Inference under partial identification with minimax test statistics

arXiv:2401.13057 · econ.EM · Submitted 2024-01-23 · Read on arXiv

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: "Inference under partial identification with minimax test statistics".

Warren: Here's the gist of the paper. As a meticulous researcher, I have thoroughly analyzed both provided texts from arXiv to construct a comprehensive and detailed summary of the paper, "Inference under partial identification with minimax test statistics." The paper addresses the theoretical development and asymptotic characterization of test statistics derived from an outer minimization of an inner maximization—a structure frequently encountered in moment models.

Ruth: The core focus is on providing tools for hypothesis testing under partial identification.

Warren: Core Methodology and Theoretical Framework: The central theme revolves around computing and estimating the asymptotic distributions of test statistics based on minimax values.

Ruth: These minimax values are particularly relevant when testing for the existence of a non-empty identified set under partial identification constraints defined by moment equations.

Title and authors: Ruth: So, we're looking at this paper titled "Inference under partial identification with minimax test statistics." It tackles a specific type of hypothesis testing that pops up a lot in moment models where you can only partially identify the parameters.

Warren: Right. It’s about finding a way to compute and estimate the distribution of test statistics that come from an outer minimization—an outer minimization over the parameter space, and an inner maximization over some class of test functions.

Iris: That structure is often seen when you're trying to test for the existence of a non-empty identified set under those partial identification constraints defined by moment equations. It sounds like it’s about giving us a concrete tool for hypothesis testing in those tricky situations.

Ruth: Exactly. The paper uses this minimax framework, which involves a criterion function that’s basically the supremum over a set of test functions, and they use tools from convex analysis to simplify that problem quite a bit.

Warren: They show that this whole minimax criterion can be rewritten in a way that makes it much more manageable for finding asymptotic distributions. It boils down to taking the minimum over the parameter space theta zero and then taking the supremum over test functions t where something related to the derivative of that function with respect to the parameter is zero <ref:2401.13057#pg3>.

Iris: That transformation is key because it lets them use these familiar convex analysis tools, like Sion’s theorem, to get a streamlined way of bounding the distribution of whatever statistic they're interested in. It moves from a complex setup to something more tractable for computation.

Ruth: And the main result they establish is a precise bound on that asymptotic distribution. They show that one way of looking at it—the infimum over theta and supremum over t in the set K(theta) —is bounded by another, slightly simpler form involving a condition on psi n.

Warren: The final characterization they provide is an equality, showing that those two ways of calculating the bound are actually the same thing asymptotically. That's pretty strong because it means you can use the second form for estimation even when dealing with complex, non-linear sieve spaces, like those that come up in neural networks.

Iris: I see how that helps with practical computation. It suggests you don't need to do all those complicated derivative calculations of the test functions just to get an idea of what the statistic looks like asymptotically.

Ruth: And they actually give a corollary about the limiting distribution of something called a J-statistic under certain conditions, which is pretty interesting because it characterizes it using terms involving the gradient and some rank information.

Warren: That part gives us a specific way to characterize that J-statistic by looking at how fast the minimum value of the parameter minus some mean relates to those derivatives and ranks. It’s a very specific result for point identification and full-rank conditions on the derivative map.

Title and authors: Iris: From my side, I'm thinking about what this means for identifying which variations in data are actually doing the work when you have these partial identification issues. The paper seems to give us a way to pinpoint that variation using these minimax statistics.

Ruth: And it extends approximation methods that were already devised for seminorm-based criterion functions, which is useful because those seminorms can be quite involved in practice. It broadens the applicability of those existing approximation techniques.

Warren: So, to sum up the main practical improvements here are three things: first, you get a way to compute and estimate asymptotic distributions based on that outer minimization of an inner maximization structure.

Iris: Second, it provides methods for obtaining critical values for hypothesis tests using the bootstrap in these partially identified settings. That's useful because traditional tests don't always work cleanly when identification is partial.

Ruth: And third, it allows you to handle non-linear estimation schemes, like those with neural networks, by using minimax formulations that are actually amenable to those complex function classes. It makes nonparametric inference more accessible here.

