Irrational Random Utility Models
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: "Irrational Random Utility Models".
Title and authors: Ruth: We’re starting with the title and who wrote this, "Irrational Random Utility Models". It sounds very focused on a specific type of choice behavior >
Warren: The authors are Daniele Calipari and Henrik Petri. They seem to be coming from a background where they can bridge the gap between theoretical models and actual data analysis >
Iris: I've looked at their setup, and they introduce these new bounds that are analogous to the relationship between joint and marginal distributions, but they call them correlation bounds >
Ruth: That’s what I’m seeing. They are using these correlation bounds to define when a stochastic choice function is actually an irrational random utility model >
Warren: So, instead of just looking at the aggregate numbers, the paper suggests we need to check if those underlying preference relationships satisfy these specific correlation bounds for it to be considered irrational >
Iris: That’s right. They show that a stochastic choice function is an RNUM if and only if it satisfies these correlation bounds >
The paper's summary: Ruth: So, what’s the core message here about this paper? It’s about showing that even though we think RUMs are the gold standard for rational choice, they aren't the only way to explain aggregate data >
Warren: They show that aggregate choices can be stochastically rational even if every single person in that population is behaving irrationally, as long as their irrationalities are sufficiently uncorrelated >
Iris: The paper introduces the concept of I-RUMs, which are RUMs that display this property where aggregate data looks rational despite individual irrationality >
Ruth: And to capture this intuition, they introduce a measure called I rho, which they define as the maximal fraction of irrational decision-makers in any RCM representation of that model >
Warren: That I rho is important because it tells you the maximum level of irrationality you can have while still maintaining that aggregate data looks stochastically rational >
The paper's improvements: Iris: One improvement they bring in is characterizing Irrational Random Utility Models as those that are both RUM and RNUM, and they show this only exists if the probability distribution on preferences is sufficiently spread out >
Ruth: That’s a big condition. They say specifically, for a RUM with support on two preference distributions P1 and P2 to be an I-RUM, the probabilities must be exactly one half for each distribution, so mu(P1) equals mu(P2) which is one over two >
Warren: That makes sense mathematically. It’s like they have to be balanced in some way for this cancellation of irrationalities to happen properly >
Conclusion: Ruth: So wrapping up, the main implication here is that we might not need people to be perfectly rational at the individual level if their mistakes cancel out when you look at large groups >
Warren: And they give us a way to check if aggregate choice functions actually come from a rational population by seeing if they satisfy those correlation bounds, which is Theorem three point four >
Iris: And for anyone trying to test the limits of this, Proposition six point one is useful because it connects having an RCM representation with having an alpha-RNUM representation for some positive alpha >
Ruth: I think the final point they make about approximation being possible is interesting, showing that any RUM can be gotten close to an I-RUM as you increase the number of alternatives >
Warren: So, it's about mapping out the boundary between what we assume is rational and what might actually be happening at the individual level in these models called Irrational Random Utility Models >
WZB Berlin · University of Gothenburg
econ.TH
Submitted: 2024-03-15
Updated: 2026-11-16
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: The Random Utility Model (RUM) is shown to be observationally equivalent to a population of irrational decision-makers if their preferences are sufficiently uncorrelated…
Key concepts
- Random Utility Model (RUM)
- A model describing choices where the utility of an option depends on a random variable. The paper questions if this model necessarily implies rational choice behavior, suggesting it might be compatible with irrational preferences under certain conditions.
- Random Non-Utility Model (RNUM)
- A contrast to RUMs, RNUMs are characterized as a specific type of Random Choice Model (RCM). They are defined by satisfying certain correlation bounds, which helps distinguish them from standard utility-based models.
- Irrationality Measure I(ρ)
- This concept quantifies how 'irrational' a specific RUM ($ ho$) is. It equals the maximum fraction of irrational decision-makers that could possibly be present in any Rational Choice Model (RCM) representation of that RUM, indicating the model's susceptibility to irrationality.
- Correlation Bounds
- These are mathematical constraints used to define and characterize RNUMs and I-RUMs. They highlight the relationship between rational and irrational choice functions, serving as a necessary condition for certain models to exhibit irrational representations.
