📈 economics

Nonparametric Bayesian Inference for Partially Identified Discrete Response Models

This paper proposes a nonparametric Bayesian inference framework for partially identified discrete response models that utilizes Gaussian process priors to directly infer identified sets from conditional choice probabilities, offering a flexible, computationally efficient approach that ensures posterior consistency under both correct specification and misspecification without requiring covariate discretization or moment conversion.

Original authors: Elie Tamer, Christopher D. Walker

Published 2026-08-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Elie Tamer, Christopher D. Walker

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of economics, researchers often try to understand how people make choices when faced with uncertainty, such as deciding whether to enter a market, switch jobs, or buy a product. These decisions are rarely simple; they depend on hidden factors, the actions of others, and complex rules that govern the environment. When economists build mathematical models to explain these behaviors, they sometimes hit a wall: the data available is not enough to pinpoint a single, exact answer for how the world works. Instead of finding one specific number that represents a key rule, the data only allows them to narrow the possibilities down to a range of plausible values. This is known as partial identification. It is like trying to locate a ship in a foggy ocean; you cannot see the exact coordinates, but you can be certain the ship is somewhere within a specific circle. The challenge for statisticians has been how to measure the size and shape of that circle with precision, especially when the rules governing the choices change depending on the specific circumstances of each person or situation.

For decades, the standard way to handle this problem involved a series of compromises. To make the math workable, researchers often had to simplify the real world by grouping continuous variables, like income or age, into broad, artificial buckets. They would also have to convert complex, situation-specific rules into simpler, one-size-fits-all summaries. While this made the calculations possible, it often threw away valuable details, blurring the very boundaries of the circle they were trying to draw. A new approach developed by Elie Tamer of Harvard University and Christopher D. Walker of Duke University offers a different path. They have created a method that allows researchers to work directly with the messy, continuous reality of the data, preserving the fine details of how choices are made without forcing them into artificial boxes. Their work provides a way to draw the boundaries of uncertainty with much greater clarity, revealing a sharper picture of economic behavior than was previously possible.

The core of Tamer and Walker's innovation is a shift in perspective. Instead of trying to guess the hidden structural rules directly, they focus on the observable outcome: the probability that a person will make a specific choice given their circumstances. They treat this probability as a flexible, unknown shape that can be learned from the data. Using a powerful statistical tool called a nonparametric Bayesian framework, they build a model that learns the shape of these probabilities directly from the observations. Imagine a flexible sheet that can stretch and bend to fit the contours of the data perfectly, rather than a rigid mold that forces the data into a pre-set shape. Once the model has learned the most likely shape of these choice probabilities, it uses a set of logical rules to determine what the range of possible answers for the hidden economic parameters must be. This process transforms the problem from guessing a hidden number into mapping a known shape, which is a much more stable and reliable task.

One of the most significant advantages of this method is that it does not require the researcher to chop up continuous data into discrete groups. In previous methods, if a researcher wanted to study how a continuous variable like price affected a decision, they might have to divide prices into low, medium, and high categories. This coarse grouping often led to a loss of information, making the final range of answers wider and less precise. Tamer and Walker's approach handles the continuous nature of the data naturally. It respects the smooth transitions in the real world, allowing the model to detect subtle shifts in behavior that would be invisible if the data were forced into buckets. This means the resulting range of plausible answers is tighter and more informative, giving a clearer view of the underlying economic forces at play.

The researchers tested their method using computer simulations that mimicked real-world economic scenarios, such as dynamic choices where a person's current decision depends on their past actions. They generated thousands of fake datasets with known underlying rules and then applied their new method to see if it could recover the correct range of answers. The results were consistent and encouraging. As the amount of data increased, the model's estimate of the range of possible answers became increasingly concentrated around the true values. The method successfully identified the correct boundaries of uncertainty, even when the data was complex and the rules were not fully known. Furthermore, the researchers showed that their approach could also detect when a model was fundamentally flawed. If the rules being tested were incompatible with the data, the method would signal this by producing an empty range, acting as a reliable diagnostic tool to warn researchers that their assumptions were wrong.

