Vous êtes sur la page 1sur 25

RAND Journal of Economics

Vol. 44, No. 4, Winter 2013


pp. 585609
A model of ops
Patrick Hummel

John Morgan

and
Phillip C. Stocken

A rm surveys a large number of consumers, some of whom sincerely report their tastes and
others of whom report strategically. It makes product decisions using the sample mean of survey
responses. When rms and consumers agree on the fraction of sincere consumers, information loss
is severe, and many products are ops as they poorly match consumer tastes. When beliefs differ,
however, equilibrium is in linear strategies, and information aggregates. Despite this, ops still
arise. A rm, however, can solve the ops problem by limiting the effect of strategic consumers.
Binary surveys offer one such solution.
1. Introduction
Many of todays most successful rms pride themselves on the practice of data-driven
decision making, the extensive use of statistical data to inform key management decisions.
1
One
of the most famous examples is Capital One, which ran a vast experimentation programconsisting
of hundreds of different offers to consumers to determine product offerings. In large part, this
practice is a response to the vast harvest of data made possible by the information technology
revolution. Managers can now access copious and detailed information about nearly every aspect
of the rm. Nowhere has this change been more profound than in new product introductions.
A century ago, introducing new products was a matter of intuition, experience, and gut feel.
Statistics were often viewed with disdain. The essence of this spirit is nicely captured in Henry
Fords famous quip that If I had asked my customers what they wanted, they would have said a
faster horse.
2
Perhaps not surprisingly, many product introductions turned out to be ops. Even

Google Inc.; phummel@google.com.

University of California-Berkeley; rjmorgan@haas.berkeley.edu.



Dartmouth College; phillip.c.stocken@dartmouth.edu.
We are indebted to Ali Hortacsu (editor), David Laibson, Don Lehmann, Preston McAfee, Michael Schwarz, Joel Sobel,
Eric van den Steen, Miguel Villas-Boas, and the anonymous referee for helpful remarks, Dana Sisak for outstanding
assistance, and seminar participants at Harvard University for comments. The second author gratefully acknowledges the
nancial assistance of the National Science Foundation.
1
Brynjolfsson, Hitt, and Kim(2011) report that data-driven decision making rms enjoy 5%6%higher productivity
than rms that do not use data-driven decision making.
2
As quoted by Kelley and Littman (2005).
Copyright
C
2014, RAND. 585
586 / THE RAND JOURNAL OF ECONOMICS
as late as 1982, around 46% of commercialized products were deemed to be failures within one
year (Barczak, Grifn, and Kahn, 2009).
Fast forward to today. In a world of data-driven decision making, product design now
routinely incorporates a wealth of information, including large-scale surveys, focus groups, and
test markets. Yet despite all of this additional information, the extensive statistical training of
managers, and the ease and ubiquity of number crunching software like Excel, the percentage
of commercialized products judged to be failures has hardly budged, amounting to around 40%
of products in 2004 (see Barczak, Grifn, and Kahn, 2009).
Turning this ever-expanding mass of data into a usable form is a critical problem. Managers
must rely on summary statistics, such as the sample mean, because poring over the data line
by line is simply infeasible. Firms routinely correct survey data for selection bias, such as the
undersampling of cell phone users but rarely correct for strategic misrepresentation on the part
of respondents. At rst glance, this seems entirely appropriate. If some survey respondents
exaggerate their tastes toward more extreme positions, the effect is similar to measurement
errorthe noise of responses increases but, with a large enough sample, this is of no particular
consequence. Moreover, strategic respondents probably represent only a small fraction of the
population, further limiting their inuence.
In this article, we show that both arguments are fallacious: A rm relying on the sample
mean of a large survey when developing a product will make systematic errors, even when only a
small fraction of the populace is strategic and when the survey fully aggregates information. The
reason is that, though misrepresentation by strategic respondents washes out ex ante, it produces
a systematic effect at the interim stage once consumer tastes are realized. To see why, consider
a snack that can be, to varying degrees, sweet or savory. Ex ante, strategic consumers are as
likely to exaggerate toward savory as toward sweet. However, if the average consumers tastes
tilt in the savory direction (relative to the managers prior beliefs), then exaggerations will be
systematically biased in the savory direction. In ignoring strategic motives, the result will be
product ops despite vast amounts and rigorous use of data. This systematic bias also implies that
strategic respondents need not (and indeed will not) put forth extreme exaggerations to achieve
their objectives.
This is not to say that strategic misrepresentation is the only (or perhaps even the main)
driver of ops. Products fail for many reasons, including managerial hubris (the movie Ishtar),
ill-considered brand extensions (Bic disposable panties), or unwise cost-cutting (Pontiac Aztek).
3
Our broader point is that, even with limitless amounts of data, data-driven decision making can
come to grief by failing to account for the motives of those producing the data. The same point
applies to corporate prediction markets, sales forecasts based on reports by downstream clients,
or any other data where strategic motives play a role. It also applies outside of rm settings, such
as with politicians using polls to determine the appropriate policy or position to take on an issue.
The mechanism we describe illustrates why such carefully tested poll driven policies or positions
can still prove unpopular, particularly with hot-button issues like abortion or gun control where
strategic responses play a signicant role. The ops problem, our particular focus, illustrates this
issue starkly and in terms of a critical management decision.
In the model, a rm seeks to design an ideal productone that matches the tastes of a
representative consumer. Consumer tastes share a systematic component but are idiosyncratic
as well. The rm conducts an open-ended survey, which we model as a costless information
transmission game, to determine tastes. Although some (perhaps most) consumers simply report
their tastes honestly, others are strategic and tailor their reports to try to obtain a product that
matches their own tastes rather than those of the representative consumer. The results of the
survey are then summarized by the sample mean of responses. When rms and consumers share
common beliefs about the number of strategic consumers, strategic misrepresentation, even from
3
See Dederer (2006) and Maricano (2009) for further details.
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 587
a small fraction of respondents, is enough to destroy information aggregation in large surveys.
Indeed, product choice is unresponsive to the survey in the unique monotone equilibrium.
When priors differ, however, information aggregation is still possible and, when the com-
ponents of taste are normally distributed, equilibrium has a remarkably simple form: Strategic
consumers and rms use linear strategies that are uniquely determined. Even so, differing priors
can lead to perverse effects in how well the product matches consumer tastes. The match quality
depends on two ratios. The rst is the noise-signal ratiothe ratio of the variance of the idiosyn-
cratic to the systematic taste component. The second is the calibration-alignment ratiothe error
in the rms prior about the fraction of strategic consumers relative to the disagreement in priors
between rms and consumers. Survey performance declines in both ratios.
Although it is apparent that being poorly calibrated harms the rm, the model has more
subtle implications as well. In particular, a survey conducted by a well-calibrated rm will still
perform poorly if there is little disagreement between the rm and consumers. This is true even
in the limit as the rms priors become perfectly calibrated. Idiosyncratic noise rarely plays a role
in asymptotic statistics but, when survey respondents are strategic, adversely affects large-survey
performance as well.
The consequences of these distortions can be considerable. Under a wide range of parameter
values, a rm relying on the sample mean of a large survey would be better served by eschewing
surveys entirelyeven when survey costs are negligible.
Properly constructed surveys, however, can still be tremendously useful. The key to effective
survey design is to limit the scope for strategic consumers to inuence outcomes. Binary surveys,
such as Facebooks Like button, offer one such avenue. By simply asking consumers for the
direction in which their tastes differ relative to some status quo, misreporting incentives are
eliminated. A second way is to use the median of reported consumer tastes rather than the more
usual sample mean. Although both of these approaches have poor properties in small surveys
they sacrice information and do not necessarily produce sincere survey responsesthey are
extremely effective in large surveys, producing truthful reporting and optimal product choices.
Thus, the model provides some justication for the use of Likert-type scales or other simple
reporting devices rather than more involved and in-depth survey procedures or complicated
scoring rules.
The remainder of the article is as follows. Section 2 sketches the model, reviews the literature,
and describes our contribution. Section 3 establishes that, when all strategic players share the
same prior beliefs, information aggregation from a survey is impossible and information loss is
severe. Section 4 shows that, when we allow for the possibility of heterogeneous prior beliefs,
information aggregation is restored. Reporting strategies are simple and intuitive, consisting of
linear strategies, and in large surveys, there is a unique responsive linear equilibrium. Section 5
examines the performance of products designed using survey information. We offer conditions
under which these products are likely to be opsdespite information aggregation. Indeed, for
a wide range of parameter values, a rms products would be more successful were it to forego
market surveys altogether rather than conduct (and act upon) a large survey. This is the case even
if it does not cost the rm anything to conduct the survey. Section 6 shows how a rm can solve
the ops problem by gathering less information from consumers. Section 7 concludes. The proofs
to some propositions are contained in the Appendix.
2. Model
A rm is introducing a new product, and cost considerations limit it to a single offering.
The key decision is the type of product to introduce. Let y R denote a product type. One can
think of y as summarizing all aspects of the product: its technical specications, design features,
color, avor, etc. A product will be successful if it appeals to consumer tastes, which are unknown
to the rm. There is a continuum of potential consumers for the product. Each consumer i has
taste t
i
= +
i
where represents a systematic taste component and
i
an idiosyncratic taste
C
RAND 2014.
588 / THE RAND JOURNAL OF ECONOMICS
component. The systematic component consists of a single draw from a normal distribution
with mean zero and variance
2

. Let F denote the cumulative distribution of with associated


density f. The idiosyncratic component is drawn independently for each consumer from a normal
distribution with mean zero and variance
2

. Consumers only know their own taste t


i
and not the
separate components and
i
. The distributions of the taste components are, however, commonly
known by the rm and the consumers. For these taste distributions, E [|t
i
] = at
i
, where
a
2