Terminology
Summary
The Random Utility Model (RUM) is shown to be observationally equivalent to a population of irrational decision-makers if their preferences are sufficiently uncorrelated, which questions the rational interpretation of RUMs <ref:2403. p. 1 3
In a nutshell, a population of rational decision-makers is, in aggregate, observationally equivalent to a population of irrational
ones if, and only if, their preferences are sufficiently uncorrelated. The paper begins with the characterization of irrational representations, which, in seemingly sharp contrast to RUMs, we call random non-utility models (RNUMs).Where the paper says it
Characterization of Irrational Representations
The paper introduces random non-utility models (RNUMs) as a contrast to RUMs, characterizing them as an RCM with support in C excluding P <ref:2403.10208#pg10>. A key observation is that the relationship between ρ and µ is analogous to the one between joint and marginal distributions and is governed by the Fréchet bounds <ref:2403. p. 8 Second, although
the latter result is a corollary (being the intersection of RUMs and RNUMs), there are some
important differences between the two results that we highlight in section 4. A key observation in our characterization of RNUMs, is that the relationship between ρ
and µ is analogous to the one between joint and marginal distributions and, as a result,
it is governed by the Fr´echet bounds. We will apply a variation of
the Fr´echet bounds that we call ”correlation bounds”. This is to highlight the role of the
correlation between rational and irrational choice functions that will be soon apparent (see
lemma 3.Where the paper says it
Where the paper says it
Main Results on I-RUMs
Theorem 3.4 states that a stochastic choice function is an RNUM if and only if it satisfies the correlation bounds <ref:2403. p. 1 This result is
interesting per s´e as it opens the door to irrational representation theorems in which aggregate choices can be rationalized by some well-known class of (behavioral) choice functions. Then, in our main theorem, we overlap rational (RUM) and irrational (RNUM) representations to characterize the set of I-RUMs using the correlation bounds. 3
In a nutshell, a population of rational decision-makers is, in aggregate, observationally equivalent to a population of irrational
ones if, and only if, their preferences are sufficiently uncorrelated. The paper begins with the characterization of irrational representations, which, in seemingly sharp contrast to RUMs, we call random non-utility models (RNUMs). Introduction
The random utility model (RUM) is the most renowned model of stochastic choice and
it is the leading notion of stochastic rationality in economics (McFadden & Richter (1990),
McFadden (2006)): choice probabilities satisfy the RUM hypothesis if they result from the
aggregation of the choices of rational - preference maximizing - decision-makers. These
aggregate choices are normally referred to as stochastically rational.Where the paper says it
Where the paper says it
Where the paper says it
Comparative Statics on Irrationality
The paper introduces the concept of I(ρ) as a measure of irrationality, which equals the maximal fraction of irrational decision-makers in any RCM representation <ref:2403.10208#pg21>. Proposition 5.6 shows that max µ∈∆(P) ρ=ρµ µ(C / P) = I(ρ), meaning I(ρ) is the maximal fraction of irrational decision-makers in any RCM representation µ of ρ <ref:2403. p. 9 5 −
x
2
0
cs 0. 5 −
x
2
0
cf 1 0 0
sf 0 1 0
First, note that any RCM representation µ of this data must be such that the fraction
of rational decision-makers is at least x, more specifically, with preference cP1sP1f. However,
they are conceptually different. The requirement supp(µ) ⊆ P implies that each decision-maker is rational
throughout all choices, while as shown in Example 1, the same ρ may also be represented by irrational
decision-makers.Where the paper says it
Where the paper says it
Limits of Non-Falsifiability
The paper discusses the non-falsifiability of the rational interpretation of RUMs, showing that almost all RUMs are compatible with a fraction of irrational individuals <ref:2403.10208#pg7>. Proposition 6.1 establishes that for a RUM ρ, having an RCM representation µ with µ(C / P) > 0 is equivalent to having an α-RNUM representation for some α > 0 <ref:2403. p. 1 This result is
interesting per s´e as it opens the door to irrational representation theorems in which aggregate choices can be rationalized by some well-known class of (behavioral) choice functions. Then, in our main theorem, we overlap rational (RUM) and irrational (RNUM) representations to characterize the set of I-RUMs using the correlation bounds. First, we introduce a class of representations in
which only a fraction α of decision-makers are irrational (α-RNUMs). Leveraging this result,
we show that almost all RUMs can be represented by a population where a strictly positive
fraction α of decision-makers are irrational, and in fact, that all RUMs with full support
(e.Where the paper says it
Where the paper says it
Asymptotic Approximation
