Identification of Rational Inattention-Additive Random Utility Models

arXiv:2611.14752 · econ.EM, stat.ME · Submitted 2026-11-10 · 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: "Identification of Rational Inattention-Additive Random Utility Models".

Warren: Here's the gist of the paper. This paper introduces Rational Inattention-Additive Random Utility Models (RI-ARUMs), a novel class of econometric models based on Matˇejka and McKay (2015) that incorporates additive unobserved heterogeneity to analyze behavioral biases in decision-making processes The paper shows that some systematic components of the model are nonparametrically identified even if the distributions of unobserved variables and the prior of decision makers are both unknown to the Econometrician Model Framework and Components The authors propose a general class of RI-ARUMs as an econometric toolkit for analyzing empirical behavioral biases The model extends the classic representative agent model by allowing Decision Makers (DMs) to have knowledge about parts of the uncertain payoffs, replacing the unconditional prior with a conditional one based on existing knowledge This results in three key components modeled additively in observed and unobserved heterogeneity: "The conditional prior of the payoffs, the cost of information, and the realised payoffs" The model contains a special case corresponding to the classic full-information ARUM or a degenerate prior Theoretical Insights into Choice Probabilities The authors provide theoretical new insights into features and properties of rational inattention discrete choice models by analyzing conditional choice probabilities (CCPs) using convex optimization theory They note that optimal CCPs may be situated on the boundary of the probability simplex, which is important for subsequent identification analysis The optimal signal rule is derived by maximizing an objective function involving expected payoffs and information cost The resulting optimal CCPs are characterized by a multinomial mixed–logit model where the integrated CCPs enter as alternative-specific weights Nonparametric Identification Results The paper develops nonparametric identification results for the unknown conditional prior, cost of information, and realized payoffs under four scenarios regarding unobserved heterogeneity In the most benevolent sub-scenario—where DMs have correct beliefs about payoffs—strong identification results are achieved In a more testing environment where the Econometrician has no knowledge of the DM’s prior, only certain features of the model can be identified The identification results extend existing nonparametric identification results for ARUMs to allow for rational inattention Identification with Unobserved Heterogeneity When unobserved heterogeneity is added to all three components (payoffs, information cost, and prior), the authors rely on the identity involving the composite mapping Π which is integrated with respect to the distribution of unobserved variables Fε They show that for certain components like u1, identification can be achieved under specific assumptions by leveraging properties of this integrated mapping However, identification of u0 in the case where the prior is completely unknown and unrestricted remains difficult without very strong restrictions on the deviation of subjective beliefs from the physical law Identification Strategies for Structural Functions The paper presents general results for identifying composite functions of the form Φ(w, z) = Ψ(ν(w), z) under increasing number restrictions on the model The final identification result shows that under specific support conditions and normalisations, structural functions like u0 are identified on certain index regions The analysis demonstrates that while some components like u1 can be identified even with unobserved heterogeneity, identification of u0 is much more challenging unless the prior is not too badly misspecified Finally, the paper concludes that identifying the DM’s prior πV1 remains largely unachievable without very strong restrictions Based on the discrete choice rational inattention model, a flexible class of econometric models is developed with a comprehensive identification analysis showing what can and cannot be identified under different scenarios The authors establish that functions u0, u1, and λ are identified even when unobserved heterogeneity is present under suitable assumptions However, identifying the prior πV1 remains largely unachievable without very strong restrictions References Allen R.

Ruth: and Rehbeck J.

Warren: (2019) Identification with additively separable heterogeneity.

Ruth: Econometrica, 87(3):1021–1054 Allen R.

Title and authors: Ruth: So we’re looking at this paper now, "Identification of Rational Inattention-Additive Random Utility Models." It’s pretty dense, but the core idea is bringing additive random utility models and rational inattention together.

Warren: Yeah, it builds on Matˇejka and McKay's work from two thousand fifteen but adds this additive unobserved heterogeneity <ref:2611.14752#pg1>. The main thing they’re showing is that even if you don't know the distribution of the unobserved variables or what the decision makers’ prior belief is, you can still identify some parts of the model.

Iris: Exactly, and that's where it gets interesting for us as econometricians. They are proposing this general class of models called RI-ARUMs as a toolkit for analyzing behavioral biases in decision-making processes.

Ruth: So, what does that actually mean when we translate it into something we can use? It seems like they're tackling the problem of how people make choices when they aren't perfectly informed about what’s on the other side of those options.

Warren: They do that by modeling three additive components: the conditional prior, the cost of information, and then the realized payoffs. That structure is key to how they build their toolkit for empirical work.

Iris: And they also dive into choice probabilities using convex optimization theory to find out what the optimal choice probabilities look like in this rational inattention setting.

The paper's summary: Ruth: In essence, the paper is filling a gap in the literature because there wasn't much theoretical work on the econometrics of rational inattention discrete choice models before this.

Warren: Right, and they lay out four different scenarios for how unobserved heterogeneity can be present. The most benevolent one is when decision makers have correct beliefs about the payoffs, which gives them strong identification results.

Iris: And then there’s the more testing environment where the econometrician has no knowledge of the prior belief at all, and in that case, they’re only able to identify certain features of the model. That’s a crucial limitation for practical application.

