Does it pay to reduce your customers wait?

An empirical
industrial organization study of the fast-food drive-thru
industry based on structural estimation methods.
Gad Allon
Kellogg School of Management, 2001 Sheridan Road Evanston , IL 60208 , [email protected]
Awi Federgruen
Columbia, Graduate School of Business, [email protected]
Margaret Pierson
Columbia, Graduate School of Business, [email protected]
August 5, 2009
In many service industries, companies compete with each other on the basis of the waiting time their customers experi-
ence, along with other strategic instruments such as the price they charge for their service. The objective of this paper is to
conduct, what we believe to be the rst, empirical study of an important industry to test whether and to what extent waiting
time performance measures impact different rms market shares and price decisions. We report on a large scale empirical
industrial organization study in which the demand equations for fast-food drive-thru restaurants in Cook County are esti-
mated based on so-called structural estimation methods. Our results conrm the belief expressed by industry experts, that
in the fast-food drive-thru industry customers trade off price and waiting time. More interestingly, our estimates indicate
that consumers attribute a very high cost to the time they spend waiting.
1. Introduction
In many service industries, companies compete with each other on the basis of the waiting time their cus-
tomers experience, along with other strategic instruments such as the price they charge for their service.
In their popular textbook Competing Against Time, Boston Consulting Group partners Stalk and Hout
(1990) documented how time base competition has been reshaping global markets. Often, specic waiting
time standards or guarantees are advertised. For example, Ameritrade has increased its market share in the
on line discount brokerage market by guaranteeing that trades take no more than 10 seconds to be exe-
cuted; the guarantee is backed up with a complete waiver of commissions in case the time limit is violated.
This has led most major on line brokerage rms (E-trade, Fidelity) to offer and aggressively advertise even
more ambitious waiting time standards. Various call centers promise that the customer will be helped within
1
Pierson et al Waiting Time in Fast Food
2
one hour, say, possibly by a callback. See Allon and Federgruen (2007) for a longer list of examples. In
other industries, average waiting times are monitored by independent organizations; examples include the
airline industry where independent government agencies (e,g., the Aviation Consumer Protection Division
of the Department of Transportation) as well as Internet travel services (e.g., Expedia) report, on a ight
by ight basis, the average delay and percentage of ights arriving within 15 minutes of schedule. Another
example is the fast food industry, the prime focus of this paper, where industry trade organizations such as
Quick Service Restaurants, publicize yearly surveys of the average waiting time experienced at the various
fast food chains. While the vast majority of fast food outlets are owned by independent franchisees, and
while these franchisees select their own prices, a common waiting time standard is chosen by each of the
chains. Moreover, chains invest heavily to shave seconds off the waiting times, clearly believing that their
market shares are very sensitive to the relative waiting times experienced.
Beginning with the seminal paper by Naor (1969), stochastic models for service systems have been ana-
lyzed in which the expected demand volume depends on either the steady state waiting time or the specic
waiting time to be experienced by a customer seeking service. Additional early examples in the economics
literature include Luski(1976), Levhari and Luski (1978) and Devany and Savings (1983). In the opera-
tions management literature, it has been conventional to represent a service systems demand process as an
exogenous stochastic process, the characteristics of which do not depend on the waiting time experienced
or any other controllable attribute such as the price charged. However, starting with Lee and Cohen(1985),
Mendelson (1985) and Mendelson and Whang(1990), (queueing) models have started to acknowledge that
potential customers, even in a monopoly setting, are able to choose whether to buy the service or not. In
the presence of multiple service providers, they face the additional choice of whom to patronize and select
by trading off relative waiting times, along with prices and other attributes. More recently a series of the-
oretical models, for example Lederer and Li (1997), Cachon and Harker (2002) and Allon and Federgruen
(2007) have been proposed to analyze a general market for an industry of competing service providers.
Firms differentiate themselves by their price levels and the waiting time their customers experience, as well
as different attributes not determined directly through competition such as the convenience of the pick-up
Pierson et al Waiting Time in Fast Food
3
process, or the accuracy with which the service process is delivered. Customers select a specic rm by
trading off these three categories of service attributes.
To date, all contributions to the literature on service competition have been theoretical, with numerical
investigations conned to small hypothetical examples. The objective of this paper is to conduct, what we
believe to be the rst, empirical study of an important industry to test whether and to what extent wait-
ing time performance measures impact different rms market shares and price decisions. Most executives
realize that time is money for the consumer, but it is unclear how much money and how the exchange rate
varies with other factors such as location, brand etcetera. We also characterize how the price equilibrium
