. That means the flexibility doesn't slow things down asymptotically under those conditions.
Iris: And they provide a modified bootstrap method that’s first-order valid if you have a consistent Hessian estimate, which helps in getting reliable inference even when the estimators are non-smooth. They also show size control for scalar parameters, meaning you can get accurate confidence intervals along sequences where the parameter difference is of the same order as the estimate itself.
Ruth: So, overall, this paper gives us tools to analyze maximum score estimators in complex data settings and provides a way to use modified bootstrap intervals that are asymptotically exact under certain conditions. It's a solid piece for understanding how these models behave when we move beyond simple assumptions.
Warren: It’s definitely a deep dive into the mechanics of these estimators, showing how the rates separate—like that cube-root term in two-way clustering—but ultimately proving that with high-level conditions, you can still get uniform control over interval sizes.
Iris: And we should remember their limitation: they are dealing with flexible semiparametric models, and while they manage the asymptotic variance well under non-degenerate cluster variances, the exact nature of the convergence in the presence of multiway dependence is still quite intricate.
Ruth: Exactly. They show us that even when things get complicated, we can still make progress by using modified bootstrap methods to ensure our inferences are robust and controlled. We’re looking forward to seeing how this framework helps us model more complex economic situations next.
econ.EM, math.ST, stat.TH
Submitted: 2026-04-11
Updated: 2026-11-12
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: The paper investigates the asymptotic properties of maximum score estimators in settings characterized by multiway dependence and non-smooth M-estimators…
Key concepts
- Maximum Score Estimator
- This is a method used to estimate parameters in complex models. It involves finding the maximum value of a score function derived from the model's structure. The paper analyzes its convergence rates and asymptotic distribution under different dependence structures, showing it can be asymptotically normal.
- Multiway Dependence
- This refers to situations where data points are not independent but exhibit complex relationships across multiple dimensions or clusters. The paper investigates how this type of dependence affects the estimator's performance, distinguishing between i.i.d. sampling and clustered settings.
- Modified Bootstrap Procedure
- This is a statistical resampling technique used to estimate the sampling distribution of an estimator without knowing its exact theoretical distribution. The paper shows this modified bootstrap can reproduce the Gaussian limit exactly when cluster-level variance is positive definite, ensuring accurate inference.
- Cube-root Rate Convergence
- This describes a specific, slower rate at which an estimator converges to its true value when sampling independently (i.i.d.). The paper contrasts this with the faster $n^{1/2}$ rate achieved under multiway dependence, highlighting how dependence structures change the convergence speed.
Terminology
Summary
The paper investigates the asymptotic properties of maximum score estimators in settings characterized by multiway dependence and non-smooth M-estimators, drawing upon a general theory for such estimators. p. 4
Under i.i.d. sampling, the maximum score estimator converges at a cube-root rate to a non-Gaussian limit. p. 0 However, under multiway dependence and low-level conditions, it exhibits asymptotic normality at a parametric rate in the smallest number of clusters and achieves rates comparable to regular smooth parametric estimators when cluster-level variance is non-degenerate. p. 0 The result stems from the general theory for non-smooth M-estimators, which allows for a modified bootstrap to be first-order valid given a consistent Hessian estimate. p. 0 For two-way clustering, a cube-root term may persist; under high-level conditions, symmetric bootstrap intervals maintain uniform control over size for suitable scalar parameters. p. 0
The maximum score estimator is asymptotically normal at rate n 1/2, where n is the smallest number of clusters in any dimension, provided the cluster-level variance sigma of the scores is positive definite. p. 0 When multiway dependence and non-degenerate cluster-level variances (sigma 0) are present, the flexible semiparametric model of maximum score does not lose its rate of convergence relative to regular smooth parametric models; specifically, estimation errors are of the same order, and flexibility only affects the asymptotic variance. p. 1
For K=2, a second term competes with the Gaussian one. p. 1 No cluster shock smooths the cell-level component of the criterion, introducing a cube-root term of order (N 1N 2)-1/3 for an N 1 times N 2 array. p. 1 The convergence rate separates: when row and column scores vanish, maximum score converges at (N 1N 2)-1/3 under Theorem 3 conditions, while a sample mean whose row and column components also vanish, but whose cell-level component has positive variance, converges at the rate (N 1N 2)-1/2. p. 1
The modified bootstrap with multiplier weights exactly reproduces the Gaussian term but inflates the cube-root term by 21/3. p. 2 Theorem 5 demonstrates that for a scalar parameter theta 0, symmetric bootstrap intervals achieve asymptotic coverage of at least 1-alpha along sequences where a'(theta bn-theta 0) is of the same order as theta bn-theta 0.