/
_

+
2

_
.
To determine the product to offer, a rm can conduct market research by asking consumers
about their tastes. This research is sufciently inexpensive that the rm can survey an arbitrarily
large number of consumers. Specically, the rm randomly selects n consumers from the popu-
lation and asks each of them to provide a costless report m
i
R. This research can be thought of
as a combination of market surveys, focus groups, taste tests, and so on. The information gleaned
from these reports is then summarized by M, the sample mean of the reports.
4
The rm makes
its decision based on this information.
In addition to differing in their tastes, consumers also differ in how they respond to the rm.
A strictly positive fraction of consumers are sincere. These consumers report their type honestly
when asked, that is, sincere consumer i reports m
i
= t
i
. The remaining consumers are strategic.
These consumers choose a report m
i
that maximizes their payoffs, recognizing the connection
between their response and the product the rm ultimately offers. Specically, suppose that
strategic consumers suffer quadratic losses in the difference between the product offered by the
rm and their tastes; that is, the payoffs for these consumers are
u = (y t
i
)
2
.
Notice that the individual report, m
i
, does not enter directly into payoffs; it only enters to the
extent it inuences y.
The rm, however, wants a product with broad appeal and suffers quadratic losses as well.
5
Formally, suppose that the rms payoff is
v = (y )
2
.
Of course, the rm does not know , nor is it known individually by any of the consumers.
If all consumers were sincere, the rm could determine with arbitrary precision through its
research. The rm, however, cannot commit to the product it will offer as a function of the
information received, M. Instead, the rm will select the project optimally given its posterior
beliefs about given M; that is, the rm will choose the product
y (M) = E [|M] .
Firms and consumers may differ in their beliefs about the number of consumers who are
sincere. Specically, when a rm surveys n consumers, the rm believes k
n
of these consumers
are sincere. In contrast, strategic consumers believe that h
n
consumers are sincere for the same
size survey. Thus, the rm and consumers potentially have heterogeneous prior beliefs about the
number of sincere consumers surveyed. As we mainly will be concerned with large surveys, it
is useful to dene = lim
n
k
n
/n and = lim
n
h
n
/n. We also allow for the possibility that
neither set of beliefs may be correct. Suppose further that the true number of sincere consumers
in a survey of size n is s
n
, and let = lim
n
s
n
/n > 0. Thus, and correspond to the limiting
4
Having the rm base its product choice on a summary statistics of the survey rather than the underlying microdata
greatly facilitates the analysis. It is also consistent with practice. Firms typically employ external research agencies to
conduct marketing research. Owing to privacy concerns, their analyses tend to be generic rather than tailored to the
rm (e.g., Deshpande and Zaltman, 1982, 1984; Moorman, Zaltman, and Deshpande, 1992). Further, the use of classical
statistics to analyze market survey data is ubiquitous (e.g., Sawyer and Peter, 1983; Lehmann, Gupta, and Steckel, 1998).
5
It is straightforward to identify conditions on demand where prot maximization is equivalent to minimizing a
quadratic loss function. See Alonso and Matouschek (2008) for details.
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 589
beliefs of the rm and consumers, respectively, though coincides with the true fraction of
sincere consumers in the limit.
6
The modelling choice where the rm and strategic consumers
may differ in their views about the fraction of sincere consumers, but where there is agreement
among the strategic consumers, represents a minimal departure from the common priors model.
It also aids in the mathematical tractability of the model while still highlighting the key economic
effects of differing priors.
7
We study weak perfect Bayesian equilibria of the model. In these equilibria, (i) the rm
and consumers use Bayes Rule wherever possible in forming beliefs, and (ii) messages sent by
strategic consumers and product choices made by the rm must be best responses given their
beliefs. This equilibrium concept is standard for cheap talk games, such as our setting.
Because much of the analysis concerns what happens in an equilibrium of a large survey,
it is useful to dene this precisely. An equilibrium of a survey of n consumers consists of a
reporting strategy
_
m
i,n
(t
i
)
_
n
i =1
and a product type strategy y
n
(M
n
). For strategic consumers and
the rm, these strategies comprise mutual best responses. Reporting strategies map tastes to
survey responses, whereas product type strategies map the sample mean of a size n survey to
product designs. Call this equilibrium
n
. When we study an equilibrium of a large survey, we
are interested in the properties of