responds to changes in the waiting standards. Additionally, we wish understand whether the trend in various
industries to continuously improve waiting times and service levels can be explained on (game-)theoretical
grounds. To answer these questions, a competition model, derived from underlying consumer choice mod-
els, has been estimated using detailed empirical data.
There are several reasons, to our knowledge, why a voluminous array of theoretical models has not
been complemented by any empirical studies. First, it is very difcult to access data regarding customer
waiting times, in particular when seeking to quantify the waiting time experience at all competing service
providers. Yet, while absolute waiting times at a given rm might explain the rms demand volume in
a monopoly setting, it is the relative waiting times at various competing providers which, along with the
rms other strategic choices, explain ultimate consumer choices and hence, realized market shares. Second,
it is, typically very hard, if not impossible, to collect data on sales volumes by the competing rms. Such
data are sometimes accessible for consumer products, such as automobiles, pharmaceuticals or consumer
goods sold in supermarkets (In the automobile industry, such data are generally, although not universally,
available at the dealer level, because cars need to be registered and registration data are publicly accessible;
similarly, every prescription has a prescription number which is entered into a nationwide data base, and
which identies the prescribing physician. As to goods sold at supermarkets, marketing rms like IRI and
Nielsen sell sales data collected from scanners). In contrast, in the service industry, it is even rarer that sales
volumes can be gathered and rms are reluctant to provide the information, considering it of the highest
strategic value. Indeed, sales volumes were unavailable in our context as well. Instead, we were able to infer
Pierson et al Waiting Time in Fast Food
4
them by estimating the parameters in the system of equations characterizing the unique equilibrium in the
above mentioned competition model. In other words, instead of computing the equilibrium from a set of
estimated demand functions and cost structures, the demand functions are backed out from the equilibrium
conditions, with the help of the observed equilibrium. This technique has been applied in a number of
economics studies, e.g., Feenstra and Levinsohn(1995) and Thomadsen (2005) but, to our knowledge, not
in the operations management literature.
To conduct our empirical tests we have selected the drive-thru fast-food industry. The fast-food industry
realized over a hundred billion $ in sales in 2007 with hamburger sales representing 73% of the market.
McDonalds alone reported in excess of $28 billion in revenues, representing 46% of the quick service
burger market. The drive-thru sector accounts for about 70% of the industrys sales, a 10% increase from 6
years ago Hughlett (Nov 28, 2008). As reported by QSR (Quick Service Restaurant) Magazine, the afore-
mentioned industrys trade publication, rms invest heavily to improve customer waiting and service times
and the accuracy with which orders are lled. As far as the former is concerned, seconds are perceived to
matter signicantly. In the most recent (2008) survey of QSR of the top 25 chains, it was found that the
best performing chain in terms of average waiting times (Wendys) was twice as fast as the least perform-
ing chain, 13% better than the second performing chain (Bojangles) and 17% better than McDonalds, the
industry leader. Moreover, Wendys improved their waiting time by over 3% from the QSR survey results
3 years earlier. Many fast food chains have installed timer systems. These let an operator know how many
cars visited the drive-thru at various times of the day, and, in addition, the average time customers spend in
the drive-thru and which point had the longest wait time. Consumer surveys by QSR focus on the maximum
waiting time customers are willing to tolerate. Recently, in a Chicago Tribune article, R. Craig Coulter,
chief scientist at HyperActive Technologies, a restaurant software rm, stated that People decide whether
to come to your restaurant based on how long the drive-thru line is Hughlett (Nov 28, 2008). In the same
article, the president of data management at Restaurant Technologies, states that theres an industry maxim
that for every seven-second reduction in drive-thru service time, sales will increase 1% over time.
We report on a large scale empirical industrial organization study in which the demand equations for
Pierson et al Waiting Time in Fast Food
5
restaurants are estimated based on so-called structural estimation methods. As mentioned, the demand equa-
tions are inferred from the equations characterizing the unique equilibrium in the competition model which
results from a detailed consumer choice model and a cost structure reecting a broad category of queueing
systems. More specically, as in most service industries, it appears impossible to obtain outlet specic sales
data because the rms keep this information as propriety data. Building on the framework of Feenstra and
Levinsohn (1995) and Berry et al. (1995), we accommodate the absence of demand data with three assump-
tions: (1) The consumers attribute a utility level to each potential outlet, which depends stochastically on
price, waiting time, the distance to the outlet and various other outlet specic characteristics; similarly,
customers assign a utility level to the no-purchase option, which depends stochastically on the consumers
gender, race, age bracket and occupational status. (2) Firms encounter a cost structure which is afne in the
sales volume, with random noise terms for the marginal costs; this cost structure applies to many queueing
