SSRN 4508227

Choice Models and Permutation Invariance: Demand Estimation in
Differentiated Products Markets

Amandeep Singh* 1 , Ye Liu†1 , and Hema Yoganarasimhan‡1
1
University of Washington
February 16, 2024
Abstract
Choice modeling is at the core of understanding how changes to the competitive landscape affect
consumer choices and reshape market equilibria. In this paper, we propose a fundamental characteri-
zation of choice functions that encompasses a wide variety of extant choice models. We demonstrate
how non-parametric estimators like neural nets can easily approximate such functionals and overcome
the curse of dimensionality that is inherent in the non-parametric estimation of choice functions. We
demonstrate through extensive simulations that our proposed functionals can flexibly capture underlying
consumer behavior in a completely data-driven fashion and outperform traditional parametric models.
As demand settings often exhibit endogenous features, we extend our framework to incorporate esti-
mation under endogenous features. Further, we also describe a formal inference procedure to construct
valid confidence intervals on objects of interest like price elasticity. Finally, to assess the practical appli-
cability of our estimator, we utilize a real-world dataset from Berry et al. (1995). Our empirical analysis
confirms that the estimator generates realistic and comparable own- and cross-price elasticities that are
consistent with the observations reported in the existing literature.
Keywords: Choice Models, Demand Estimation, Permutation Invariance, Set Functions, Neural net-
works
* Email: [email protected]
†
Email: [email protected]
‡
Email: [email protected]
Electronic copy available at: https://ssrn.com/abstract=4508227

1 Introduction
Demand estimation is a critical component in fields of marketing, operations, and economics, enabling
practitioners to model consumer choice behavior and understand how consumers react to changes in a mar-
ket. This understanding helps policymakers and businesses to make informed decisions, whether it be about
launching new products, adjusting pricing strategies, or analyzing the consequences of mergers (Nevo, 2000;
Petrin, 2002; Nevo, 2003; Wollmann, 2018). Over the years, various approaches, both parametric and non-
parametric, have been developed to address the complexities inherent in demand estimation.
Parametric methods, based on logit or probit assumptions, often require strong assumptions about the
underlying choice process, limiting their ability to capture the true complexity of consumer preferences.
As such, the substantive and policy (counterfactual) implications from such models could be largely biased
and/or misleading. Nevertheless, they have remained popular because of a few reasons – (1) simplicity,
interpretability, and scalability to a large product set, (2) the ability to model counterfactual demand in
situations involving new product introductions or mergers, and (3) the ability to handle endogenous product
features.
Non-parametric methods, on the other hand, offer a more flexible approach to demand estimation, al-
lowing for more nuanced representations of consumer preferences without restrictive assumptions about
the underlying distributions of observables and unobservables. However, despite their potential advantages,
non-parametric approaches face significant challenges that have limited their widespread adoption in prac-
tice. Firstly, a common limitation across all non-parametric demand estimation methods is the “curse of
dimensionality,” where the computational complexity of estimating choice functions increases exponen-
tially with the number of products. Additionally, a large subset of these models, specifically those that are
‘black box’ in nature and lack additional economic structure, are unable to perform counterfactual predic-
tions, which are crucial for important tasks such as policy simulations. Furthermore, most non-parametric
methods cannot properly handle endogenous product features. This inability to account for endogeneity and
scalability constraints, along with the drawback of limited counterfactual analysis, significantly undermines
their utility in practice, where such capabilities are often the primary objective of demand estimation.
In this paper, we make significant strides in bridging the gap between the flexibility of non-parametric
methods and the tractability of parametric models. We achieve this by introducing a fundamental character-
ization of choice models. Our work begins by considering a broad set of choice functions, encompassing
most existing choice modeling approaches in the empirical literature. We demonstrate that most choice
functions exhibit specific symmetry properties. We then leverage recent advances in computer science and
mathematics literature (for instance, Han et al. (2019); Zaheer et al. (2017); Wagstaff et al. (2019)) to char-
acterize these functions. Our characterization allows us to build on the strengths of the non-parametric
approaches (e.g., not assuming a model of consumer behavior) and overcome the challenges associated with
them. First, it addresses the curse of dimensionality in choice systems and enables the flexible estimation
of choice functions via non-parametric estimators. By leveraging the inherent permutation-invariant struc-
ture of choice models, our characterization can close the estimator’s generalization gap (for instance, see
√
Sannai and Imaizumi (2019) in the context of neural networks) by a factor of J! (where J is the num-
ber of products). Second, we recognize that real-world demand systems often contain unobserved demand

shocks correlated with observable product features, such as prices. These shocks can lead to endogeneity
issues, which can bias the estimated choice functions if not properly addressed. To tackle this challenge, we
extend our framework to accommodate endogeneity. Third, our proposed approach successfully estimates
counterfactual demand for scenarios that involve changes in the product set (e.g., new product introduction,
mergers, product exit). Finally, we build upon recent advances in automatic debiased machine learning and
provide an inference procedure for constructing valid confidence intervals on objects of interest, such as the
average effect of price.
We demonstrate the effectiveness of our proposed approach using a series of numerical simulations. We
consider a variety of data-generating processes in our simulations – multinomial logit model with linear
utility, random coefficients logit with linear utility, random coefficient logit with non-linear utilities, and
a setting where some consumers have inattention (i.e., ignore certain products in the market). Across all
these scenarios, we show that our approach can predict market shares, own-, and cross-price elasticities with
relatively high accuracy, even though we do not make any assumptions about consumer behavior. Further,
even when the underlying DGP is complex, our model can recover market shares and elasticities similar to
oracle estimators (that are assumed to know the true DGP). This allows researchers and managers to use our
approach in general-purpose situations without making ad-hoc assumptions about consumer behavior.
Next, we consider counterfactual analyses and empirically show that our model can generate accurate
counterfactuals in the case of new product introductions. Finally, we showcase the performance of the auto-
matic debiasing procedure and show that we can provide consistent inference and confidence intervals over
the average effect of price on demand across all the products. In sum, our extensive numerical simulations
cover a wide range of DGPs and show that our approach – (1) is able to accurately predict market shares
and price elasticities for a variety of scenarios, (2) can generate realistic counterfactual predictions in cases
where the product set changes, and (3) can provide inference over economic objects of interest.
Finally, to showcase the effectiveness and applicability of our approach to real-life datasets, we use
the Berry et al. (1995) automobile dataset and estimate the price elasticities using our non-parametric es-
timator (while correcting for price endogeneity). The results of our analysis align with existing literature
and demonstrate the practical utility of our approach, underscoring its potential for adoption in real-world
demand estimation settings. In summary, given the theoretical, practical, and computational advantages of
our approach, we expect it to be easily applicable to a wide variety of demand estimation problems in both
research and practice.
2 Literature Review
2.1 Parametric Discrete Choice Models
Discrete choice models play a crucial role in various fields, including economics, marketing, and operations
management, as they describe decision-making processes when individuals face multiple alternatives. These
models typically rely on a random utility maximization assumption, which can be traced back to Thurstone
(1927) and Marschak (1959).
Over time, various choice models have emerged under different specifications for the density of unob-

served utility, following the general framework in Marschak (1959). The (Multinomial) logit model, first
proposed by Luce (1959), is widely used for capturing systematic taste variance based on observed char-
acteristics across alternatives. However, the logit model assumes that error terms are independent of each
other and of the characteristics of the alternative. Additionally, the independent extreme value distribution
results in the irrelevant alternatives (IIA) property, implying proportional substitution across alternatives.
To enable more flexible substitution patterns, Generalized Extreme Value (GEV) models were intro-
duced, allowing for correlated unobserved utility across alternatives. The most commonly used GEV model
is the nested logit model (Train et al., 1987; Forinash and Koppelman, 1993). In nested logit models, the
unobserved error of alternatives within a nest is specified as correlated, while the marginal distribution of
each unobserved error remains as a univariate extreme value.
The mixed logit model (McFadden and Train, 2000), also known as random coefficient logit (RCL),
is an even more versatile model that allows for randomness in both unobserved factors and coefficients of
observed characteristics. The model was first applied by Boyd and Mellman (1980) and Cardell and Dun-
bar (1980), with random coefficients typically specified as normal or lognormal (Ben-Akiva et al., 1993;
Mehndiratta, 1996; Revelt and Train, 1998) but also other distributions such as uniform and triangular
(Greene and Hensher, 2003; Train, 2001). When random coefficients follow a mixture distribution, the
model becomes the well-known mixed logit latent class model.
Aside from logit family models (MNL, GEV, mixed logit) that have an extreme value distributed random
component, the probit model Hausman and Wise (1978) assumes a jointly normal distribution for the error
term. This allows for any pattern of substitution and can handle random taste variation. Blanchet et al.
(2016) proposed a model where substitution between alternatives is a state transition in a Markov chain,
which can approximate MNL, probit, and mixed logit models.
As the decision space grows, assuming that all alternatives are considered becomes unrealistic, leading
to research in consideration sets and consumer search (Honka et al., 2019; Jiang et al., 2021). These models
typically impose strong parametric assumptions on the decision rule, including whether decisions are simul-
taneous or sequential, the stopping rule for searches, the size of the consideration set, and the functional
form of the match value.
Further, real-world demand systems often contain unobserved demand shocks correlated with observable
product features, such as prices. These shocks can lead to endogeneity issues, which can bias the estimated
parameters if not properly addressed. Thus, various approaches have been proposed to address these issues.
For instance, Berry et al. (1995) proposed a generalized method of moments-based estimator to estimate a
random-coefficients logit model of demand using instrumental variables. Similarly, Petrin and Train (2010)
demonstrated the application of control functions to resolve endogeneity in the random coefficient logit
model of demand.
Two primary concerns related to discrete choice models are the need for correct model specification and
the distribution of unobserved factors. To address these issues, nonparametric demand estimation models
have been developed.

2.2 Nonparametric Demand Estimation
While parametric discrete choice models make assumptions about the distribution of unobserved variables
and specify functional forms for utilities, recent research has developed more flexible semi-parametric and
nonparametric approaches for demand estimation. These newer methods ease some of the restrictive as-
sumptions and improve the computational efficiency of choice models while retaining some structure.
Early semi-parametric work (Manski, 1987; Lewbel, 2000; Honoré and Kyriazidou, 2000; Abrevaya,
2000) focused on relaxing the distribution of random shocks in individual-level binary choice models. More
recent research (Khan et al., 2021; Shi et al., 2018; Pakes and Porter, 2022) has extended this approach to
relax the parametric assumptions on random shocks in individual-level multinomial choice models, allow-
ing for increased flexibility, such as individual-level fixed effects. Fox and Gandhi (2016); Briesch et al.
(2010); Allen and Rehbeck (2019); Fosgerau and Kristensen (2021); Chitla et al. (2022); Lu et al. (2023);
Wang (2023) concentrate on nonparametric identification and estimation of distributions of heterogeneous
unobservables, like random coefficients, in various demand models, relaxing assumptions about heterogene-
ity distribution. More recent work has looked at keeping the indirect utility or choice probability functions
largely unspecified while retaining parametric error terms. For instance (Bentz and Merunka, 2000; Wang
et al., 2020; Han et al., 2022; Sifringer et al., 2020; Wong and Farooq, 2021; Aouad and Désir, 2022) use
neural networks to characterize the indirect utility while retaining the logit error structure. Our work, in
comparison, provides a completely flexible mapping from observed product and consumer characteristics to
observed demand without relying on any assumptions regarding the choice making process.
Extant research in nonparametric methods aims to completely avoid any parametric assumptions to re-
move any potential source of misspecification. Berry and Haile (2014, 2020) presented the identification
results for nonparametric estimation of demand from market-level and individual-level data, respectively.
Studies by Hausman and Newey (2016); Blundell et al. (2017); Chen and Christensen (2018) focus on
individual-level data. In particular, Hausman and Newey (2016) employs a nonparametric approach to esti-
mate consumer surplus bounds, while Blundell et al. (2017) introduces a method for consistently estimating
demand functions with nonseparable unobserved taste heterogeneity, subject to the shape restriction im-
posed by the Slutsky inequality. Chen and Christensen (2018) concentrates on nonparametric instrumental
variables and inference in individual-level data. This paper, alongside Compiani (2022) and Tebaldi et al.
(2023), examines market-level data. Compiani (2022) develops a nonparametric method based on Bern-
stein polynomials, drawing on the identification result of Berry and Haile (2014). Meanwhile, Tebaldi et al.
(2023) proposes a technique for deriving nonparametric bounds on demand counterfactuals and applies it to
California’s health insurance market.
The common challenge across all extant nonparametric demand estimation work has been that the com-
putational complexity of estimating nonparametric demand models increases exponentially with the number
of products. For instance, Compiani (2022) could estimate their nonparametric model for just two products.
Similarly, Cai et al. (2022) needed data on more than 100,000 markets to reasonably estimate demand with
50 products. To put this in perspective, economic datasets are much smaller; for instance, the dataset in
Berry et al. (1995) had 150 products across only 20 markets. This curse of dimensionality has been the
primary obstacle preventing the widespread adoption of nonparametric methods in demand estimation. To

overcome this barrier, our work demonstrates how to exploit the inherent permutation invariant structure of
choice functions to break this curse of dimensionality and flexibly estimate demand in markets with large a
number of products. In addition, our work is also related to recent work in marketing Wei and Jiang (2022)
exploring the use of neural networks to estimate parameters of structural models.
Finally, our work leverages recent work in mathematics and computer science (for instance, Han et al.
(2019); Zaheer et al. (2017); Wagstaff et al. (2019)), which investigate the universal approximation of sym-
metric and antisymmetric functions, offering fundamental characterizations for functions defined on sets.
We build on this literature to characterize a general class of choice functions in demand systems.
3 Theory
3.1 Choice Models
In this section, we provide a general characterization of consumer choice functions. In particular, we focus
on a scenario where researchers have access only to aggregate market-level demand data, while individual-
level choices and characteristics remain unobserved. Aggregate demand models have been extensively stud-
ied in marketing and economics (Berry et al., 1995; Besanko et al., 1998; Sudhir, 2001; Chintagunta, 2001;
Albuquerque and Bronnenberg, 2009; Compiani, 2022), and are particularly useful when individual-level
demand data is not available.
Suppose consumers in a market t face an offer set St that can comprise any subset of Jt distinct products
({1, 2, . . . , Jt }). We use uijt to represent the index tuple {Xjt , pjt , Iit , εijt }, where Xjt ∈ Cd denotes d
non-price features belonging to some countable universe Cd ; pjt ∈ C denotes the price of the product;
Iit ∈ Cl denotes demographics of consumer i in market t, we assume there are l features and belong to
some countable universe Cl , and εijt denotes random idiosyncratic components pertinent to consumer i for
product j in market t that are not unobservable to the researcher but observable to consumers.
Definition 1 (Choice Function). Given the offer set St ⊂ {1, 2, 3, . . . , Jt }, we define a function π : {uijt :
j ∈ St } → R|St | that maps a set of index tuples {uijt }j∈St to a |St |-dimensional probability vector. Each
element in the π(·) vector represents the probability of consumer i choosing product j in market t.
Here we present a very general characterization of choice functions that maps the observable and un-
observable components of product and individual characteristics to observed choices through some choice
function π. Note that, traditionally, uijt is a scalar that represents utility in choice models. However, in our
framework, uijt does not necessarily represent utility. Further, we have not yet imposed any assumption on
π, i.e., how consumers make choices.
We now specify a set of assumptions on the model and data-generating process below.
Assumption 1 (Exogeneity). The unobserved error term εijt is independent and identically distributed (i.i.d.)
across all products. This can be expressed as follows:
P(εijt | X·t , p·t ) = P(εijt )