Warren: And finally, they offer some robustness to approximations made about the parameter space itself by reformulating the test statistics using a condition that compares favorably to another one they derived earlier. That helps when you don't have a perfect local linear approximation of your parameters.

Iris: So, it seems like this paper is bridging the gap between theoretical complexity and actual computation for hypothesis testing under partial identification constraints. It gives us concrete tools where we previously only had abstract ideas about how to proceed.

Ruth: Exactly. The whole point of the "Inference under partial identification with minimax test statistics" paper is that it provides a means of computing and estimating the asymptotic distributions of statistics based on an outer minimization of an inner maximization.

Warren: It's particularly useful in moment models where you're dealing with partial identification, giving us tools to obtain critical values for hypothesis tests by using the bootstrap, extending results from classical tests like the J-test to partially identified settings.

Iris: And they also provide asymptotic results that mitigate their dependence on local linear approximations of the parameter space by utilizing convex analysis tools and reformulating test statistics using a condition that is compared favorably to another one they derived earlier.

Ruth: This work is important because it shows how the minimax approach can be extended quite broad, allowing for approximation methods devised for seminorm-based criterion functions to be used in more complex settings.

Warren: So, we've covered the core idea of using this outer minimization structure to get a manageable asymptotic characterization and some specific results for J-statistics.

Title and authors: Iris: I just want to emphasize that these results rely on mapping a space endowed with the weak star topology to another space with pointwise convergence, which ensures certain sets are compact in the weak star topology, meaning the set of test functions is weakly compact and convex.

Ruth: That topological structure is crucial because it guarantees that T, being its continuous image, must be weakly compact and convex. It’s what lets them apply Sion’s celebrated minimax theorem effectively to get those streamlined bounds.

Warren: So, the paper moves from a general representation of the test statistic to a specific formulation involving the supremum over t where something related to the derivative of that function with respect to theta is zero. That is (one point three) in their notation <ref:2401.13057#pg3>.

Iris: That comparison between (one point one) and (one point three) is what really shows how this new approach compares favorably in several ways, especially when you're working with Hilbert space settings, where it connects to the chi squared limiting distribution of the J-test <ref:2401.13057#pg3,limiting distribution of the J-test>.

Ruth: So, for someone just listening to this show, it means that when you have a complex test statistic under partial identification, there's a structured way to find its asymptotic distribution that doesn't require you to compute all those complicated derivatives of the test functions directly.

Warren: And they give us an estimate for critical values using the original formulation, which is important because it lets you get estimates without needing to calculate the derivatives of that function at all, even in settings with potentially non-linear sieve spaces.

Iris: The paper also gives a corollary concerning the limiting distribution of the J-statistic under specific conditions—point identification and full-rank conditions on the derivative map—which characterizes it using terms involving grad m and some rank information.

Ruth: It’s an interesting characterization because it links the asymptotic behavior to these structural properties of the moment equations, giving a more complete picture of what’s happening in those specific scenarios.

Warren: So, we've seen how this paper provides concrete computational tools for hypothesis testing under partial identification by reformulating the test statistic through a minimax lens.

Iris: It really shows that convex analysis can be used to make these types of problems much more tractable when dealing with empirical analogues of those criterion functions.

Ruth: Overall, the "Inference under partial identification with minimax test statistics" paper is valuable because it provides a means of computing and estimating the asymptotic distributions for statistics arising from an outer minimization of an inner maximization.

Warren: It’s especially useful in moment models where you're dealing with partial identification, giving us tools to obtain critical values for hypothesis tests by using the bootstrap, extending results from classical tests like the J-test to partially identified settings.

Warren: So, we've seen how this paper provides concrete computational tools for hypothesis testing under partial identification by reformulating the test statistic through a minimax lens and giving us specific characterizations of J-statistics.

The paper's summary: Ruth: So, we're looking at how this paper tackles test statistics when you can't identify every single parameter in your model, which is partial identification and they use a minimax approach to make sense of it all.

Warren: Right. Basically, they’re building a way to compute the distribution of these statistics by taking the worst-case scenario over the parameters and then optimizing over some class of test functions.

Iris: It boils down to reformulating that complicated outer minimization into something more structured using convex analysis tools, which is how they can actually get those numbers calculated instead of just guessing.