models used to describe the service process such as M/M/1 systems or open Jackson networks. (2) Firms
adopt the (unique) Nash equilibrium in the price competition model which results from the above consumer
choice model and the outlets cost structure.
The rst assumption is used to derive the relationships between prices, service levels, and sales quan-
tities. Based on the second and third assumptions, these relationships are subsequently used to derive the
rms Nash equilibrium conditions to jointly estimate the parameters of the indirect utility functions of the
consumers as well as the parameters of the rms cost structure. Our estimation method is a Generalized
Method of Moments (GMM), as opposed to more standard maximum likelihood estimators for systems
of non-linear equations. This is done to avoid making specic distributional assumptions about the error
terms, and to circumvent the fact that these error terms are correlated with the explanatory variables The
data we use consist of the prices, service levels, locations and outlet distributions of the top fast-food ham-
burger chains in Cook County, Illinois. To accurately represent the impact travel distances and demographic
features may have on the consumer choices, we divide the county into more than 1300 so-called tracts, a
geographic unit used by the U.S. Census Bureau, with an average area of only 1.2 square miles, and employ
the demographic composition of the population in each tract.
Pierson et al Waiting Time in Fast Food
6
We estimate the model parameters and, in particular, focus on the impact that pricing and waiting deci-
sions have on the demand for the products at each location. We also focus on how various environmental
factors (such as the restaurant location and the chain identity) impact on the equilibrium choices. We use the
tted models to conduct counter-factual experiments that demonstrate how the geographic, brand identity
and demographic properties jointly determine the prices and service levels. Some of the principal insights
obtained are that customers attribute a large cost to waiting, and that differences in waiting time standards
play a more signicant role in explaining differences in sales volumes than price differentials. When waiting
time standards and prices are determined sequentially in a two stage competition game, we observe that the
competitive dynamics drive the rms to reduce waiting time standards to their minimum levels, as dictated
by organizational or technological considerations.
In summary, the main contribution of this paper is that, to our knowledge, it represents the rst to esti-
mate how sales volumes for a service organization depend on all price and waiting times of all competing
providers within a reasonable geographic distance, as well as other attributes (e.g, brand-specic character-
istics). In particular, we appear to be the rst to study how customers trade off prices and waiting times in
an empirical market-based study, concluding for the fast food industry that consumer attribute a value to
their wait time of at least $40/hr. We conrm that a 7 second reduction results, on average, in a 1% market
share increase. However, for a large chain like McDonalds, the increase is by more than 3% and sales go
up by 15%. Our model explains the continuing trend of all chains investing heavily to reduce their waiting
time standards.
The remainder of the paper is organized as follows: Section 2 provides a review of the relevant literature.
Section 3 develops our consumer choice and competition model. Section 4 describes the many data sources
employed and the approach we adopted to collect them. Section 5 is devoted to a description of the Gener-
alized Moment Method as applied to our model. Section 6 describes the estimation results and completes
the paper with a discussion of the above mentioned counter-factual studies.
2. Literature Review
The literature on competition in service industries dates back to the late 1970s. As mentioned in the intro-
duction, Luski (1976) and Levhari and Luski (1978) were the rst to model competition between service
Pierson et al Waiting Time in Fast Food
7
providers. The latter paper addresses a duopoly where each of the rms acts as an M/M/1 system, with given
identical service rates. In this model, customers select their service provider strictly on the basis of the full
price, dened as the direct price plus the expected steady state waiting time multiplied with the waiting
time cost rate. The question whether a price equilibrium exists in this model remained an open question,
until, for the basic model with a uniform cost rate, it was recently resolved in the afrmative by Chen and
Wan (2003). These authors show, however, that the Nash equilibrium may fail to be unique. More recent
variants of the Levhari and Luski models include Li and Lee (1994), Armony and Haviv (2001) and Wang
and Olsen (2008).
Loch (1991) considers a variant of the Luski model in which the service times of the two providers
have a general, though still identical, distribution, i.e. in which each provider is modeled as an M/G/1
system. Assuming that the total demand rate for service is given by a general function of the full price, the
author shows that a symmetric equilibrium pair of prices exists, irrespective of whether the two rms target
prices directly (Bertrand competition), or indirectly, via demand rates (Cournot competition). Lederer and
Li (1997) generalize Loch (1991) to allow for an arbitrary number of service providers and a nite number
of customer classes, each with a given waiting cost rate.
In the above papers, rms compete in terms of their price (only), with xed exogenously specied capac-
ity levels (7 in Chen and Wan relaxes this assumption; see below). Several other papers assume, alterna-
tively, that prices are xed while rms compete in terms of their capacity levels. Kalai et al. (1992) consider,