Ruth: The paper suggests that their approach provides a general blueprint for taking these complex structural representative agent models and actually turning them into data analysis tools.

Warren: And one of the main contributions is showing that some systematic components of these primitives are nonparametrically identified even if you don't know the distributions or the prior belief.

Iris: That’s a big deal because it means we don't always need perfect knowledge of everything to do identification in these complex models, provided you stick to certain assumptions.

The paper's improvements: Ruth: Now let's talk about what they suggest improving or extending from the original model setup. They focus on how you can identify specific parts of the model under different conditions.

Warren: They develop nonparametric identification results for the unknown conditional prior, the cost of information, and realized payoffs across four scenarios involving unobserved heterogeneity. That’s a lot to manage.

Iris: The paper explores how they handle this uncertainty when unobserved heterogeneity is added to all three components—the payoffs, the information cost, and the prior belief. They rely on an identity involving a composite mapping called Pi which is integrated over the distribution of those unobserved variables.

Ruth: And while that integrated mapping helps identify some things like u1 under specific assumptions, they admit that identifying u0 in the case where the prior is completely unknown remains difficult without very strong restrictions on how much subjective beliefs can deviate from physical laws.

Warren: They also present general results for identifying composite functions of a form called Phi of w and z, which means they can identify structural functions like u0 on certain index regions under increasing number restrictions.

Iris: The final point is that they conclude that identifying the decision maker’s prior, Pi V1, remains largely unachievable without very strong restrictions on how far the subjective beliefs misspecify the actual law.

Conclusion: Ruth: So to wrap up, this paper introduces a flexible class of econometric models with a comprehensive identification analysis showing what can and cannot be identified under different scenarios for rational inattention.

Warren: The main implication is that they’ve established that functions u0, u1, and lambda are actually identified even when unobserved heterogeneity is present under suitable assumptions.

Iris: That’s a strong statement because it means you get identification for those specific components even with uncertainty about the environment. But identifying the prior belief Pi V1 still needs very strict restrictions to work reliably.

Ruth: So, they provide a comprehensive analysis showing what works and what doesn't in this framework, which is really helpful for anyone trying to build these models from scratch.

Warren: It sets a clear direction for future research into how we can make these RI-ARUMs more robust when applied to real data where priors are messy.

Iris: Yeah, it’s a solid paper that gives us the tools to test the limits of rational inattention models and see where the identification breaks down.

Dennis Kristensen, Ruolei Zhu

University College London

econ.EM, stat.ME

Submitted: 2026-11-10

Updated: 2026-11-10

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

Importance score: 75/100

The gist: This paper introduces Rational Inattention-Additive Random Utility Models (RI-ARUMs), a novel class of econometric models based on Matˇejka and McKay (2015) that incorporates additive unobserved…

Key concepts

Rational Inattention
This concept describes a decision-maker's tendency to focus their attention on only a subset of available options rather than considering all possibilities. This leads to making decisions based on incomplete information, as the agent cannot process every piece of information simultaneously.
Additive Random Utility Models (ARUMs)
These are econometric models that extend standard choice models by incorporating unobserved heterogeneity in three key areas: the payoffs themselves, the cost of obtaining information, and the decision-maker's prior beliefs. This allows for a more flexible analysis of complex behavioral patterns.
Nonparametric Identification
This is a method used to identify unknown parts of a model without assuming specific functional forms for those unknown components. The paper demonstrates that under certain conditions, key parameters like the conditional prior and realized payoffs can be identified even if their underlying distributions are completely unknown.
Conditional Choice Probabilities (CCPs)
CCPs represent the probability that a decision-maker chooses one option over another given their current knowledge. The paper uses convex optimization theory to show how optimal CCPs are determined, which helps in deriving rules for identifying model components.

Terminology

Summary

This paper introduces Rational Inattention-Additive Random Utility Models (RI-ARUMs), a novel class of econometric models based on Matˇejka and McKay (2015) that incorporates additive unobserved heterogeneity to analyze behavioral biases in decision-making processes The paper shows that some systematic components of the model are nonparametrically identified even if the distributions of unobserved variables and the prior of decision makers are both unknown to the Econometrician

Model Framework and Components

The authors propose a general class of RI-ARUMs as an econometric toolkit for analyzing empirical behavioral biases The model extends the classic representative agent model by allowing Decision Makers (DMs) to have knowledge about parts of the uncertain payoffs, replacing the unconditional prior with a conditional one based on existing knowledge This results in three key components modeled additively in observed and unobserved heterogeneity: The conditional prior of the payoffs, the cost of information, and the realised payoffs The model contains a special case corresponding to the classic full-information ARUM or a degenerate prior

Theoretical Insights into Choice Probabilities

The authors provide theoretical new insights into features and properties of rational inattention discrete choice models by analyzing conditional choice probabilities (CCPs) using convex optimization theory They note that optimal CCPs may be situated on the boundary of the probability simplex, which is important for subsequent identification analysis The optimal signal rule is derived by maximizing an objective function involving expected payoffs and information cost The resulting optimal CCPs are characterized by a multinomial mixed–logit model where the integrated CCPs enter as alternative-specific weights