To make this powerful method practical for everyday use, the authors developed a specific computational recipe based on a technique called Gaussian processes. This is a way of modeling smooth, continuous functions that allows the computer to update its beliefs efficiently as new data arrives. The beauty of their implementation is that it breaks the complex problem into smaller, independent pieces that can be solved simultaneously. This parallel processing makes the method fast and scalable, capable of handling large datasets without getting bogged down in computational complexity. The researchers demonstrated that their approach works not just for simple choices, but also for more complicated scenarios involving aggregated data, such as market shares, and even for outcomes that are continuous rather than discrete.

The implications of this work extend beyond just a new statistical trick; it represents a fundamental change in how economists can think about uncertainty. By removing the need for artificial simplifications, the method allows for a more honest and detailed exploration of economic models. It respects the richness of the data and the complexity of human behavior, providing a framework where the uncertainty is quantified with precision rather than obscured by approximation. The authors have shown that it is possible to have a fully Bayesian approach—one that updates beliefs in a coherent, logical way—without sacrificing the flexibility needed to model the real world. This opens the door for researchers to ask more nuanced questions and get more precise answers, moving the field of economics closer to a true understanding of how people make decisions in an uncertain world. The work stands as a testament to the power of combining deep theoretical insight with practical computational innovation to solve long-standing problems in statistical inference.

Technical Summary: Nonparametric Bayesian Inference for Partially Identified Discrete Response Models

1. Problem Statement and Motivation
Many economic models involving discrete responses—such as games with multiple equilibria, semiparametric choice models with fixed effects, dynamic models with state dependence, and models with limited consideration—result in set-identified structural parameters. In these settings, the identified set $\Gamma_I$ is not a single point but a set of parameters consistent with the data and model restrictions.

A common feature across these models is that the identified set is defined by an unrestricted reduced-form conditional probability mass function (PMF), denoted $p(X)$ . Specifically, $\Gamma_I = \{ \gamma : f(X, p(X), \gamma) \geq 0 \text{ a.s.} \}$ , where $\gamma$ represents structural parameters (e.g., payoff parameters) and $f$ is a known vector of functions. Existing frequentist approaches to inference in such partially identified models often face significant practical hurdles:

They frequently require converting conditional moment inequalities (CMIs) into unconditional moments by discretizing continuous covariates or selecting specific instruments.
These "ad hoc" choices can lead to a loss of identifying information (reduced sharpness) or sensitivity to instrument selection under misspecification.
Linear system formulations often presume finite support for covariates to generate a finite set of inequalities, ignoring the potential for richer identification from continuous covariates.

This paper proposes a nonparametric Bayesian inference framework that directly targets the reduced-form conditional PMF $p(X)$ , thereby avoiding the need to discretize covariates or select unconditional moments.

2. Methodology

2.1. The Inference Framework
The core observation is that the identified set $\Gamma_I$ is a functional of the reduced-form parameter $p$ . The authors propose a two-step Bayesian approach:

Nonparametric Bayesian Inference for $p$ : A prior is placed directly on the infinite-dimensional space of conditional PMFs $p$ . The posterior distribution $\Pi(p | Y, X)$ is derived based on the observed data $(Y, X)$ .
Mapping to the Identified Set: The posterior for the identified set, $\Pi(\Gamma_I | Y, X)$ , is induced by mapping draws from the posterior of $p$ through the correspondence defining the identified set. If $\gamma \in \Gamma_I$ if and only if $f(X, p(X), \gamma) \geq 0$ , then a draw $p^{(s)}$ from the posterior of $p$ yields a draw $\Gamma^{(s)}_I = \{ \gamma : f(X, p^{(s)}(X), \gamma) \geq 0 \}$ .

This framework nests conditional moment inequalities and linear systems with unknown coefficients as special cases. It treats the identified set as a random closed set, allowing for the derivation of a posterior capacity functional that fully characterizes the uncertainty about the set.

2.2. Implementation: Logistic Stick-Breaking Gaussian Processes
To implement this framework computationally, the authors utilize a flexible class of priors based on Gaussian Processes (GPs) combined with a stick-breaking parametrization of the categorical distribution.