again, a duopoly with Poisson arrivals and exponential service times. A xed customer population joins,
upon arrival, a single queue from which they are served on a FIFO basis by the rst available server. (When
a customer arrives to an empty queue, he is randomly assigned to one of the two providers). In this model
asymmetric Nash equilibria of service rate pairs may arise. Gilbert and Weng (1998) and Cachon and Zhang
(2007) extended the above model to allow for routing probabilities that depend on the providers service
rates according to more general (allocation) schemes.
DeVany and Savings (1983) are the rst to address a richer type of competition in which rms compete
with several rather than a single strategic instrument. This paper addresses a variant of the Levhari and
Luski model with an arbitrary number of identical rms who simultaneously choose a price and service
Pierson et al Waiting Time in Fast Food
8
rate. All customers share the same waiting cost rate , but the total demand volume in the industry is given by
a general function of the lowest full price. The authors establish the existence of a symmetric equilibrium.
Cachon and Harker (2002) and So (2000) analyzed the rst models in which customers consider addi-
tional criteria beyond the lowest full price when choosing a service provider. Both conned themselves,
again, to M/M/1 service providers. Cachon and Harker (2002) considered the case of two rms where
demand rates are given as either linear or MultiNomial Logit functions of the two full prices. So (2000)
considers an arbitrary number of competing rms and a different class of so-called attraction models. In an
attraction model, each rms market share is proportional to an attraction value, specied as a function of
several of the rms attributes. The competition model derived and estimated in this paper is a generalization
of the attraction model, in which the parameters in the attraction value functions are treated as random. In
So (2000), the logarithm of the rms attraction value is specied as a common linear combination of the
logarithm of the price and the logarithm of the waiting time standard, plus a rm dependent constant. This
specication continues to imply that the price and waiting time are aggregated into a single, albeit, different
full price measure. Allon and Federgruen (2006,2007) appear to be the rst models to treat the price and
waiting time standard as completely independent rm attributes which different customers may trade off
in different ways. Nevertheless, Allon and Federgruen (2007)s paper conned itself to systems of demand
rates that are linear in the prices and to M/M/1 service providers while Allon and Federgruen (2006) studies
more general demand models such as attraction models and allow for more general queueing facilities. We
refer to these two papers as well as Hassin and Haviv (2003) for a recent survey text on queueing models
with competition.
Many service processes are provided via call centers. Here, customers are known to be very sensitive
to their waiting times, which is why such centers are designed and staffed to meet specic service level
agreements (SLAs), see Hasija et al. (2007) for a recent survey of such agreements. However, virtually
all planning models in this vast literature assume that demand processes are exogenous inputs, or, at best,
dependent on service charges. We refer to Gans et al. (2003) for an excellent tutorial. When describing future
challenges in this area, the authors single out determining a better understanding of customer behavior
(7.3) and the need to model and estimate multiple levels of equilibria. Beyond the levels mentioned
Pierson et al Waiting Time in Fast Food
9
there, we suggest the desirability of models incorporating the competition effects of service levels provided
by the call centers of competing service providers. Another stream of papers, in particular Hall and Porteus
(2000) and Gans (2002), models the competition between services providers selecting a distribution for
the (non-congestion related) quality of service, based on specic consumer choice models. Gans et al.
(2007) describes an empirical study to test these models, based on laboratory experiments, as opposed to
econometric eld studies, as in this paper.
The above reviewed literature is based on the observation that rms compete along the service level
dimension as well as anecdotal evidence that customers value waiting time when making decisions regard-
ing their preferred service provider. Our study complements this literature by estimating the parameters
used and assumed by these models. The approach we use to estimate the impact of waiting times, prices,
geographic dispersion, chain attributes, and socio-economic factors on demand follows the work by Bres-
nahan (1987), Berry (1994), Berry et al. (1995). These authors demonstrate how to estimate consumer
choice models and cost structures in oligopolistic markets with differentiated goods using aggregate con-
sumer level data and structural models of competition. (Berry et al. (1995) applied this approach to study
price competition in the US automobile industry.) The estimation method, based on the (GMM) allows for
prices to be determined endogenously (as the equilibrium of an underlying competition model), rather than
being selected exogenously. It has the additional advantage of avoiding distributional assumptions for the
error terms in the equations to be estimated. The general approach, posits a distribution of consumer prefer-
ences for the competing goods, based on their attributes. The preferences are aggregated into a market level
demand system that, when combined with assumptions on cost and price-setting behavior, allows one to
estimate the parameters. In the above papers, market shares are observed. Feenstra and Levinsohn (1995),