Nonparametric Identification Results

The paper develops nonparametric identification results for the unknown conditional prior, cost of information, and realized payoffs under four scenarios regarding unobserved heterogeneity In the most benevolent sub-scenario—where DMs have correct beliefs about payoffs—strong identification results are achieved In a more testing environment where the Econometrician has no knowledge of the DM’s prior, only certain features of the model can be identified The identification results extend existing nonparametric identification results for ARUMs to allow for rational inattention

Identification with Unobserved Heterogeneity

When unobserved heterogeneity is added to all three components (payoffs, information cost, and prior), the authors rely on the identity involving the composite mapping Π which is integrated with respect to the distribution of unobserved variables Fε They show that for certain components like u1, identification can be achieved under specific assumptions by leveraging properties of this integrated mapping However, identification of u0 in the case where the prior is completely unknown and unrestricted remains difficult without very strong restrictions on the deviation of subjective beliefs from the physical law

Identification Strategies for Structural Functions

The paper presents general results for identifying composite functions of the form Φ(w, z) = Ψ(ν(w), z) under increasing number restrictions on the model The final identification result shows that under specific support conditions and normalisations, structural functions like u0 are identified on certain index regions The analysis demonstrates that while some components like u1 can be identified even with unobserved heterogeneity, identification of u0 is much more challenging unless the prior is not too badly misspecified Finally, the paper concludes that identifying the DM’s prior πV1 remains largely unachievable without very strong restrictions

Conclusion

Based on the discrete choice rational inattention model, a flexible class of econometric models is developed with a comprehensive identification analysis showing what can and cannot be identified under different scenarios The authors establish that functions u0, u1, and λ are identified even when unobserved heterogeneity is present under suitable assumptions However, identifying the prior πV1 remains largely unachievable without very strong restrictions

References

Allen R. and Rehbeck J. p. 2

Where the paper says it

(2010), and Matzkin (2012) and Matzkin (2019). Finally, Allen and Rehbeck (2019) provides identification results for an extended version of ARUM that allows for perturbations of the random utilities, while Allen and Rehbeck (2025) identify random-coefficient distributions in latent utility models using restrictions on the mean demand mapping.

p. 2
Where the paper says it

(2010), and Matzkin (2012) and Matzkin (2019). Finally, Allen and Rehbeck (2019) provides identification results for an extended version of ARUM that allows for perturbations of the random utilities, while Allen and Rehbeck (2025) identify random-coefficient distributions in latent utility models using restrictions on the mean demand mapping.

(2019) Identification with additively separable heterogeneity. p. 2
Where the paper says it

Related results for discrete choice with multidimensional unobserved heterogeneity include Matzkin (2007a), Berry and Haile (2009), Briesch et al. (2010), and Matzkin (2012) and Matzkin (2019).

Econometrica, 87(3):1021–1054

Allen R. and Rehbeck J. (2025) Identification of random coefficient latent utility models. p. 2

Where the paper says it

(2010), and Matzkin (2012) and Matzkin (2019). Finally, Allen and Rehbeck (2019) provides identification results for an extended version of ARUM that allows for perturbations of the random utilities, while Allen and Rehbeck (2025) identify random-coefficient distributions in latent utility models using restrictions on the mean demand mapping.

Quantitative Economics, 16(4):1223–1265

An Y., Hu Y., and Xiao R. (2021) Dynamic decisions under subjective expectations: A structural analysis. p. 3

Where the paper says it

Similarly, An et al. (2021) study dynamic discrete-choice models where agents have subjective and potentially incorrect transition beliefs.

Journal of Econometrics, 222(1):645–675

Berry S. T. and Haile P. p. 2

Where the paper says it

(2013), Fox and Gandhi (2016), and Fox (2021). Related results for discrete choice with multidimensional unobserved heterogeneity include Matzkin (2007a), Berry and Haile (2009), Briesch et al.

p. 2
Where the paper says it

(2013), Fox and Gandhi (2016), and Fox (2021). Related results for discrete choice with multidimensional unobserved heterogeneity include Matzkin (2007a), Berry and Haile (2009), Briesch et al.

p. 2
Where the paper says it

(2013), Fox and Gandhi (2016), and Fox (2021). Related results for discrete choice with multidimensional unobserved heterogeneity include Matzkin (2007a), Berry and Haile (2009), Briesch et al.

A. (2009) Nonparametric identification of multinomial choice demand models with heterogeneous consumers. p. 3
Where the paper says it

Liao (2024) considers partial identification of a Berry–Levinsohn–Pakes (BLP) demand model, where consumers are rationally inattentive. For related nonparametric identification results in differentiatedproducts demand models using market-level and individual-level data, respectively, see Berry and Haile (2014) and Berry and Haile (2024).

Discussion paper, Cowles Foundation, Yale University

Berry S. T. and Haile P. A. (2014) Identification in differentiated products markets using market level data. p. 3

Where the paper says it

For related nonparametric identification results in differentiatedproducts demand models using market-level and individual-level data, respectively, see Berry and Haile (2014) and Berry and Haile (2024). The framework of these papers is different from ours since in theirs the Econometrician observes aggregate market shares across different markets, while we consider the case of individual choices in a single population.