Stick-Breaking: The PMF $p(x) = (p_1(x), \dots, p_K(x))$ is reparametrized using weights $q_k(x) \in [0, 1]$ such that $p_k(x)$ is a function of $q_1(x), \dots, q_k(x)$ .
GP Priors: The weights $q_k(x)$ are modeled as $\Lambda(B_k(x))$ , where $\Lambda$ is the logistic function and $B_k$ are independent Gaussian processes with mean $\mu_k$ and covariance $\kappa_k$ .
Computation: This parametrization allows the use of Pólya-Gamma data augmentation (Polson et al., 2013). This technique transforms the likelihood into a form where the GP priors become conditionally conjugate. Consequently, posterior sampling can be performed via a Gibbs sampler with closed-form updates.
- The blocks of the GP corresponding to different outcomes are independent under the posterior, allowing for parallel sampling.
- Once draws of $p$ are obtained, draws of the identified set are computed by solving the corresponding optimization problem (e.g., linear programming or minimum distance) for each draw. This step is parallelizable and does not require discretizing the covariates.

3. Key Theoretical Results

The paper establishes rigorous posterior consistency guarantees for the identified set.

Consistency under Correct Specification: The authors prove that if the model is correctly specified (i.e., the true PMF $p_0$ yields a non-empty identified set $\Gamma_I(p_0)$ ), the posterior distribution of the identified set concentrates around the true identified set $\Gamma_I(p_0)$ in the Hausdorff distance as the sample size $n \to \infty$ .
- This result relies on the posterior for $p$ concentrating around $p_0$ and the mapping from $p$ to $\Gamma_I$ satisfying lower hemicontinuity and uniform convergence conditions.
- The paper verifies these conditions for the proposed Logistic Stick-Breaking Gaussian Process (LSBGP) priors, showing that the posterior for $p$ converges to the truth in the empirical supremum norm at the minimax optimal rate under standard smoothness assumptions.
Consistency under Misspecification: The framework addresses model misspecification (where $\Gamma_I(p_0) = \emptyset$ ).
- The posterior probability that the identified set is empty is shown to be a consistent diagnostic for misspecification.
- The authors introduce a pseudo-identified set $\tilde{\Gamma}_I$ , defined as the set of minimizers of the criterion function (or the set where the criterion is zero after subtracting its infimum). They prove that the posterior for $\tilde{\Gamma}_I$ is consistent for the pseudo-true set $\tilde{\Gamma}_I(p_0)$ even when the model is misspecified.
Extensions: The theoretical results and implementation details are shown to extend to:
- Aggregated discrete responses (e.g., market shares), where data follows a multinomial distribution.
- Continuous response variables, where the identified set is defined by conditional moment inequalities involving continuous outcomes.

4. Contributions and Significance

The paper claims several significant contributions to the literature on partial identification and Bayesian econometrics:

Direct Handling of Conditional Restrictions: Unlike existing methods that often require converting conditional moments to unconditional moments or discretizing continuous covariates, this framework targets the conditional PMF directly. This preserves the sharpness of the identified set and avoids the loss of information associated with binning or instrument selection.
Computational Tractability: By leveraging the stick-breaking representation and Pólya-Gamma augmentation, the authors provide a computationally attractive method where posterior sampling is performed in closed form. The independence of the stick-breaking components allows for efficient parallelization.
Theoretical Rigor in Infinite Dimensions: The paper provides a general posterior consistency theorem for set-valued parameters derived from infinite-dimensional reduced-form parameters. It carefully verifies the necessary conditions (uniform convergence, lower hemicontinuity) for a specific class of nonparametric priors (GPs), addressing concerns about the "curse of dimensionality" and the delicacy of posterior consistency in functional spaces.
Robustness to Misspecification: The framework offers a principled Bayesian approach to handling model misspecification, providing both a diagnostic (probability of an empty set) and a valid inference target (the pseudo-identified set) when the model is incorrect.
Flexibility: The approach accommodates a wide range of models, including conditional moment inequalities, linear systems, dynamic panel models, and aggregated data, without requiring structural changes to the inference procedure.

In summary, the paper proposes a unified, flexible, and computationally feasible Bayesian framework for inference in partially identified discrete response models, bridging the gap between the flexibility of nonparametric frequentist methods and the coherent uncertainty quantification of Bayesian inference, all while avoiding the discretization of covariates.

Technical Summary: Nonparametric Bayesian Inference for Partially Identified Discrete Response Models

More like this