This assumption implies that the error term εijt is not correlated with any of the observed variables X·t and
p·t . As such, it precludes the possibility of endogenous prices and/or marketing-mix variables, as is common
in observational data. We start with the basic case with exogenous covariates in this section and later in §3.2,
we relax this assumption and allow for endogenous covariates.
Assumption 2 (Identity Independence). For any product j ∈ St and any market t, we assume the choice
function π does not depend on the identity of the product (jt). That is:
πijt ({uikt }k∈St ) = πijt (uijt , {uikt }k∈St ,k̸=j ) = π(uijt , {uikt }k∈St ,k̸=j )
This assumption implies two things: first, the functional form of the choice probability for different products
and markets is the same; second, for any market-level heterogeneity (e.g., in the distribution Ft (Iit , εijt )),
we can include them in Xjt as features. Intuitively, this assumption suggests that conditional on product and
consumer features and the unobserved error term, the choice probabilities are not functions of the identities
of the products themselves.
Assumption 3 (Permutation Invariance). The choice function π is invariant under any permutation function
σj () that rearranges the indices of the competitors of product j, such that:
πijt = π(uijt , {uiσj (k)t }k∈St ,k̸=j )
In this assumption, we state that the choice function for product j is invariant to all permutations of its
competitors. This implies that the individual’s choice for product j is not affected by the order or identity of
the other products in the market, and it only depends on the set of competitors’ characteristics.
Since researchers only observe aggregate data, we next define the aggregate demand function. In aggre-
gate demand settings, individual-level choices are not observable and only aggregate demand is observable.
It is often the case that the market-specific individual features are not observable and are assumed to be
exogenously drawn from some distribution F(mt ), where mt represents the market-level characteristics.
For the sake of notional simplicity, we let mt to be the same across all markets. One can easily incorporate
market-specific user demographics in the choice function. Thus the demand of product j in market t denoted
by πjt can be expressed as follows:
Z Z
πjt = πijt ({uikt }k∈St )dF(mt )dG(εijt ), (1)
where G(εijt ) denotes the CDF of unobserved errors εijt . Since uijt is determined by {Xjt , pjt , Iit , εijt }
and Iit , εijt are integrated out in a market. Hence, we can express πjt as a function of only the observable
product characteristics –
πjt = g(Xjt , pjt , {Xkt , pkt }k∈S,k̸=j ). (2)
Lemma 1. For any choice function that satisfies Assumption 1 and 3, the aggregate demand function is
permutation invariant.
This permutation invariance of the aggregate demand function exists because, under the exogeneity

assumption, the aggregate demand function is simply the sum (or integral) of individual choice functions that
are themselves invariant to permutation. Hence, changes to the order of competitors have no impact on the
aggregated result. When the assumption of exogeneity is not satisfied, the aggregate demand function does
not retain the permutation invariance, even though the individual-level choice function exhibits permutation
invariance. We will return to this issue in §3.2.
Our Assumptions 2 (identity independence) and 3 (permutation invariance) are fairly standard in the
choice modeling literature, although they might not always be explicitly stated as such. Table 1 summarizes
models that satisfy these assumptions. Please see Web Appendix A for detailed derivations of how these
models satisfy permutation invariance.
Table 1: Choice Models Satisfying Identity Independence and Permutation Invariance
Choice Model Literature

Multinomial Logit Model McFadden et al. (1973)
Mixed Logit Model McFadden and Train (2000)
Nested Logit Model Train et al. (1987)
Random Coefficients Nested Logit Grigolon and Verboven (2014)
Generalized Extreme Value (GEV) Model Train (2009)
Probit Model Hausman and Wise (1978)
Latent Class Logit Model Kamakura and Russell (1989)
Markov Chain Choice Model Blanchet et al. (2016)
Customer Inattention Based Models Goeree (2008), Turlo et al. (2023), Compiani
(2022) and Joo (2023)
Customer Search Models Mehta et al. (2003)
Theorem 1. For any offer set St ⊂ {1, 2, 3, . . . , Jt }, if a choice function π : {uijt : j ∈ St } → R|St | where
uijt represents the index tuple {Xjt , pjt , Iit , εijt } satisfies Assumption 1, 2 and 3, then there exists suitable
ρ, ϕ1 and ϕ2 such that
X
πjt = ρ(ϕ1 (Xjt , pjt ) + ϕ2 (Xkt , pkt )),
k̸=j,k∈St
Proof: See Web Appendix B.

This result is the generalization of the results shown in Zaheer et al. (2017) and can be shown following
similar arguments. The above result is very powerful and has two important takeaways: (i) The input space
of the choice function does not grow with the number of products in the assortment. Rather, the input space
of the choice function (i.e., ϕ1 and ϕ2 ) grows only as a function of the number of features of the products in
consideration, and (ii) the same transformations (ρ, ϕ1 , and ϕ2 ) remain valid for all offer sets, denoted by S,
irrespective of their size. This property allows us to easily simulate the demand and entry of new products
or changes in market structure, as one does with traditional parametric models. As an example, assuming
vjt is the utility of product
" j at#market t, for the multinomial
" # logit model one possible set of transformations
exp(vjt ) 0
could be ϕ1 (vjt ) = and ϕ2 (vkt ) = that generate two-dimensional vectors, and the
0 exp(vkt )

" #!
ϕ1 (vjt )
function ρ = Pϕ1 (vjt ) operates on these vectors. 1
P ϕ1 (vjt )+ ϕ2 (vkt )
k̸=j,k∈St ϕ2 (vkt )
k̸=j,k∈St
3.2 Endogenous Covariates

In this section, we relax the exogeneity assumption and handle the potential endogeneity issue that is com-
monplace in demand settings. Note that, when the price (or other product characteristics or market-mix
variables, such as promotions, correlate with unobserved variables (εijt ), Assumption 1 (exogeneity) is
compromised. As a result, it becomes infeasible to integrate out εijt in the aggregate demand function, as
we did in Equation (1). This means that the aggregate demand function loses its property of permutation
invariance with respect to the observable characteristics of competitors. To address this, we build on the
approach developed in Petrin and Train (2010) to allow for endogenous observable features. Without loss
of generality, we assume that price pjt ∈ C is the endogenous variable and all other characteristics of the
product Xjt ∈ Cd are exogenous variables. i.e.,
E[pjt · εijt ] ̸= 0 and E[Xjt · εijt ] = 0.
Given valid instruments IVjt , we can express pjt as
pjt = γ (Xjt , IVjt ) + µjt . (3)
At this point, no specific assumptions are made regarding the function γ. However, in the subsequent
inference section, we will discuss that the estimator of γ must be estimable at n−1/2 in order to construct
valid confidence intervals. Next, to address the issue of price endogeneity, we impose a mild restriction on
the space of choice functions we consider.
Assumption 4 (Linear Separability). The unobserved product characteristics can be expressed as the sum of
an endogenous (CF) and exogenous component
εijt = CF (µjt ; λ) + ε̃ijt , (4)
where E[pjt · ε̃ijt ] = 0.

This assumption implies that, after controlling for µjt using the control function CF , the endogenous
variable pjt is uncorrelated with the error term εijt in the model, thus it becomes exogenous. Then, we can
re-write the index tuple uijt as
uijt = {Xjt , pjt , CF (µjt ; λ) + ε̃ijt }, (5)

h i
such that E ε̃ijt |(Xjt , pjt , µjt ) = 0

1
P exp(vjt ) 0 P exp(vjt )
ρ ϕ1 (vjt ) + k̸=j ϕ2 (vkt ) = ρ + P = ρ =
0 k̸=j exp(vkt ) k̸=j exp(vkt )
exp(vjt )
P
exp(vjt )+ k̸=j exp(vkt )

uijt represents the index tuple {Xjt , pjt , Iit , εijt } satisfies Assumptions 2 to 4. Then under the condition of
knowing the true function (γ0 ) of γ, there exists suitable ρ, ϕ1 and ϕ2 such that
X
πjt = ρ(ϕ1 (Xjt , pjt , µjt (γ0 )) + ϕ2 (Xkt , pjt , µkt (γ0 ))),
k̸=j,k∈S
The result follows straightforwardly from the observation that after controlling for CF (µjt ; λ) the un-
observable component ε̃ is exogenous. This implies the aggregate demand function is invariant under any
permutation applied to the competitors of product j. The result demonstrates that endogeneity can be ad-
dressed by using the residuals from Equation (4) as an additional set of features along with observable
product characteristics.
3.3 Inference
This paper aims to estimate choice functions flexibly using non-parametric estimators. However, often in
social science contexts, researchers and managers are also interested in conducting inference over some
economic objects. Note that because non-parametric regression functions are estimated at a slower rate
compared to parametric regressions, it is often infeasible to construct confidence intervals directly on the
estimated π̂. However, it is generally possible to perform inference and construct valid confidence intervals
for specific economic objects that are functions of π. In this section, we will provide an example of one such
important economic object and demonstrate how to construct valid confidence intervals for it. This will be
done by leveraging the recent advances in automatic debiased machine learning as shown in the works of
Ichimura and Newey (2022); Chernozhukov et al. (2022b,a, 2021), and others. However, unlike existing
automatic debiased machine learning setups we also have to account for an additional first-stage estimator
γ̂.
In demand estimation, researchers are often interested in estimating the average effect of a price change
on the demand for a product, as it can significantly influence market dynamics, pricing strategies, and
regulatory decisions. To proceed with our analysis, let wjt = (yjt , pjt , Xjt , {pkt , Xkt }k̸=j ) and zjt =
(pjt , Xjt , {pkt , Xkt }k̸=j ) represent the variables associated with product j in market t. Here, pjt ∈ C
denotes the observed prices, Xjt ∈ Cd−1 represents other product characteristics, and yjt ∈ R refers to
the observed demand for product j in market t, such as market shares or log shares. Note that either the
observed price (pjt ) or other characteristics (Xjt ) could be endogenous. For simplicity and without loss of
generality, we focus on pjt as the endogenous variable in the following analysis.
The average effect of a price change2 can be expressed as the difference between the demand function
πjt (·; γ) evaluated at the original price pjt and at the price incremented by ∆pjt , given by the following
expression:
m(wjt , π(·; γ)) = π(pjt + ∆pjt , Xjt , {pkt , Xkt }k̸=j ); γ) − π(pjt , Xjt , {pkt , Xkt }k̸=j ); γ).
2
The expression for the average effect of a price change can be adapted to represent average price elasticity by placing the
known and fixed value of ∆pjt in the denominator.

The parameter of interest, θ0 , is the expected value of this price change effect over the true population
distribution3 of wjt , which can be calculated as:
θ0 = E[m(wjt , π(·; γ))] = E[π(pjt + ∆pjt , Xjt , {Xkt }k̸=j ; γ) − π(pjt , Xjt , {Xkt }k̸=j ; γ)].
In summary, the average effect of a price change on demand, denoted by θ0 , is calculated by evaluating
the difference between the demand function at the original price and at the price incremented by ∆pjt , and
then computing the expected value of this difference.
In practice, we estimate θ0 by computing its empirical analog using the estimated demand function π̂
and first-stage estimator γ̂, i.e.,
n
1X
θ̂ = m(wjt , π̂(zjt ; γ̂)), (6)
n
t=1
where n is the number of observations. When parametric methods are employed to estimate π̂ and γ̂, the
√ √
estimator for θ̂ is generally n-consistent, assuming that the model is correctly specified. However, n-
consistency may not hold when non-parametric estimators are used, particularly if the first-order bias does
√
not vanish at a rate of n. Irrespective of the method used to estimate π, this is often the case, as flexible
estimation of π always requires some form of regularization and/or model selection. Debiasing techniques
are required to mitigate the effects of regularization and/or model selection when learning flexible demand
models. These approaches can help improve the performance of the estimator and facilitate valid inference
with θ̂. We therefore adapt recent debiasing techniques developed in recent automatic debiased machine
learning literature (see Chernozhukov et al. (2022b)). Specifically, we will focus on problems where there
exists a square-integrable random variable α0 (z) such that ∀ ||γ − γ0 || small enough –
E[m(wjt , π(zjt ; γ))] = E[α0 (zjt )π(zjt ; γ)]

∀π with E[πjt (zjt ; γ)2 ] < ∞
By the Riesz representation theorem, the existence of such α0 (zjt ) is equivalent to E[m(wjt , π(zjt ; γ))]
being a mean square continuous functional of π(zjt ; γ). Henceforth, we refer to α0 (z) as Riesz representer
(or RR). Newey (1994) shows that the mean square continuity of E[m(wjt , πjt (zjt ; γ))] is equivalent to
the semiparametric efficiency bound of θ0 being finite. Thus, our approach focuses on regular functionals.
Similar uses of the Riesz representation theorem can be found in Ai and Chen (2007), Ackerberg et al.
(2014), Hirshberg and Wager (2020), and Chernozhukov et al. (2022b) among others. The debiasing term
in this case takes the form α(zjt )(yjt − π(zjt ; γ)). To see that, consider the score m(wjt , π(zjt ; γ)) +
α(zjt )(yjt − π(zjt ; γ)) − θ0 . It satisfies the following mixed bias property:
E[m(wjt , π(zjt ; γ)) + α(zjt )(yjt − π(zjt ; γ)) − θ0 ]

= −E [(α(zjt ) − α0 (zjt )) (π(zjt ) − yjt )] .
This property implies double robustness (Robins et al., 1994; Funk et al., 2011) of the score. That is, if either
3
We assume the data reflects the true population.
10

α(zjt ) is correctly estimated, which would mean α(zjt ) − α0 (zjt ) = 0, or π(zjt ) is correctly estimated,
implying π(zjt ) − yjt = 0, then the term (α(zjt ) − α0 (zjt ))(π(zjt ) − yjt ) will be zero. This results in the
score going to zero, thereby making the estimator consistent for θ0 . A debiased machine learning estimator
of θ0 can be constructed from this score and first-stage learners πb and αb. Let En [·] denote the empirical
1 Pn
expectation over a sample of size n, i.e., En [xi ] = n i=1 xi . We consider:
θb = En [m(wjt ; π b(zjt )(yjt − π

b) + α b(zjt ))].
The mixed bias property implies that the bias of this estimator will vanish at a rate equal to the product of
the mean-square convergence rates of α
b and π
b. Therefore, in cases where the demand function π can be
estimated very well, the rate requirements on αb will be less strict, and vice versa. More notably, whenever
√ √
the product of the mean-square convergence rates of α b and fb is larger than n, we have that n θb − θ0
converges in distribution to centered normal law N 0, E ψ0 (wjt )2 , where

ψ0 (wjt ) := m (wjt ; π0 ) + α0 (zjt ) (yjt − π0 (zjt )) − θ0 ,
as proven formally in Theorem 3 of Chernozhukov et al. (2022b). Results in Newey (1994) imply that
E ψ0 (wi )2 is a semiparametric efficient variance bound for θ0 , and therefore the estimator achieves this

bound.
Theorem 3. [Chernozhukov et al. (2021)] One can view the Riesz representer as the minimizer of the loss
function: h i
α0 = arg minE (α(zjt ) − α0 (zjt ))2
α
= arg minE α(zjt )2 − 2α0 (zjt )α(zjt ) + α0 (zjt )2

α
= arg minE α(zjt )2 − 2m(wjt ; α) ,

α
In our earlier discussions, we employed the moment function of π, whereas in Theorem 3, we focus
on the moment function of α. This shift is justified by the Riesz Representation Theorem, which implies
E[m(wjt ; π)] = E[α0 (zjt )π(zjt )]. Given that π can represent any function, substituting α for π is permis-
sible, thereby validating the transition from the second to the third line in Theorem 3. We use the above
theorem to flexibly estimate the RR. The advantage of this approach is that it eliminates the need to derive
an analytical form for the RR estimator, allowing it to be addressed as a simple computational optimization
problem.
Theorem 4. [Chernozhukov et al. (2021)] Let δn be an upper bound on the critical radius (Wainwright
(2019)) of the function spaces:
{z 7→ ζ (α(z) − α0 (z)) : α ∈ An , ζ ∈ [0, 1]} and