The asymptotic normality under multiway dependence, established by Proposition 1, yields: n 1/2(theta bn - theta 0) N(0, V), where V = A-1 10 sigma A-1 10, sigma = P K k=1 lambda k E[delta k delta'k], and A 0 0 is given in (B.8). The bootstrap procedure yields: n 1/2(theta e*n - theta bn) p N(0, V) conditionally given W n. p. 3
In the non-standard asymptotics for K=2, Theorem 3 approximates the law of the estimator as that of a sum of a Gaussian vector and an independent Chernoff-type vector, regardless of which term dominates. p. 2 Theorem 4 confirms that the modified bootstrap reproduces the Gaussian term exactly but inflates the cube-root term by kappa 1/3 = 21/3 about 1. p. 226. Theorem 5 establishes size control for a scalar parameter irrespective of degeneracy, showing intervals are asymptotically exact when non-degenerate in direction a and conservative when n c n,a > 0. p. 2
The asymptotic distribution in the non-degenerate case is described by Theorem 3: r n(theta bn - theta 0) A-1 10G + Z, with G about N(0, omega) independent of Z. p. 0 The bootstrap reproduces this: r n(theta e*n - theta bn) p A-1 10G + kappa 1/3 Z conditionally given W n, with kappa = 2. p. 1
The convergence of the scaled estimator is described by Lemma 18: (a) Under Theorem 3 conditions, theta bn - theta 0 = O P(s n) and theta bn - theta 0 - A-1 10 psi n = o P(s n c 3/4). p. 0 For K=2, the unsmoothed top-level component adds a term | theta bn 1/2I n-1/2 to the bound, yielding the smaller term of order I n-1/3 studied in Section 2.3.
The asymptotic distribution in the non-degenerate case is described by Theorem 3: r n psi n, W n (G,W) in R d times l infinity(K), and delta 0 n to infinity E u,u' in K |u-u'|W n(u) - W n(u') = 0. The convergence of the scaled estimator is described by Lemma 19: (b) If weights are of the multiplier type in Assumption 4, then r n delta e* n, W* n p (G, kappa 1/2W) conditionally given U n, with kappa = 2. p. 8 Size control is established by Theorem 5: intervals are asymptotically exact when non-degenerate in direction a and conservative when n c n,a > 0. p. 2
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved asymptotic normality of maximum score estimators under multiway dependence by showing convergence to a limit that is a convolution of Gaussian and Chernoff-type laws for K=2, which is more accurate than the cube-root rate convergence under i.i.d. sampling where it converges at rate N 1/3.
-
Enhanced statistical inference via the modified bootstrap procedure, which is shown to reproduce the Gaussian limit exactly when cluster-level variance Σ is positive definite, providing asymptotically exact symmetric intervals even in regimes where the cube-root term remains significant for K=2.
-
Development of non-standard asymptotics that control size uniformly across all regimes (fixed Σ > 0 to degenerate cases), which allows for reliable confidence interval construction using symmetric bootstrap intervals when the parameter difference is of the same order as the parameter estimate itself.
-
Implementation of a modified bootstrap procedure using multiplier weights, which inflates the cube-root term by a factor of 2/3, ensuring that comparison inequalities remain conservative rather than invalid, thus providing robust coverage guarantees in complex dependence structures.
-
Integration of local VC-type conditions and maximal inequalities for the top-level component to prove size control for scalar parameters regardless of degeneracy, leading to asymptotically exact intervals when the Gaussian term vanishes in a specific direction.
Sources
- Cross-Fitting-Free Debiased Machine Learning with Multiway Dependence
- Analytic inference with two-way clustering
Related papers
- Stationarity with Occasionally Binding Constraints
- Backtesting High-Dimensional Expected Shortfall
- Higher-Order Neyman Orthogonality in Moment-Condition Models
- Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models
- Debiased Conditional Independence Tests for Causal Effects
- Sharp Testable Implications of Encouragement Designs