were the rst to demonstrate how this estimation framework can be used in the absence of quantity data. As
mentioned in the Introduction, we face the same challenge since in the fast food industry, sales data are not
reported and treated as strategic and proprietary information. See also Dube et al. (2008).
More recent work by Davis (2006) and Thomadsen (2005a) incorporated geography in the BLP frame-
work. More specically, Thomadsen (2005a) studies the impact of ownership structure on prices in the
Pierson et al Waiting Time in Fast Food
10
fast food industry. Thomadsen (2005a) uses this method to establish that the impact of mergers in such an
industry can be large, but the impact of mergers decreases as the merging outlets are further apart.
Our study is related to the recent empirical literature in operations management. To our knowledge,
most of this literature focuses on consumer products rather than services. See Chen et al. (2005), Chen
et al. (2007), Gaur et al. (2005) and Olivares and Cachon (2007) and references therein for surveys of this
literature.
3. The Model
In this section we develop the competition model representing the competitive interdependencies and inter-
actions between the outlets in our geographic region (i.e. Cook county). The model combines two sub-
models
(a) a consumer choice model which determines how many of the residents of the region choose, for any
given lunch or dinner meal, to go to a fast food establishment, and among those, how many select a specic
outlet, and
(b) a model to represent the variable cost cost structure of the different outlets as a function of its sales
volume and service level, i.e. its waiting time standard.
Combining the two sub-models permits us to derive the outlets prot functions. We then characterize the
equilibrium behavior in the price competition model under given waiting time standards, as specied by six
major chains operating in the selected geographical region. We showthat the model has a unique equilibrium
which is the unique solution of a non-linear systemof equations. It is this systemof equations which permits
us to estimate the parameters that describe the consumer choice model and associated demand functions,
as well as the parameters in the cost structure. We conclude the section with a description of a plausible
model for the joint determination of all outlets prices and all chains waiting time standards. The model
embeds the aforementioned price competition model as the second game (with the outlets as independent
players) of a two-stage game, where the rst stage game has the chains as competing players selecting the
waiting time standards. As mentioned in the introduction, in the fast food industry, waiting time standards
are selected and prescribed by the chains: however, price decisions are relegated to the independent outlets
Pierson et al Waiting Time in Fast Food
11
to avoid illegal forms of price xing, if for no other reason. As franchising became popular in the sixties,
the US courts began to limit the types of pricing restrictions chains can impose on their franchises. Even
the specication of maximum retail prices has become illegal, by the Supreme Court ruling in Albrecht vs.
Herald (1968). (Indeed, we have observed signicant price differences among outlets of the same chain,
see Table 3 in Section 4.) In general, one may envision other ways in which the two sets of strategic
choices (price and waiting time standard) are selected: for example, settings where all choices are made
simultaneously and those where prices are selected in advance of waiting time standards. See Allon and
Federgruen (2007) for a systematic comparison of the three types of competition models, in the context of
demand functions that are linear in the prices. Indeed, these authors show that the equilibria arising in these
three models are, in general, different and, under mild conditions, more aggressive service levels emerge
in equilibrium when set in advance of prices, as opposed to being chosen simultaneously with the prices.
However, in the fast-food, only the waiting time rst-price second competition model is applicable.
3.1. The Consumer Choice Model
Demand for fast-food meals at each outlet is specied by a discrete choice model. Consumers choose either
to purchase a specic lunch or dinner meal from one of the fast-food outlets or to consume an outside good.
Consumers assign a utility value to each outlet as well as to the no-purchase option, specied as a linear
function of the price, waiting time, distance, chain identity, observable (to the modeler) attributes of the
outlet, and various demographic factors including the consumers gender, race, age bracket and occupational
status. Each of these utility equations contains an additional random noise term. It is natural to assume that
customers make their choices in two stages: (i) they rst decide whether to dine at a fast food outlet as
opposed to alternatives, such as eating at home or a different type of restaurant, and (ii) assuming the rst
question is answered in the afrmative, which of the various outlets in the region to patronize. We model
the two stage choice process by assuming that the (potential) customer attributes a utility value to the no-
purchase option which depends on his or her demographic attributes. The customer also assigns a utility
value to each of the outlets in the region that depends on the attributes of the outlet and the chain it belongs
to. The customer purchases a meal at one of the fast food outlets if and only if the highest of the outlets
Pierson et al Waiting Time in Fast Food
12
utility values is in excess of that of the no-purchase option; in this case the meal is consumed at the outlet
with the highest utility value.
Formally, the conditional indirect utility of consumer i from fast food outlet j is specied as follows:
U
i,j
=X