{w 7→ ζ (m(w; α) − m (w; α0 )) : α ∈ An , ζ ∈ [0, 1]}
and suppose that for all f in the spaces above: ∥f ∥∞ ≤ 1. Suppose, furthermore, that m satisfies the
11

mean-squared continuity property:
h 2 i 2
E m(w; α) − m w; α′ ≤ M α − α′ 2
for all α, α′ ∈ An and some M ≥ 1. Then for some universal constant C, we have that w.p. 1 − ζ :
M log(1/ζ)
α − α0 ∥22 ≤ C(δn2 M +
∥b
n
2
+ inf ∥α∗ − α0 ∥2
α∗ ∈An
The critical radius has been widely studied in various function spaces, such as high-dimensional linear
functions, neural networks, and superficial regression trees, often showing δn = O dn n−1/2 , where dn

represents the effective dimensions of the hypothesis spaces (Chernozhukov et al. (2021)). In our research,
we focus on applying Theorem 3 from an application standpoint to neural networks.
To that end, we make the following assumptions.
Assumption 5. (i) α0 (z) is bounded, (ii) ∀ ||γ − γ0 || small enough, E[(y − π0 (zjt ; γ))2 |zjt ] is bounded, and
(iii) E[m(wjt , π0 (zjt ; γ0 ))2 ] < ∞.
These assumptions are standard regularity conditions used in the automatic machine learning literature.
p p √
Assumption 6. i) ∀ ||γ − γ0 || small enough ||π̂(; γ) − π0 (; γ)|| → − 0 and ||α̂ − α0 || →
− 0; ii) n||α̂ −
p √ p
α||||(π̂(; γ) − π0 (; γ)|| →
− 0; iii) α̂ is bounded; (iv) n||γ̂ − γ0 || →
− 0
Intuitively these assumptions mean that (i) the estimator of both π and α should be consistent for values
of γ in a close enough neighborhood of γ0 . Further, it requires that the product of mean square error of α̂
√
and mean square error of π should vanish at n− rate. This can be achieved if both these terms converge
at least at n−1/4 rate. Finally, we also assume that the first stage estimator γ̂ is estimable at n−1/2 rate. This
limits the class of functions one can use to estimate γ.
Assumption 7. m(w, π) is linear in π and there is C > 0 such that
|E [m(w, π) − θ0 + α0 (z)(y − π(z; γ))]| ≤ C ∥π − π0 ∥2
Proposition 1. If Assumptions 5-7 are satisfied then for V = E[{m(w, π0 (z; γ0 )) − θ0

+α0 (z)(y − π0 (z; γ0 ))}2 ],
√ D p
n(θ̂ − θ0 ) −
→ N (0, V ), V̂ →
− V.
We show the proof in Web Appendix B. This theorem shows that if γ̂ is estimable at a fast enough
rate one can still construct valid confidence intervals for θ̂. This result can be shown following similar
arguments as in Chernozhukov et al. (2022a). Finally, we note that while the above arguments focus on the
estimation of the average effect of a price change on demand, we can follow the same arguments to derive
inference results for other economic quantities of interest, e.g., the effect of changing some product features
on demand.
12

4 Estimation Procedure
Based on theoretical results presented above, we now outline an estimation procedure for both the choice
function (π) and the average effects of price changes (θ).
Consider a dataset, where {yt , zt , IVt }nt=1 are independently and identically distributed. Here, yt is the
vector of market shares in market t, zt is J × (d + 1) matrix of product features and IVt is the J-dimension
vector of instrumental variables.
• Stage 0 (Data partition): We randomly split the observed markets into L folds such that the data
Dl := {yt , zt }t∈l , where l denotes the lth partition. Note that all the observations for one market are
always in one fold.
• Stage 1 (Estimate γ̂): For each fold l, we estimate γl by regressing the endogenous variable on the
exogenous instruments on the left out data Dlc := {yt , zt }t∈l
/ . We then use the cross-fitting technique,
same as Chernozhukov et al. (2021), to calculate the residual µ̂l of fold l with estimated γ̂l on Dlc .
• Stage 2 (Estimate π̂ and θ̂):
– Stage 2a (Estimation): In the second stage, for each fold l, we estimate both the choice function
(π̂) and the Riesz estimator (α̂) on the left out data Dlc := {yt , zt }t∈l
/
1 X X
π̂l = arg min P [(yjt − π(zjt ; γ̂))2 ] (7)
π∈F t∈Dlc Jt
t∈Dlc j∈Jt
1 X X
α(zjt )2 − 2m(wjt ; α) .

α̂l = arg min P (8)
α∈A t∈Dlc Jt
t∈Dlc j∈Jt
Based on Theorem 2, instead of directly estimating the function π, we decompose the estimation
into three sub-components: ρ, ϕ1 , and ϕ2 . Specifically, for each component of our model (ϕ1 ,
ϕ2 , and ρ), we use a standard 3-layer neural network4 , and this is implemented without further
hyperparameter tuning. We implement ReLU activation function at each layer as it is standard
in feedforward designs due to simplicity and computation efficiency in gradients. Figure 1 il-
lustrates how we pass the data to the neural networks. Specifically, we pass the focal product’s
characteristics (price pjt and other product features Xjt ) and the residuals (µ̂jt ) estimated from
the first stage regression to the ϕ1 . In parallel, we pass all the products’ characteristics of the
other products in the same market (pjt and Xkt ) and the corresponding residuals (µ̂kt ) to the
same ϕ2 , and then sum the output up. The output of ϕ1 and ϕ2 have the same data structure
(e.g., a 64-dimension vector). Next, we pass the summation of the output of ϕ1 and ϕ2 to a third
neural network ρ. The output of ρ is a scalar which represents the market share of the focal
product jt.
4
For ϕ1 and ϕ2 , each of the three layers consists of 64 neurons, with the output vector also featuring 64 neurons. For ρ, the
layers are configured with 300, 100, and 64 neurons for the first, second, and third layers, respectively.
13

Figure 1: Illustration of Neural Network Architecture
We use the same neural network structure (with the three sub-components same as ϕ1 , ϕ2 , and
ρ) to estimate α. The only difference is that the loss function of α is not based on the difference
between the observed and the predicted market share as in Equation 7. Instead, the loss function
is based on the squared difference between α and the moment function of α as stated in Equation
8 and Theorem 3.
– Stage 2b (Cross-fitting): Now we again use the cross-fitting technique to reduce the bias when
estimating θ̂. Specifically, we use the estimators (π̂ and α̂) estimated on Dlc to estimate the
θ̂l of l. By applying cross-fitting, we ensure that the nuisance functions and the parameters are
estimated on separate, non-overlapping datasets. This approach diminishes the risk of overfitting
and enhances the robustness of our estimation. And finally, to estimate θ̂, we randomly select
one observation t∗ in each market t and average it out across all folds. Thus the estimator for θ0
and its variance can be given as follows:
L
1XX
θ̂ = {m (wt∗ , π̂ℓ ) + α̂ℓ (zt∗ ) (yt∗ − π̂ℓ (zt∗ ; γ̂))} (9)
n c
ℓ=1 t∈Dℓ
L
1XX 2
V̂ = ψ̂t∗ ℓ , ψ̂t∗ ℓ = m (wt∗ , π̂ℓ ) − θ̂ + α̂ℓ (zt∗ ) (yt∗ − π̂ℓ (zt∗ ; γ̂)) (10)
n c
ℓ=1 t∈Dℓ
5 Numerical Experiments
We now present a series of simulation studies that establish the numerical performance of our approach.
First, in §5.1, we examine the predictive performance of our model on a series of models, including stylized
discrete choice models with linear utilities as well as more general models that allow non-linear utilities
14

and realistic consumer behaviors such as inattention. Next, in §5.2, we present numerical experiments
that demonstrate our approach’s ability to simulate counterfactuals. Finally, in §5.3, we demonstrate the
applicability of the inference procedure proposed in §3.3.
5.1 Predictive Performance

In this section, we show that our approach can recover demand and elasticity estimates for a wide variety of
settings without the knowledge of the underlying choice model and/or making parametric assumptions on
consumers’ behaviors. Before describing the simuations, we first describe the metrics used for comparing
the performance of different estimators and the benchmark estimation approaches used.
First, to assess the predictive performance of our approach, we focus on three quantities of interest:
• Market share (π̂jt )
∂ π̂ /π
• Own price elasticity ( ∂pjt jt
jt /pjt
)
∂ π̂ /π
jt jt
• Cross-price elasticity ( ∂pk̸=jt /pk̸=jt )
In each simulation, we compare the performance of our model with the predictive performance of four
baseline models:
• Multinomial Logit model (MNL)
• Random Coefficient Logit model (RCL)
• A standard neural network-based non-parametric method (NP).
Here we simply use all the features of all the products in the market as input and predict a market share
vector without using the permutation invariance property or correcting for endogeneity. This is similar
to the approaches used by Gabel and Timoshenko (2022) and Cai et al. (2022), which simply use a large
neural network for demand prediction. For this baseline, we tune the hyperparameters of the neural
network, including the number of layers, number of nodes in each layer, learning rate, and the number
of epochs using 5-fold cross-validation for each data generation. We detail the space of hyperparameters
in Web Appendix C. We also apply the ReLU activation for each layer.
• A Mean Predictor (MP).
Here we predict a uniform market share for all observed products within a market, excluding the outside
option. This predicted share is set to the average of all observed market shares in that market5 . This
serves as a baseline estimator as it does not account for individual product characteristics, providing
a benchmark for the simplest prediction scenario. Additionally, comparing the performance of other
models with MP also enables us to quantify the natural decrease in the MAE and RMSE when the
number of products increases because the magnitude of the market shares decreases as the number of
products increases6 .
5
For instance, consider a market with three observed products having market shares of 0.2, 0.3, and 0.4, respectively. This
implies the outside option holds a market share of 0.1. In such a scenario, the MP would predict the market share for each of the
three products to be 0.3, which is the average ((0.2 + 0.3 + 0.4) / 3).
6
For instance, in a market with 100 products, the MAE and RMSE for any estimator are expected to be quite low. Comparing
with the MP in such situations helps establish a baseline for MAE or RMSE.
15

5.1.1 Multinomial Logit and Random Coefficient Logit with Linear Utility
We first consider the two standard Data Generating Processes (DGPs) used in the demand estimation lit-
erature that use linear utility-based choice models – (1) Multinomial Logit model, and (2) the Random
Coefficient Logit model. For both cases, we consider a setting with 10 products (J = 10), 100 markets
(T = 100), one price feature, and 10 non-price features (d = 10). We define the utility uijt that consumer i
in market t derives from product j as the following linear function:
uijt = αi pjt + βi Xjt + εijt , (11)
where εijt represents an independently and identically distributed (iid) Type-I extreme value across products
and consumers. Xjt ∈ Cd denotes the non-price features of the product. αi , βi are the model coefficients,
which are kept constant for all consumers in the MNL, while in the RCL, they are normally distributed across
consumers. The probability distribution of features and coefficients used are shown in Web Appendix D.
Also, the mean utility from the outside option is normalized to 0.
M N L and the market
We denote the market share of product j in market t generated from MNL by πjt
RCL . For each market, we generate the market shares of each product by
share generated from RCL by πjt
simulating N = 10, 000 individual choices and aggregating by each market as shown below.
N
MNL 1 X
πjt = 1(uijt = max(uikt )) (12)
N k∈St
i
N
RCL 1 X exp(αi pjt + βi Xjt )
πjt = P (13)
N 1 + k∈St exp(αi pkt + βi Xkt )
i
Note that for MNL, instead of simulating each individual’s choice probability, we simulate each individual’s
choice based on the utility maximization principle. This approach ensures that when we use MNL (true
model) for estimation, it does not reproduce the data perfectly.
For each DGP, we split the generated data into training data (80%) and test data (20%). We use the
training data for estimation (both our model and the benchmark models described above).7 For the predicted
market share, we present all the model results and comparisons on the test data. For the predicted own- and
cross-elasticities, we present all the model results and comparisons on the training data.8
Tables 2a shows the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) in the predicted
market share (π̂jt ) for our approach as well as the baseline models.We see that when the true model is MNL,
7
In our simulations, we train both our model and NP with the log market shares (log(πjt )). The performance metrics reported
for predicted market shares are computed based on the exponential values of the predicted log market shares, bringing these metrics
back to market shares (πjt ). The performance metrics reported for elasticities are computed based on the relative change of the
predicted log market shares (∂log(πjt )) divided by the percentage change of the price (∂pjt /pjt for own-elasticity and ∂pkt /pkt
∂ π̂ /πjt
for cross-elasticity), which is equivalent to the elasticity calculated directly using the market share ( ∂pjtjt /pjt
for own-elasticity and
∂ π̂jt /πjt
∂pk̸=jt /pk̸=jt
for cross-elasticity).
8
The reason we only use test data to report predicted market share accuracy is to demonstrate the model’s predictive perfor-
mance on unseen data. In contrast, we use training data to report accuracy in elasticities to mimic the real empirical setting where
we use full data to estimate elasticity.
16

Table 2: Baseline Predictive Performance: Market Shares, Own-Elasticity, and Cross-Elasticity
(a) Market Shares (π̂jt )

# True Model J T d Our Model MNL RCL NP MP No. Obs.
MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE
0 MNL 5 100 10 0.0534 0.0834 0.0078 0.0105 0.0082 0.0052 0.1269 0.2364 0.2312 0.2220 2000
1 MNL 10 20 10 0.0585 0.1086 0.0040 0.0131 0.0089 0.0134 0.1129 0.2191 0.1365 0.2948 800
2 MNL 10 100 10 0.0333 0.0591 0.0044 0.0039 0.0026 0.0053 0.1181 0.1717 0.1422 0.1503 4000
3 MNL 10 200 10 0.0307 0.1346 0.0032 0.0102 0.0034 0.0197 0.1096 0.2170 0.1416 0.2106 8000
4 MNL 20 100 10 0.0194 0.0765 0.0015 0.0077 0.0023 0.0068 0.0707 0.2242 0.0768 0.2201 8000
5 RCL 5 100 10 0.0240 0.0314 0.0307 0.0382 0.0033 0.0042 0.0456 0.0583 0.0538 0.0656 2000
6 RCL 10 20 10 0.0206 0.0281 0.0270 0.0343 0.0034 0.0044 0.0540 0.0612 0.0418 0.0525 800
7 RCL 10 100 10 0.0171 0.0231 0.0262 0.0326 0.0025 0.0033 0.0458 0.0583 0.0413 0.0514 4000
8 RCL 10 200 10 0.0141 0.0187 0.0252 0.0318 0.0032 0.0039 0.0431 0.0559 0.0412 0.0513 8000
9 RCL 20 100 10 0.0099 0.0140 0.0262 0.0281 0.0018 0.0024 0.0390 0.0489 0.0276 0.0354 8000
∂ π̂ /π
(b) Own-Elasticity ( ∂pjt jt
jt /pjt
)
# True Model J T d Our Model MNL RCL NP No. Obs

MAE RMSE MAE RMSE MAE RMSE MAE RMSE
0 MNL 5 100 10 0.2588 1.1815 0.1414 1.1232 0.1757 1.1322 0.4554 1.2115 8000
1 MNL 10 20 10 0.3523 1.3557 0.1967 1.2970 0.2585 1.3118 1.0057 1.5316 3200
2 MNL 10 100 10 0.3346 1.4327 0.2066 1.3851 0.2150 1.3863 0.9266 1.5857 16000
3 MNL 10 200 10 0.3245 1.4131 0.2007 1.3842 0.2357 1.3876 0.8266 1.5768 32000
4 MNL 20 100 10 0.4146 1.6596 0.3305 1.6203 0.3570 1.6229 1.0151 1.8353 32000
5 RCL 5 100 10 0.1189 0.2310 0.1474 0.3735 0.0125 0.0305 0.1802 0.3145 8000
6 RCL 10 20 10 0.1799 0.3729 0.2039 0.5326 0.0365 0.1196 0.3768 0.3254 3200
7 RCL 10 100 10 0.1498 0.2862 0.2154 0.5531 0.0224 0.0685 0.2987 0.3643 16000
8 RCL 10 200 10 0.1209 0.2416 0.2188 0.5512 0.0233 0.0732 0.2464 0.4241 32000
9 RCL 20 100 10 0.1658 0.3533 1.4591 1.7099 0.0429 0.1319 0.4555 0.4741 32000
∂ π̂ /π
(c) Cross-Elasticity ( ∂pk̸=jt jt
jt /pk̸=jt
)
# True Model J T d Our Model MNL RCL NP No. Obs