= lim
n

n
for some convergent sequence
n
. Of particular
interest is whether an equilibrium of a large survey fully aggregates information, that is, if it
is possible to infer the state, , perfectly through the sample mean of a large survey. Formally,
we say that information aggregates in an equilibrium of a large survey if M plim
n
M
n
is a
deterministic function of that is an invertible mapping from M onto .
Unlike standard cheap talk models, in our setting there is no commonly known inherent
conict between senders and receivers of information. An individual consumers bias is not
known to the rm. Moreover, because has zero mean, then consumers and the rm agrees ex
ante on the ideal product. Subsequent disagreement arises from two sources. First, though each
consumer knows precisely his or her ideal product, the rm cares about the average consumer,
and thus, while in expectation preferences are aligned, for each taste realization, preferences
are misaligned. The second source of disagreement stems from differing prior beliefs about the
fraction of sincere consumers in the survey. This leads to differing views as to howthe rmshould
use the data from the survey in its product design. This source of disagreement is designed to
capture the differing experiences and views of the rm and strategic consumers about the veracity
of survey responses. As we shall see, this can produce situations where the survey fully aggregates
available information yet produces opsproducts whose characteristics are far from the tastes
of the average consumer.
Literature review. Our model of survey research contributes to the vast cheap talk literature
dating back to Crawford and Sobel (1982). In this literature, there are typically one or a fewexperts
providing information to a decision maker.
8
Tension in these models stems fromthe experts bias
the difference between the experts and decision-makers preferences as to the appropriate action.
9
Although this bias is mainly assumed to be common knowledge, a few papers study the case
where it is privately known (e.g., Morgan and Stocken, 2003; Li and Madar asz, 2008). Moreover,
the exact bias of the expert ( in our model) is unknown to all parties. Finally, on average, the
experts, which we term consumers in our model, are unbiased. Disagreement arises from the rm
6
In the limit of a large survey, our model is strategically equivalent to one in which a consumer is sincere with
probability , the rm believes a consumer is sincere with probability , the strategic consumers believe a consumer
is sincere with probability , and each consumers type is an independent and identically distributed draw from these
distributions.
7
Relaxing this assumption introduces a new factor in the rms inference problemthe rm can glean additional
information about through the interaction of a consumers beliefs and tastes as manifested in the sample mean. This
indirect effect substantially complicates the analysis without changing the main economic forces in the model.
8
See Farrell and Rabin (1996) for an excellent survey of this literature. Grossman and Helpman (2001) offer a
comprehensive perspective on costly and costless information transmission in political settings.
9
Gordon (2010) offers a general treatment for how sender bias translates into equilibrium actions.
C
RAND 2014.
590 / THE RAND JOURNAL OF ECONOMICS
anchoring on its prior beliefs over rather than preference misalignment. In our model, none of
the consumers is an expert in the sense of knowing the state realization precisely; hence our
focus on the informational properties of large surveys.
This focus also places our work in the much older literature on information aggregation.
Dating back to Condorcet (1763; see Young, 1988), this literature mostly examines on the
performance of election mechanisms in aggregating preference information of voters.
10
The
main nding here is that elections perform extremely well in this regard (e.g., Feddersen and
Pesendorfer, 1997). Our model departs from this framework in several ways. First, the rm
chooses fromamong a continuumof possibilities rather than making a binary choice. Second, and
more importantly, the rmcannot commit to its choice ahead of time, unlike a voting rule. Instead,
it collects consumer taste information and then uses it optimally, given its beliefs. In that respect,
our model is closer to studies of polling (e.g., Meirowitz, 2005; Morgan and Stocken, 2008).
Conict in these models again arises from constituents privately known ideological bias rather
than anchoring on prior beliefs. These models highlight that, when the message space is binary,
information aggregation is possible but truth telling is not generally consistent with equilibrium.
By contrast, we show that, when the message space is unrestricted and priors are homogeneous,
information aggregation is not possible (Proposition 1). Consistent with the intuition from these
earlier models, however, we show how the rm can still benet from message space restrictions
(Proposition 7).
We make two key behavioral departures fromstandard models. First, we introduce behavioral
types in the formof sincere consumers. When these types rst appeared in the cheap talk literature
(e.g., Sobel, 1985; Benabou and Laroque, 1992), their main role was to affect the dynamics of the
game. In particular, using intuition akin to Kreps and Wilson (1982) and Milgrom and Roberts
(1982), strategic types imitate behavioral types to gain an advantage in the dynamic game. Our
model is static, so such a role is absent. Instead, behavioral types smooth the posterior beliefs of
the rm following any message. This eliminates equilibria sustained by the freedom to specify
out of equilibrium beliefs (e.g., Krishna and Morgan, 2001) while allowing equilibrium reporting
strategies that are continuous in the consumers type. Continuous reporting strategies also have the
property that equilibriuminformation aggregation is nowpossible. Absent such types, equilibrium
reporting strategies and posterior beliefs are typically discontinuous.
More recent work has studied the role of behavioral types in static settings. One branch of
this literature explores the implications of (possibly) naive receivers on signaling (e.g., Ottaviani
and Squintani, 2006; Kartik, Ottaviani, and Squintani, 2007). Like our article, Kim and Pogach
(2009) study the implications of a possibly naive sender on information transmission. Other work
allows for naivete on the part of the sender and receiver (Chen, 2011). A fundamental difference
between these settings and the present article is that they study situations where there is only a
single sender, whereas our main concern is with information transmission when there are a large
number of senders.
Our second departure is to allow for the possibility that the rm and consumers have
heterogeneous priors over the fraction of behavioral types in the population. Although it has
long been recognized that players might hold differing prior beliefs and that this view is not
inconsistent with the assumption that players are rational (e.g., Harsanyi, 1968), the common
priors assumption is more or less standard. However, there has recently been renewed interest in
the effects of heterogeneous priors on player behavior (see Brunnermeier and Parker, 2005; Van
den Steen, 2010). Heterogeneous priors have been shown to rationalize a variety of apparently
irrational behaviors, including overcondence in choice of actions (Van den Steen, 2004), the
pop typical of rst-day trading following an initial public offering (Morris, 1996), the winners
curse arising in private value settings (Compte, 2002), and speculative bubbles in asset prices
10
The Condorcet Jury Theorem claims that if voters behave nonstrategically, then the majority in a dichotomous
election is more likely than a single individual to make a correct decision, and this likelihood increases as the size of the
group increases (see Young, 1988).
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 591
(Scheinkman and Xiong, 2003).
11
More recently, Van den Steen (2010) shows how relaxing the
common priors assumption creates benets to rm integration despite complete contracting.
In our model, heterogeneous priors likewise lead to a form of overcondencerms and
consumers simultaneously believe they are obtaining an ideal product (in expectation) even
though their preferences differ. Relaxing the common priors assumption in a cheap talk context
changes the informational properties of equilibrium starkly: Under common priors, information
loss is an inevitable consequence of equilibrium whereas with heterogeneous priors, it is not.
Kawamura (2011) independently shows that truth telling is not an equilibrium in a setting
in which a utilitarian planner uses a survey to recover private taste parameters from a nite
population with uncorrelated tastes, which consist of iid draws from a uniform distribution.
Several differences between the two articles are worth noting. First, in our setting, tastes are
correlated; and, indeed, discovering the underlying state is the heart of the decision-makers
problem. Second, we study asymptotic properties of surveys in a large population. Kawamuras
model, however, is uninteresting in large population settings as the particular realizations of tastes
are irrelevant to the decision-makers choice. Third, we obtain a much stronger result than the
absence of truth telling: the information-loss problem is extremely severe under common prior
beliefs about the distribution of types in the population (Proposition 3)a phenomenon that
does not occur in Kawamuras setting. Perhaps most importantly, unlike Kawamura we focus on
the case of heterogeneous priors and show that information does aggregate, though this is not
necessarily to the benet of the decision maker.
Finally, we contribute to the literature on survey research. Within economics, the literature
largely focuses on a monopolist conducting a survey to ascertain an uncertain demand curve
(see, e.g., Manning, 1979; Venezia, 1984). The focus is on the costs and benets of the survey
rather than the sincerity of survey respondents. Within marketing, this literature recognizes the
possibility that consumers might not answer sincerely when asked about issues they nd sensitive,
such as questions dealing with sex or health (see Fowler, 1993; Groves, 1989; Lehmann, Gupta,
and Steckel, 1998; Hartman, Doane, and Woo, 1991). Nevertheless, the literature mostly ignores
this possibility, analyzing data under the assumption that all consumers are sincere (e.g., Churchill
and Iacobucci, 2005; Lehmann, Gupta, and Steckel, 1998; Sawyer and Peter, 1983). In contrast, we
examine how the interaction between strategic rms and respondents impacts the informational
properties of surveys and the quality of subsequent decision making.
3. Homogeneous priors
Compared with standard cheap talk models, our model makes two behavioral departures.
First, we assume that a fraction of consumers are sincere and merely report their tastes honestly
when asked. Second, we assume that rms and strategic consumers may differ in their views
about the fraction of sincere consumers in the survey population. In this section, we admit only
the rst of these departures and explore the implications of the sincere consumers on the quality
of market research.
Straightforward intuition would suggest that, the larger the fraction of sincere consumers, the
greater the informational benet to the rm. Clearly, if all consumers were sincere, the rm could
infer the systematic taste component, , arbitrarily preciselythat is, information aggregates.
When none of the consumers are sincere, it cannot. Thus, one might surmise that the quality of
information increases with the fraction of sincere consumers. Indeed, this would be the case if the
behavior of strategic consumers remained xed; however, their behavior adjusts to the fraction
of sincere types in the population. We show that, even if the fraction of strategic consumers is
small, their inuence on information aggregation is large. Specically, when strategic consumers
are present, information loss is inevitableeven in large surveys. Formally,
11
See also Admati and Peiderer (2004) as well as Che and Kartik (2009) for applications of overcondence in
cheap talk settings.
C
RAND 2014.
592 / THE RAND JOURNAL OF ECONOMICS
Proposition 1. Under homogeneous priors, there does not exist an equilibrium in which informa-
tion aggregates.
Proposition 1 reveals a key consequence of disagreement between strategic consumers and
the rm. Although, on average, strategic consumers share the same preferences as the rm,
individually, they do not. Thus, each strategic consumer sends a report so that, in expectation,
she gets her ideal product, t
i
. At the same time, the rm uses its research to design a product
that converges in probability to the tastes of the average consumer, . However, as Proposition 1
shows, these two objectives are mutually inconsistent when information aggregates. This intuition
takes no advantage of the particular structure of the taste realizations and, indeed, the normality
assumptions on the realized state and taste variables are not needed for Proposition 1 to hold.
What is required is that the rms response to any survey realization is well behaved (in the sense
that y (M) is differentiable in M) so that strategic consumers and the rm are at loggerheads in
terms of satisfying their objectives. For this, it sufces that at least one consumer is sincere and
the taste realizations for consumers have positive density over the real line.
Of course, the fact that information does not fully aggregate does not imply that informa-
tion loss is necessarily severe. For instance, in the Crawford and Sobel (1982) framework, the
misalignment of preferences between sender and receiver ensures that some information is lost,
but the loss becomes minimal as preferences become aligned. One might have a similar intuition
in our modelthe larger the fraction of sincere consumers, the smaller the information loss. To
the contrary, we show that, even in the limit as the fraction of sincere consumers goes to one,
information loss remains severe.
To formalize this idea, we study equilibria that are well behaved in the following sense:
Fix an equilibrium and let y () and M () denote the equilibrium product choice and sample
mean in state in the limit as the number of consumers surveyed goes to innity. We say that
an equilibrium is monotone if y () and M () are (weakly) monotone in . Monotone equilibria
capture the following idea: When the systematic component of tastes shifts upward, the rm
and the typical consumer prefer a higher product type to a lower one, and this is reected in
the equilibrium product type. Monotone equilibria capture this idea but are weakerif product
types were decreasing in , this relation, too, would satisfy monotonicity, but would make little
economic sense.
We rst show that a monotone equilibrium always exists. Specically,
Proposition 2. The following is a monotone equilibrium: Strategic consumers choose the message
m
i
= t
i
k
n
/(n k
n
)
and the product design is unresponsive to M; that is, for all M
y (M) = 0.
Proposition 2 is the survey analog to the familiar babbling equilibria that arise in costless
signaling games. In those games, all agents are assumed to be strategic; and, therefore, a babbling
strategy corresponds to sending messages that are independent of the senders type. Here, the
messages of strategic senders are correlated with their types; however, these messages are designed
to perfectly signal jam the informative messages of the sincere senders. Notice, in particular, that
the strategy of strategic respondents varies with the number of sincere consumers in the population.
Normally, babbling equilibria hold little interest as there are more informative equilibria
provided that the divergence in incentives between the incentives of the sender and receiver is not
too great. Our next proposition shows that, for all parameter values under homogeneous priors,
there is a unique outcome of any monotone equilibrium: the product type is always unresponsive to
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 593
the results of the survey. Thus, not only is information aggregation impossible under homogeneous
priors, credible information transmission is impossible as well. Formally,
Proposition 3. In any monotone equilibrium, the rms product choice is unresponsive to the
survey; that is, y (M) = 0 for all M.
Proposition 3 is, essentially, an unravelling result. The impossibility of simultaneously
satisfying the objectives of the rm and strategic consumers leads to an arms race whereby
strategic consumers send ever more extreme messages which, in turn, leads the rm to ever more
discount the survey results. This process equilibrates only when the rm is totally unresponsive
to survey results.
Nonmonotone equilibria. Proposition 3 points out that, in any monotone equilibrium,
the product type is unresponsive to the results of the survey. A natural question is whether the
rm might do better with a nonmonotone equilibrium. First note that the restriction to monotone
equilibria is stronger than what is needed for the proof. In particular, as long as M and y are
monotone for sufciently large or small values of , the proof of Proposition 3 can be adapted
to show that the product design is unresponsive in any equilibrium. The key is that strategies
eventually become monotone and, for consumers with extreme tastes, the interval in which
strategies are monotone is the only payoff relevant circumstance.
12
The remaining case is where strategies are innitely oscillatory over the entire real line. We
cannot say what the informational properties of such strategies might be, but these strategies seem
sufciently counterintuitive as to be implausible. We investigated such strategies numerically and
were not able to derive any equilibrium with these properties. A full characterization of innitely
oscillatory strategies over the real line remains for future research.
4. Heterogeneous priors
The combination of sincere and strategic consumers in the common priors model proves to
be a volatile mixture leading to severe information loss. In this section, we show that by relaxing
the common priors assumption, information aggregation is restored. There exists an equilibrium
with a simple and intuitive form in which information aggregates: Reports and products are
chosen using linear strategiesstrategic consumers use a reporting strategy t
i
, where is a
parameter, whereas the rm selects the product using the strategy M, where, again, is a
parameter.
Although the modelling of homogeneous priors is straightforward, heterogeneous priors
requires more care. Following Scheinkman and Xiong (2003) and Van den Steen (2004, 2010),
we assume that the prior beliefs of the rm and strategic consumers are commonly known; that
is, the rm and strategic consumers agree to disagree. The main practical application of the
article occurs when the rm treats all responses as being sincere, that is, = 1. Here, the priors
of strategic consumers are irrelevant to the rms product choice for any given M. So long as
strategic consumers are aware that the data will be treated in this fashion, they need only concern
themselves with their own prior beliefs. That is, strategic consumers agree to disagree with the
rm whereas the rm simply discounts the possibility of strategic consumers altogether. This
seems sensible. After all, a strategic consumer clearly knows that some fraction of the population
is strategic, so < 1, whereas the rm disregards strategic distortion of survey responses, that is,
it implicitly holds the belief that = 1. This case corresponds to most practical applications of
survey data, where no statistical corrections are made to account for the possibility of strategic
consumers.
With this background in mind, consider the case where the product is responsive to the
reports of consumers, that is, where = 0 and nite. Recall that the rm chooses product y
12
The detailed calculations are available on request.
C
RAND 2014.
594 / THE RAND JOURNAL OF ECONOMICS
such that
y (M) = E [|M] .
Given and , the rm believes that M converges in probability to +(1 ) . The rm
should then choose such that
( +(1 ) ) = ,
which yields
=
1
+(1 )
. (1)
Now consider strategic consumer i . She should choose m
i
such that
t
i
= E [y (M) |t
i
]
= E [M|t
i
] .
Given her beliefs about the populace of respondents, it follows that
t
i
=
h
n
E
_
t
j
|t
i
_
+(n h
n
1) E
_
t
j
|t
i
_
+m
i
n
,
for j = i . Then,
m
i
=
nt
i