j
D
ij
P
j
W
j
+
ij
(1)
where X
j
is a vector of dummy variables indicating the chain identity of the outlet (as well as, possibly, other
observed properties), D
ij
is the distance between consumer i and outlet j, P
j
is the price of a (standard)
meal at outlet j, and W
j
is the waiting time standard associated with outlet j; , , and are parameters
to be estimated, and
ij
is the unobserved (to the modeler) portion of the utility of individual i at outlet j
arising from factors such as the cleanliness of the outlet or the safety of the location.
The indirect utility associated with the no-purchase option is given by
U
i,0
=
0
+M
i
+
i,0
. (2)
Here, M
i
is a vector specifying the consumers age, gender, race, and whether they are at work or at home
when making the decision (occupational status);
0
and represent another set of parameters to be estimated
and
i0
denotes the unknown portion of the utility of individual i for the non-purchase option.
We distinguish among a limited number of age brackets. Therefore, there is a nite list of {1, . . . , M}
of consumer-types , combining age, gender, race and occupational status. In view of the importance of the
distances between the consumer and the various outlets, we partition our geographic region into a grid of
very small sub-areas b {1, . . . , B} =B and assume all consumers residing in a sub-area are located at the
sub areas centroid. (In our study, we use tracts, as dened by the U.S. Census, with an average of 1.2 square
miles in Cook county.) Thus, all potential consumers residing in a given sub-area b B and belonging to a
given socio-economic group m M, experience the same mean utility value for all outlets as well as the
no-purchase option.
Assuming the distributions of the random noise terms, {
ij
: j =1, . . . , J}, are Gumbel (or doubly expo-
nential), with common scale parameter , this gives rise to the following multinomial logit model in which
each outlets market share for each tract and demographic group is given by the following expression:
Pierson et al Waiting Time in Fast Food
13
S
j,b,m
(P, W, X|, , , , ) =
e
(X