0 MNL 5 100 10 0.1349 0.8115 0.0107 0.6780 0.0118 0.6807 0.1968 0.9442 32000
1 MNL 10 20 10 0.0649 0.6742 0.0043 0.5326 0.0059 0.5424 0.1527 0.7123 28800
2 MNL 10 100 10 0.0862 0.6104 0.0043 0.5142 0.0049 0.5143 0.1885 0.7488 144000
3 MNL 10 200 10 0.0901 0.6075 0.0042 0.5140 0.0047 0.5153 0.2110 0.7831 288000
4 MNL 20 100 10 0.0482 0.4665 0.0015 0.4151 0.0016 0.4157 0.1270 0.5303 608000
5 RCL 5 100 10 0.0293 0.0571 0.0492 0.0551 0.0030 0.0070 0.0617 0.0960 32000
6 RCL 10 20 10 0.0261 0.0435 0.0332 0.0455 0.0039 0.0143 0.0972 0.0791 28800
7 RCL 10 100 10 0.0257 0.0493 0.0324 0.0447 0.0028 0.0090 0.0795 0.1124 144000
8 RCL 10 200 10 0.0213 0.0417 0.0353 0.0455 0.0035 0.0097 0.0745 0.1277 288000
9 RCL 20 100 10 0.0220 0.0431 0.0204 0.0359 0.0022 0.0103 0.0715 0.0890 608000
Note: Table 2a, 2b, 2c present the baseline predictive performance for predicted market shares, own- and cross-elasticities of our
model and four baseline models. J, T, and d represent the number of products, non-price features, and markets, respectively. NP
denotes a benchmark non-parametric method, which is a standard neural network. MP denotes the mean predictor. The Mean
Absolute Error (MAE) and Root Mean Square Error (RMSE) of predicted market shares for each scenario (i.e., true model, J, d, T)
are computed using the test data from 20 iterations of data generation, while the MAE and RMSE of own- and cross-elasticities
are computed based on the training data from 20 iterations of data generation. The column titled “No. Obs.” indicates the total
number of observations for each metric. Specifically, the number of observations for market share (π̂jt ) is calculated based on
T × J × 20% (the portion of test data) ×20 (the number of draws of simulations). The number of observations for own-elasticity
∂ π̂ /πjt
( ∂pjtjt /pjt
) is calculated based on T × J × 80% (the portion of training data) ×20 (the number of draws of simulations). The
∂ π̂ /π
number of observations for cross-elasticity ( ∂pk̸=jt jt
jt /pk̸=jt
) is calculated based on T × J × (J − 1) × 80% (the portion of training
data) ×20 (the number of draws of simulations).
17

our model cannot beat RCL or MNL, which is as expected; but the error of our model is quite close to the
true model. When the true model is RCL, our model can beat MNL consistently and the performance of
our model is also close to the true model. Importantly, we find that our model consistently outperforms
the benchmark Non-Parametric (NP) method in all data generation processes. This is despite extensive
hyperparameter tuning for the NP method.
There are two key reasons why our approach outperforms the standard neural network-based non-
parametric method, especially as the number of products increases. First, unlike the standard neural net-
work, our method can circumvent the curse of dimensionality that arises with the increase in the number of
products. The standard neural network uses the stacked product features (d-dimension non-price features
Xjt and price pjt ) as input and has (J × (d + 1)) × h1 parameters in the input layer, where h1 denotes the
size of the first hidden layer.9 In contrast, our model only uses product features as the input (with dimension
d + 1); see Figure 1. Therefore, the parameters for our model do not scale with the number of products.
Thus, as the number of products increases, our method is able to exploit this information to improve its
performance, whereas the standard NP is unable to do so. Second, our model can leverage product-level
market-share data more effectively. Note that one observation in the standard NN consists of one market,
whereas one observation for our method consists of one product in a market. Hence, given data on T mar-
kets, the number of samples available for the NP method is T , whereas the sample available for our model
is T × J. Together, these strengths of our approach lead to significantly better performance compared to a
naive neural network.
We observe both sources of performance improvement in the numerical simulation results in Table 2a.
As we vary the number of products (5, 10, and 20), the MAE of the predicted market shares from our model
decreases monotonically (see simulation numbers 0, 2, and 4 for MNL, and 5, 7, and 9 for RCL). In contrast,
the performance of the benchmark non-parametric estimator deteriorates as the number of products increases
and becomes even worse than the Mean Prediction for the 20 product cases (simulation numbers 4 and 9).
These findings demonstrate the overfitting problems and the curse of dimensionality issues discussed above.
Indeed, this limitation has also been theoretically established by Sannai and Imaizumi (2019), who showed
that for neural networks that do not take into account the inherent invariance structure, the generalization
√
gap increases in proportion to the number of possible permutations, which is J! in our case. Further, we
examine the performance of our model and the other benchmarks by varying the number of markets (20,
100, and 200) while keeping the number of products constant at 10; see simulations 1, 2, and 3 for MNL and
6, 7, and 8 for RCL. Although both our model and the NP method show improved performance with more
markets, the non-parametric estimator is more adversely affected by a decrease in market numbers due to a
more significant reduction in its sample size. This is particularly problematic in scenarios with one market,
as the NP method becomes infeasible for estimation for only one sample.
jt jt ∂ π̂ /π
Finally, in Tables 2b and 2c, we show the predictive performance of own-elasticity ( ∂pjt /pjt ) and cross-
∂ π̂ /π
elasticity ( ∂pk̸=jt jt
jt /pk̸=jt
), respectively. Again, we find that our model consistently outperforms the NP method
in all scenarios for both own- and cross-elasticity predictions. When the true model is MNL, our model un-
9
In this section, we consider only cases where there are no correlated unobservables. When there are potential endogeneity
concerns, then we can include a residual µjt estimated from a first-stage regression, and then size of the input layer becomes
J × (d + 2). We consider settings with endogeneity in Web Appendix §E and in the experiments on inference in §5.3.
18

derperforms compared to RCL, given that RCL inherently captures the substitution pattern among products
in MNL10 . When the true model is RCL, our model is the closest one to the true model. Unlike the market
share predictions, the accuracy of own-elasticity predictions does not exhibit a clear monotonical improve-
ment as the number of products increases. We observe a similar pattern even for the true model–the accuracy
of own-elasticity decreases as the number of products increases. This suggests that it is not a deficiency of
our model but due to the inherent complexities in estimating own-elasticities in markets with many prod-
ucts. In the prediction of cross-elasticity, our model shows a monotonical improvement in accuracy with
an increase in the number of products. Additionally, for both own- and cross-elasticities, as the number of
markets increases, the performance of our model is better due to the increase in the sample size.
So far, in the above simulations, we did not consider any endogenous explanatory variables. As dis-
cussed in §3.2, our approach can easily account for endogeneity following the estimation steps in §4. For
interested readers, we present a set of numerical experiments with endogeneous explanatory variables in
Web Appendix §E. The key takeaway from this analysis is that ignoring endogeneity can lead to significant
biases in the estimates of own- and cross-price elasticities. Therefore, in the application to real data, we take
care of endogeneity carefully; please see §6 for further details.
5.1.2 Random Coefficient Logit with Non-Linear Utility
In the previous section, we focused on linear utility specifications and standard choice behaviors. Recent
literature has highlighted that the oversight of non-linear relationships between features and utilities can
introduce biases in the estimates (Allenby et al., 2004). Conversely, non-parametric estimators are adept
at capturing these non-linear patterns directly from the data. As a result, there has been a growing trend
towards the adoption of non-linear utility functions. Therefore, we now focus on data generated from a
random coefficient logit model with non-linear transformations applied to observable features. We consider
a case with two product features – price and a non-price feature x. We apply a non-linear transformation
g(x). Following Bakhitov and Singh (2022),we consider two functions for g(x):
a. log(): g(x) = log(|16x − 8| + 1)sign(x − 0.5)
b. sin(): g(x) = sin(x)
The log transformation is common in empirical studies, to capture a diminishing sensitivity of a feature
on the market share. The sine transformation captures periodic or cyclical effects. For example, when the
feature represents the time of year, normalized from 0 to 1, then using sine transformation can effectively
capture the seasonal variations in consumer preferences.
The utility that consumer i in market t derives from product j then has the following non-linear form:
uijt = αi pjt + βi g(Xjt ) + εijt . (14)
The marketshares then follow a similar structure to that from Equation (13).
10
In an extreme case, when the variance of random coefficients is zero, RCL is equivalent to MNL.
19

As before, we generate data using this model and estimate the marketshares and elasticities using both
our approach and the baseline models. When estimating the baseline MNL and RCL models, we assume
that the researcher does not have knowledge of the non-linearities in the utility function and hence uses the
simple linear utility(s) in their estimation (as shown in Equation (11)). The results for the predicted market
∂ π̂ /π ∂ π̂ /π
shares (π̂jt ), own-price elasticity ( ∂pjt jt
jt /pjt
jt
) and cross-elasticity ( ∂pk̸=jt jt
/pk̸=jt ) are presented in Tables 3a, 3b,
and 3c, respectively.
Regarding the MAE of predicted market shares, our model surpasses the RCL model by a factor of
8X and 4X across transformations (a) and (b), respectively. Similarly, considering the MAE of predicted
own-elasticity in transformations (a) and (b), our model outperforms RCL by factors of 20X and 2.5X,
respectively. For the MAE of predicted cross-elasticity, our model is 2X and 1.5X superior to RCL across
transformations (a) and (b), respectively. It’s worth noting that while our model consistently outperforms
the NP method across metrics, the NP method still shows better performance than both RCL and MNL
∂ π̂ /π
in terms of MAE and RMSE for estimated own-elasticity ( ∂pjt jt
jt /pjt
), underscoring the strengths of neural
network-driven approaches in navigating non-linearities.
Table 3: Predictive Performance in Non-Linear Utility: Market Shares, Own-Elasticity, and Cross-Elasticity

# True Model Our model MNL RCL NP Mean No. Obs.
0 RCL-log() 0.0025 0.0063 0.0358 0.0361 0.0213 0.0309 0.0588 0.1235 0.0836 0.1401 200
1 RCL-sin() 0.0029 0.0046 0.0281 0.0340 0.0102 0.0172 0.0315 0.0449 0.0388 0.0527 200
∂ π̂ /π
jt /pjt
)
# True Model Our Model MNL RCL NP No. Obs

0 RCL-log() 0.0566 0.1523 5.4278 2.8249 1.1961 2.1135 0.6119 0.9085 16000
1 RCL-sin() 0.0609 0.2820 0.6229 1.1350 0.1777 0.4246 0.4057 1.0787 16000
∂ π̂ /π
jt /pk̸=jt
)
# True Model Our Model MNL RCL NP No. Obs

0 RCL-log() 0.0150 0.0543 0.0436 0.1389 0.0372 0.4414 0.2552 0.5444 144000
1 RCL-sin() 0.0226 0.1047 0.0471 0.1794 0.0354 0.1751 0.1448 0.3357 144000
Note: This table presents the results when we add non-linear transformation in data generation. We generate using the Random
Coefficient Logit (RCL) model, with 10 products and 100 markets, while only considering a single non-linearly transformed
feature, which is the price.
5.1.3 Models with Consumer Inattention and Consideration Set Formation
Finally, we consider a scenario where consumers do not pay attention to all the products and/or are not fully
informed of all the alternatives in the choice set. Recent literature has shown that this is often the case in
many empirical settings (Goeree, 2008; Gabaix, 2019; Honka et al., 2019; Abaluck et al., 2020; Compiani,
2022). However, such cases violate a standard assumption of the choice model: that consumers are informed
and consider all options when they make purchase decisions. In some parametric models (e.g., Van Nierop
20

et al., 2010), this issue is managed by constructing a consumer-level consideration set. However, considera-
tion sets are usually unobserved in data; so these approaches often require assumptions on how consideration
sets are formed, which might not always be appropriate or reflective of actual consumer behavior. Another
way to manage this issue is to model search costs, i.e., allow consumers to ignore certain products because
search is costly (e.g., Weitzman, 1978; Mehta et al., 2003; Hortaçsu and Syverson, 2004; Kim et al., 2010).
Similarly, it also requires researchers to specify how search cost enters the utility function and decision
process. In contrast to these models, our approach refrains from making any parametric assumptions, which
allows for a potentially more flexible representation of consumer behavior in the case of consumer attention.
To demonstrate how our model can capture the inattentive behavior, we look at a scenario where consumers
are inattentive and deviate from the traditional random coefficient logit model.
1
We consider a simple model of inattention where a portion of consumers (1 − 1+p ht
) ignore the product
1
with the highest price, following Compiani (2022). In other words, the consideration set of 1 − 1+pht
consumers in market t excludes the highest-price product h. So when the price increases, the portion of
inattentive consumers increases. Suppose there is only one feature, price, then, the choice probability of the
highest price product h in the market t is:
1 exp(αi pht )
P .
1 + pht k exp(αi pkt )
The choice probability of other products j ̸= h are given by:
1 exp(αi pjt ) 1 exp(αi pjt )

P + (1 − )P .
1 + pht k exp(αi pkt ) 1 + pht k̸=ht exp(αi pkt )
We first present detailed results on the own- and cross-elasticity for the two-product case in Figure 2.
We consider the number of markets to be 1,000 so that we can observe enough variance in our data. Figure
2a shows how the estimated own-elasticity for the highest-priced product (which is ignored by a subset of
consumers), for a range of prices. Note that in our model, when the price is higher, the portion of inattentive
consumers is higher. Thus when we change the price, the change in market share is smaller than the case
without inattention. While our model and the fully NP model are able to capture this pattern, both the
parametric models (MNL and RCL) are unable to do so. Figure 2b shows how the estimated cross-elasticity
of the other products vary with the price of the highest-priced product. Similarly, due to the ignorance
of inattention, both MNL and RCL overestimate the magnitude of the elasticity of the other product. In
contrast, our model and the fully NP are able to capture the true cross-price elasticity and are close to the
true model.
Next, we show a more comprehensive set of results for all three metrics (market-shares, own-, and
cross-price elasticities) when there are more products (2, 5, and 10) and fewer markets (100) in Table 4a,
4b, and 4c. We find that our approach consistently outperforms RCL and MNL, as expected. Further,
the performance of our model improves as the number of products increases while that of the NP model
monotonically decreases with the number of products (for the reasons discussed in §5.1.1).
In summary, we find that our approach adapts well even as the underlying model of consumer behavior
21

∂ π̂ /π
(a) Own-elasticity ( ∂pjt jt
jt /pjt
) in consumer inattention
∂ π̂ /π
(b) Cross-elasticity ( ∂pk̸=jt jt
jt /pk̸=jt
) in consumer inattention
Figure 2: Elasticity Effects in Consumer Inattention

1
Note: Figure 2 illustrates how different models perform when there is 1 − 1+p ht
consumers who ignore the product with the
highest price. We simulate market shares in the case of 2 products, 1000 markets, and 1 feature (price). Due to the existence of
inattention, when the price is higher, the portion of inattentive consumers is higher. Thus when we change the price, the change in
market share is smaller than the case without inattention.
22

Table 4: Consumer Inattention - Model Performance of Predicted Market Shares, Own-Elasticity, and Cross-
Elasticity

# J Our Model MNL RCL NP Mean No. Obs
0 2 0.0047 0.0198 0.0590 0.0753 0.0316 0.0439 0.0076 0.0273 0.1840 0.2044 40
1 5 0.0064 0.0139 0.0203 0.0252 0.0178 0.0226 0.0250 0.0345 0.0780 0.1022 100
2 10 0.0033 0.0068 0.0137 0.0166 0.0072 0.0100 0.0258 0.0365 0.0492 0.0656 200
∂ π̂ /π
jt /pjt
)
# J Our Model MNL RCL NP No. Obs

0 2 0.0609 0.5190 0.6758 1.5698 0.3929 1.4598 0.0573 0.8804 3200
1 5 0.0978 1.8579 0.6917 1.9614 0.3753 1.9782 0.4288 1.9546 8000
2 10 0.0708 2.4273 0.8306 2.1031 0.1842 2.2787 0.5464 2.1450 16000
∂ π̂ /π
jt /pk̸=jt
)
# J Our Model MNL RCL NP No. Obs

0 2 0.0419 4.3553 0.1911 8.8197 0.3350 8.5791 0.0378 5.6814 3200
1 5 0.0486 4.7745 0.0827 5.8357 0.0503 5.8404 0.1897 5.5601 32000
2 10 0.0212 3.8765 0.0476 4.0961 0.0146 4.1113 0.1676 4.1198 144000
Note: These tables present the MAE and RMSE for predicted market shares, own-elasticity, and cross-elasticity under consumer
inattention across scenarios with 2, 5, and 10 products. The market share and elasticity are simulated assuming a portion of
consumers (1 − 1+p1j ), jht is the index of the highest-price product in market t) ignore the product with the highest price. Each
ht
scenario is fixed at 100 markets and 1 feature (price) to maintain consistency with the RCL baseline.
23

changes without the need to impose any specific assumptions on consumer decision-making.
5.2 Counterfactual Analysis

In general, a key advantage of parametric models like MNL and RCL, or a structural approach to consumer
behavior is their ability to predict outcomes in counterfactual scenarios outside the distribution of the data
used for estimation. In contrast, fully non-parametric approaches like the naive neural network model that
simply uses all product features as inputs cannot be used for counterfactual predictions11 . By leveraging
the choice invariance property, our method adds some structure to the neural network architecture; and as
a result, accommodates certain types of counterfactual models. In particular, we focus on a specific type
of counterfactual that is of interest to firms and policy-makers – demand estimation when the choice set
changes; e.g., through the introduction of a new product, the exit of an existing product, or the merger of
two firms (Nevo, 2000; Petrin, 2002; Nevo, 2003; Wollmann, 2018). Our model can easily handle such
counterfactuals since the choice function is specified as a function of the focal product features and the
features of a set of competing products (see Theorem 1 and Figure 1). In contrast, estimating counterfac-
tual demand when the choice set changes is infeasible with a standard neural network estimator due to its
structural constraints on the input space. A change in the choice set would result in the change of size of the
input vector, making such estimation infeasible.
To showcase the capability of our model to estimate counterfactuals, we consider a setting where a new
product is introduced to the market. For comparison, we will only consider an MNL and RCL estimator
since it is infeasible for the NP method to estimate the counterfactual. We use two data generation processes
– Multinomial Logit (MNL) and Random Coefficient Logit (RCL), the same as §5.1.1 with 100 markets, 10
products (and one outside option)m and 10 features. For each of these two cases, we consider a counterfac-
tual where an 11th product is introduced to each market. The observable characteristics of the new product
are simulated from the same distribution as other products.
In Table 5, we present the error in the estimated market share of all products when a new product is
introduced to the market. We find that our model does quite well. When the true model is MNL, both MNL
and RCL outperform our method, similar to what we find in §5.1.1. Our model also outperforms the MNL
when the underlying data generation process is RCL, and produces results comparable to the true model.
Overall, these results suggest that our approach can be used for counterfactual estimation, which is often a
key focus of many substantive studies in marketing and economics.
5.3 Inference and Coverage Analysis

We now demonstrate the performance of the debiasing and inference procedure discussed in §3.3. The ob-
jective is to demonstrate the validity of the estimated confidence intervals. To this end, we estimate the
average effect of a 1% change in own price on demand over all products (θ̂) and compute the corresponding
11
Both our model and fully NP can do counterfactuals when shocks only result in the change in features {Xjt , Pjt }. For
example, consider a choice model where ranking is a feature. In such a scenario, researchers would like to see how the demand
would change if the ranking policy is changed. To estimate this counterfactual scenario, one would update the value assigned to the
ranking feature within a model. Therefore, both our model and fully NP are capable of doing such analysis, offering insights into
how changes in specific features could impact demand.
24

Table 5: New Product Demand Estimation - Predicted Market Share (π̂jt )
True Model Our Model MNL RCL No. Obs.