h
n
E
_
t
j
|t
i
_
(n h
n
1) E
_
t
j
|t
i
_
.
Substituting in the strategic respondents strategy m
i
= t
i
and rewriting yields
t
i
n
=
t
i

h
n
n
E
_
t
j
|t
i
_

(n h
n
1)
n
E
_
t
j
|t
i
_
.
In a large survey, we then have
0 =
t
i

E
_
t
j
|t
i
_
(1 ) E
_
t
j
|t
i
_
.
Recall that E
_
t
j
|t
i
_
= E [|t
i
] = at
i
. Solving for , we have
=
_
1

a
_
1
(1 ) a
. (2)
Equilibrium then consists of simultaneously solving equations (1) and (2), which yields:
Proposition 4. Suppose that priors are heterogeneous, that is, = . Then the following comprises
the unique responsive linear equilibrium in a large survey:
m
i
= t
i
y = M,
where =
(a)
a(1)(1)
and =
(a(1)(1))
a()
.
Furthermore, information aggregates in this equilibrium: M converges in probability to the
invertible function () = ( +(1 ) ) .
In our model, the combination of normally distributed tastes and quadratic payoffs leads to
simple linear strategies.
13
While these conditions sufce to ensure the existence of a responsive,
linear equilibrium, they are not necessary. Two ingredients are essential for the result. First, a
13
This is not true of all cheap talk models with these features. For instance, in a version of our model where there
is only a single survey respondent, and this person is known to be strategic, there is no responsive linear equilibrium.
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 595
rm must choose a product design to E [|M] and strategic consumers must try to induce a
product design equal to their tastes. Quadratic loss functions, as well as other specications of
preferences, produce marginal conditions where this is the case.
14
The second key ingredient is
that E [|t ] is of the form t , where (0, 1) is a xed coefcient. Although this holds under
normality, other distributions also have this property. Indeed, for any distribution of having
positive density over R, there exists a distribution of such that this condition is satised.
15
Information aggregates under heterogeneous priors because, when rms and consumers have
differing beliefs about the fraction of sincere consumers in the population, they can simultaneously
believe that their objectives are being met. In particular, given beliefs , the rm is choosing its
product to match its perception of the average type in the population, . At the same time,
given beliefs , strategic consumers respond to the survey condent that they will get their most
preferred product in expectation.
The strategies contained in Proposition 4 are not the only linear equilibrium. Indeed, the
strategies offered in Proposition 2 also comprises a linear equilibrium, and, because the proof
makes no use of the particular beliefs held by strategic respondents, the proof works under
heterogeneous priors as well. Because this equilibrium is unresponsive to the underlying state, we
discard it fromthe subsequent analysis. Obviously, ops can arise when the rmobtains no usable
information froma survey. The more interesting case is when it does receive (and act upon) usable
information.
To summarize, despite the apparent complexity of the model, equilibrium can be strikingly
simpleconsumers and the rm play linear strategies and information aggregates in a large sur-
vey. Moreover, the equilibrium is unique in the class of linear strategies (using the responsiveness
renement). This is in sharp contrast to standard models of strategic information transmission in
which information often does not aggregate, strategies are typically partitional, and equilibrium
multiplicity is commonly a problem. One might then be tempted to conclude that, with heteroge-
neous priors, a rms decision about whether to conduct a large survey or to forego surveying has
a clear answerin favor of the large survey. As we will show in the next section, the presence of
heterogeneous priors makes this determination less obvious than it rst appears.
5. Flops
In this section, we investigate the quality of product designs when a rm conducts a
large survey. Earlier we saw that information aggregated in a linear equilibrium, and thus, one
might have some measure of condence that large surveys would produce correct product
designs. Our main result of this section is to identify conditions where, instead of producing good
designs, acting on a large survey produces ops with high probability. Put differently, information
aggregation does not inoculate a rm from producing a design that is ill suited to the average
consumer.
We rst consider what would seem to be an obvious question: Should a rm undertake a
large survey in designing its product? Recall that when a rm designs based on a large survey, its
expected payoffs are
V = (M )
2
,
and because M converges in probability to ( +(1 ) ) , then, using the expressions for
and given in Proposition 4, it follows that for large n, the rms payoffs converge in probability
to
V =
_
( ) a +
( ) a
1
_
2

2
14
In particular, the Taylor rule, a commonly used objective function in macroeconomics, has this property.
15
Specically, when and are both drawn from the same distribution, it may be readily veried that E [|t ] = t /2
for all t.
C
RAND 2014.
596 / THE RAND JOURNAL OF ECONOMICS
=
_

_
2

2
. (3)
Notice that the rms expected payoffs depend on a pair of ratios. We term the rst expression,
|( ) / ( )|, the calibration-alignment ratio. The numerator reects how well calibrated
the rms beliefs are about the fraction of strategic consumers in the population. The denominator
reects the alignment of priors between the rm and strategic consumers. We term the second
expression,
2

/
2

, the noise-signal ratio. It is the ratio of the variance in the idiosyncratic com-
ponent of consumers tastes compared to the variance of the systematic component. Equation (3)
reveals that, despite information aggregation, the product design does not converge to the tastes
of the average consumer, . Otherwise, this expression would be zero. Thus, unless there is no
variance in the idiosyncratic component of consumers tastes or the rms priors are perfectly
calibrated with the truth, the rms product designs meet the needs of the market only imperfectly.
Now suppose that the rm eschews surveys entirely. Lacking any additional information on
which to make its design decision, the rm opts for a design that matches its prior beliefs about
, that is, y = E [] = 0. In this case, for every realization of , such a rm suffers losses of
V
0
=
2
. (4)
Comparing equations (3) and (4), it may be readily seen that performing a large survey is superior
if and only if

< 1. (5)
Thus, we have established that:
Proposition 5. A rm benets from a large survey if and only if the magnitudes of the calibration-
alignment ratio and the noise-signal ratio are sufciently small. Formally, a large survey is better
than no survey if and only if equation (5) holds.
Proposition 5 reveals that, even when information aggregates in a survey, rm management
may be better off ignoring this information entirely and designing based on gut feel. This is more
likely to be the case for products that have a large idiosyncratic taste component or in settings
in which rms are relatively less sure about the fraction of strategic consumers. In technology
industries where the product is novel and difcult for consumers to envision ahead of time, one
might expect that the noise-signal ratio will be large. Likewise, for the design of intermediate
inputs in the supply chain, survey respondents will be other downstream businesses. In this case,
both the rm and the consumers probably share similar beliefs about the fraction of strategic
types in the population and hence the calibration-alignment ratio is likely to be high. In either
case, a rm might be better off not surveying at all.
When do product designs turn out to be ops? To examine this question, dene a product as
a op if it is more than z > 0 standard deviations away from the tastes of the average consumer,
. The probability that a design is a op following a large survey is then
Pr[op] = Pr[|M | > z

],
and, because M converges in probability to ( +(1 ) ) , it then follows (after some algebra)
that
Pr [op] = 2
_

z
_
, (6)
where () is a cumulative distribution function of a standard normal random variable. Notice
that the chance of a op again depends on the calibration-alignment ratio, the noise-signal ratio,
and z, the tolerance for missing the ideal product, .
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 597
FIGURE 1
PROBABILITY OF A FLOP AS A FUNCTION OF , THE PRODUCT OF THE CALIBRATION-ALIGNMENT
RATIO AND NOISE-SIGNAL RATIO
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
4 3 2 1 0
Probability
of
flop

Probability of a f lop
for a large survey
Probability of a f lop
without a survey
When the rm eschews surveys entirely, the probability of a op is simply
Pr [op] = 2(z) . (7)
The comparison between equations (6) and (7) is straightforward. Once again, a rm reduces
the chance of a op by conducting a large survey (relative to not surveying) if and only if
|( ) / ( )|
2

/
2

> 1, which is equivalent to the condition in (5) derived under payoff


maximization. Thus, we have shown:
Proposition 6. Choosing whether to conduct a large survey so as to maximize the rms expected
payoff is equivalent to choosing whether to conduct a large survey to reduce the probability of a
op.
In other words, payoff maximization and op minimization lead to the same decision as to
when to conduct a large survey. Proposition 6 implies that, when the rm wishes to maximize
payoffs, it can delegate the product design decision to a product manager who might have very
different concernssay, avoiding a failed productand still achieve this objective.
We have emphasized the critical role played by the calibration-alignment and noise-signal
ratios in whether a design turns out to be a op. It is useful to dene to be the product of these
two ratios. Formally, = |( ) / ( )|
2

/
2

. If we let the op threshold be z = 1/2


(i.e., a product is deemed a op if it falls outside of a one standard deviation band around ),
then we can compute the probability of a op as a function of . Figure 1 depicts this plot. The
gure highlights that, unless is very low, the chance of ops is substantial. Indeed, even when
is only 0.75, the majority of products are ops. At the point of indifference between surveying
or not, the chance of a op is about 62 percent. Surveying when is higher than this indifference
point further exacerbates the ops problem. The main point of the gure is that, unless tastes
are highly correlated (i.e.,
2

is small) and the rm is well calibrated (i.e., is close to ), the


probability of producing a op is surprisingly high despite information aggregation.
Conventional wisdom suggests that large surveys are most important when tastes exhibit
large idiosyncratic variation (e.g., Lehmann, Gupta, and Steckel, 1998). Under the standard view,
the size of the survey helps to wash out this variation, thus revealing the systematic component
of tastes. Our results suggest the opposite conclusion. When idiosyncratic taste variation is large,
surveys perform poorly and ops are likely. Indeed, in the limit, as
2