The paper concludes by showing that all RUMs can be approximated arbitrarily close by an I-RUM when the set of alternatives X is sufficiently large <ref:2403.10208#pg24>. Proposition 6.4 provides a bound on this distance, stating that for every RUM ρ there is an I-RUM ρ′ with ρ - ρ' ≤ 1/2 X − X − 1 <ref:2403.10208#pg42>. This bound decreases exponentially as the number of alternatives increases <ref:2403. p. 0 ∗ We would like to thank the following for helpful comments and suggestions: Dilip Ravindran, Victor
Aguiar, Jose Apesteguia, Miguel Ballester, Christian Basteck, Valentino Dardanoni, Tilman Fries, Gilat
Levy, Paola Manzini, Marco Mariotti, Yusufcan Masatlioglu, Ran Spiegler, Ran Shorrer, J¨org Stoye, and
Ferdinand Vieider; the participants of the 10th BRIC conference 2024, and internal seminars at the WZB
and Humboldt University. 1
arXiv:2403.Where the paper says it
Discussion on Individual vs Stochastic Rationality
The analysis reveals a tension between stochastic and individual rationality, showing that some non-rationalizable stochastic choice functions may need more rational choice functions than some rationalizable ones do <ref:2403.10208#pg25>. This is illustrated by an example where a stochastic choice function with concentrated probability mass that is not a RUM implies both a RUM and an RNUM representation do not exist, and that both rational and irrational decision-makers are present <ref:2403. p. 2 Knowing only the aggregate choices,
she is oblivious to the individual rationality of the decision-makers, and so she searches
for an irrational interpretation similar to the one in Example 1. We next argue that an
irrational representation does not exist.Where the paper says it
Model Specific Results
For the Logit Model, it is shown that a sufficient condition for a Luce stochastic choice function to have an irrational representation is u(a) ≤ (K − 2)u(b) for all a, b ∈ X <ref:2403.10208#pg29>. This condition intuitively shows that even the strictest notion of stochastic rationality is susceptible to irrational representations whenever the utilities are not sufficiently far apart <ref:2403.10208#pg29>. The paper also extends these arguments to other aligned influential models such as the additive perturbed utility model and the simple scalability model <ref:2403. p. 0 IRRATIONAL RANDOM UTILITY MODELS∗
DANIELE CALIARI† & HENRIK PETRI‡
Abstract. The Random Utility Model (RUM) is the leading model to represent the aggregate choices of a heterogeneous population of preference maximizers.Where the paper says it
Auxiliary Results
The appendix provides auxiliary results, including Carathéodory’s theorem for convex analysis <ref:2403.10208#pg32>, and the characterization of RUMs using Block-Marschak polynomials <ref:2403. p. 5 This paper nicely complements our analysis
for two reasons: (i) its focus on identification instead of falsifiability; and (ii) on rational (or
minimally irrational) decision-makers. We also contribute to the longstanding literature on the characterizations of the RUM
(Block & Marschak (1960), Falmagne (1978), Barber´a & Pattanaik (1986), McFadden &
Richter (1990), Gilboa (1990), Fiorini (2004), McFadden (2006), Stoye (2019)), and its
special cases such as single-crossing RUM (Apesteguia et al. By considering irrational decision-makers, our main result shows that the common interpretation of stochastically rational data is sometimes undermined. This observation questions the falsifiability of the rational interpretation of the RUM hypothesis, i. ∗ We would like to thank the following for helpful comments and suggestions: Dilip Ravindran, Victor
Aguiar, Jose Apesteguia, Miguel Ballester, Christian Basteck, Valentino Dardanoni, Tilman Fries, Gilat
Levy, Paola Manzini, Marco Mariotti, Yusufcan Masatlioglu, Ran Spiegler, Ran Shorrer, J¨org Stoye, and
Ferdinand Vieider; the participants of the 10th BRIC conference 2024, and internal seminars at the WZB
and Humboldt University. 1
arXiv:2403. This result is
interesting per s´e as it opens the door to irrational representation theorems in which aggregate choices can be rationalized by some well-known class of (behavioral) choice functions. Then, in our main theorem, we overlap rational (RUM) and irrational (RNUM) representations to characterize the set of I-RUMs using the correlation bounds. This result is
interesting per s´e as it opens the door to irrational representation theorems in which aggregate choices can be rationalized by some well-known class of (behavioral) choice functions. Then, in our main theorem, we overlap rational (RUM) and irrational (RNUM) representations to characterize the set of I-RUMs using the correlation bounds. (2003): ”revealed preference tests are unlikely to reject the integrability conditions [aggregate rationality] for such data but it is not clear that we would wish to characterize them as the outcome of
a ’rational’ procedure. ”
5A RUM with full support has strictly positive probability mass on every preference. ∗ We would like to thank the following for helpful comments and suggestions: Dilip Ravindran, Victor
Aguiar, Jose Apesteguia, Miguel Ballester, Christian Basteck, Valentino Dardanoni, Tilman Fries, Gilat