MAE RMSE MAE RMSE MAE RMSE
MNL 0.0234 0.0644 0.0041 0.0095 0.0023 0.0045 22,000
RCL 0.0186 0.0145 0.0265 0.0331 0.0023 0.0031 22,000
Note: This table presents the MAE and RMSE of predicted market shares of all products in the market when a new product enters.
We simulate market shares as in our baseline scenario (10 products and one outside option, 100 markets, 10 non-price features)
when an eleventh product is introduced.
confidence intervals of this effect. The difference between this section and §5.1.1 lies in both the estima-
tors and the methods. In terms of estimators, in §5.1.1, we predict the market share (π̂jt ), own-elasticity
∂ π̂ /π
jt jt ∂ π̂ /π
jt jt
( ∂pjt /pjt ) and cross-elasticity ( ∂pk̸=jt /pk̸=jt ) for individual products. In contrast, the object of interest in this
section is the average effect of price on demand across all products (θ̂). As a result, in §5.1.1, we did not use
the debiasing techniques that we apply here. It is important to emphasize that, in our approach, constructing
a confidence interval is viable only for aggregate measures, not for individual observations.
To simulate the data, we consider a random coefficient logit model of demand with 3 products across
100 markets. We set the true model parameters to be βik ∼ N (1, 0.5), αi ∼ N (−1, 0.5). The effect of a
1% increase in a product’s price is given by
θ0 = E[m(wjt , π)] = E[π(pjt ∗ (1.01), Xjt , {pkt , Xkt }k̸=j ) − π(pjt , Xjt , {pkt , Xkt }k̸=j )],
As discussed earlier, one way to estimate this effect is to compute the sample analog of this using the
estimated π̂, such that θ̂ = n1 ni=1 m(wjt , π̂). However, as we pointed out earlier, the distribution of θ̂
P
might not be asymptotically normal. To demonstrate this, in Figure 3a, we display the histogram of the
estimated effect across 50 random samples by using the plug-in method. We standardize the estimates by
subtracting the mean and then dividing by the standard deviation and plot them against the standard normal
distribution. As one can observe the distribution appears multi-modal and deviates from a standard normal
distribution. Next, we use our proposed debiased estimator and plot the standardized estimates across 50
samples of draws in Figure 3b. The resultant distribution with the debiased estimator is much closer to a
standard normal. Finally, we calculate the 95% confidence intervals using our debiased estimator across
50 random draws. In Table 6, we report the bias (mean absolute error from all draws) and the coverage
i.e., the percentage of times the true parameter is covered in the estimated confidence intervals. We find
that bias across both data-generating processes (RCL and MNL) is notably low, reflecting only a -0.0001
difference from the true effect. The coverage rate of the 95% confidence interval in our corrected model
is 90%, indicating good coverage. This shows that our debiased estimator can be used to conduct valid
inference in finite samples.
6 Emprical Data Analysis: US Automobile Data (1971 - 1990)

In this section, we apply our model to a real-world dataset. We use the “US Automobile Data (1971 –
1990)” from Berry et al. (1995). The dataset features cars in the US market from 1971 to 1990, with each
25

(a) Distribution of Estimated Average Effect of Price (b) Distribution of Estimated Average Effect of Price
Change with Plug-in Change with the Debiased Estimator
Figure 3: Distribution of Estimated Average Effect of Price Change

Note: The figure shows the distribution of the standardized plug-in and debiased estimators of the average effect of 1% change in
price on demand. To simulate the data, we consider 3 products across 100 markets with 5 non-price features using RCL for 50
samples. For each sample, we set the true model parameters to be βik ∼ N (1, 0.5), αi ∼ N (−1, 0.5). Figure 3a displays the
distribution of the estimated effect with the Plug-in method and Figure 3b shows the result when employing the debiased
estimator.
Table 6: Inference Coverage Analysis
True Model True Effect Bias 95% CI Cov.

RCL -0.0013 -0.0001 90%
MNL -0.0016 -0.0001 90%
Note: This table presents the bias and coverage rate of 95% confidence interval using our debiased estimator of the average effect
of 1% change in price on demand. To simulate the data, we consider 3 products across 100 markets with 5 non-price features using
RCL and MNL for 50 samples of draws. For each RCL sample, we set the true model parameters to be βik ∼ N (1, 0.5),
αi ∼ N (−1, 0.5). For each MNL sample, we set the true model parameters to be βik = 1, αi = −1.
year regarded as a market. The number of cars varies from 86 to 150 each year. For each car, the dataset
provides information such as the car’s name, the manufacturing company, factory region, market share,
price, and four exogenous car characteristics: horsepower, space, mileage per dollar, and the presence of an
air conditioning device.
Even though the dataset is relatively small, it presents three key challenges that make it difficult to
use naive non-parametric estimators: (i) the dataset features markets with more than 100 products and
only 20 markets in total, (ii) the product assortment in each year or market varies; (iii) the feature “price”
has the endogeneity issue, which was not considered in numerical experiments in §5. In this section, we
demonstrate the use of our estimator, which is capable of effectively addressing such challenges posed in
real-world datasets.
For each component of our model (ϕ1 , ϕ2 , and ρ), we use a standard 3-layer neural network, and this
is implemented without further hyperparameter tuning. We implement ReLU activation function at each
layer. This architecture is the same as the one we used in our numerical experiments. For comparison, we
replicate the random coefficient logit model (with only the demand side) used by Berry et al. (1995) using
the Python package pyblp (Conlon and Gortmaker, 2020). In our replication, we allow for heterogeneity
in random coefficients across all variables. Our findings show that the estimates obtained from our model
26

∂ π̂ /π ∂ π̂ /π
(a) Estimated own-Elasticity ( ∂pjt jt
jt /pjt
)without IV (b) Estimated own-Elasticity ( ∂pjt jt
jt /pjt
) with IV
Figure 4: Elasticity Estimation Comparison

∂ π̂ /π
Figure 4 presents the estimated own-elasticity ( ∂pjt jt
jt /pjt
) of our model without IV and with IV. The x-axis represents the price of
the focal product, while the y-axis shows the products’ own-elasticities. Each point corresponds to a product in a market, resulting
in 2,217 observations. We report the estimated elasticity based on the same price variation used in the BLP paper (a 1,000-dollar
change). Although we cannot ascertain the true value of own-elasticity, it is widely accepted that own-elasticity should generally
be negative for most, if not all, products. In Figure 4a, we observe that the majority of low-priced products (priced below 5,000
dollars) exhibit positive estimated own-elasticity, demonstrating the existence of the endogeneity.
are comparable to the random coefficient logit estimation presented in Berry et al. (1995).
We estimate our model both without and with consideration of endogeneity. To address endogeneity,
we utilize three sets of IVs – (i) the sum of characteristics of all car models, excluding the product in focus,
produced by the same firm in the same year; (ii) the sum of characteristics of all car models, excluding the
product in focus, produced by rival firms in the same year; and (iii) cost shifters, which encompass the wage
and exchange rate prevalent in the year and region where the factory is located. The utilization of traditional
BLP-style instruments, as discussed by Gandhi and Houde (2019), can be problematic due to their relative
weakness, often resulting in considerable bias in the estimation of parameters. These issues are significantly
exacerbated in non-parametric models. Thus, to counter potential concerns related to weak instruments, we
employ a machine-learning-based IV methodology (MLIV) as proposed by Singh et al. (2020). We detail
the estimation procedure and results using BLP style IVs in Web Appendix F.
∂ π̂ /π
In Figure 4, we present the estimated own-elasticity ( ∂pjt jt
jt /pjt
) of our model without IV and with IV.
The x-axis represents the price of the focal product, while the y-axis shows the product’s own-elasticity.
Each point corresponds to a product in a market, resulting in 2,217 observations. We report the estimated
elasticity based on the same price variation used in the BLP paper (a 1,000-dollar change). In Figure 4a,
we observe that the majority of low-priced products (priced below 6,000 dollars) exhibit positive estimated
own-elasticity, demonstrating the existence of the endogeneity. We also notice that this issue is attenuated
when we correct for endogeneity (see Figure 4), i.e., which suggests that our approach is able to handle
situations with endogenous features.
jt ∂ π̂ /π
jt jt jt ∂ π̂ /π
We report the own-elasticity ( ∂pjt /pjt ) and cross-elasticity ( ∂pk̸=jt /pk̸=jt ) estimated in our model and
random coefficient logit model with a sample of 13 cars in the 1990 market in Tables 7 and 8. The sample
of 13 cars is the same as the one reported in Berry et al. (1995). Overall, our results are very similar and
27

∂ π̂ /π
comparable to Berry et al. (1995). We also plot the distributions of the estimated own-elasticity ( ∂pjt jt
jt /pjt
)
∂ π̂ /π
jt jt
and cross-elasticity ( ∂pk̸=jt /pk̸=jt ) obtained from our model and the BLP model in Figure 5. The filled
areas in the violin plots represent the complete range of the elasticities, while the text labels next to the
line indicate the mean values. The estimated mean own- and cross-elasticities appear to be similar between
our model and the BLP model, though our model exhibits a larger RMSE in the estimated elasticity values
compared to the BLP model.
(a) Own-Elasticity Estimation (Our Model vs. BLP (b) Cross-Elasticity Estimation (Our Model vs. BLP
Model) Model)
Figure 5: Elasticity Estimation Comparison

Note: Figure 5 illustrates the distributions of the estimated own- and cross-elasticities obtained from our model and the BLP
model. The filled areas in the violin plots represent the complete range of the elasticities, while the text labels indicate the mean
values.
We further estimate the average own-elasticity (θ̂) for high-priced, medium-priced, and low-priced cars
and construct a confidence interval for each category using our inference procedure. We present our result
in Table 9. Both our model and the BLP model indicate that the average own-elasticity is highest (in terms
of the absolute value of own-elasticity) for high-priced cars and lowest for low-priced cars. Moreover, the
95% confidence intervals for all three categories do not include zero. This also demonstrates the efficiency
of our model even when there is a limited sample of only 20 observations.
The empirical analysis demonstrates the applicability and effectiveness of our model in a real-world set-
ting, addressing challenges such as limited sample size, variability in product assortments, and endogeneity.
The comparable results with established econometric models, such as BLP model, help in validating the
robustness and reliability of our approach.
7 Conclusion
Choice models are fundamental in understanding consumer behavior and informing business decisions.
Over the years, various methods, both parametric and non-parametric, have been developed to represent
consumer behavior. While parametric methods, such as logit or probit-based models, are favored for their
simplicity and interpretability, their restrictive assumptions can limit their ability to fully capture consumer
preferences’ intricacies. On the other hand, non-parametric methods offer a more flexible approach, but
28

Acura BMW Buick Cadillac Chevy Ford Ford Honda Lexus Lincoln Mazda Nissan Nissan
Legend 735i Century Seville Cavalier Escort Taurus Accord LS400 Town Car 323 Maxima Sentra
Acura Legend -5.6060 0.1993 0.2198 0.2221 0.0632 0.2317 0.2199 0.2337 0.2144 0.2187 0.2595 0.2354 0.2595
BMW 735i 0.4095 -6.1528 0.3653 0.4093 0.0352 0.3525 0.3655 0.3807 0.4120 0.4084 0.3547 0.4161 0.3547
Buick Century 0.1234 0.1020 -6.1023 0.1213 0.0376 0.1294 0.1215 0.1205 0.1175 0.1191 0.1400 0.1299 0.1400
Cadillac Seville 0.2895 0.2179 0.2631 -7.2896 0.0647 0.2692 0.2631 0.2566 0.2810 0.2818 0.2892 0.2849 0.2892
Chevy Cavalier 0.0142 -0.0013 0.0202 0.0167 -1.3447 0.0171 0.0202 0.0353 0.0126 0.0167 0.0354 0.0291 0.0355
Ford Escort 0.0410 0.0245 0.0353 0.0413 -0.0148 -1.8519 0.0353 0.0520 0.0392 0.0413 0.0494 0.0513 0.0494
Ford Taurus 0.1166 0.0914 0.1188 0.1160 0.0183 0.1199 -6.1473 0.1258 0.1066 0.1157 0.1451 0.1290 0.1451
Honda Accord 0.0975 0.0647 0.0968 0.1006 -0.0003 0.0945 0.0969 -5.7438 0.0914 0.1006 0.1166 0.1140 0.1165
29
Lexus LS400 0.3357 0.2606 0.3136 0.3325 0.0822 0.3235 0.3137 0.3126 -6.8791 0.3271 0.3495 0.3348 0.3494
Lincoln Town Car 0.2681 0.2310 0.2548 0.2656 0.0713 0.2663 0.2548 0.2648 0.2586 -5.3996 0.3009 0.2719 0.3009
Mazda 323 0.0361 0.0212 0.0272 0.0326 -0.0127 0.0249 0.0272 0.0363 0.0323 0.0326 -2.6589 0.0404 0.0357
Nissan Maxima 0.1579 0.1367 0.1589 0.1555 0.0425 0.1670 0.1589 0.1689 0.1484 0.1534 0.1884 -7.2216 0.1884
Nissan Sentra 0.0386 0.0239 0.0304 0.0384 -0.0202 0.0294 0.0304 0.0496 0.0375 0.0383 0.0439 0.0470 -1.8754
Table 7: Estimated own- and cross-elasticities of a sample of automobile data using our method
Note: This table presents the estimated own- and cross-elasticity of a sample of 13 cars in the 1990 market using our model. The selected cars are the same as Berry et al. (1995)
reports. Each entry with row index i and column index j gives the percentage change in demand divided by the percentage change in price (based on $1,000 change in the price of i).