Econometrica, 82(5):1749–1797

Berry S. T. and Haile P. A. (2024) Nonparametric identification of differentiated products demand using micro data. p. 3

Where the paper says it

Liao (2024) considers partial identification of a Berry–Levinsohn–Pakes (BLP) demand model, where consumers are rationally inattentive. For related nonparametric identification results in differentiatedproducts demand models using market-level and individual-level data, respectively, see Berry and Haile (2014) and Berry and Haile (2024).

Econometrica, 92(4):1135–1162

Bertoli S., Moraga J. F.-H., and Guichard L. (2020) Rational inattention and migration decisions. p. 1

Where the paper says it

(2020), Brown and Jeon (2024), Joo (2023), Natan (2025), and Turlo et al. However, there is precious little theoretical work on the econometrics of rational inattention discrete choice models.

Journal of International Economics, 126:103364

Briesch R., Chintagunta P., and Matˇejka R. p. 0

Where the paper says it

Identification of Rational Inattention–Additive Random Utility Models Dennis Kristensen∗ Ruolei Zhu† November 13, 2026 Abstract This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matˇejka and McKay (2015) with additive unobserved heterogeneity incorporated.

L. (2010) Nonparametric discrete choice models with unobserved heterogeneity. p. 2
Where the paper says it

Related results for discrete choice with multidimensional unobserved heterogeneity include Matzkin (2007a), Berry and Haile (2009), Briesch et al. (2010), and Matzkin (2012) and Matzkin (2019).

Journal of Business & Economic Statistics, 28(2):291–307

Brown Z. Y. and Jeon J. p. 1

Where the paper says it

(2020), Caplin et al. (2020), Brown and Jeon (2024), Joo (2023), Natan (2025), and Turlo et al.

(2024) Endogenous information and simplifying insurance choice. p. 2
Where the paper says it

However, there are two important differences: First, the mixture weights are endogenous and depend on the underlying prior in a highly complex manner. In contrast, the distribution of random coefficients is primitive and the conditional logit specification does not contain these endogenous unconditional choice probabilities.

Econometrica, 92(3):881–911

Caplin A., Csaba D., Leahy J., and Nov O. p. 0

Where the paper says it

14752v1 [econ. EM] 10 Nov 2026

(2020) Rational inattention, competitive supply, and psychometrics. p. 1
Where the paper says it

(2020), Brown and Jeon (2024), Joo (2023), Natan (2025), and Turlo et al. However, there is precious little theoretical work on the econometrics of rational inattention discrete choice models.

The Quarterly Journal of Economics, 135(3):1681–1724

Caplin A., Dean M., and Leahy J. (2019) Rational inattention, optimal consideration sets, and stochastic choice. p. 1

Where the paper says it

Under rational inattention, DMs are uncertain about the payoffs of the available alternatives and gathering information about these is costly. DMs then have to decide on the optimal trade-off between expected utility and information cost, resulting in stochastic choice behaviour also from the DMs’ perspective.

The Review of Economic Studies, 86(3):1061–1094

Caplin A., Dean M., and Leahy J. (2022) Rationally inattentive behavior: Characterizing and generalizing shannon entropy. p. 5

Where the paper says it

Upon receiving signal S = s from a given signal rule, the DM updates her beliefs and forms a posterior of V given S = s and V0 = v0 denoted PV |S,V0 (·|s, v0), which enables her to compute expected payoffs EPV |S,V0 [Vj |S = s, V0 = v0] for j ∈ J . However, there is an information cost λ·I(PS|V,V0 ) of receiving a given type of signal, where I(PS|V,V0 ) quantifies the amount of information she receives, as measured by Shannon mutual information, I(PS|V,V0 |V0 = v0) = H(πV |V0 (·|v0)) − EPS|V0 H(PV |S,V0 (·|S, v0)) , with H(·) denoting the Shannon entropy and PS|V0 (s|v0) := R V PS|V,V0 (s|v, v0)dπV |V0 (v|v0) is the marginal distribution of S|V0 = v0.

Journal of Political Economy, 130(6):1676–1715

Chen X., Hansen L. P., and Hansen P. G. (2024) Robust inference for moment condition models without rational expectations. p. 3

Where the paper says it

Chen et al. (2024) consider models where agents’ subjective probabilities differ from the DGP and study how this affects identification and construct robust confidence sets for set-identified functionals.

Journal of Econometrics, 243(1–2):105653

Dunker F., Hoderlein S., and Kaido H. (2013) Random coefficients in static games of complete information. p. 2

Where the paper says it

However, there are two important differences: First, the mixture weights are endogenous and depend on the underlying prior in a highly complex manner. In contrast, the distribution of random coefficients is primitive and the conditional logit specification does not contain these endogenous unconditional choice probabilities.

Working paper, Centre for Microdata Methods and Practice (cemmap)

Fosgerau M. and Kristensen D. p. 0

Where the paper says it

Identification of Rational Inattention–Additive Random Utility Models Dennis Kristensen∗ Ruolei Zhu† November 13, 2026 Abstract This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matˇejka and McKay (2015) with additive unobserved heterogeneity incorporated.