, a op results almost
surely. The reason that idiosyncratic taste variation matters is that, because the rm gives little
weight to individual survey responses (relative to its prior beliefs), strategic respondents offer
C
RAND 2014.
598 / THE RAND JOURNAL OF ECONOMICS
more extreme messages to shift the rms product design to their ideal. The result of this tension
is that, in the limit, strategic respondents effectively jam the signals of sincere respondents and
the rm becomes hypersensitive to small changes in M and responds accordingly. As the rm
is not perfectly calibrated, this hypersensitivity magnies the rms errors in design, inevitably
resulting in a op.
Our results also highlight the special role played by assuming priors are homogeneous.
Under this assumption, one comes to the pessimistic conclusion that surveys are never helpful.
When priors are not homogeneous, there is scope for information revelation in large surveys.
Moreover, the greater the disagreement in the priors of the rm and the strategic types, the
more scope there is for surveys to be helpful. One might be tempted to equate disagreement in
the sense of differences between and with misalignment of preferences. In most models of
strategic information transmission, such misalignment is harmful to information transmission
(e.g., Crawford and Sobel, 1982). Here, disagreement is helpful in that it reduces the incentives
for strategic types to distort and for rms to respond to these distortions. Indeed, our results show
that, all else equal, when disagreement is large, ops are unlikely.
Outsized role of strategic consumers. The ops problem is caused by distortions in the
reports of strategic consumers; that is, on their outsized role in determining survey results.
One might have the intuition that strategic consumers wield this inuence by choosing either
very large or very small survey responses, (i.e., large equilibrium values of ). Were this the
case, then one possible solution to the ops problem would be to constrain the space of reports
through upper and lower bounds. The next example shows that, even when strategic consumers
only modestly exaggerate their tastes, aggregate survey results can be sufciently distorted that
ops are a signicant problem; thus, capping the reporting space is not a panacea.
Example 1. Suppose = 1, = 3/4, = 7/10,
2

= 2, and
2

= 1. Strategic types exaggerate


modestly, sending a report m = 3t . The chance of a product that falls outside of a one standard
deviation band around is 40.5% in a large survey, so ops are a signicant possibility.
The relatively poor calibration of the rm in Example 1 is obviously helpful in generating
the result. A more subtle intuition might be that, when the rm is well calibrated, extreme
exaggeration is required to produce ops. Our next example shows that exaggeration increases
only modestly and ops arise fairly frequently even when the rm is closely calibrated to the true
fraction of strategic consumers.
Example 2. Suppose = 1, = = 0.95,
2

= 5, and
2

= 2. Strategic types send a report


m = 9t . The chance of a product that falls outside of a one standard deviation band around is
21.1% in a large survey.
Again, it is the miscalibration rather than the exaggeration by strategic consumers that is at
the heart of the ops problem. One might think that the problem essentially disappears when the
rm is arbitrarily well calibrated. Our next result shows that this is not the case. Recall that the
chance of a op depends on the calibration-alignment ratio rather than solely on the calibration
level itself. With this in mind, the next example shows that, even when nearly all consumers are
sincere, the chance of ops remains stubbornly high.
Example 3. Suppose = 1 and strategic consumers are well calibrated, that is, = < 1. In the
limit as all consumers become sincere, the probability of a op becomes
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 599
lim
1
2
_

1
1

z
_
= 2
_

z
_
> 0,
and is independent of along the limiting path.
The key insight fromthis example is that the probability of a op depends on the noise-signal
ratio regardless of how well the rm is calibrated. Thus, the presence of even a small fraction of
strategic consumers fundamentally changes the quality of rm decisions compared to the case in
which these consumers are absent.
6. Avoiding ops
The inuence of strategic consumers stems, in part, from the fact that the rm is observing
only a summary statistic of the survey data, the sample mean, rather than the data itself. For
normally distributed tastes, the sample mean is a sufcient statistic when all consumers are
sincere, but not when there are strategic consumers. Typically, one needs to turn to scoring rules
or other exotic mechanisms to induce truthful revelation and thereby avoid ops. Here we show
that no such complexity is required.
Instead, a rm need only adjust its choice of summary statistics in interpreting the results of
survey data. Absent strategic respondents, there are many other summary statistics that the rm
might use to infer tastes and, indeed, in a large survey the rm ought to be indifferent among these
alternatives. In this section, we show that, in the presence of strategic respondents, the seemingly
innocuous choice of summary statistics has profound effects on incentives for truthful reporting
and hence on the informational properties of surveys. We consider two such statistics and show
how, merely through the judicious use of summary statistics, the rm can solve the ops problem.
Binary surveys. Suppose the rm continues to solicit open-ended responses to its survey
and consumers still offer costless messages in R; however, now the rm bases its designs only on
the statistic, the fraction of responses where m 0. In the limit, this is also a sufcient statistic
for the rmto deduce the state froma large sample of sincere consumers with normally distributed
tastes. Here we will show that the incentive properties of this statistic are vastly superior to using
the more familiar sample mean as the basis for decision making.
Proposition 7. When the rmuses a statistic as the basis for its decisions, then truthful revelation
is an equilibrium in a large survey. The statistic solves the ops problem.
The proof is by construction. First, suppose that the rm acts as though all surveyed con-
sumers were truthful. Under this hypothesis, in a large survey, will converge in probability to
1
_

_
.
Hence, in equilibrium, the rms product will converge in a large survey to
y () =

