Model Restrictiveness in Functional and Structural Settings

This paper extends the restrictiveness measure to functional and structural econometric settings using Gaussian process priors, demonstrating that models appear more restrictive over continuum domains than finite observations, establishing restrictiveness as the limit of noise-free learning curves, and revealing how endogeneity and discrepancy function choices fundamentally alter model rankings.

The Gist

Imagine you are a detective trying to solve a mystery. You have a suspect (an economic model) who claims to know exactly how the world works. But how do you know if this suspect is actually smart and insightful, or just a "yes-man" who agrees with everything you say?

This paper introduces a new way to measure how "picky" or "restrictive" an economic model is. Think of it as a test to see how much the model refuses to believe, even when the data tries to convince it otherwise.

Here is the breakdown of their new method using simple analogies:

1. The Old Way vs. The New Way: The "Menu" Analogy

The Old Way (Finite Sets):
Previously, to test a model, researchers would give it a small, fixed menu of questions (like 25 specific lottery games). If the model got them right, it seemed flexible. If it got them wrong, it seemed too rigid.

• The Problem: It's like testing a chef by only asking them to cook three specific dishes. If they can't cook a fourth dish you didn't ask for, you don't know if they are a bad chef or just a chef who only knows those three recipes.

The New Way (Functional/Continuum):
The authors say, "Let's stop testing on a small menu. Let's test the model on the entire infinite universe of possible dishes."

• The Analogy: Instead of asking the chef to cook 25 specific meals, we ask them to describe how they would cook any meal imaginable.
• The Finding: When you test a model on this "infinite menu," it turns out to be much more restrictive than we thought. Models that seemed flexible on a small list are actually very rigid when faced with the whole world. They rule out more possibilities than we realized.

2. The "Rigid vs. Flexible" Ruler

The paper creates a ruler to measure Restrictiveness.

• High Restrictiveness: The model is like a strict librarian. It says, "I only accept books written in this specific font, with this specific cover, and this specific plot." It rules out a huge amount of the library. This is good for theory (it's disciplined) but risky if the real world is messy.
• Low Restrictiveness: The model is like a chaotic hoarder. It says, "I'll accept anything!" It fits the data perfectly but tells you nothing new because it doesn't rule out anything.

The Twist: The paper shows that when you measure this "picky-ness" over the whole universe of possibilities (not just a sample), the "strict librarians" look even stricter.

3. Structural Models: The "Puzzle with Hidden Pieces"

Economics often deals with "Structural Models"—these are like complex puzzles where some pieces are hidden (endogeneity). For example, in a market, price affects how much people buy, but how much people buy also affects the price. It's a chicken-and-egg problem.

The authors show how to measure restrictiveness even when the puzzle has these hidden, tricky pieces.

• The Finding: When you add the "hidden pieces" (like using special tools called Instruments to solve the chicken-and-egg problem), the models become much more restrictive.
• The Metaphor: Imagine a detective who can solve a crime with just a few clues (easy). But if you force them to solve it while also proving they weren't the one who stole the evidence (a hard constraint), their ability to "guess" the solution drops. They have to be much more precise. The paper finds that adding these economic constraints (like endogeneity) makes models significantly more disciplined.

4. What NOT to Use: The "Wrong Ruler"

The paper warns against using certain standard math tools (like GMM or Rademacher complexity) to measure this.

• The Analogy: It's like trying to measure the "spiciness" of a curry using a ruler. You might get a number, but it doesn't tell you what you actually care about (heat).
• The Solution: They argue you must choose a "discrepancy function" (a ruler) that makes sense for the specific economic question. If you want to know how well a model predicts prices, your ruler should measure price errors, not some abstract math complexity.

5. The "Learning Curve" Connection

The authors connect this idea to how machines learn.

• The Analogy: Imagine a student taking a test.
  • Completeness: How many questions did they get right? (Did they capture the data?)
  • Restrictiveness: How many questions did they refuse to guess on because their theory said those answers were impossible? (Did they rule out nonsense?)
• The paper shows that "Restrictiveness" is essentially the limit of how well a model can learn if there is no noise in the world. It measures the pure "brainpower" or structural logic of the model, stripped of random luck.