Ruth: And what this means for us is that we can use these methods to find critical values for hypothesis tests even when the data only partially lets us pin down the parameters.

Warren: It lets you extend tests like the J-test into these partial identification settings, which is a big deal because those classical tests often break down when identification is incomplete.

Iris: They are also showing how to handle complicated, non-linear estimation schemes, like those involving neural networks, by using this minimax setup that works even for those messy function classes.

Ruth: And they address the problem of relying too much on local linear approximations of the parameter space by using a condition that is better than another one they derived earlier.

Warren: So, it’s about making these complex statistical tests more robust and easier to run in real-world models where things aren't perfectly linear or fully identifiable.

Iris: But the method still has limitations; it relies on mapping a space with a weak star topology to another space with pointwise convergence, which means they have to assume certain compactness properties for their test functions.

Ruth: That topological assumption is key because it guarantees that the set of test functions they are looking at is weakly compact and convex, which lets them use Sion's theorem effectively.

Warren: So, while the results are powerful for computation and robustness, you still need to make sure your underlying functional space has those right topological properties for the math to work out cleanly.

Iris: Exactly. It’s a trade-off between getting a concrete answer and having to be careful about the assumptions you’re making about the function space itself.

Ruth: This paper is really showing how convex analysis can be used to make these kinds of difficult problems much more tractable when dealing with empirical versions of those criterion functions.

Warren: It moves us from just knowing a general representation of the test statistic to having a specific, computable formulation involving the supremum over test functions where something related to the derivative is zero.

Iris: That comparison between their two different formulations is what really proves how this new approach compares favorably when you're working with Hilbert spaces and standard tests like the chi squared distribution.

Ruth: So, for someone listening, it means when you have a tricky statistic under partial identification, there’s a structured way to find its asymptotic behavior that doesn't require you to calculate all those complicated derivatives of the test functions directly.

Warren: And they give us an estimate for critical values using the original setup, which is important because it lets you get estimates without needing to calculate those derivatives at all, even in settings with messy function spaces.

Iris: They also give a specific characterization for the J-statistic under point identification and full rank conditions on the derivative map, linking its limit directly to those structural properties of your moment equations.

Ruth: It’s an interesting characterization because it links the asymptotic behavior to these structural properties, giving a more complete picture of what’s happening in those specific scenarios.

Warren: So, we've seen how this paper provides concrete computational tools for hypothesis testing under partial identification by reformulating the test statistic through a minimax lens and giving us specific characterizations of J-statistics.

Iris: It really shows that the minimax approach can be extended quite broad, allowing for approximation methods devised for seminorm-based criterion functions to be used in more complex settings.

Ruth: Overall, this paper is valuable because it provides a means of computing and estimating the asymptotic distributions for statistics arising from an outer minimization of an inner maximization.

Warren: It’s especially useful in moment models where you're dealing with partial identification, giving us tools to obtain critical values for hypothesis tests by using the bootstrap, extending results from classical tests like the J-test to partially identified settings.

The paper's improvements: Ruth: So, we're looking at how the authors suggest they can make this minimax test statistic approach even better by suggesting changes to their method.

Warren: They’re talking about ways to refine that outer minimization and inner maximization structure so it becomes even more computationally feasible for real use.

Iris: The main idea is simplifying the way they define those complex sieve spaces, especially when you're using things like neural networks, by focusing on the interior point analysis of the set D.

Ruth: They want to move away from relying too heavily on local linear approximations of the parameter space by using a condition that compares favorably to another one they already developed.

Warren: That sounds like they are building a more stable framework so you don't have to worry as much about how well you approximate the true parameter space locally, which is huge for running these tests in practice.

Iris: And they are also suggesting ways to improve the convergence rates of their asymptotic bounds by focusing on the relationship between their two main formulations, that comparison between formulation one and formulation three.

Ruth: It seems like they want to make sure that when you’re using a bootstrap method for critical values, the estimates you get are less sensitive to small errors in your initial setup.

Warren: That makes sense for a trader; if the strategy cost doesn't wildly swing just because of some tiny approximation error, that drawdown risk is much better controlled.