Acura BMW Buick Cadillac Chevy Ford Ford Honda Lexus Lincoln Mazda Nissan Nissan
Legend 735i Century Seville Cavalier Escort Taurus Accord LS400 Town Car 323 Maxima Sentra
Acura Legend -5.4677 0.0489 0.0205 0.1029 0.0221 0.0220 0.0143 0.1477 0.1503 0.0273 0.0013 0.1359 0.0039
BMW 735i 0.1267 -9.8502 0.0156 0.1058 0.0122 0.0121 0.0057 0.1313 0.1546 0.0278 0.0006 0.1375 0.0022
Buick Century 0.0184 0.0054 -5.1978 0.0124 0.1153 0.1043 0.1475 0.1982 0.0165 0.0800 0.0076 0.0379 0.0174
Cadillac Seville 0.1293 0.0513 0.0175 -6.6819 0.0151 0.0150 0.0099 0.1393 0.1576 0.0271 0.0008 0.1406 0.0027
Chevy Cavalier 0.0131 0.0028 0.0766 0.0071 -3.1163 0.1421 0.0849 0.2608 0.0086 0.0404 0.0100 0.0395 0.0241
Ford Escort 0.0137 0.0029 0.0726 0.0074 0.1487 -3.0590 0.0603 0.2781 0.0090 0.0263 0.0106 0.0419 0.0258
Ford Taurus 0.0048 0.0007 0.0554 0.0027 0.0479 0.0326 -4.0258 0.0727 0.0017 0.1779 0.0026 0.0122 0.0057
Honda Accord 0.0387 0.0133 0.0582 0.0291 0.1151 0.1173 0.0568 -4.3399 0.0409 0.0297 0.0081 0.0618 0.0200
30
Lexus LS400 0.1350 0.0536 0.0166 0.1126 0.0130 0.0130 0.0045 0.1400 -7.4316 0.0243 0.0006 0.1464 0.0024
Lincoln Town Car 0.0087 0.0034 0.0286 0.0069 0.0217 0.0135 0.1693 0.0362 0.0086 -5.6139 0.0011 0.0123 0.0024
Mazda 323 0.0114 0.0020 0.0743 0.0056 0.1476 0.1494 0.0679 0.2723 0.0063 0.0304 -2.8631 0.0390 0.0254
Nissan Maxima 0.1008 0.0393 0.0314 0.0831 0.0493 0.0500 0.0271 0.1749 0.1209 0.0286 0.0033 -4.7872 0.0086
Nissan Sentra 0.0140 0.0031 0.0707 0.0078 0.1471 0.1504 0.0617 0.2763 0.0095 0.0271 0.0105 0.0422 -3.1799
Table 8: Estimated own and cross-elasticities of a sample of automobile data using the BLP model
Note: This table presents the estimated own- and cross-elasticity of a sample of 13 cars in the 1990 market using the BLP model. The selected cars are the same as Berry et al. (1995)
reports. Each entry with row index i and column index j gives the percentage change in demand divided by the percentage change in price (based on $1,000 change in the price of i).

BLP Model Our Model No. Obs.
Mean Estimate Mean Estimate (95% Confidence Interval)
High -5.6705 -4.1922 (-6.2204, -2.1641) 20
Medium -3.7354 -3.7215 (-4.9260, -2.5169) 20
Low -2.9174 -2.0697 (-3.1980, -0.9415) 20
Table 9: Estimates of Average Own-Elasticity

Note: This table presents the estimated average own-elasticity for cars across various price categories. We categorize cars with a
price over $20k as “high-priced”, cars priced between $8k and $20k as “medium-priced”, and all other cars as “low-priced”. For
each category, we randomly select one car of the category from each market as one observation. For the BLP model, we calculate
the average own-elasticity of the sampled cars as the mean estimate. For our model, we estimate the average own-elasticity and
construct the confidence interval following our inference procedure.
they often suffer from the “curse of dimensionality”, where the complexity of estimating choice functions
escalates exponentially with an increase in the number of products.
In this paper, we propose a fundamental characterization of choice models that combines the tractability
of traditional choice models and the flexibility of non-parametric estimators. This characterization specif-
ically tackles the challenge of high dimensionality in choice systems and facilitates flexible estimation of
choice functions. Through extensive simulations, we validate the efficacy of our model, demonstrating its
superior ability to capture a range of consumer behaviors that traditional choice models fail to capture.
We also show how to address the endogeneity issue and estimate counterfactuals in our characterization.
Furthermore, leveraging the recent strides in the automatic debiased machine learning literature, we offer
an inference procedure that constructs confidence intervals on relevant objects, such as price elasticities.
Finally, we apply our method to the automobile dataset from Berry et al. (1995). Our empirical analysis
affirms that our model produces results that align well with the extant literature.
Our paper opens many avenues for future research. We focus on using neural network-based estimators.
However, estimators, such as Gaussian processes and Gradient boosting-based estimators can be adopted to
estimate the proposed functionals. Also, we believe more experimentation needs to be done on the neural
network design side.
Competing Interests Declaration

Author(s) have no competing interests to declare.
References
J. Abaluck, G. Compiani, and F. Zhang. A method to estimate discrete choice models that is robust to
consumer search. Technical report, National Bureau of Economic Research, 2020.
J. Abrevaya. Rank estimation of a generalized fixed-effects regression model. Journal of Econometrics, 95
(1):1–23, 2000.
D. Ackerberg, X. Chen, J. Hahn, and Z. Liao. Asymptotic efficiency of semiparametric two-step gmm.
Review of Economic Studies, 81(3):919–943, 2014.
31

C. Ai and X. Chen. Estimation of possibly misspecified semiparametric conditional moment restriction
models with different conditioning variables. Journal of Econometrics, 141(1):5–43, 2007.
P. Albuquerque and B. J. Bronnenberg. Estimating demand heterogeneity using aggregated data: an appli-
cation to the frozen pizza category. Marketing Science, 28(2):356–372, 2009.
R. Allen and J. Rehbeck. Identification with additively separable heterogeneity. Econometrica, 87(3):1021–
1054, 2019.
G. M. Allenby, T. S. Shively, S. Yang, and M. J. Garratt. A choice model for packaged goods: Dealing with
discrete quantities and quantity discounts. Marketing Science, 23(1):95–108, 2004.
A. Aouad and A. Désir. Representing random utility choice models with neural networks. arXiv preprint
arXiv:2207.12877, 2022.
E. Bakhitov and A. Singh. Causal gradient boosting: Boosted instrumental variable regression. In Proceed-
ings of the 23rd ACM Conference on Economics and Computation, pages 604–605, 2022.
M. Ben-Akiva, D. Bolduc, and M. Bradley. Estimation of travel choice models with randomly distributed
values of time. Transportation Research Record, 1413:88–97, 1993.
Y. Bentz and D. Merunka. Neural networks and the multinomial logit for brand choice modelling: a hybrid
approach. Journal of Forecasting, 19(3):177–200, 2000.
S. Berry, J. Levinsohn, and A. Pakes. Automobile prices in market equilibrium. Econometrica: Journal of
the Econometric Society, pages 841–890, 1995.
S. T. Berry and P. A. Haile. Identification in differentiated products markets using market level data. Econo-
metrica, 82(5):1749–1797, 2014.
S. T. Berry and P. A. Haile. Nonparametric identification of differentiated products demand using micro
data. Technical report, National Bureau of Economic Research, 2020.
D. Besanko, S. Gupta, and D. Jain. Logit demand estimation under competitive pricing behavior: An
equilibrium framework. Management Science, 44(11-part-1):1533–1547, 1998.
J. Blanchet, G. Gallego, and V. Goyal. A markov chain approximation to choice modeling. Operations
Research, 64(4):886–905, 2016.
R. Blundell, J. Horowitz, and M. Parey. Nonparametric estimation of a nonseparable demand function under
the slutsky inequality restriction. Review of Economics and Statistics, 99(2):291–304, 2017.
J. H. Boyd and R. E. Mellman. The effect of fuel economy standards on the us automotive market: an
hedonic demand analysis. Transportation Research Part A: General, 14(5-6):367–378, 1980.
R. A. Briesch, P. K. Chintagunta, and R. L. Matzkin. Nonparametric discrete choice models with unobserved
heterogeneity. Journal of Business & Economic Statistics, 28(2):291–307, 2010.
Z. Cai, H. Wang, K. Talluri, and X. Li. Deep learning for choice modeling. arXiv preprint arXiv:2208.09325,
2022.
N. S. Cardell and F. C. Dunbar. Measuring the societal impacts of automobile downsizing. Transportation
Research Part A: General, 14(5-6):423–434, 1980.
32

S. Chatterjee and J. Jafarov. Prediction error of cross-validated lasso. arXiv preprint arXiv:1502.06291,
2015.
X. Chen and T. M. Christensen. Optimal sup-norm rates and uniform inference on nonlinear functionals of
nonparametric iv regression. Quantitative Economics, 9(1):39–84, 2018.
V. Chernozhukov, W. K. Newey, V. Quintas-Martinez, and V. Syrgkanis. Automatic debiased machine
learning via neural nets for generalized linear regression. arXiv preprint arXiv:2104.14737, 2021.
V. Chernozhukov, J. C. Escanciano, H. Ichimura, W. K. Newey, and J. M. Robins. Locally robust semipara-
metric estimation. Econometrica, 90(4):1501–1535, 2022a.
V. Chernozhukov, W. K. Newey, and R. Singh. Automatic debiased machine learning of causal and structural
effects. Econometrica, 90(3):967–1027, 2022b.
P. K. Chintagunta. Endogeneity and heterogeneity in a probit demand model: Estimation using aggregate
data. Marketing Science, 20(4):442–456, 2001.
S. Chitla, S. Jagabathula, and A. Venkataraman. Nonparametric demand estimation in the presence of
unobserved factors. Available at SSRN, 2022.
G. Compiani. Market counterfactuals and the specification of multiproduct demand: A nonparametric ap-
proach. Quantitative Economics, 13(2):545–591, 2022.
C. Conlon and J. Gortmaker. Best practices for differentiated products demand estimation with PyBLP. The
RAND Journal of Economics, 51(4):1108–1161, 2020.
C. V. Forinash and F. S. Koppelman. Application and interpretation of nested logit models of intercity mode
choice. Transportation research record, (1413), 1993.
M. Fosgerau and D. Kristensen. Identification of a class of index models: A topological approach. The
Econometrics Journal, 24(1):121–133, 2021.
J. T. Fox and A. Gandhi. Nonparametric identification and estimation of random coefficients in multinomial
choice models. The RAND Journal of Economics, 47(1):118–139, 2016.
M. J. Funk, D. Westreich, C. Wiesen, T. Stürmer, M. A. Brookhart, and M. Davidian. Doubly robust
estimation of causal effects. American journal of epidemiology, 173(7):761–767, 2011.
X. Gabaix. Behavioral inattention. In Handbook of behavioral economics: Applications and foundations 1,
volume 2, pages 261–343. Elsevier, 2019.
S. Gabel and A. Timoshenko. Product choice with large assortments: A scalable deep-learning model.
Management Science, 68(3):1808–1827, 2022.
A. Gandhi and J.-F. Houde. Measuring substitution patterns in differentiated-products industries. NBER
Working paper, (w26375), 2019.
M. S. Goeree. Limited information and advertising in the us personal computer industry. Econometrica, 76
(5):1017–1074, 2008.
W. H. Greene and D. A. Hensher. A latent class model for discrete choice analysis: contrasts with mixed
logit. Transportation Research Part B: Methodological, 37(8):681–698, 2003.
33

L. Grigolon and F. Verboven. Nested logit or random coefficients logit? a comparison of alternative discrete
choice models of product differentiation. Review of Economics and Statistics, 96(5):916–935, 2014.
J. Han, Y. Li, L. Lin, J. Lu, J. Zhang, and L. Zhang. Universal approximation of symmetric and anti-
symmetric functions. arXiv preprint arXiv:1912.01765, 2019.
Y. Han, F. C. Pereira, M. Ben-Akiva, and C. Zegras. A neural-embedded discrete choice model: Learning
taste representation with strengthened interpretability. Transportation Research Part B: Methodological,
163:166–186, 2022.
J. A. Hausman and W. K. Newey. Individual heterogeneity and average welfare. Econometrica, 84(3):
1225–1248, 2016.
J. A. Hausman and D. A. Wise. A conditional probit model for qualitative choice: Discrete decisions
recognizing interdependence and heterogeneous preferences. Econometrica: Journal of the econometric
society, pages 403–426, 1978.
D. A. Hirshberg and S. Wager. Debiased inference of average partial effects in single-index models: Com-
ment on wooldridge and zhu. Journal of Business & Economic Statistics, 38(1):19–24, 2020.
E. Honka, A. Hortaçsu, and M. Wildenbeest. Empirical search and consideration sets. In Handbook of the
Economics of Marketing, volume 1, pages 193–257. Elsevier, 2019.
B. E. Honoré and E. Kyriazidou. Panel data discrete choice models with lagged dependent variables. Econo-
metrica, 68(4):839–874, 2000.
A. Hortaçsu and C. Syverson. Product differentiation, search costs, and competition in the mutual fund
industry: A case study of s&p 500 index funds. The Quarterly journal of economics, 119(2):403–456,
2004.
H. Ichimura and W. K. Newey. The influence function of semiparametric estimators. Quantitative Eco-
nomics, 13(1):29–61, 2022.
Z. Jiang, T. Chan, H. Che, and Y. Wang. Consumer search and purchase: An empirical investigation of
retargeting based on consumer online behaviors. Marketing Science, 40(2):219–240, 2021.
J. Joo. Rational inattention as an empirical framework for discrete choice and consumer-welfare evaluation.
Journal of Marketing Research, 60(2):278–298, 2023.
W. A. Kamakura and G. J. Russell. A probabilistic choice model for market segmentation and elasticity
structure. Journal of marketing research, 26(4):379–390, 1989.
S. Khan, F. Ouyang, and E. Tamer. Inference on semiparametric multinomial response models. Quantitative
Economics, 12(3):743–777, 2021.
J. B. Kim, P. Albuquerque, and B. J. Bronnenberg. Online demand under limited consumer search. Market-
ing science, 29(6):1001–1023, 2010.
A. Lewbel. Semiparametric qualitative response model estimation with unknown heteroscedasticity or in-
strumental variables. Journal of econometrics, 97(1):145–177, 2000.
Z. Lu, X. Shi, and J. Tao. Semi-nonparametric estimation of random coefficients logit model for aggregate
34

demand. Journal of Econometrics, 2023.
R. D. Luce. On the possible psychophysical laws. Psychological review, 66(2):81, 1959.
C. F. Manski. Semiparametric analysis of random effects linear models from binary panel data. Economet-
rica: Journal of the Econometric Society, pages 357–362, 1987.
J. Marschak. Binary choice constraints on random utility indicators. 1959.
D. McFadden and K. Train. Mixed mnl models for discrete response. Journal of applied Econometrics, 15
(5):447–470, 2000.
D. McFadden et al. Conditional logit analysis of qualitative choice behavior. 1973.
S. R. Mehndiratta. Time-of-day effects in inter-city business travel. University of California, Berkeley, 1996.
N. Mehta, S. Rajiv, and K. Srinivasan. Price uncertainty and consumer search: A structural model of
consideration set formation. Marketing science, 22(1):58–84, 2003.
A. Nevo. Mergers with differentiated products: The case of the ready-to-eat cereal industry. The RAND
Journal of Economics, pages 395–421, 2000.
A. Nevo. New products, quality changes, and welfare measures computed from estimated demand systems.
Review of Economics and statistics, 85(2):266–275, 2003.
W. K. Newey. The asymptotic variance of semiparametric estimators. Econometrica: Journal of the Econo-
metric Society, pages 1349–1382, 1994.
A. Pakes and J. Porter. Moment inequalities for multinomial choice with fixed effects. Quantitative Eco-
nomics, Forthcoming, 2022.
A. Petrin. Quantifying the benefits of new products: The case of the minivan. Journal of political Economy,
110(4):705–729, 2002.
A. Petrin and K. Train. A control function approach to endogeneity in consumer choice models. Journal of
marketing research, 47(1):3–13, 2010.
D. Revelt and K. Train. Mixed logit with repeated choices: households’ choices of appliance efficiency
level. Review of economics and statistics, 80(4):647–657, 1998.
J. M. Robins, A. Rotnitzky, and L. P. Zhao. Estimation of regression coefficients when some regressors are
not always observed. Journal of the American statistical Association, 89(427):846–866, 1994.
A. Sannai and M. Imaizumi. Improved generalization bound of permutation invariant deep neural networks.
2019.
X. Shi, M. Shum, and W. Song. Estimating semi-parametric panel multinomial choice models using cyclic
monotonicity. Econometrica, 86(2):737–761, 2018.
B. Sifringer, V. Lurkin, and A. Alahi. Enhancing discrete choice models with representation learning.
Transportation Research Part B: Methodological, 140:236–261, 2020.
A. Singh, K. Hosanagar, and A. Gandhi. Machine learning instrument variables for causal inference. In
Proceedings of the 21st ACM Conference on Economics and Computation, pages 835–836, 2020.
35