1
(1 )
= .
It remains to show that strategic consumers can do no better than respond truthfully. We
provide an intuitive argument below and relegate the formal argument, which is more involved,
to the Appendix. A strategic individual faces a choice between reporting m 0 and m < 0. The
particular value of m in each region is irrelevant. Given taste t
i
, a strategic consumer expects
to take on the value at
i
; thus, in a large survey, when t
i
> 0, a strategic consumer expects that
the design will converge to at
i
< t
i
. As the equilibrium design is lower than her ideal design,
she strictly prefers to report m = t
i
rather than any value of m < 0 because reporting m < 0 will
C
RAND 2014.
600 / THE RAND JOURNAL OF ECONOMICS
only lead to a design that is further from her ideal in expectation. Similarly, when t
i
< 0, the
equilibrium design is higher than her ideal design, so she strictly prefers to report m = t
i
rather
than any value of m 0 since reporting m 0 will only lead to a design that is further from her
ideal in expectation. Therefore, truth telling is incentive compatible for strategic consumers, and
the rms expectations about consumer reports are conrmed. Thus, we have identied a fully
revealing equilibrium, regardless of calibration, alignment, or the noise-signal ratio.
Proposition 7 points out the crucial connection between the data processing strategy of the
rm and incentives for the strategic consumers. The key to solving the ops problem is for the
rm to focus on the fraction of consumers sending positive messages. A simple way a rm might
gather this data is by using blunter survey instruments, such as Likert-type scales. In terms of
the model, were the rm to survey consumers using only a simple binary survey instrument, then
the equilibrium we constructed in Proposition 7 would remain valid and the rm could end up
making better decisions than if it used a more comprehensive or open-ended survey instrument.
In short, when some consumers are strategic, the rm can learn more by asking less in a large
survey.
Sample median. The binary survey instrument works by not allowing strategic individuals
to exaggerate their type, but at the cost of coarsening the information gleaned from surveys when
there are only a small number of respondents. A different way one might solve the exaggeration
problem is by trimming the summary statistic, so it is less sensitive to outliers. Utilizing the
median, rather than the sample mean, represents the most stringent trimming strategy available
to the rm. To analyze this problem, it is helpful to discretize the action spaces of the players and
allow the grid over the set of possible actions to become arbitrarily ne. Specically, suppose
that the action spaces of rms and consumers are constrained to a grid ranging from l to l. Let
g > 0 denote the neness of the grid. There are 2l/g +1 actions available. Thus, the set of actions
is given by m {l, l + g, l +2g, .. , 3g, 2g, g, 0, g, 2g, 3g, . . . l 2g, l g, l} and
similarly for y. Note that as g becomes arbitrarily small and l becomes arbitrarily large, the action
space becomes arbitrarily close to the action space in the continuum model.
Consider a survey consisting of an odd number of individuals (so the median is unambigu-
ously dened) where the rm bases its product choice on the median report.
16
We will show that
truth telling comprises an equilibrium in a large survey utilizing this technology. As the message
space and the type space no longer coincide, we need to dene truth telling. We will say that a
truthful response is one where the message that is closest to the consumers taste is selected. As
above, we will assume that sincere consumers are truthful and consider the incentives of strategic
consumers. A truth-telling equilibrium is an equilibrium where all consumers engage in truthful
strategies and the rmis choosing a product optimally under the assumption of truthful responses.
Proposition 8. For any grid with neness g and length 2l, when the rm selects its product on the
basis of the median report, truth telling is an equilibrium for a sufciently large survey.
Why does trimming using the median lead to truth telling? The report of a strategic type
is relevant only to the extent that it is the median report and thus, in determining whether to
distort a report, a strategic consumer conditions on that event. In a small survey, the rm will still
anchor on its prior beliefs, and therefore there is still some benet from limited exaggeration.
As the survey size increases, however, less weight is placed on the rms prior belief, and
therefore there is little incentive to distort conditional on being the median report. The discrete
action space ensures that there is a sufciently large survey whereby this slight misalignment of
preferences between the rm and the strategic consumer ceases to matter and reporting becomes
truthful.
16
The results also hold if there is an even number of respondents.
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 601
Now one might conjecture that the same effect will also hold when the sample mean is
the summary statistic. The key difference is that, when the median statistic is used, strategic
consumers only affect the product choice conditional on being the median report, whereas when
the sample mean is used, their reports always have a small effect on the product choice. Thus,
under the median, strategic types have a small chance of having a large effect on the product
design, while with the sample mean, they have a large (probability one) chance of having a small
effect on the product design. In the former instance, truth telling is optimal, whereas in the latter
instance, it is not, because the product design does not necessarily coincide with the tastes of the
strategic consumer.
Because Proposition 8 holds for any grid with neness g and length 2l, it follows that, as we
let the length become arbitrarily long (i.e., l ) and the gaps become arbitrarily small (i.e.,
g 0), truth telling remains an equilibrium for large surveys along this sequence. Moreover, as
g 0 and l , the action space converges to the countably innite analog of the continuum
action model. While this is not a proof that truth telling or near truth telling is an equilibrium in
the continuum model, it is suggestive.
Finally, while Proposition 8 was derived under the assumption that the systematic and
idiosyncratic components of consumer tastes are normally distributed with known mean and
variance, the result also holds under more relaxed assumptions. In particular, so long as the
idiosyncratic taste component is symmetrically distributed around zero, then the systematic taste
component can be drawn from any distribution and truth telling will still hold. Thus, the rm
need only have minimal knowledge of the data generating process giving rise to consumer tastes
for ops to be avoided using this statistic. Moreover, note that truth telling is an equilibrium
for a sufciently large sample irrespective of the rms beliefs about the fraction of sincere and
strategic consumers, and in particular, irrespective of whether the rm is well calibrated.
Discussion. All of these solutions to the ops problemrely on two key ingredients. First, the
method used by the rm to analyze the survey data must be common knowledge; and second, the
rm must somehow commit to its information-processing strategy. There are various practical
means by which the rm might achieve both goals simultaneously. For instance, the use of
lightweight feedback devices such as the Like button on Facebook ensures that the rm will
have only binary data on which to make decisions and signals quite effectively to consumers that
only binary responses can be used in formulating product designs. Likert-type scales commonly
used in gathering survey data play a similar role.
To see what can go wrong when either of these ingredients is absent, consider a situation
where a rm gathers data from a limited sample and uses trimmed means, perhaps even the
median, in determining product designs. If strategic consumers fail to anticipate that only the
median will be used, they will have no incentive to tell the truth. Even if the rm somehow
conveys its intent to use the median as the sole statistic and consumers respond truthfully, the
rm will be sorely tempted to use all of the sample data to glean additional information about the
state. Strategic consumers can anticipate this and will no longer wish to answer truthfully. Thus,
the rm needs to somehow commit. In principle, one way of doing this would be to hire a third
party to conduct the survey, retain the raw data, and only report the relevant statistic to the rm.
The rm might still be tempted to ask the third party for the raw data, but mechanism design
solutions like nancial hostages (Aghion, Dewatripont, and Rey, 1994) can mitigate this concern.
If there is some friction, such as a hassle cost of transferring data or a small fee for the transfer,
then, provided the sample is sufciently large, the rm again will not be tempted because the
incremental gains from this information become arbitrarily small in a large survey. While formal
mechanism design solutions appear rare in practice, fee agreements for additional data requests
are quite common.
17
17
For example, surveys conducted by the large survey rm Knowledge Networks often contain these provisos.
C
RAND 2014.
602 / THE RAND JOURNAL OF ECONOMICS
7. Conclusion
The science of market research has progressed rapidly since the time of the quintessential
op, the Ford Edsel (Carlson, 2007). Now, at relatively modest cost, rms can (and do) conduct
large surveys of consumers tastes. This data is critical to nal product offerings. Yet, despite
these gains, products that prove to be ops are commonplaceas any visit to a local movie theater
will attest.
When rms rely on the sample mean in making product design decisions, they expose
themselves to the risk that sophisticated consumers will answer survey questions strategically
rather than honestly. As a consequence, relying on the sample mean of an open-ended survey,
even when the sample size is large, by no means guarantees the avoidance of ops. The effects of
these strategic consumers on survey data depend crucially on the beliefs held by consumers and
rms about their prevalence. When beliefs are common, the presence of even a small fraction
of strategic respondents makes information aggregation impossible. In contrast, when beliefs
differ, large surveys can still aggregate information. Moreover, equilibrium reporting strategies
are simple and intuitivethere is a unique, responsive equilibrium where reporting strategies are
linear in tastes. The key intuition is that heterogeneous prior beliefs allow strategic consumers
and rms to simultaneously believe that their objectives are being met. That is, all parties believe
that the survey will produce their ideal product even though views as to what this product is
differ.
Even when information aggregates, ops remain a distinct possibility. We show that the
chance of a op depends on two ratios: the noise-signal ratio, which measures the relative amount
of idiosyncratic taste variation among consumers, and the calibration-alignment ratio, which
measures the accuracy of the rms views about the fraction of strategic consumers relative to
the degree to which priors are aligned. Absent strategic consumers, the noise-signal ratio would
have no bearing on the quality of a large survey in that idiosyncratic taste variation washes
out of the analysis. With sophisticated consumers, a higher noise-signal ratio leads to larger
errors in the rms assessment of the tastes of the average consumer and, as a consequence, more
opseven in the limit. The calibration-alignment ratio reveals that rms are better served when
their beliefs as to the fraction of strategic consumers are well calibrated with the true fraction and
worse off when consumers share the rms beliefs. Because the chance of a op depends on the
calibration-alignment ratio, ops still arise even as a rm becomes well calibrated. When these
ratios are large and the rm uses the sample mean of the survey responses to make a decision,
a rm is better off eschewing surveys entirely rather than relying on large surveys. This is true
even if surveys are costless to conduct.
Our results highlight the hazards of using open-ended surveys and standard summary mea-
sures, such as the sample mean, to determine product choices. This is not to say that surveys
cannot play an important role in product design provided they are well designed. We offer two
simple variations in survey design that solve the ops problem. Offering a binary survey where
individuals can only report the direction of their tastes relative to the status quo completely elim-
inates the incentives for misrepresentation. Alternatively, extreme trimming of outliers through
the use of the median report can also align incentives in a large survey. Both designs capture
enough information so that the rm can make optimal product choices. By asking for or using
less information in its survey, the rm ends up obtaining more and better information. Thus, our
model suggests an incentive rationale for the widespread use of Likert-type scales and trimming
in survey design.
More broadly, though many rms embrace surveys as a central input in their decision
making (e.g., Deshpande and Zaltman, 1982; 1984), this article highlights the importance of
careful survey design and analysis of the data. Sometimes survey designers worry about choosing
questions to extract as much information as possible without worrying too much about incentives
to misrepresent. Likewise, statistical techniques for analyzing survey data often fail to correct for
strategic misrepresentation. However, consumers are not mere passive data points nor draws from
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 603
some unbiased urn. They have their own preferences, their own agendas, and will pursue these if
given the opportunity. Seemingly well-meaning but naive procedures can easily give rise to ops
whereas procedures that seem less effective at extracting information, in fact, can be powerful
tools for making correct decisions.
We have couched our results in terms of surveys rms conduct to guide their product
offerings, but our model is also relevant to political decision making. Polling data increasingly
guides the formulation of policy. If anything, the problem of strategic survey respondents is
even more pronounced when one is dealing with emotionally charged issues such as abortion,
immigration, and the environment. Our results offer a cautionary tale for policy makers seeking
to govern or design policies merely to reect the will of the people as measured by survey
data. Product ops are unfortunate, but policy ops can have far graver consequencesnot just
to individual politicians, but to society at large.
Appendix
The following lemma proves useful in establishing Propositions 1 and 3.
Lemma 1. In any equilibrium, y (M) is differentiable and therefore nite for all M.
Proof. To establish the result, rst, x a value M for the realized sample mean. Dene X (M) =
_
(m
1
, m
2
, . . . , m
n
) :
1
n

m
i
= M
_
. Suppose, without loss of generality, that i = 1 is a sincere consumer. Then it is
obvious that X (M) +(n, 0, . . . , 0) X (M +) for all , where X (M +) is analogously dened.
Claim. For all , X (M +) = X (M) +(n, 0, . . . , 0) .
To establish the claim, suppose, without loss of generality, that i = k, k +1, . . . , n are strategic consumers for
some value k and x their reports at m

i
. Call this vector m
k
. Now dene a set
X (M| m
k
) =
_
(m
1
, m
2
, . . . , m
k1
) :
1
n
_
k1

i =1
m
i
+ m
k
I
T
_
= M
_
=
_
(m
1
, m
2
, . . . , m
k1
) :
k1

i =1
m
i
= nM m
k
I
T
_
,
where I
T
denotes the transpose of the unit vector of size n (k 1). Similarly, dene
X (M +| m
k
) =
_
(m
1
, m
2
, . . . , m
k1
) :
k1

i =1
m
i
= n (M +) m
k
I
T
_
.
To establish the claim, consider an arbitrary element of the set X (M +| m
k
). A generic element of this set satises
k1

i =1
m
i
= n (M +) m
k
I
T
,
or, equivalently,
m
1
+
k1

i =2
m
i
= n (M +) m
k
I
T
m
1
n +
k1

i =2
m
i
= nM m
k
I
T
.
However, the right-hand side of this expression denes a generic element of X (M| m
k
). Thus, we have shown that for
every m

X (M +| m
k
), there exists an element in the set X (M| m
k
) such that adding n to m
1
yields m

. Because this
holds for all m
k
, the claim then follows.
Next, we will show that y (M) is differentiable and nite. Recall that
y (M) = E [|M] ,
where
E [|M] =
_
mX(M)
( m|) f () d
_
mX(M)
( m|) f () d
,
C
RAND 2014.
604 / THE RAND JOURNAL OF ECONOMICS
where m is a generic 1 n vector of messages in X (M) and is the likelihood of the realization of messages m given .
Similarly, when the realized sample mean is M +, we obtain
y (M +) = E [|M +]
=
_
mX(M+)
( m|) f () d
_
mX(M+)
( m|) f () d
.
Now, using the fact that X (M +) = X (M) +(n, 0, . . . , 0), we can equivalently express this as
E [|M +] =
_
mX(M)
( m +(n, 0, . . . , 0) |) f () d
_
mX(M)
( m +(n, 0, . . . , 0) |) f () d
.
In addition, because the likelihood function, , is differentiable in its arguments, it then follows that E [|M +] (and
hence, y (M +)) is differentiable and nite. Q.E.D.
Proof of Proposition 1. First, note that information does not aggregate for a survey of nite size since, even if all reporting
were truthful, the survey information would not reveal the state. It remains to show that there does not exist an equilibrium
for a large survey in which information aggregates.
Suppose by means of contradiction that there exists a convergent sequence of equilibria
n
such that lim
n

n
=

where information aggregates. Assume, without loss of generality, that consumer i = 1 is a strategic consumer. Her
expected payoff from reporting a message m
1
is given by
E