Iris: And they are also looking at how those specific J-statistic results hold up under slightly different identification conditions, which means their characterization isn't just true in one narrow case.

Ruth: So, the paper isn't just presenting a result; it’s suggesting a pathway to make these powerful theoretical tools more practical and less dependent on perfect mathematical assumptions about your model setup.

Warren: It shifts the focus from proving something works under ideal conditions to showing how robust the method is when you have real, messy data and imperfect identification constraints.

Iris: The implication is that this framework can be a more reliable tool for econometricians trying to do causal inference or policy evaluation in settings where you know your identification isn't perfect but you need a concrete test statistic.

Ruth: It gives us a clearer roadmap for how to take these abstract minimax ideas and turn them into tools we can actually deploy when the parameters are hard to pin down.

Warren: So, the future work seems to be about taking these refined conditions and seeing how they perform when you apply them to different kinds of large scale moment models, testing if it scales well with larger datasets.

Conclusion: Ruth: So, to wrap up, this paper on "Inference under partial identification with minimax test statistics" shows how you can compute asymptotic distributions for these tricky tests using an outer minimization of an inner maximization structure.

Warren: It’s really about giving us concrete tools for hypothesis testing when you only have partial information about your model's parameters.

Iris: The core idea is using convex analysis to take that complex optimization and turn it into a manageable form that we can actually estimate with bootstrap methods.

Ruth: This means we can get critical values for those tests even in moment models where identification isn't perfect, which is a big step for real-world economics.

Warren: It’s useful because it extends tests like the J-test into these partially identified settings, which is a huge deal when you can't perfectly pin down all the parameters.

Iris: The method still has limitations though; it relies on those weak star topology assumptions to ensure the set of test functions is compact and convex, which means you have to be careful about your underlying functional space.

Ruth: That’s a fair point; we need those topological conditions for the Sion theorem to give us those streamlined bounds we're looking for.

Warren: So, while it gives us a solid computational framework, you still have to check if your specific moment model fits the required topological structure for the math to hold up.

Iris: Exactly. It’s a trade-off between getting a concrete answer and having to be careful about those strict assumptions you're making about the function class itself.

Ruth: This paper is important because it shows how we can use this minimax approach to make inference more robust when dealing with complex, non-linear estimation schemes, like those involving neural networks.

Warren: It moves us from just knowing a general representation of the test statistic to having a specific, computable formulation that works even in those messy function classes.

Iris: And they also provide that specific characterization for the J-statistic under point identification and full rank conditions on the derivative map, which links its limit directly to your moment equations.

Ruth: It’s an interesting link because it shows how the asymptotic behavior connects to the structural properties of your model's constraints.

Warren: So, we've seen how this paper provides concrete computational tools for hypothesis testing under partial identification by reformulating the test statistic through a minimax lens and giving us specific characterizations of J-statistics.

Iris: It really shows that convex analysis can be used to make these kinds of difficult problems much more tractable when dealing with empirical versions of those criterion functions.

Ruth: Overall, "Inference under partial identification with minimax test statistics" provides a means of computing and estimating the asymptotic distributions for these tricky statistics using an outer minimization structure.

Warren: It’s especially useful in moment models where you're dealing with partial identification, giving us tools to obtain critical values by using the bootstrap, extending results from classical tests like the J-test to partially identified settings.

Iris: I just want to say that this work is a strong step forward because it bridges the gap between abstract theoretical complexity and actually being able to run these kinds of tests with real data.

Ruth: It really does show that when you have a complex test statistic under partial identification, there's a structured way to find its asymptotic distribution without having to compute all those complicated derivatives directly.

Warren: That’s the practical side—it lets you get estimates for critical values without needing those derivative calculations, which saves time and effort in the trading world.

Iris: So next time you're dealing with a tricky moment model where identification is only partial, this paper gives you a specific path to follow for your inference.

Ruth: We’ll keep an eye on how these new characterizations of J-statistics play out in larger, more complex economic data sets.

Warren: And we’ll be looking at how much computational cost those methods actually add when you move from theory to running a big simulation.