Levy, Paola Manzini, Marco Mariotti, Yusufcan Masatlioglu, Ran Spiegler, Ran Shorrer, J¨org Stoye, and
Ferdinand Vieider; the participants of the 10th BRIC conference 2024, and internal seminars at the WZB
and Humboldt University. 1
arXiv:2403. Leveraging this result,
we show that almost all RUMs can be represented by a population where a strictly positive
fraction α of decision-makers are irrational, and in fact, that all RUMs with full support
(e. Logit model) have this property. We, therefore, go
on to show that the fraction of irrational decision-makers that are compatible with the RUM
hypothesis is larger the less constrained the set of possible irrational behaviors is. These
results have important consequences for empirical applications where irrational behavior is
unlikely to be extreme (i. Hence, more than 49% of decision-makers must choose s in cs, which is
incompatible with an aggregate share of 49% choosing s in cs. By considering irrational decision-makers, our main result shows that the common interpretation of stochastically rational data is sometimes undermined.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Concluding Remarks
In concluding remarks, the authors touch on a different and more positive viewpoint of their paper borrowed from Becker (1962) <ref:2403.10208#pg32>. They show that rational aggregate choices may be individually fully irrational but also that the aggregation of individually irrational behaviors may well be stochastically rational and that individuals with correlated mistakes can never be translated into rational aggregate behavior <ref:2403. p. 2 These well-known irrational behaviors are summarized by the
choices c1 and c2 in the table. A population in which 1/2 of the decision-makers choose
according to c1 and c2 will induce the same aggregate data as the original population of
rational decision-makers. 3
In a nutshell, a population of rational decision-makers is, in aggregate, observationally equivalent to a population of irrational
ones if, and only if, their preferences are sufficiently uncorrelated. The paper begins with the characterization of irrational representations, which, in seemingly sharp contrast to RUMs, we call random non-utility models (RNUMs).Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved systems can determine if aggregate choice probabilities result from a population of rational preference maximizers by checking if they satisfy correlation bounds, as stated in Theorem 3.4:
A stochastic choice function rho is an RNUM if and only if it satisfies the correlation bounds.
-
Improved systems can characterize the set of irrational representations by identifying conditions under which an aggregate choice function has a random non-utility model (RNUM) representation, which involves checking for the satisfaction of correlation bounds (Theorem 3.4).
-
Improved systems can distinguish between stochastic choice functions that are rationalizable and those that are not, based on whether they satisfy the correlation bounds:
A stochastic choice function rho is an RNUM if and only if it satisfies the correlation bounds.
-
Improved systems can estimate the maximal fraction of irrational decision-makers compatible with a given aggregate data by calculating I(rho), defined as
I(rho) = 1 - max P in P C rho P,
which equalsthe maximal fraction of irrational decision-makers in any RCM representation mu of rho.
-
Improved systems can assess the limits of non-falsifiability for RUMs by checking Proposition 6.1, which states that "rho has an RCM representation mu with mu(C C P) > 0
is equivalent to
rho has an alpha-RNUM representation for some alpha > 0." -
Improved systems can quantify the distance between a given stochastic choice function and the set of Irrational Random Utility Models (I-RUMs) using Proposition 6.4: "For every RUM rho there is an I-RUM rho' with rho - rho' <= 1/2 X - X - 1."
-
Improved systems can identify the necessary and sufficient conditions for a stochastic choice function to be an I-RUM by checking the correlation bounds:
Theorem 4.3. Let rho be a RUM, then it is an I-RUM if and only if it satisfies the correlation bounds.
-
Improved systems can detect when the rational interpretation of the RUM hypothesis may fail even when aggregate choices are stochastically rational, by noting that
some non-rationalizable stochastic choice functions may need more rational choice functions than some rationalizable ones do.
Abstract
The Random Utility Model (RUM) is the leading model to represent the aggregate choices of a heterogeneous population of preference maximizers. We show that if (and only if) preferences are sufficiently uncorrelated, RUM choices can also be generated by a population of decision makers who do not maximize any preference. In proving this result, we also characterize the general class of choices generated by such irrational populations, with applications beyond the RUM framework. We discuss the relevance of our results for the falsifiability of the rational interpretation of RUMs, the inference of individual rationality from aggregate choices, and the nature of welfare judgments.