j
D
bj
P
j
W
j
)/
e
(M+
0
)/
+

J
t=1
e
(X

t
D
bt
P
t
W
t
)/
, j =1, . . . , J; b =1, . . . , B; m=1, . . . , M.
(3)
Without loss of generality and to simplify the notation we may choose the units in which the utility
levels are measured such that the scale parameter =1. Also, as is evident from (3), the market shares are
invariant to a common additive shift in the utility measures. Therefore, again without loss of generality, we
may specify the choice of the parameters such that
max
_
X

: X

=X

j
for some j =1, . . . , J
_
=1 (4)
(For example, in our study, the X -vector is a unit vector which identies the chain to which the outlet
belongs: thus (4) simplies to
max
max
j=1,...,J

j
=1, a constraint which is easily incorporated into the
optimization problem that is solved in our (GMM)-estimation method, see 5.)
Multiplying the market shares with the number of consumers in each geographic region b and demo-
graphic group m allows us to specify expected aggregate sales in an outlet as a function of the various
parameters {, , , , } in the utility equations:
Q
j
(P, W, X|, , , ) =

m
h(b, m)S
j,b,m
(P, W, X|, , , , ) (5)
Here h(b, m) denotes the number of residents (or commuters) in sub-area b who belong to demographic
group m.
3.2. The outlets cost structure
We assume the outlets cost structure expressed as a function of its expected sales volume is afne with an
intercept that is proportional with the reciprocal of the waiting time standard:
C
j
(Q
j
) =(c
k(j)
+
j
)Q
j
+[d
k(j)
+u
j
]/W
j
, j =1, . . . , J (6)
Here
k(j) = the index of the chain to which outlet j belongs, j =1, . . . , J
c
k
= the average variable food, labor and equipment cost rate per customer for an outlet of chain k,k =1, . . . , K
Pierson et al Waiting Time in Fast Food
14
d
k
= the average variable capacity cost rate for an outlet of chain k,k =1, . . . , K

j
= a noise term, denoting the difference between outlet j

s variable cost rate and the norm for his chain, j =1, . . . , J
u
j
= a noise term, denoting the difference between outlet js variable capacity cost rate and the norm for his chain,
j =1, . . . , J
Each outlets marginal cost rate as well as the capacity cost rate are equal to a chain-specic cost plus a zero-
mean unobserved component. This specication is supported by the franchisers effort to create a uniform
customer experience across their outlets, by standardizing the equipment, preparation process, and food
components used at each of its outlets. The unobserved shock to the cost rate comes from outlet specic
conditions. For example, a smaller kitchen could create crowding and reduce efciency.
The afne cost structure in (6) arises in several queueing models which may describe the service pro-
cess of an outlet. For example, the structure in (6) arises in an M/M/1 system, where the waiting time
standard W denotes the expected total sojourn time in the drive-thru queue and the variable capacity cost
is assumed to be proportional with the service rate. More realistically, a fast food service process could
be represented as a Jackson (queueing) network. A food order may travel along a path of service stages,
from order taking to the cooking of the hamburgers, assembly of the cooked burgers with the side dish
and required drink and back to the drive-thru counter. Allon and Federgruen (2006) have shown that the
cost structure in (6) applies to a general Jackson network, assuming the variable capacity costs are pro-
portional with the service rates installed at the various nodes of the network. Thirdly, the service process
may be best described as a GI/GI/s system, with an arbitrary renewal arrival process, arbitrary service
time distribution and a team of s parallel servers. If the consumer is particularly focused on the delay expe-
rienced in the drive-thru queue and if W denotes a given fractile of the delay distribution, then the cost
structure in (6) arises as a close approximation, see Allon and Federgruen (2006). This approximation is
based on so-called exponential approximations for the tail probability of steady-state delays, also referred
to as Cramer-Lundberg approximations. The exponential approximation states the existence of constants
a, >0 such that Pr[D>x] ae
x
, i.e. lim
x
e
x
Pr[D>x] =0. This identity is, in fact, exact, rather
Pierson et al Waiting Time in Fast Food
15
then an asymptotically correct approximation when the service time distribution is exponential, i.e in the
case of a GI/M/s system.
We refer to Allon and Federgruen (2006) for additional queueing models resulting in afne
cost structures of type (6). Allon and Federgruen (2006) also show that an even larger set
of queueing models give rise to cost functions of the type C
j
(Q
j
) = (c
k(j)
+
j
)Q
j
+
_
(d
(1)
k(j)
+u
(1)
j
)Q
2
j
+(d
(2)
k(j)
+u
(2)
j
)Q
j
/W
j
+(d
(3)
k(j)
+u
(3)
j
)/W
2
j
for parameters {d
(1)
k
, d
(2)
k
, d
(3)
k
: k
1, . . . , K} and mean zero noise terms {u
(1)
j
, u
(2)
j
, u
(3)
j
}. Our estimation method can be adapted to this more
general cost structure.
3.3. The Price Competition Model
We are now ready to analyze the price competition model which arises when all waiting time standards
have been specied. We assume, that every outlet is independently owned. However, our methodology is
readily adapted if various outlets are jointly managed by the same franchisee, an issue centrally considered
in Thomadsen (2005a). In view of (6)