K. Sudhir. Competitive pricing behavior in the auto market: A structural analysis. Marketing Science, 20
(1):42–60, 2001.
P. Tebaldi, A. Torgovitsky, and H. Yang. Nonparametric estimates of demand in the california health insur-
ance exchange. Econometrica, 91(1):107–146, 2023.
L. L. Thurstone. A law of comparative judgment. Psychological review, 34(4):273, 1927.
K. Train. A comparison of hierarchical bayes and maximum simulated likelihood for mixed logit. University
of California, Berkeley, pages 1–13, 2001.
K. E. Train. Discrete choice methods with simulation. Cambridge university press, 2009.
K. E. Train, D. L. McFadden, and M. Ben-Akiva. The demand for local telephone service: A fully discrete
model of residential calling patterns and service choices. The RAND Journal of Economics, pages 109–
123, 1987.
S. Turlo, M. Fina, J. Kasinger, A. Laghaie, and T. Otter. Discrete choice in marketing through the lens of
rational inattention. 2023.
E. Van Nierop, B. Bronnenberg, R. Paap, M. Wedel, and P. H. Franses. Retrieving unobserved consideration
sets from household panel data. Journal of Marketing Research, 47(1):63–74, 2010.
E. Wagstaff, F. Fuchs, M. Engelcke, I. Posner, and M. A. Osborne. On the limitations of representing
functions on sets. In International Conference on Machine Learning, pages 6487–6494. PMLR, 2019.
M. J. Wainwright. High-dimensional statistics: A non-asymptotic viewpoint, volume 48. Cambridge uni-
versity press, 2019.
A. Wang. Sieve blp: A semi-nonparametric model of demand for differentiated products. Journal of Econo-
metrics, 235(2):325–351, 2023.
S. Wang, Q. Wang, and J. Zhao. Deep neural networks for choice analysis: Extracting complete economic
information for interpretation. Transportation Research Part C: Emerging Technologies, 118:102701,
2020.
Y. Wei and Z. Jiang. Estimating parameters of structural models using neural networks. USC Marshall
School of Business Research Paper, 2022.
M. Weitzman. Optimal search for the best alternative, volume 78. Department of Energy, 1978.
T. G. Wollmann. Trucks without bailouts: Equilibrium product characteristics for commercial vehicles.
American Economic Review, 108(6):1364–1406, 2018.
M. Wong and B. Farooq. Reslogit: A residual neural network logit model for data-driven choice modelling.
Transportation Research Part C: Emerging Technologies, 126:103050, 2021.
M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. R. Salakhutdinov, and A. J. Smola. Deep sets. Advances
in neural information processing systems, 30, 2017.
36

Appendices
A Identity Independence and Permutation Invariance of Traditional Choice

Models
In this Web Appendix, we show why choice models listed in Table 1 satisfy the identity independence and
the permutation invariance assumption. Identity independence is achieved when all the information about a
product utilized in the model is included in its features. To show that these models satisfy the permutation
invariance assumption, we first permutate any two competitor products associated with a focal product. We
then illustrate that such permutations do not alter the demand for the focal product, thereby establishing the
models’ compliance with the permutation invariance assumption.
Since all models require parametric assumptions about the utility function, for the sake of simplicity in
notation, we assume that the utility of a product is a parametric function of its observed features Xjt and
price pjt , determined by a set of parameters β unless otherwise specified. Mathematically, we express the
utility as:
vjt = f (Xjt , pjt ; β)
A.1 MNL
Assuming there are J products, the market share of each product under the MNL model is
exp(f (Xjt , pjt ; β))

πjt = g(Xjt , pjt , {Xkt , pkt }k∈S,k̸=j ) = P . (A1)
exp(f (Xjt , pjt ; β) + k∈S,k̸=j exp(f (Xkt , pkt ; β))
If we permute two products k1 and k2 in the set S\{j}, the only change is in the order
Pof terms in the summa-
tion in the denominator of the market share. Since addition is commutative, the sum k∈S,k̸=j exp(f (Xkt , pkt ; β))
remains the same regardless of the order of k1 and k2 . Therefore, the value of πjt remains unchanged,
demonstrating that the MNL model satisfies permutation invariance.
A.2 Mixed Logit

Assuming there are J products, the market share of each product under the RCL model is
Z
exp(f (Xjt , pjt ; βi ))
πjt = g(Xjt , {Xkt }k∈S,k̸=j ) = P dF (βi ) (A2)
exp(f (Xjt , pjt ; βi ) + k∈S,k̸=j exp(f (Xkt , pkt ; βi ))
, where F (βi ) is the CDF of random coefficient βi . Similar to MNL, if we permute the position of any other
two products k1 , k2 ∈ S\{j}, it only affects the sequence of the summation in the denominator, thus the
demand of product j remains the same. Therefore, it satisfies the permutation invariance.
A.3 Nested Logit

In a nested logit model, products are grouped into nests to account for the correlation in unobserved factors
among products within the same nest. Consider a scenario where there are J products categorized into L
nests. The market share of a product j in a nest l is calculated taking into account this nesting structure. A
key aspect of showing permutation invariance in the Nested Logit model is to include the nest affiliation of
each product as a feature. This is represented by an L-dimensional 0-1 vector Njt ∈ R1×L , which denotes

the affiliation of product jt to one of the L nests. Each element of the vector corresponds to a specific nest,
indicating the product’s membership in that nest. Specifically,
πjt = g(Xjt , pjt , Njt , {Xkt , pkt , Nkt }k∈S,k̸=j ). (A3)
Then, the Nested Logit model can be represented as:
πjt = P (jt|l)P (l) (A4)

v
exp Njtjtσ
P (j|l) = P (A5)
vkt
k∈Nl exp Njt σ
P
vkt
exp Nl σ log k∈Nl exp Nkt σ
P (l) = P P (A6)
L vkt
m=1 exp Nm σ log k∈Nm exp Nm σ
Where:
• Nl is the set of products in nest l.
• σ ∈ RL is an L-dimentional vector where the lth element represents the scale parameter for nest l
which captures the correlation in unobserved utilities among products in the same nest. Therefore, the
scale parameter of product jt, which belongs to nest l, can be expressed as Njt σ.
• Nm denotes a binary vector of dimension L, in which the mth element is one and all other elements
are zero. Therefore, Nm σ is the scale parameter of the nest m.
It is intuitive that when we swap the position of any two products, the probability of choosing a nest
(P (l)) would not change because both the denominator and the numerator only depend on the nest affilia-
tions. Moreover, P (j|l) also satisfies permutation invariance because of the additive nature of the denom-
inator in its formula, where the sum over all products in nest l remains constant despite any permutation
of product positions within the nest. The permutation invariance property could be easily extended to that
random coefficient nested logit following the same logic as random coefficient logit.
A.4 Latent Class Logit Model

In a latent class logit model, products can be divided into multiple latent classes and each latent class has
a distinct set of parameters. Assume there are c latent classes, the utility of a product in class c is denoted
by vjt,c = f (Xjt , pjt ; βc ). The probability of one product belonging to a class c is determined by the linear
combination of Zjt with parameters αc .
πjt = g(Xjt , pjt , Zjt , {Xkt , pkt , Zkt }k∈S,k̸=j )

X exp(f (Xjt , pjt ; βc )) exp(αc Zjt ) (A7)
= P P
c
exp(f (Xjt , pjt ; βc )) + k∈S,k̸=j exp(f (Xkt , pkt ; βc ) c exp(αc Zjt )
P
Similarly, the k∈S,k̸=j exp(f (Xkt , pkt ; βc ) is commutative thus the demand remains the same when
we alter the order of any two products except jt. Therefore, latent class models satisfy the permutation
invariance.
ii

A.5 Consumer Inattention and Search Model
In this part, we specifically demonstrate the permutation invariance property of the consumer inattention
model that we used in the numerical experiments. Our analysis could easily be extended to other models
based on consumer inattention in the literature, such as Goeree (2008), Turlo et al. (2023), Compiani (2022)
and Joo (2023), as well as search models, such as Mehta et al. (2003).
In our simulation, we generate the market shares of products by considering a segment of consumers
who ignore the highest-priced product. Here we generalize the specification by assuming the portion of the
inattentive consumers is a function of observed features of the highest-priced product 1 − s(Xht ), where h
denotes the highest-priced product. A consumer forms their consideration set with the products they pay
attention to. For products in their consideration set, they use the logit function as the decision rule. We allow
for each consumer to have a set of random coefficients. Then the choice probability of the highest-priced
product becomes
s(Xht )exp(βi Xht )

πht = g(Xht , {Xkt }k∈S,k̸=h ) = P (A8)
f (Xjt , pjt ; βi ) + k∈S,k̸=h f (Xkt , pkt ; βi )
When we permutate the order of any two competitor products, the choice probability of the highest-price
product remains the same. The choice probability of other products can be expressed as
πjt = g(Xjt , {Xkt }k∈S,k̸=j )

s(Xht )exp(f (Xkt , pjt ; βi )) (1 − s(Xht ))exp(f (Xkt , pjt ; βi ))
= P + P
exp(f (Xkt , pjt ; βi )) + k∈S,k̸=j f (Xkt , pkt ; βi ) exp(f (Xkt , pjt ; βi )) + k∈S,k̸=j,h f (Xkt , pkt ; βi )
(A9)
Similar to the highest-priced product, the choice probability for other products remains unaffected when the
order of competitor products is altered. This is because the feature values of the most expensive product
Xh do not change even though the index h may change. Therefore, consumer inattention models are also
permutation invariant.
A.6 Other Models

We have illustrated that the Multinomial Logit model, Nested Logit model, Mixed Logit model, and Latent
Class Logit model all exhibit the property of permutation invariance. It becomes apparent that other models,
which employ a similar random utility framework but differentiate in the probability distribution of the error
term, likewise possess this permutation invariance property.
For example, this permutation invariance property extends Generalized Extreme Value (GEV) models
(Train, 2009), where the error term follows the generalized extreme value distribution, and the Probit model
(Hausman and Wise, 1978; Chintagunta, 2001), where the error term follows a normal distribution. Al-
though there is no straightforward closed-formed representation of choice probability for these models, the
choice probability can be expressed as
πijt = Pr(uijt ≥ uikt , ∀k ̸= j, k ∈ St ), (A10)
where uijt represents the consumer i’s utility of product j in market t. Since uikt for any k ∈ St is
parameterized by its own feature (Xkt ), permuting any two k ̸= j, k ∈ St , does not affect the choice
probability of πijt as well as the aggregate demand πjt . Indeed, this shows that any models building on the
random utility framework with the decision rule as stated in Equation A10 are permutation invariant.
The permutation invariance property also extends to other recently developed choice models, such as
Markov Chain Choice model(Blanchet et al., 2016). The choice probability of a product jt is modeled as a
iii

Markov Chain, which consists of two parts: arrival probability λjt and transition probability pjk,t , defined
as
λjt = f (j, St ), (A11)


1,
 if j = 0 and k = 0,
f (j,St \{j})−f (k,St )
jk,t = f (j,St ) , if j, k ∈ St , k ̸= j, (A12)

0, otherwise.

Given that both λjt and pjk,t only rely on jt, St and St \ {j}, permuting the order of any two products
that are not the focal product jt in the market t does not change both arrival probability and transition
probability. Therefore, the choice probability does not change.
B Proof of Main Results

uijt represents the index tuple {Xjt , pjt , Iit , εijt } satisfies Assumption 1, 2 and 3, then there exists suitable
ρ, ϕ1 and ϕ2 such that X
πjt = ρ(ϕ1 (Xjt , pjt ) + ϕ2 (Xkt , pkt )),
k̸=j,k∈St
P
Proof. The sufficiency follows by observing that the function πjt = ρ(ϕ1 (Xjt , pjt )+ k∈S\{j} ϕ2 (Xkt , pkt ))
satisfies assumption 2 and 3. To prove necessity, first consider E = {2n | n ∈ N} and O = {2n+1 | n ∈ N}
as the set of all even and odd natural numbers, respectively. Next, to show that all choice functions sat-
isfying assumptions 2 (identity independence) and 3 (permutation invariance) can be decomposed in the
above manner, we begin by noting that there must be a mapping from the elements to the set of even
number and odd numbers respectively, since the elements belong to a countable universe Ck . Let these
mappings be denoted by ce : Ck → E and co : C k → O. Now if we let ϕ (X ) = 4−ce (Xjt ,pjt ) and
1 jt
o
ϕ2 (Xkt , pkt ) = 4−c (Xkt ,pkt ) then ϕ1 (Xjt , pjt ) + k∈St \{j} ϕ2 (Xkt , pkt ) constitutes an unique represen-
P
tation for every product j andcompeting assortment St \ {j}. Now afunction ρ : R → R can always be
P
constructed such that πjt = ρ ϕ1 (Xjt , pjt ) + k∈S\{j} ϕ2 (Xkt , pkt ) .
Proposition 1. If Assumptions 5-7 are satisfied then for V = E[{m(w, π0 (z; γ0 )) − θ0

+α0 (z)(y − π0 (z; γ0 ))}2 ],
√ D p
n(θ̂ − θ0 ) −→ N (0, V ), V̂ →
− V.
Proof. To show the asymptotic normality we will first verify the Assumptions 1-3 of Chernozhukov et al.
(2022a), from now on CEINR, with g(w, π(z; γ), θ) = m(w, π(z; γ)) − θ and ϕ(w, π(z; γ), α(z), θ) =
p
α(z) · (y − π(z; γ)). Using Taylor series expansion, Assumption 6 and ||π̂(z; γ) − π0 (z; γ)|| →
− 0 we have,
iv

Z
||g(w, π̂(zi ; γ̂), θ0 ) − g(w, π0 (zi ; γ0 ), θ0 )||2 P0 (dw)
Z
= ||m(w, π̂(zi ; γ̂)) − m(w, π0 (zi ; γ0 ))||2 P0 (dw)
Z
≤ C ||π̂(zi ; γ̂) − π0 (zi ; γ0 )||2 P0 (dw)
Z
≤ C ||π̂(zi ; γ̂) − π̂(zi ; γ0 ) + π̂(zi ; γ0 ) − π0 (zi ; γ0 )||2 P0 (dw)
By the triangle inequality

Z
≤C ||π̂(zi ; γ̂) − π̂(zi ; γ0 )||2 P0 (dw)
Z
+ C ||π̂(zi ; γ0 ) − π0 (zi ; γ0 )||2 P0 (dw)
Z
p
+ C ||π̂(zi ; γ̂) − π̂(zi ; γ0 )||||π̂(zi ; γ0 ) − π0 (zi ; γ0 )||P0 (dw) →
− 0 (A13)
The first term converges in probability to 0 by Taylor series expansion.

p
Also by Assumption 5 i) and ii), and as just showed ||π̂(z; γ̂) − π0 (z; γ0 )|| →− 0,
Z Z
||ϕ(w, π̂(z; γ̂), α0 , θ0 ) − ϕ(w, π0 , α0 , θ0 )|| P0 (dw) = ||α0 (z)(π0 (z; γ0 ) − π̂(z; γ̂))||2 P0 (dw)
2
Z
≤ C ||(π0 (z; γ0 ) − π̂(z; γ̂))||2 P0 (dw)
p
≤ C||π̂(z; γ̂) − π0 (z; γ0 )||2 →
− 0
(A14)
p
Also by Assumption 5 i) and ||α̂ − α0 || →
− 0, we have,
Z Z
2
||ϕ(w, π0 (z; γ0 ), α̂, θ̃) − ϕ(w, π0 (z; γ0 ), α0 , θ0 )|| P0 (dw) = || (α̂(z) − α0 (z)) (y − π0 (z; γ0 )) ||2 P0 (dw)
Z
≤ C ||α̂ − α0 ||2 P0 (dw)
p
≤ C||α̂ − α0 ||2 →
− 0
(A15)
This satisfies Assumption 1 parts i), ii), and iii) of CEINR.