(2020) Identification of a class of index models: A topological approach. p. 1
Where the paper says it

This is done under the following 1Although our class of models shares the same name as the one proposed by Fosgerau et al. (2020), ours is fundamentally different from theirs.

The Econometrics Journal, 24(1):121–133

Fosgerau M., Melo E., De Palma A., and Shum M. (2020) Discrete choice and rational inattention: A general equivalence result. p. 3

Where the paper says it

Finally, as part of our identification analysis, we develop a number of master theorems that provide general conditions for identification of composite mappings, where the ”outer” mapping is one-to-one on a set that is allowed to vary with other covariates. This result extends existing results found in the literature such as Allen and Rehbeck (2019) and Fosgerau and Kristensen (2020), and should be useful in showing identification in other complex models of discrete choice as noted earlier.

International economic review, 61(4):1569–1589

Fox J. T. (2021) A note on nonparametric identification of distributions of random coefficients in multinomial choice models. p. 2

Where the paper says it

Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al. (2013), Fox and Gandhi (2016), and Fox (2021).

Annals of Economics and Statistics, (142):305–310

Fox J. T. and Gandhi A. p. 2

Where the paper says it

Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al. (2013), Fox and Gandhi (2016), and Fox (2021).

(2016) Nonparametric identification and estimation of random coefficients in multinomial choice models. p. 2
Where the paper says it

Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al. (2013), Fox and Gandhi (2016), and Fox (2021).

RAND Journal of Economics, 47(1):118–139

Fox J. T., Kim K., Ryan S. P., and Bajari P. (2012) The random coefficients logit model is identified. p. 2

Where the paper says it

However, there are two important differences: First, the mixture weights are endogenous and depend on the underlying prior in a highly complex manner. In contrast, the distribution of random coefficients is primitive and the conditional logit specification does not contain these endogenous unconditional choice probabilities.

Journal of Econometrics, 166(2):204–212

Gautier E. and Kitamura Y. p. 2

Where the paper says it

, Matzkin (1991, 1992, 1993, 1994, 2007b) and references therein. Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al.

(2013) Nonparametric estimation in random coefficients binary choice models. p. 2
Where the paper says it

, Matzkin (1991, 1992, 1993, 1994, 2007b) and references therein. Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al.

Econometrica, 81(2):581–607

Ichimura H. and Thompson T. p. 2

Where the paper says it

, Matzkin (1991, 1992, 1993, 1994, 2007b) and references therein. Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al.

S. (1998) Maximum likelihood estimation of a binary choice model with random coefficients of unknown distribution. p. 2
Where the paper says it

However, there are two important differences: First, the mixture weights are endogenous and depend on the underlying prior in a highly complex manner. In contrast, the distribution of random coefficients is primitive and the conditional logit specification does not contain these endogenous unconditional choice probabilities.

Journal of Econometrics, 86(2):269–295

Joo J. (2023) Rational inattention as an empirical framework for discrete choice and consumer-welfare evaluation. p. 1

Where the paper says it

(2020), Brown and Jeon (2024), Joo (2023), Natan (2025), and Turlo et al. However, there is precious little theoretical work on the econometrics of rational inattention discrete choice models.

Journal of Marketing Research, 60(2):278–298

Liao M. (2024) Identification of a rational inattention discrete choice model. p. 0

Where the paper says it

Identification of Rational Inattention–Additive Random Utility Models Dennis Kristensen∗ Ruolei Zhu† November 13, 2026 Abstract This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matˇejka and McKay (2015) with additive unobserved heterogeneity incorporated.

Journal of Econometrics, 240(1):105670

Matˇejka F. and McKay A. p. 0

Where the paper says it

Identification of Rational Inattention–Additive Random Utility Models Dennis Kristensen∗ Ruolei Zhu† November 13, 2026 Abstract This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matˇejka and McKay (2015) with additive unobserved heterogeneity incorporated.

(2015) Rational inattention to discrete choices: A new foundation for the multinomial logit model. p. 2
Where the paper says it

Our model and techniques share features with this literature, but rational inattention introduces additional challenges that require new tools and results. Similar to random– coefficient ARUM, the CCPs that the DM faces in our RI-ARUM model have a mixed– logit form due to the uncertainty regarding payoffs.

American Economic Review, 105(1):272–98

Matzkin R. L. (1991) Semiparametric estimation of monotone and concave utility functions for polychotomous choice models. p. 2

Where the paper says it

, Matzkin (1991, 1992, 1993, 1994, 2007b) and references therein. Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al.

Econometrica, pages 1315–1327

Matzkin R. L. (1992) Nonparametric and distribution-free estimation of the binary threshold crossing and the binary choice models. p. 2

Where the paper says it

, Matzkin (1991, 1992, 1993, 1994, 2007b) and references therein. Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al.

Econometrica, 60:239–270

Matzkin R. L. (1993) Nonparametric identification and estimation of polychotomous choice models. p. 2

Where the paper says it

, Matzkin (1991, 1992, 1993, 1994, 2007b) and references therein. Identification of random-coefficient distributions in discrete choice models is considered by Ichimura and Thompson (1998), Gautier and Kitamura (2013), Dunker et al.