Summary: Why Does This Matter?

In the past, economists might have chosen a model because it fit their specific dataset well. This paper says: "Wait, let's check how 'picky' that model is across the entire universe of possibilities."

• For Researchers: It helps them choose models that aren't just "data-fitting machines" but actually have strong, disciplined theories.
• For the Real World: It reveals that many popular economic models are actually much more rigid and specific than we thought. This is good news because it means they are making stronger, more testable predictions, but it also means we need to be careful not to force the real world into a box that is too small.

In a nutshell: This paper gives economists a better, more honest ruler to measure how much structure their theories really impose on the world, revealing that our models are often stricter (and more interesting) than we previously believed.

Technical Summary

Here is a detailed technical summary of the paper "Model Restrictiveness in Functional and Structural Settings" by Fudenberg, Gao, and You.

1. Problem Statement

Economic models are designed to be restrictive, ruling out data patterns that conflict with theoretical priors. However, researchers often lack a quantitative measure to assess how much structure a specific model imposes relative to plausible alternatives.

• Limitations of Existing Measures: Previous work (Fudenberg, Gao, and Liang, 2026) measured restrictiveness over finite sets of observations. This approach fails to capture the full functional flexibility of models defined over continuum domains.
• Structural Complexity: Standard econometric tools (like GMM criteria or Rademacher complexity) are often unsuitable for measuring restrictiveness because they conflate estimation noise, instrument choice, or worst-case adversarial bounds with the intrinsic structural content of the model.
• The Gap: There is a need for a framework that extends restrictiveness to:
  1. Functional settings (infinite-dimensional spaces).
  2. Structural models involving endogeneity, instrumental variables (IV), multiple equilibria, and semiparametric components.

2. Methodology

The authors extend the restrictiveness framework using Bayesian nonparametrics and Gaussian Processes (GPs).

A. Core Definition

Restrictiveness ($r$) is defined as the expected approximation error of a model class $F_\Theta$ relative to a flexible "eligible" set $F$, normalized by a baseline model $F_{base}$.
$$r(F_\Theta; F, d) = \frac{E_{\lambda_F}[d(F_\Theta, f)]}{E_{\lambda_F}[d(F_{base}, f)]}$$
Where:

• $f \sim \lambda_F$: A pseudo-true prediction rule drawn from an evaluation distribution over the eligible set.
• $d$: A discrepancy function measuring the distance between prediction rules.
• $d(F_\Theta, f) = \inf_{f_\theta \in F_\Theta} d(f_\theta, f)$: The best approximation error the model can achieve against $f$.

B. Functional Settings (Continuum Domains)

To handle infinite-dimensional spaces, the authors operationalize $\lambda_F$ using Gaussian Process (GP) priors.

• Sampling: They sample functions from a GP (e.g., Matérn 3/2 kernel) to represent the "pseudo-truth."
• Constraints: To ensure economic validity (e.g., monotonicity in demand), they employ constrained GP sampling techniques (e.g., transforming outputs or using virtual points).
• Discrepancy: They advocate for continuous discrepancy functions (e.g., squared $L_2$ norm) rather than binary or capacity-based metrics.

C. Structural Models

The framework is adapted for three complex structural scenarios:

1. Reduced-Form Representation: For models admitting a reduced form (e.g., additive error models), restrictiveness is defined on the space of conditional distributions induced by the structural parameters.
2. Instrumental Variables (IV) & Endogeneity: For models identified via instruments (e.g., BLP models), the eligible set is constrained by moment conditions. The authors show that the infinite-dimensional optimization over nuisance functions can often be reduced to finite-dimensional optimization over structural parameters.
3. Multiple Equilibria & Semiparametrics:
   • Multiple Equilibria: The reduced form becomes a correspondence. Restrictiveness is defined by the infimum approximation error over all possible equilibrium selection rules.
   • Semiparametric: The model includes infinite-dimensional nuisance parameters ($h$). The definition integrates over the space of these nuisance functions to measure the restriction imposed solely by the structural parameters.