_
(y (M) t
1
)
2
|t
1
_
.
Therefore, for nite n, the consumer sends a message m
1
satisfying
max
m
1

_
y
_
m
1
+

n
i =2
m
i
n
_
t
1
_
2
d H
_
n

i =2
m
i
|t
1
_
,
where H
_
n
i =2
m
i
|t
1
_
is the cumulative distribution function of the sum of the other consumers messages conditional
on t
1
. Differentiating with respect to m
1
and using the rst-order condition yields
_

_
y
_
m
1
+

n
i =2
m
i
n
_
t
1
_
y

_
m
1
+

n
i =2
m
i
n
_
d H
_
n

i =2
m
i
|t
1
_
= 0,
for the equilibrium m
1
. By Lemma 1, we know that the derivative y

() is nite for all M. By denition, information


aggregation implies that plim
n

n
i =1
m
i
/nis an invertible function M(). Optimality on the part of the rm implies
that y (M ()) = in the limit of an arbitrarily large survey.
Thus, taking limits as n , the rst-order condition converges to
_

( t
1
) y

(M ()) f (|t
1
) d = 0, (A1)
where f (|t ) is the probability density function corresponding to the posterior beliefs of a consumer with type t . A
necessary condition for equation (A1) to hold is that
_

x
( t
1
) y

(M ()) f (|t
1
) d
exists and is nite for all x. Thus, we may rewrite equation (A1) as
_
t
1

( t
1
) y

(M ()) f (|t
1
) d +
_

t
1
( t
1
) y

(M ()) f (|t
1
) d = 0,
where both expressions are nite.
Case 1: Suppose that y

(M ()) > 0 almost everywhere. Then limit as t


L
, the left-hand side of equation (A1)
converges to:
lim
t
1

_
t
1

( t
1
) y

(M ()) f (|t
1
) d + lim
t
1

_

t
1
( t
1
) y

(M ()) f (|t
1
) d
= lim
t
1

_
t
1

( t
1
) y

(M ()) f (|t
1
) d
< 0,
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 605
where the rst equality follows from the fact that, since
_

t
1
( t
1
) y

(M ()) f (|t
1
) d exists and is nite for all t
1
,
observe lim
t
1

t
1
( t
1
) y

(M ()) f (|t
1
) d = 0. The inequality follows from the fact that < t
1
for all values of
the integrand and y

(M ()) > 0.
Case 2: Suppose that y

(M ()) < 0 almost everywhere. An analogous argument establishes that equation (A1) cannot
hold.
Therefore, we may conclude that for t
1
sufciently large (small), a strategic consumers rst-order condition cannot
be satised and hence information aggregation is impossible. Q.E.D.
Proof of Proposition 2. Because the rms product does not vary with M, the strategic respondents strategy is consistent
with equilibrium play. It remains to show that there is no protable deviation for the rm. Suppose, without loss of
generality, that the rm believes that consumers i = 1, 2, . . . , k are sincere while the remainder are strategic; in this
proof, the subscript n on the variable k has been dropped to streamline the exposition. Given the rms beliefs,
M =
t
1
+ +t
k
n
+
(t
k+1
+ + t
n
)
n
can be written as
M =
k
n
+

1
+ +
k
n
+
(n k)
n
+
(
k+1
+ +
n
)
n
,
Using the fact that = k/ (n k) yields
M =

1
+ +
k
n
+
(
k+1
+ +
n
)
n
.
Taking expectations yields
E [|M] = E [] = 0.
Thus, y = 0 is optimal. Notice that the result is independent of the beliefs of strategic consumers, and, therefore, applies
regardless of whether the rm and the strategic consumers have homogeneous or heterogeneous priors as to the fraction
of strategic consumers in the population. Q.E.D.
Proof of Proposition 3. Suppose, to the contrary, that y (M ()) = 0 for a positive measure of values of . Consider the
strategic consumers limiting rst-order condition, which can be written as
_

(E [|M ()] t
1
) y

(M ()) f (|t
1
) d = 0,
where, by Lemma 1, we know that y

(M ()) is nite. Recall that there exists a xed value a E [|t ] /t , and x b such
that b (a, 1). We then may rewrite the rst-order condition as
_
bt
1

(E [|M ()] t
1
) y

(M ()) f (|t
1
) d +
_

bt
1
(E [|M ()] t
1
) y

(M ()) f (|t
1
) d = 0.
Now, taking limits as t
1
yields
lim
t
1

_
bt
1

(E [|M ()] t
1
) y

(M ()) f (|t
1
) d
+ lim
t
1

_

bt
1
(E [|M ()] t
1
) y

(M ()) f (|t
1
) d
= lim
t
1

_
bt
1

(E [|M ()] t
1
) y

(M ()) f (|t
1
) d.
The equality follows from the fact that, since the integrand is everywhere bounded, the second term becomes negligible
in the limit.
Now, recall that, for all bt
1
, E [|M ()] E [| bt
1
]. Furthermore, for t
1
sufciently large, E [| bt
1
] <
t
1
. Together, these inequalities imply that, for sufciently large t
1
, E [|M ()] < t
1
for all bt
1
.
Case 1: Suppose that y

(M()) 0. Then
lim
t
1

_
bt
1

(E [|M ()] t
1
) y

(M ()) f (|t
1
) d 0,
since the integrand is everywhere nonpositive. Thus, the rst-order condition can only be satised when y

(M()) = 0.
C
RAND 2014.
606 / THE RAND JOURNAL OF ECONOMICS
Case 2: Suppose that y

(M()) 0. An analogous argument shows that the rst-order condition can only hold when
y

(M()) = 0.
Therefore, in any equilibrium, y

(M()) = 0. As the rmmust be acting optimally when choosing a product design,


it follows that y(M()) = 0, that is, the product design is unresponsive to the survey results. Q.E.D.
Proof of Proposition 7. Recall that f (|t ) is dened to be the probability density function corresponding to the posterior
beliefs of a consumer with type t ; these beliefs are normally distributed with mean at and variance a
2

. Under the
putative truth-telling equilibrium, the consumer believes that the product design converges to . Suppose that t < 0. If
the consumer changes her message from m < 0 to m 0, this will change the mean product design by some positive
amount. The difference in consumers payoff from reporting truthfully compared to a deviation where m > 0 is
u (t, m < 0) u (t, m > 0) =
_

( t )
2
f (|t ) d
_

( +q (, n) t )
2
f (|t ) d,
where q(, n) denotes the change in the product design as a consequence of a consumer changing her message. Below,
we will show that this expression is positive; that is, the consumer does not benet from a deviation.
We rst show that for large n,
q (, n) =

_,
where is the cumulative distribution function of a standard normal random variable and

its associated density. To see


this, note that when a consumer changes her message and reports m 0, the fraction of positive messages changes from
to +1/n, where is the fraction of positive messages reported when all consumers are truthful. When all consumers
are truthful in equilibrium, the rmchooses a project that equals y =

1
(). In contrast, when a consumer changes her
message and reports m 0, the rm chooses a project equal to y

1
( +1/n). Thus, when a consumer changes
her message, it changes the product by an amount equal to

1
( +1/n)
1
()
_
. For large n, the change in the
rms product converges to
q (, n) =

n
d
d
_

1
()
_
=

(
1
())
=

_.
It now follows that the change in the consumers utility is given by
_

_
( t )
2
+( +q (, n) t )
2
_
f (|t ) d
=
_

_
q (, n)
2
+2q (, n) ( t )
_
f (|t ) d
=
_

_
_
_
_

_
_
_
2
+
2

( t )
n

_
_
_
1

2
_
a
2

exp
_

( at )
2
2a
2

_
d
=
_

_
_
_
_

n
_
1

2
exp
_

1
2
_

_
2
__
_
_
_
_
2
1

2
_
a
2

exp
_

( at )
2
2a
2

_
d (A2)
+
_

( t )
n
_
1

2
exp
_

1
2
_

_
2
__
1

2
_
a
2

exp
_

( at )
2
2a
2

_
d,
where the second equality follows after substituting in for q and the third equality follows from expanding the expression.
Consider each term separately. Observe that the rst term in (A2 ) can be written as
_

_
_
_
_
_
_
_
_

n
_
1

2
exp
_

1
2
_

_
2
__
_
_
_
_
2
1

2
_
a
2

exp
_

( at )
2
2a
2

_
_
_
_
_
d
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 607
=
_

2
_

n
2

a
exp
_
1

_
1
1
2a
_

2
+

t
1
2
a

t
2
_
_
d,
which, for all t < 0, converges to innity when a > 1/2 because (1 1/ (2a)) > 0 and is positive when a 1/2.
Observe that the second term in (A2) can be written as
_

( t )
n
_
1

2
exp
_

1
2
_

_
2
__
1

2
_
a
2

exp
_

( at )
2
2a
2

_
d
=
_

2
n
t

a
exp
_

2
a ( at )
2
2a
2

_
d
=
2
n

a
exp
_
t
2
a
2
2 (1 a)
2

__

( t ) exp
_
(1 a)
_

at
1a
_
2
2a
2

_
d
=
2

2
n
_

2

1 a
exp
_
t
2
a
2
2 (1 a)
2

__

( t )