Journal of Econometrics, 58(1-2):137–168

Matzkin R. L. (1994) Restrictions of economic theory in nonparametric methods. p. 2

Where the paper says it

Our identification results extend existing nonparametric identification results for ARUMs to allow for rational inattention; see, e. , Matzkin (1991, 1992, 1993, 1994, 2007b) and references therein.

In Engle R. F. and McFadden D. p. 0
Where the paper says it

1 Introduction Many economic decision problems are discrete in nature. Additive Random Utility Models (ARUMs) (McFadden et al.

L., editors, Handbook of Econometrics, volume 4, pages 2523–2558 Elsevier

McFadden D. (1981) Econometric models of probabilistic choice. p. 0

Where the paper says it

Identification of Rational Inattention–Additive Random Utility Models Dennis Kristensen∗ Ruolei Zhu† November 13, 2026 Abstract This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matˇejka and McKay (2015) with additive unobserved heterogeneity incorporated.

Structural analysis of discrete data with econometric applications, 198272

McFadden D. et al. (1973) Conditional logit analysis of qualitative choice behavior. p. 2

Where the paper says it

However, there are two important differences: First, the mixture weights are endogenous and depend on the underlying prior in a highly complex manner. In contrast, the distribution of random coefficients is primitive and the conditional logit specification does not contain these endogenous unconditional choice probabilities.

Natan O. R. (2025) Choice frictions in large assortments. p. 1
Where the paper says it

(2020), Brown and Jeon (2024), Joo (2023), Natan (2025), and Turlo et al. However, there is precious little theoretical work on the econometrics of rational inattention discrete choice models.

Marketing Science, 44(3):593–625

Sørensen J. R.-V. and Fosgerau M. p. 1

Where the paper says it

We analyse the conditional choice probabilities (CCPs) that the DM faces as functions of the underlying primitives using tools from convex optimisation theory. As such this part of the paper shares ideas and techniques with Sørensen and Fosgerau (2022) who use similar tools in their analysis of CCPs in ARUM.

(2022) How McFadden met Rockafellar and learned to do more with less. p. 8
Where the paper says it

The first-moment restriction is imposed for convenience and can be relaxed. Following the normalisation approach of McFadden (1981) and Sørensen and Fosgerau (2022), one could instead work with a centred version of Q0 , obtained by subtracting its value at a fixed interior probability vector.

Journal of Mathematical Economics, 100:102629

Turlo S., Fina M., Kasinger J., Laghaie A., and Otter T. (2025) Discrete choice in marketing through the lens of rational inattention. p. 1

Where the paper says it

(2020), Brown and Jeon (2024), Joo (2023), Natan (2025), and Turlo et al. However, there is precious little theoretical work on the econometrics of rational inattention discrete choice models.

Quantitative Marketing and Economics, 23(1):45–104

Wang H. (2026) Estimating a rational inattention discrete choice model: A semiparametric approach. p. 0

Where the paper says it

Identification of Rational Inattention–Additive Random Utility Models Dennis Kristensen∗ Ruolei Zhu† November 13, 2026 Abstract This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matˇejka and McKay (2015) with additive unobserved heterogeneity incorporated.

Job market paper, Vancouver School of Economics, University of British Columbia, November 3, 2026

Widder D. V. (1941) The Laplace Transform. Princeton University Press, Princeton, NJ

Zhu R. (2026a) Demand estimation for rationally inattentive consumers. p. 3

Where the paper says it

In contrast, our results do not require any such restrictions. Liao (2024) considers partial identification of a Berry–Levinsohn–Pakes (BLP) demand model, where consumers are rationally inattentive.

manuscript

Zhu R. (2026b) Estimation of rational inattention–additive random utility models. p. 0

Where the paper says it

Identification of Rational Inattention–Additive Random Utility Models Dennis Kristensen∗ Ruolei Zhu† November 13, 2026 Abstract This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matˇejka and McKay (2015) with additive unobserved heterogeneity incorporated.

manuscript

Proofs of Section 2

Proof of Lemma 2.1 For each fixed v0, this is the standard Shannon rational inattention problem with prior πV V0 (·v0). p. 4

Where the paper says it

In Section 2, we introduce a novel class of conditional rational inattention discrete choice models for a given population of DMs and analyse the properties of the CCPs coming out of this class of models. In Section 3, we formulate our RI–ARUM and set up the general identification problem that the Econometrician faces.

The result therefore follows from Theorem 1 of Matˇejka and McKay (2015); see in particular their Equation (13), which gives the generalized multinomial logit form of the state-contingent choice probabilities Proof of Lemma 2.2 The reduced maximisation problem is Lemma 2 of Matˇejka and McKay (2015), applied conditional on V0 = v0. p. 7
Where the paper says it

Matˇejka and McKay (2015) and Caplin et al. We will now analyse the maximisation problem (2.

Substituting (2. p. 6
Where the paper says it

Instead, we first show that P 0∗ (v0) can be expressed as the solution to a maximisation problem implied by the original one: Replacing P 0∗ j (v0) by a fixed candidate P 0 (v0) in Equation (2. 3) and then substituting the resulting expression (2.