Next, consider

ˆ
∆(w) := ϕ w, π̂(z; γ̂), α̂, θ̃ − ϕ w, f0 , α̂, θ̃ − ϕ (w, π̂(z; γ̂), α0 , θ0 ) + ϕ (w, f0 , α0 , θ0 )
= − [α̂(z) − α0 (z)] [π̂(z; γ̂) − π0 (z; γ)] .
Then by the Cauchy-Schwartz inequality, and Assumptions 6 i) and ii)

h i Z
ˆ
E ∆(w) = − [α̂(z) − α0 (z)] [(π̂(z; γ̂) − π(z; γ))] P0 (dz)
(A16)
1
≤ ∥α̂ − α0 ∥ ∥(π̂(z; γ̂) − π(z; γ))∥ = op √
n
Also since α̂(z) and α(z) is bounded, we have

Z Z
2
ˆ
∆(w) P0 (dw) = [α̂(z) − α0 (z)]2 [(π̂(z; γ̂) − π0 (z))]2 P0 (dz)
Z (A17)
p
≤ C [(π̂ − π0 (z))]2 P0 (dz) −→ 0
Thus Equation A16 and Equation A17 verify Assumption 2 i) of CEINR.

Next Assumption 3 of CEINR is satisfied through Assumption 7. Thus Assumptions 1-3 of CEINR are
satisfied. Thus asymptotic normality follows by Lemma 15 of CEINR and the Lindberg-Levy central limit
theorem.
p R 2 p
Finally, we know θ −→ θ0 . And thus we have, g w, π̂(z; γ̂), θ̃ − g (w, π̂(z; γ̂), θ0 ) P0 (dw) =−→
0
To get the second conclusion, we need to show that V̂ is a consistent estimator of V . To show this, we
closely follow Chernozhukov et al. (2021). Let ψi = ψ0 (wi ) and consider
n n n n
1X 2 1 X 2 2 X 1X 2
V̂ = ψ̂i = ψ̂i − ψi + ψ̂i − ψi ψi + ψi
n n n n
i=1 i=1 i=1 i=1
hence, by re-arranging the terms and Cauchy-Schwarz inequality,

v v
n n n 2 u n 2 u n
1 X 2 2 X 1 X u1 X u1 X
V̂ −V = ψ̂i − ψi + ψ̂i − ψi ψi ≤ ψ̂i − ψi +2t ψ̂i − ψi t ψi2 .
n n n n n
i=1 i=1 i=1 i=1 i=1
Using the triangle inequality, for i ∈ Iℓ ,
2 4
X 3
X
ψ̂i − ψi ≤C Rij = C Rij + op (1)
j=1 j=1
where
Ri1 = [m (wi , π̂ℓ (zi ; γ̂ℓ )) − m (wi , π0 (zi ; γ0 ))]2 ,
Ri2 = α̂ℓ2 (zi ) [π̂ℓ (zi ; γ̂ℓ ) − π0 (zi ; γ0 )]2 ,
Ri3 = [α̂ℓ (zi ) − α0 (zi )]2 [yi − π0 (zi ; γ0 )]2 ,
2
Ri4 = θ̂ − θ0 .
p
We already showed consistency, so Ri4 → 0.
vi

Let I−ℓ denote observations not in Iℓ . By Markov’s inequality, for some δ > 0,
2
1 Pn
n
! E n i=1 ψ̂i − ψi
1 X 2
P ψ̂i − ψi > δ ≤
n δ
i=1
Note that the cross-fitting allows us to write

" n  
L XX 3 L 3
#
1 X 2 C X X nℓ X
E ψ̂i − ψi ≤ E Rij  + op (1) = C E [E [Rij | I−ℓ ]] + op (1).
n n n
i=1 ℓ=1 i∈Iℓ j=1 ℓ=1 j=1
We already showed,
Z
E [Ri1 | I−ℓ ] = [m (wi , γ̂ℓ ) − m (wi , γ0 )]2 F0 (dw) = op (1)
Next by triangle inequality, we have
Z
E [Ri2 | W−l ] = α̂l2 [π̂l (zi ; γ̂l ) − π0 (zi ; γ0 )]2 F0 (dz)
Z
= [α̂l + α0 − α0 ]2 [π̂l (zi ; γ̂l ) − π0 (zi ; γ0 )]2 F0 (dz)
Z
≤ [α̂l − α0 ]2 [π̂l (zi ; γ̂l ) − π0 (zi ; γ0 )]2 F0 (dz)
Z
+ [α0 ]2 [π̂l (zi ; γ̂l ) − π0 (zi ; γ0 )]2 F0 (dz)
Z
p
≤ Op (1) [π̂ℓ (zi ; γ̂ℓ ) − π0 (zi ; γ0 )]2 F0 (dz) → 0
Finally, we have
h h i i
E [Ri3 | I−ℓ ] = E E [α̂ℓ (zi ) − α0 (zi )]2 [yi − π0 (zi ; γ0 )]2 | zi , I−ℓ | I−ℓ
h h i i
= E [α̂ℓ (Zi ) − α0 (Zi )]2 E [yi − π0 (zi ; γ0 )]2 | Zi | I−ℓ
p
≤ C ∥α̂ℓ − α0 ∥2 → 0.
As a result,
n
1 X 2 p
ψ̂i − ψi → 0
n
i=1
p
Thus, we have V̂ −→ V
Also by Assumption 9 and iterated expectations
Z
E [Ri3 | W−ℓ ] ≤ {α̂ℓ (z) − ᾱ(x)}2 E (y − π̄(z))2 | Z = z FZ (dz)

Z
≤ C {α̂ℓ (z) − ᾱ(z)}2 Fz (dz) = C ∥α̂ℓ − ᾱ∥2 = op (1).
vii

C Hyperparameter Space for Tuning Non-parametric Estimator Benchmark
Hyperparameter Space
Number of hidden layers [3, 4, 5]
Number of nodes in each layer [64, 128, 256]
Learning rate [1e-2, 1e-3, 1e-4]
Number of epochs [1, 2, 4]
Table A1: Hyperparameter Space for Tuning Non-parametric Estimator Benchmark
D Distribution of Features and Coefficients in Numerical Experiments
MNL RCL
pjt U [0, 4] U [0, 4]
Xjt N (0, 1) N (0, 1)
Table A2: Distribution of Features
MNL RCL
αi -1 N (−1, 1)
βik 1 N (µβk , 1)
Notes : µβk ∼ N (0, 1/2d)
Table A3: Distribution of Coefficients
E An Numerical Experiment of Endogeneity

In §3.2, we show how our model handles endogeneity in theory. Here, we provide a simulation that demon-
strates the performance of our model handling endogeneity.
Let the utility uijt that consumer i in market t derives from product j as the following linear function
uijt = αi pjt + βi Xjt + γi µjt + εijt , (A18)
where µjt is the unobservable that correlates pjt and εijt is i.i.d. Type-I extreme value distributed. Specifi-
cally, without loss of generality, we specify the correlation between pjt and µjt as
pjt = δ1 IVjt + δ2 µjt , (A19)
where IVjt is the exogenous instrumental variable.

In this simulation, we first generate Xjt and µjt , following the same distribution as the Xjt in our
baseline model, and IVjt , following the same distribution as our as pjt in our baseline model, as discussed in
Web Appendix D. We next generate pjt using Equation A19. Finally, we generate market shares assuming
that the true model is RCL. For simplicity, we let δ1 = δ2 = 1. In addition, αi and βi follow the same
viii

distribution as the baseline simulations and γi follows the same distribution as βi , as stated in Web Appendix
D. We consider a case with 10 products, 100 markets, and 10 non-price features (in addition to price), the
same as our baseline simulation in §5.1.1. In the estimation, we let µjt be an unobservable. We use the OLS
linear regression as our first-stage regressor.
We consider two cases for comparison:
• Exogeneous Benchmark: Uses the exact same DGP but treat µjt as observables to researchers in
estimation. Since the coefficients of price and market shares are the same as the endogeneity case,
true elasticities are the same for these two DGPs. This gives a benchmark performance with the same
data under the assumption that endogeneity is not present.
• Ignoring Endogeneity: Directly trains the model using only the observed features Xjt and pjt with-
out considering the endogeneity problem. Note that, even though the endogeneity problem is ignored,
this does not mean the predictive performance in market shares would be low for this case because the
model could be overfitted. However, the elasticities will be biased when the endogeneity is ignored.
Following the routine in our main text, we simulate 20 times for the same DGP with different parameters
∂ π̂ /πjt
and features and report the predictive performance on market share (π̂jt ), own price elasticity ( ∂pjtjt /pjt
),
∂ π̂ /π
and cross-price elasticity ( ∂pk̸=jt jt
jt /pk̸=jt
). We report the performance of our model in Table A4. When
we use the control function to correct for endogeneity (Row 1: Our Method), our model underperforms
slightly compared to the case where µjt is treated as an observable (Row 2: Exogenous Benchmark), on
all three metrics. However, when we ignore the endogeneity issue and apply our model (Row 3: Ignore
Endogneity), although the predictive performance of market shares is not bad, the estimation of own- and
cross-elasticities are significantly biased, with the MAEs being 10–25 times higher than the cases where we
account for endogeneity. This demonstrates both the importance of accounting for endogeneity as well as
the effectiveness of our method in handling this problem in real settings.
Market shares Own-elasticity Cross-elasticity

MAE RMSE MAE RMSE MAE RMSE
Our Method 0.0256 0.0194 0.2307 0.3516 0.0669 0.1033
Exogeneous Benchmark 0.0254 0.0195 0.2290 0.3469 0.0657 0.1039
Ignore Endogeneity 0.0263 0.0202 2.3832 5.1901 1.9751 4.4306
Table A4: Model Performance of Endogeneity Case
F Details in Adopting the “MLIV” Method

Following Singh et al. (2020), we perform the steps below to construct the machine-learning-based IV
(MLIV) and use them to estimate γ̂ to control for endogeneity in prices.
• Step 1: Data Partition We randomly split the data set, D, into three separate partitions of markets,
each denoted as Dl . Each market is exclusively assigned to only one partition. For each partition, we
define its complement set, Dlc , as the subset of data in D that is not included in Dl .
• Step 2: Cross-fitting For each partition l, we first estimate a linear regression model on the com-
plement data set, Dlc , using the Lasso method with hyperparameters tuned by 3-fold cross-validation.
As discussed in section 3.3, we need the estimator of γ to converge at n−1/2 rate, a similar result
ix

that bounds the prediction error of the lasso estimator has been established in Chatterjee and Jafarov
(2015). Then, we use this trained model to predict the outcomes (prices) of the Dl . We denote the
fitted value as fˆl , which is essentially the MLIV.
• Step 3: First-stage Regression We estimate the first-stage estimator γ and residual µ using the MLIV
as the only predictor following step 1 in Section 4.
As a supplement to our main result, we also run our model using non-machine learning-based IVs.
Similar to Figure 4 in the main text, we present the estimated own-elasticity of our model without IV, with
BLP-style IVs, with differentiation IVs, and with MLIV in Figure A1. In Figure A1b, even when IVs
are applied, the persistence of many positive own-elasticities suggests the weakness of the BLP style IVs.
Furthermore, we apply the differentiation IVs (Gandhi and Houde, 2019), which use exogenous measures
of differentiation and provide a more robust instrument compared to the conventional BLP IVs. As one can
see from Figure A1c, the use of differentiation IV provides a more realistic estimation of own-elasticities,
strengthening the issue of weak instruments of the BLP style IVs. We also include the distributions of the
estimated own- and cross-elasticities obtained from our model using different sets of IVs in Figure A2.
(a) Without IV (b) BLP Style IVs
(c) Differentiation IVs (d) MLIV
Figure A1: Elasticity Estimation Comparison

Figure A1 presents the estimated own-elasticity of our model without IV, with BLP Style IVs, with differentiation IVs and with
MLIV. The x-axis represents the price of the focal product, while the y-axis shows the product’s own-elasticity. Each point
corresponds to a product in a market, resulting in 2,217 observations. We report the estimated elasticity based on the same price
variation used in the BLP paper (a 1,000-dollar change).
In addition, we also perform a weak instrument test on both BLP Style IVs and the MLIV and report the
F-statitics and p-value in Table A5. Both BLP Style IVs and MLIV pass the weak instrument tests.

(a) Own-Elasticity Estimation (Our Model vs. BLP Model)
(b) Cross-Elasticity Estimation (Our Model vs. BLP Model)
Figure A2: Elasticity Estimation Comparison

Note: Figure A2 illustrates the distributions of the estimated own- and cross-elasticities obtained from our model (using different
sets of IVs) and the BLP model. The filled areas in the violin plots represent the complete range of the elasticities, while the text
labels indicate the mean values.
xi

F-statistic P-value
BLP Style IVs 241.5 < 1e-8
MLIV 280.9 <1e-8
Table A5: Weak Instrument Test
xii

SSRN 4508227

Uploaded by

Copyright:

Available Formats

SSRN 4508227

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN 4508227

Uploaded by

Copyright:

Available Formats

Choice Models and Permutation Invariance: Demand Estimation in

Differentiated Products Markets

February 16, 2024

Electronic copy available at: https://ssrn.com/abstract=4508227

Electronic copy available at: https://ssrn.com/abstract=4508227

Electronic copy available at: https://ssrn.com/abstract=4508227

Electronic copy available at: https://ssrn.com/abstract=4508227

Electronic copy available at: https://ssrn.com/abstract=4508227

P(εijt | X·t , p·t ) = P(εijt )

Electronic copy available at: https://ssrn.com/abstract=4508227

πijt = π(uijt , {uiσj (k)t }k∈St ,k̸=j )

Electronic copy available at: https://ssrn.com/abstract=4508227

Table 1: Choice Models Satisfying Identity Independence and Permutation Invariance

Choice Model Literature

Proof: See Web Appendix B.

Electronic copy available at: https://ssrn.com/abstract=4508227

3.2 Endogenous Covariates

E[pjt · εijt ] ̸= 0 and E[Xjt · εijt ] = 0.

Given valid instruments IVjt , we can express pjt as

pjt = γ (Xjt , IVjt ) + µjt . (3)

εijt = CF (µjt ; λ) + ε̃ijt , (4)

where E[pjt · ε̃ijt ] = 0.

uijt = {Xjt , pjt , CF (µjt ; λ) + ε̃ijt }, (5)

Electronic copy available at: https://ssrn.com/abstract=4508227

Electronic copy available at: https://ssrn.com/abstract=4508227

E[m(wjt , π(zjt ; γ))] = E[α0 (zjt )π(zjt ; γ)]

E[m(wjt , π(zjt ; γ)) + α(zjt )(yjt − π(zjt ; γ)) − θ0 ]

Electronic copy available at: https://ssrn.com/abstract=4508227

θb = En [m(wjt ; π b(zjt )(yjt − π

ψ0 (wjt ) := m (wjt ; π0 ) + α0 (zjt ) (yjt − π0 (zjt )) − θ0 ,

= arg minE α(zjt )2 − 2α0 (zjt )α(zjt ) + α0 (zjt )2

= arg minE α(zjt )2 − 2m(wjt ; α) ,

{z 7→ ζ (α(z) − α0 (z)) : α ∈ An , ζ ∈ [0, 1]} and

Electronic copy available at: https://ssrn.com/abstract=4508227

|E [m(w, π) − θ0 + α0 (z)(y − π(z; γ))]| ≤ C ∥π − π0 ∥2

Proposition 1. If Assumptions 5-7 are satisfied then for V = E[{m(w, π0 (z; γ0 )) − θ0

Electronic copy available at: https://ssrn.com/abstract=4508227

• Stage 2 (Estimate π̂ and θ̂):

Electronic copy available at: https://ssrn.com/abstract=4508227

Electronic copy available at: https://ssrn.com/abstract=4508227

5.1 Predictive Performance

Electronic copy available at: https://ssrn.com/abstract=4508227

uijt = αi pjt + βi Xjt + εijt , (11)

Electronic copy available at: https://ssrn.com/abstract=4508227

(a) Market Shares (π̂jt )

# True Model J T d Our Model MNL RCL NP No. Obs

# True Model J T d Our Model MNL RCL NP No. Obs

Electronic copy available at: https://ssrn.com/abstract=4508227

Electronic copy available at: https://ssrn.com/abstract=4508227

5.1.2 Random Coefficient Logit with Non-Linear Utility

a. log(): g(x) = log(|16x − 8| + 1)sign(x − 0.5)

b. sin(): g(x) = sin(x)

uijt = αi pjt + βi g(Xjt ) + εijt . (14)

Electronic copy available at: https://ssrn.com/abstract=4508227

(a) Market Shares (π̂jt )

# True Model Our Model MNL RCL NP No. Obs

# True Model Our Model MNL RCL NP No. Obs

5.1.3 Models with Consumer Inattention and Consideration Set Formation

Electronic copy available at: https://ssrn.com/abstract=4508227