D. Critique of Existing Metrics

• Rademacher/VC Dimension: The authors argue these encode a specific discrepancy function (correlation-based) that leads to degeneracy: all finite-dimensional falsifiable models appear "fully restrictive" (value = 1) in the limit. This is an artifact of the metric, not the model.
• GMM Criteria: GMM objective functions are unsuitable as discrepancies because they depend on the choice of instruments and weighting matrices, and they measure moment violations rather than predictive distance. They argue moment conditions should define the eligible set, not the discrepancy function.

E. Theoretical Connection

The paper establishes that restrictiveness equals the normalized limit of the noise-free average-case learning curve. It represents the asymptotic approximation error achievable by the model class relative to a baseline, isolating structural flexibility from estimation noise.

3. Key Contributions

1. Extension to Functional Spaces: Moving from finite-sample evaluations to continuum domains using GP priors, revealing that models appear significantly more restrictive when evaluated over the full functional space.
2. Structural Econometrics Framework: Providing a rigorous definition of restrictiveness for models with endogeneity, IVs, multiple equilibria, and semiparametric components.
3. Substantive Modeling Decision: Arguing that the choice of the discrepancy function ($d$) is a critical modeling decision that determines the interpretation of restrictiveness, rather than a technical detail to be inherited from machine learning capacity measures.
4. Computational Feasibility: Demonstrating that restrictiveness can be computed using standard replication pipelines by adding a GP prior and a discrepancy function.

4. Empirical Results

The authors apply the framework to three economic problems:

A. Risk Preferences (CPT vs. Disappointment Aversion)

• Setup: Re-evaluating Cumulative Prospect Theory (CPT) and Disappointment Aversion (DA) over the entire space of binary lotteries (continuum) rather than a finite set of 25 lotteries.
• Finding: Restrictiveness is uniformly higher in the continuum setting.
  • Full CPT restrictiveness: 0.56 (New) vs. 0.28 (Old).
  • The relative ranking of parameters remains similar, but the absolute structural content is much larger than previously estimated.
  • Some models previously considered "interior" to the Pareto frontier (CPT($\delta$), DA($\eta$)) are now found to be dominated.

B. Multinomial Choice (Exogenous)

• Setup: Comparing Multinomial Logit (MNL), Nested Logit (NL), and Mixed Logit (MXL) using the Nevo (2000) cereal dataset.
• Finding:
  • Theoretical vs. Practical Flexibility: While MXL is theoretically flexible, its standard parametric implementation (normal random coefficients) makes it nearly as restrictive as NL.
  • Driver of Restrictiveness: Restrictiveness is driven almost entirely by the mean utility functional form. When the mean utility is allowed to be nonparametric, all models become highly restrictive. When only individual heterogeneity is nonparametric, restrictiveness drops to near zero.

C. Multinomial Choice (Endogenous)

• Setup: Introducing endogenous prices using BLP-style instruments.
• Finding:
  • Moment Restrictions Dominate: The imposition of IV moment conditions drastically increases restrictiveness for all models (e.g., MXL jumps from 0.11 to 0.67).
  • Ranking Reversal: In the exogenous case, MXL and NL were less restrictive than MNL. In the endogenous case, MXL becomes the least restrictive, while NL and MNL become nearly identical in restrictiveness. The nesting parameter in NL offers no additional flexibility once endogeneity constraints are applied.

5. Significance

• Unified Evaluation Language: The paper provides a quantitative language to compare models across behavioral economics, industrial organization, and game theory, balancing empirical accuracy (completeness) against theoretical discipline (restrictiveness).
• Policy for Model Selection: It suggests that "more parameters" does not necessarily mean "less restrictive." A model with many parameters but strong structural constraints (like IVs) can be highly restrictive.
• Future Regularization: The authors propose using restrictiveness as a regularization device in model selection. Unlike AIC/BIC (which penalize parameter count), a restrictiveness-based penalty would favor models that rule out economically implausible patterns, regardless of their parameter count.
• Methodological Shift: It challenges the econometric community to move away from using estimation criteria (like GMM) as measures of model complexity and toward explicit, interpretable discrepancy functions tailored to the economic context.