3) into the statecontingent choice problem yields the reduced maximisation problem (2. p. 6
Where the paper says it

In particular, given the model allows for corner solutions, we cannot characterise P ∗ j (v0, v1) as a solution to a simple set of first-order conditions. Instead, we first show that P 0∗ (v0) can be expressed as the solution to a maximisation problem implied by the original one: Replacing P 0∗ j (v0) by a fixed candidate P 0 (v0) in Equation (2.

4) with objective (2. p. 2
Where the paper says it

When these beliefs coincide with the true distribution, the prior also describes heterogeneity across realised states. Otherwise, it need not coincide with the objective 3

5). The Kuhn–Tucker conditions for this concave problem include inequalities for alternatives assigned zero unconditional probability; compare Corollary 2 of Matˇejka and McKay (2015) and Caplin et al. p. 7
Where the paper says it

5) strictly concave on the relevant simplex, cf. Matˇejka and McKay (2015) and Caplin et al.

(2019) Existence of a solution to the Shannon problem and the reduced formulation are established in Matˇejka and McKay (2015); see their Lemma 2 and the discussion following Corollary 2 Proof of Lemma 2. p. 1
Where the paper says it

by Matˇejka and McKay (2015); see also Caplin et al. (2019, 2022).

3 Let Q0 (P0; v0, λ) be the objective function corresponding to the prior π. p. 8
Where the paper says it

3 Any additive transformation of the prior πV1 on V1 to a new prior π˜V1 corresponding to a common location shift V + 1J+1C, where C ∈ R is a random variable with E [|C|] < ∞, does not affect the DM’s strategy: P 0∗ (v0, λ) ∈ arg max P0∈∆(J ) Q 0 (P 0 ; v0, λ) = arg max P0∈∆(J ) Q˜0 (P 0 ; v0, λ). where Q0 (P 0 ; v0, λ) and Q˜0 (P 0 ; v0, λ) are the objective functions corresponding to πV1 and π˜V1 , respectively.

Replace π with a new prior π that changes V1,j to V1,j +C, where C is integrable, and let Q 0 (P0; v0, λ) be the objective function with this new prior. p. 8
Where the paper says it

3 Any additive transformation of the prior πV1 on V1 to a new prior π˜V1 corresponding to a common location shift V + 1J+1C, where C ∈ R is a random variable with E [|C|] < ∞, does not affect the DM’s strategy: P 0∗ (v0, λ) ∈ arg max P0∈∆(J ) Q 0 (P 0 ; v0, λ) = arg max P0∈∆(J ) Q˜0 (P 0 ; v0, λ). where Q0 (P 0 ; v0, λ) and Q˜0 (P 0 ; v0, λ) are the objective functions corresponding to πV1 and π˜V1 , respectively.

It satisfies Q 0 (P0; v0, λ) = E [C] + Q 0 (P 0; v 0, λ), and so arg max P Q 0(P; v 0, λ) = arg max P 0 Q(P) (v 1); Proof of Lemma 2. p. 6
Where the paper says it

2), we obtain the following characterisation of P 0∗ (v0) as the maximiser of an integrated version Q0 of the original objective function Q: Lemma 2. 2 The optimal vector of integrated CCPs satisfies P 0∗ (v0) ∈ arg max P0∈∆(J ) Q 0 (P 0 ; v0), (2.

4 Write σ(v0, v1; P 0) = 1 +X J j=1 P 0 j exp((v0,j + v1,j)/λ), so that Q 0 (P 0; v 0, λ) = λ R V1 log σ(v0, v1; P 0)dπV1 (v1). p. 6
Where the paper says it

2 The optimal vector of integrated CCPs satisfies P 0∗ (v0) ∈ arg max P0∈∆(J ) Q 0 (P 0 ; v0), (2. 4) where Q 0 (P 0 ; v0) := λ Z V log X j∈J P 0 j (v0) exp (vj/λ) !

For each v1, σ(v0, v1; P 0) is affine in P 0 and strictly positive on ∆(J). p. 8
Where the paper says it

Above additive model for the conditional prior is a special case of the general framework of the previous section. In particular, the following version of the affine independence assumption of Lemma 2.

Since the logarithm is concave, P 0 -> log σ(v0, v1; P 0) is concave, and integration preserves concavity

Proofs of Section 4

Proof of Theorem 4.1 The proof proceeds as in Fosgerau and Kristensen (2020) Proof of Theorem 4. p. 3

Where the paper says it

Finally, as part of our identification analysis, we develop a number of master theorems that provide general conditions for identification of composite mappings, where the ”outer” mapping is one-to-one on a set that is allowed to vary with other covariates. This result extends existing results found in the literature such as Allen and Rehbeck (2019) and Fosgerau and Kristensen (2020), and should be useful in showing identification in other complex models of discrete choice as noted earlier.

2 Let (˜ν, Ψ) be an observationally equivalent representation. p. 4
Where the paper says it

For any vector of probabilities over J , say, (p0, p1, . , pJ ), we let p = (p1, .

Fix z ∈ Z0 and w2 ∈ W∗ 2(z). By Assumption D. p. 0
Where the paper says it

ARUMs assume that the problem is deterministic from the decision makers’ (DMs’) perspective in the sense that the DMs have full knowledge of their payoffs of each of the discrete alternatives. At the same time, ARUMs generate stochastic choice behaviour from the viewpoint of the researchers through the assumption that a component of each payoff is unobserved by them.