2
_
a
2

1a
exp
_
(1 a)
_

at
1a
_
2
2a
2

_
d
= t
2

2
n
_

2

1 a
exp
_
t
2
a
2
2 (1 a)
2

__
a
1 a
1
_
.
Thus, the expression
_

( t )
n
_
1

2
exp
_

1
2
_

_
2
__
1

2
_
a
2

exp
_

( at )
2
2a
2

_
d
is nite when a > 1/2, is zero when a = 1/2, and is positive when a < 1/2 for all t < 0.
Combining the two terms, we observe that the sum of the terms converges to innity when a > 1/2 and is positive
when a 1/2. This indicates that regardless of whether a > 1/2 or a 1/2, a consumer with t < 0 will not choose to
deviate and send a positive message. Analogous arguments apply when t > 0. Q.E.D.
Proof of Proposition 8. We show that for any g and l, there is some sufciently large n such that it is an equilibrium for
the rm to choose a product equal to the median of the sample reports, and for the strategic consumers to send a message
equal to the point on the grid closest to their type. First note that if the strategic consumers always send a message equal
to the point on the grid closest to their type, then in the limit as the number of survey responses becomes arbitrarily large,
the probability the median of the sample reports equals the point on the grid that is closest to becomes arbitrarily close
to 1, and the best response for the rm is to choose a product that is equal to the median of the sample reports. Also note
that if the rm always chooses a product equal to the median of the sample reports, then the best response for the strategic
consumers is to send a message equal to the point on the grid closest to their type, m

. If a strategic consumer deviated


by instead sending a message m

+ j g for some positive integer j (assuming such a deviation is feasible), this change
would only change the sample median in cases in which this median was in [m

, m

+ j g) before, in which case the


deviation could only lead to a greater median than before. As the rm is choosing a product equal to the sample median,
this would lead to a product that is further from the strategic consumers taste than before. This would not be a protable
deviation. Similar reasoning shows that a strategic consumer cannot protably deviate by sending a message smaller than
m

. Thus, it is a best response for a strategic consumer to send a message equal to the point on the grid that is closest to
her type. Q.E.D.
References
ADMATI, A.R. AND PFLEIDERERR, P. Broadcasting Opinions with an Overcondent Sender. International Economic
Review, Vol. 45 (2004), pp. 467498.
AGHION, P., DEWATRIPONT, M., AND REY, P. Renegotiation Design with Unveriable Information. Econometrica, Vol.
62 (1994), pp. 257282.
ALONSO, R. AND MATOUSCHEK, N. Optimal Delegation. Review of Economic Studies, Vol. 75, (2008), pp. 259293.
BARCZAK, G., GRIFFIN, A., AND KAHN, K.B. Perspective: Trends and Drivers of Success in NPD Practices: Results of the
2003 PDMA Best Practices Study. Journal of Product Innovation Management, Vol. 26 (2009), pp. 323.
BENABOU, R. AND LAROQUE, G. Using Privileged Information to Manipulate Markets: Insiders, Gurus, and Credibility.
Quarterly Journal of Economics, Vol. 107 (1992), pp. 921958.
C
RAND 2014.
608 / THE RAND JOURNAL OF ECONOMICS
BRUNNERMEIER, M.K. AND PARKER, J.A. Optimal Expectations. American Economic Review, Vol. 95 (2005), pp. 1092
1118.
BRYNJOLFSSON, E., HITT, L.M., AND KIM, H.H. Strength in Numbers: How Does Data-Driven Decisionmaking Affect
Firm Performance? (2011). Available at SSRN: ssrn.com/abstract=1819486.
CARLSON, P. The Flop Heard Around the World. The Washington Post (September 4, 2007). Available at
www.washingtonpost.com/wp-dyn/content/article/2007/09/03/AR2007090301419.html.
CHE, Y.-K. AND KARTIK, N. Opinions as Incentives. Journal of Political Economy, Vol. 117 (2009), pp. 815860.
CHEN, Y. Perturbed Communication Games with Honest Sender and Naive Receivers. Journal of Economic Theory,
Vol. 146 (2011), pp. 401424.
CHURCHILL, G.A., JR. ANDIACOBUCCI, D. Marketing Research: Methodological Foundations, 9th ed. Mason, OH: Thomson
Learning, 2005.
COMPTE, O. The Winners Curse with Independent Private Values. Working Paper, CERAS-ENPC, 2002.
CRAWFORD, V.P. AND SOBEL, J. Strategic Information Transmission. Econometrica, Vol. 50 (1982), pp. 14311451.
DEDERER, A. In Defense Of: The Pontiac Aztek. The Truth About Cars (October 25, 2006). Available at
www.thetruthaboutcars.com/in-praise-of-the-pontiac-aztek/.
DESHPANDE, R. AND ZALTMAN, G. Factors Affecting the Use of Market Research Information: A Path Analysis. Journal
of Marketing Research, Vol. 19 (1982), pp. 1431.
. A Comparison of Factors Affecting Researcher and Manager Perceptions of Market Research Use. Journal of
Marketing Research, Vol. 21 (1984), pp. 3238.
FARRELL, J. AND RABIN, M. Cheap Talk. Journal of Economic Perspectives, Vol. 10 (1996), pp. 103118.
FEDDERSEN, T. AND PESENDORFER, W. Voting Behavior and Information Aggregation in Elections with Private Informa-
tion. Econometrica, Vol. 65, (1997), pp. 10291058.
FOWLER, F.J. Survey Research Methods. 2nd ed. Newbury Park, CA: Sage, 1993.
GORDON, S. On Innite Cheap Talk Equilibria. Working Paper, Universit e de Montr eal, 2010.
GROSSMAN, G.M. AND HELPMAN, E. Special Interest Politics. Cambridge, Mass.: The MIT Press, 2001.
GROVES, R.M. Survey Errors and Survey Costs. New York: Wiley (1989).
HARSANYI, J.S. Games with Incomplete Information Played by Bayesian Players, IIII, Part III. The Basic Probability
Distribution of the Game. Management Science, Vol. 14 (1968), pp. 486502.
HARTMAN, R.S., DOANE, M.J., AND WOO, C.-K. Consumer Rationality and the Status Quo. Quarterly Journal of
Economics, Vol. 106 (1991), pp. 141162.
KARTIK, N., OTTAVIANI, M., AND SQUINTANI, F. Credulity, Lies, and Costly Talk. Journal of Economic Theory, Vol. 134
(2007), pp. 93116.
KAWAMURA, K. A Model of Public Consultation: Why Is Binary Communication so Common? The Economic Journal,
Vol. 121 (2011), pp. 819842.
KELLEY, T. AND LITTMAN, J. The Ten Faces of Innovation: IDEOs Strategies for Beating the Devils Advocate & Driving
Creativity Throughout Your Organization. New York: Currency/Doubleday, 2005.
KIM, K.-M. AND POGACH, J. Honesty vs. White Lies.Working Paper, University of Pennsylvania, 2009.
KREPS, D. AND WILSON, R. Reputation and Imperfect Information. Journal of Economic Theory, Vol. 27 (1982),
pp. 253279.
KRISHNA, V. AND MORGAN, J. Asymmetric Information and Legislative Rules: Some Amendments. American Political
Science Review, Vol. 95 (2001), pp. 435452.
LEHMANN, D.R., GUPTA, S., AND STECKEL, J.H. Marketing Research. Reading, Mass.: Addison-Wesley, 1998.
LI, M. AND MADAR ASZ, K. When Mandatory Disclosure Hurts: Expert Advice and Conicting Interests. Journal of
Economic Theory, Vol. 139 (2008), pp. 4774.
MANNING, R. Market Research By a Monopolist: A Bayesian Analysis. Economica, New Series, Vol. 46, (1979),
pp. 301306.
MARICANO, M. Cheetos Lip Balm & More Bizarre Brand Extensions. Wall Street Journal (July 15, 2009). Available at
online.wsj.com/article/SB124690095610001477.html.
MEIROWITZ, A. Polling Games and Information Revelation in the Downsian Framework. Games and Economic Behavior,
Vol. 51 (2005), pp. 464489.
MILGROM, P. AND ROBERTS, J. Limit Pricing and Entry under Incomplete Information: An Equilibrium Analysis.
Econometrica, Vol. 50 (1982): 443459.
MOORMAN, C., ZALTMAN, G., AND DESHPANDE, R. Relationships between Providers and Users of Market Research: The
Dynamics of Trust within and between Organizations. Journal of Marketing Research, Vol. 29 (1992), pp. 314328.
MORGAN, J. AND STOCKEN, P.C. An Analysis of Stock Recommendations. RAND Journal of Economics, Vol. 34 (2003),
pp. 183203.
. Information Aggregation in Polls. American Economic Review, Vol. 98 (2008), pp. 864896.
MORRIS, S. Speculative Investor Behavior and Learning. Quarterly Journal of Economics, Vol. 111 (1996), pp. 1111
1133.
OTTAVIANI, M. AND SQUINTANI, F. Naive Audience and Communication Bias. International Journal of Game Theory,
Vol. 35 (2006), pp. 129150.
C
RAND 2014.
HUMMEL, MORGAN, AND STOCKEN / 609
SAWYER, A.G. AND PETER, J.P. The Signicance of Statistical Signicance Tests in Marketing Research. Journal of
Marketing Research, Vol. 20 (1983), pp. 122133.
SCHEINKMAN, J.A. AND XIONG, W. Overcondence and Speculative Bubbles. Journal of Political Economy, Vol. 111
(2003), pp. 11831219.
SOBEL, J. A Theory of Credibility. Review of Economic Studies, Vol. 52 (1985), pp. 557573.
VAN DEN STEEN, E. Rational Overoptimism (and Other Biases). American Economic Review, Vol. 94 (2004), pp. 1141
1151.
. Interpersonal Authority in a Theory of the Firm. American Economic Review, Vol. 100 (2010), pp. 466490.
VENEZIA, I. Optimal Investments in Market Research. European Journal of Operational Research, Vol. 18 (1984),
pp. 198207.
YOUNG, H. Condorcets Theory of Voting. American Political Science Review, Vol. 82 (1988), pp. 12311244.
C
RAND 2014.