Isaac Loh

econ.EM

Submitted: 2024-01-23

Updated: 2026-11-02

Comments: 93 pages, 8 figures. Substantially revised with new bootstrap theory, uniform validity results, and simulation evidence

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: As a meticulous researcher, I have thoroughly analyzed both provided texts from arXiv to construct a comprehensive and detailed summary of the paper…

Key concepts

Minimax Test Statistics
This involves finding the best possible test statistic by minimizing a criterion that is itself defined as the supremum over a set of candidate test functions. This minimax approach helps determine the most robust way to test hypotheses when some parameters are not fully identified.
Partial Identification
This occurs when certain parameters in a model cannot be uniquely determined from the observed data alone due to constraints imposed by moment equations. The paper focuses on how to perform valid statistical inference even when these identifications are partial.
Convex Analysis Tools (e.g., Sion’s Theorem)
These are mathematical techniques used to simplify complex optimization problems, specifically those involving minimax formulations. They allow researchers to transform difficult problems into simpler ones, making it possible to calculate the asymptotic distributions of test statistics more easily.

Terminology

Summary

As a meticulous researcher, I have thoroughly analyzed both provided texts from arXiv to construct a comprehensive and detailed summary of the paper, Inference under partial identification with minimax test statistics.

The paper addresses the theoretical development and asymptotic characterization of test statistics derived from an outer minimization of an inner maximization—a structure frequently encountered in moment models. The core focus is on providing tools for hypothesis testing under partial identification. p. 0

Where the paper says it

We provide a means of computing and estimating the asymptotic distributions of statistics based on an outer minimization of an inner maximization. Such test statistics, which arise frequently in moment models, are of special interest in providing hypothesis tests under partial identification.

Core Methodology and Theoretical Framework:

The central theme revolves around computing and estimating the asymptotic distributions of test statistics based on minimax values. These minimax values are particularly relevant when testing for the existence of a non-empty identified set under partial identification constraints defined by moment equations. p. 0

Where the paper says it

when hypothesis testing is conducted under the assumption of partial identification. When an identified set is characterized as the solution to a system of moment equations, the existence of a nonempty identified set can be consistently tested with adequate knowledge of how the moment functions will asymptotically behave thereon.

The criterion function is formulated as a supremum over a set of test functions, (theta) = t in T v(theta, t). p. 1
Where the paper says it

2 ISAAC LOH As a point of departure, we note that ℓ can oftentimes be written as a supremum of a class of test functions v indexed by t ∈ T : ℓ(θ) = sup t∈T v(θ, t). For instance, when ℓ is a norm or seminorm of a moment function of parameter θ over some Banach space, it always admits such a representation with linear t.

This formulation allows the paper to employ powerful convex analysis tools, such as Sion’s theorem, to derive simplified representations of the minimax problem. p. 1
Where the paper says it

Such a representation is especially amenable to treatment by familiar tools arising from convex analysis. In particular, we show that an application of Sion’s celebrated minimax theorem (Sion (1958)) leads to a streamlined way of bounding the distribution of such minimax test statistics from above, with asymptotic equality under certain rate or convexity conditions.

The paper demonstrates that this minimax criterion can be reformulated into a more tractable form:

sqrt n theta in theta n(theta) = theta in theta 0 t: d v(theta,t) over d theta=0

This transformation is crucial as it facilitates the computation of asymptotic distributions and the estimation of critical values via bootstrap methods, even when dealing with complex, non-linear sieve spaces for test functions, such as those arising from neural networks.

Asymptotic Characterization and Results:

The main theorem establishes a precise bound for the asymptotic distribution. Specifically, it shows that:

theta in theta 0 t in K(theta) G n(theta, t) at most theta in theta 0 t: psi n(theta,t) at most c n G n(theta, t) + o(1) G n(theta, t)

The final result characterizes the asymptotic distribution by stating the equality:

theta in theta 0 t in K(theta) G n(theta, t) = theta in theta 0 t: psi n(theta,t) at most c n G n(theta, t) + o(1)

The paper also provides a corollary concerning the limiting distribution of the J-statistic under specific conditions (point identification and full-rank conditions on the derivative map), characterizing it via:

theta in theta 0 t k M grad m rn((theta) - m(theta)) k X theta in theta 0 q chi squared k-rank(grad m)

Technical Foundations and Assumptions:

The rigorous establishment of these results relies on several key mathematical prerequisites. The analysis involves mapping a space (X times R)* endowed with the weak-* topology to a space A endowed with the topology of pointwise convergence, demonstrating that the map zeta is continuous. p. 7

Where the paper says it

4 therein shows that the lower hemicontinuity condition is fulfilled if the correspondence sending pairs (P, θ) to the tangent cone TmP (θ)C ⊂ X is merely upper hemicontinuous with respect to the weak topology on X. On the other hand, by boundedness of the maps t ∈ T , continuity of the map (P, θ, t) 7→ hmP (θ), ti may be ascertained by verifying norm-continuity of the map (P, θ, t) 7→ mP (θ), or by confining the points mP (θ) to lie in a norm-compact set and requiring (P, θ, t) 7→ mP (θ) to be merely weakly continuous.

This topological structure ensures that certain sets are compact in the weak-* topology, which implies that T, being its continuous image, must be weakly compact and convex. p. 7
Where the paper says it

4 therein shows that the lower hemicontinuity condition is fulfilled if the correspondence sending pairs (P, θ) to the tangent cone TmP (θ)C ⊂ X is merely upper hemicontinuous with respect to the weak topology on X. On the other hand, by boundedness of the maps t ∈ T , continuity of the map (P, θ, t) 7→ hmP (θ), ti may be ascertained by verifying norm-continuity of the map (P, θ, t) 7→ mP (θ), or by confining the points mP (θ) to lie in a norm-compact set and requiring (P, θ, t) 7→ mP (θ) to be merely weakly continuous.

Furthermore, the analysis utilizes concepts from functional analysis:

  1. Fenchel-Moreau Theorem: Used to show gamma tilde = gamma tilde**, leading to the definition gamma(x) = t in T t(x) for x in D.

  2. Interior Point Analysis: The analysis focuses on a point in the interior of D, where gamma tilde coincides with gamma locally, allowing for the definition of the set K about(theta) based on conditions involving grad m(theta) and the boundary d gamma(m(theta)).

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

  1. Minimax test statistic computation for partial identification: The system can compute and estimate asymptotic distributions of statistics based on an outer minimization of an inner maximization using a minimax formulation, which is useful in moment models under partial identification.

  2. Hypothesis testing under partial identification: The system can provide methods to obtain critical values for hypothesis tests by using the bootstrap, extending results from classical tests like the J-test to partially identified settings.

  3. Inference on distances to convex sets: The system can compute asymptotic distributions of distances between random vectors and convex sets in Banach spaces, including those related to shape restrictions enforced on random elements.

  4. Handling non-linear estimation schemes: The system can accommodate adversarial employment of nonconvex sieve spaces, such as neural networks, for nonparametric inference by using minimax formulations that are amenable to these complex function classes.

  5. Robustness to parameter space approximations: The system can provide asymptotic results that mitigate dependence on local linear approximations of the parameter space by utilizing convex analysis tools and reformulating test statistics using the condition (1.3) which is compared favorably to (1.1).

Abstract

In partially identified models, specification and subvector tests often minimize a measure of sample violation over the parameter space. Such statistics have the minimax form T n=r n θ inΘ t in Tv n(θ,t), which includes r n θ inΘξ(m n(θ)) for a convex loss ξ of Banach-space-valued moments m n. We characterize their limiting distribution through familiar geometric objects. The limit is an upper bound in general and exact under convexity in θ, a standard polynomial minorant condition, or continuity of the contact set. It retains the loss directions that bind locally and cannot be absorbed by moving θ, which under convex loss are the subgradients of ξ at m(θ) that are nonnegative along every first-order movement of θ. A re-optimization bootstrap common to Hong and Li (2018) and Bugni, Canay and Shi (2017) re-solves the minimization after a random perturbation and needs no estimate of the identified set, the contact set, or any derivative. We show that it consistently estimates the limit under any of the three conditions for exactness. Under a standard uniform minorant condition, it is also uniformly valid for minimax criteria and for general convex losses on Banach-valued moments. Without a uniform minorant, the two properties can be incompatible. In simulations, the re-optimization bootstrap has close to nominal size and closely tracks the sampling distribution of the statistic.

Related papers