1(i), the reference slice R(w2, z) × one belongs to W0(z') for every z' in Z∗, and V (w2, z) is open and connected Thus on this image the construction of the transformation h and the differentiation argument in the proof of Theorem 4.2 apply

Proofs of Section 5

Proof of Theorem 5.1 Apply Theorem 4. p. 9

Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

4 with w = x1, z = (x0, x2), ν = u1 and Ψ(v1, z) = P∗ (u0(x0), v1, λ(x2)) The alternative-specific restriction 42

Proof of Theorem 5.3 With w = x0 and z = (x1, x2), Theorem 2. p. 9

Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

2(ii) implies that the outer map v0 -> Ψ(v0, z) is injective on V0,reg(λ(x2)), and hence on U0(x1, x2)

Proofs of Section 5.4

Proof of Theorem 5.4 Let (x1,π, x2,π) and X0 π be as in the theorem, and set λ = λ(x2,π). p. 9

Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

By Theorem 5. p. 9
Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

p. 9
Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

1, u1(x1,π) and λ are identified on X0 1 reg and projx2 Z0 0 respectively By Theorem 5. p. 9
Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

3, u0 is identified on S (x1, x2)∈Z0 u0 X0 reg(x1, x2), which contains X0 π Since X0 π ⊆ X 0 reg(x1, π, x2, π), the CCPs Φ(x0, x1,π, x2,π) are identified and u0(x0) ∈ V0 reg(λ) for every x0 ∈ X0 π

Proofs of Section 5.2

Proof of Theorem 5.2 Under Assumption 5. p. 0

Where the paper says it

ARUMs assume that the problem is deterministic from the decision makers’ (DMs’) perspective in the sense that the DMs have full knowledge of their payoffs of each of the discrete alternatives. At the same time, ARUMs generate stochastic choice behaviour from the viewpoint of the researchers through the assumption that a component of each payoff is unobserved by them.

1 and (5.10), u1 is identified on X1,reg, cf. p. 9
Where the paper says it

This identity in turn allows us to analyse the mapping (v0, λ) 7→ P 0∗ (v0, λ). In particularly, we will show that P 0∗ (·, λ) : V0,reg 7→ int ∆(J ) is one-to-one, where V0,reg (λ) = v0 ∈ R J : P 0∗ (v0, λ) ∈ int ∆(J ) (2.

Theorem 5.5 For any (x0, x2) in Z0 π, we can then compute/identify Φ0 π(x0, x2) given by Φ0 π(x0, x2) = P∗(u0(x0), λ (x2))

Proofs of Section 5.3

Proof of Theorem 5.3 With w = x1 and z = (x1, x2), Theorem 2. p. 9

Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

2(ii) implies that the outer map v0 -> Ψ(v0, z) is injective on V0,reg(λ(x2)), and hence on U0(x1, x2)

Proof of Theorem 5.4 Let (x1,π, x2,π) and X0 π be as in the theorem, and set λ = λ(x2, π). p. 9

Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

By Theorem 5.1, u1(x1 π) and λ are identified on X0 1 reg and projx2 Z0 0 respectively By Theorem 5. p. 9
Where the paper says it

7) = v0 ∈ R J : P 0∗ (v1|v0, λ) ∈ int ∆(J ) for all v1 ∈ R J is the set of ”regular” known payoffs, where the second equality uses Eq. The following theorem first characterises V0,reg (λ), and then analyses the behaviour of P 0∗ (v0, λ) on this set: 10

3, u0 is identified on S (x1, x2)∈Z0 u0 X0 reg(x1, x2), which contains X0 π Since X0 π ⊆ X 0 reg(x1, π, x2, π), the CCPs Φ(x0, x1,π, x2,π) are identified and u0(x0) ∈ V0 reg(λ) for every x0 ∈ X0 π

**Proofs of Section 5.

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. Rational Inattention-Additive Random Utility Models (RI-ARUMs) allow econometric models to be built as a general blueprint for taking complex structural representative agent models to data, incorporating observed and unobserved heterogeneity into the primitives of the structural model.

  2. The system can perform nonparametric identification of unknown components like conditional prior, cost of information, and realized payoffs even if distributions are unknown to the econometrician under specific conditions.

  3. The improved system can identify alternative-specific covariates u1(x1) and information costs lambda(x2) using finite-difference log-odds formulas or choice probabilities at two covariate values without requiring prior identification of u1.

  4. The system can identify the common variables u0 and lambda by assuming the prior is known, allowing for the inversion of an integral equation derived from optimal integrated conditional choice probabilities.

  5. The improved system can identify unobserved heterogeneity in all three components (u0, u1, and lambda) under certain assumptions regarding the deviation of subjective beliefs from physical laws.

Abstract

This paper introduces a novel class of econometric models, called Rational Inattention-Additive Random Utility Models. It is based on the discrete choice rational inattention model of Matvejka and McKay (2015) with additive unobserved heterogeneity incorporated. We show that some of the systematic components of the primitives of the model are nonparametrically identified even if the distributions of unobserved variables and the prior of decision makers are both unknown to the Econometrician. If additional information about the prior is known to the Econometrician, even stronger identification results become available.

Related papers