j
(P, W, X, ) =(P
j
c
k(j)

j
)Q
j
(P, W, X|) [d
k(j)
+u
j
]/W
j
, j =1, . . . , J (7)
denotes rms js prot level as a function of all prices charged by the various outlets. The following theo-
rem shows that the price competition model has a unique equilibrium, which is the unique solution of the
system of equations (8), below. The theorem also shows that the equilibrium can be computed effectively
by a tat onnemont scheme. This scheme iteratively identies each rms best response to the prices of the
competing rms, see Topkis (1998) for details. The availability of a simple computational scheme, under
estimated parameter values, is important when conducting the counter factual study in Section 6.2. We refer
to Allon et al. (2009) for a proof of the following theorem.
Theorem 3.1(a) The price competition model has a unique equilibrium in the interior of the price cube

J
j=1
_
c
k(j)
+
j
, c
k(j)
+
j
+1/

. The equilibrium is the unique solution of the system of equations:

Q
j
(P, W, X|) +(P
j
c
k(j)

j
)
Q
j
(P, W, X|)
P
j
=0, , j =1, . . . , J (8)
Pierson et al Waiting Time in Fast Food
16
(b) Fix an outlet j = 1, . . . , J. Let P
j
denote the vector of prices charged by the other outlets and let
P

j
(P
j
) denote outlet js best response to the prices charged by the competing outlets. The best response
function P

j
() is increasing.
(c) For a given parameter vector , the unique equilibrium can be found by a simple tat onnemont scheme,
starting with the lower-bound price vector p
min
[c
k(j)
+
j
: j =1, . . . , J].
The following is a simple expression for the partial derivatives of the rms sales quantities with respect to
price:
Q
j
(P, W, X|)
P
j
=
B

b=1
M

m=1
h(b, m)
_
1
Sj, b, m(P, W, X|)
h(b, m)
_
S
j,b,m
(P, W, X|) (9)
All three parts of Theorem 3.1 are noteworthy, as the results apply to general price competition models
with so-called multinomial logit demand function with random coefcients, and any afne cost structure
1
.
Thomadsen (2005b), recently, identied some sufcient conditions for the existence of an equilibrium but
concludes his paper, stating While the theoremprovides a limited set of conditions under which there exists
a pure strategy equilibrium . . . , and while existence is shown for any distribution of consumers and rms,
what is lacking are proofs of uniqueness.
In matrix notation, the equilibrium conditions (8) can be stated as:
Q(P, X, W) +(P c ) =0 (10)
where is a diagonal J J matrix whose the j-th diagonal element
j,j
=
Q
j
P
j
. Essential to our estimation
method is that the vector of cost rate residuals therefore satises:
=P C +(P, X, W|)
1
Q(P, X, W|), (11)
with

=(

), and identity which follows from (10).

1
For example, in their classical paper Berry et al. (1995) note: We assume that a Nash equilibrium to this pricing game exists, and
that the equilibrium prices are in the interior of the rms strategy sets (the positive orthant). While Caplin and Nalebuff (1991)
provide a set of conditions for the existence of an equilibrium for related models of single product rms, their theorems do not
easily generalize to the multi product case. However, we are able to check numerically whether our nal estimates are consistent
with the existence of an equilibrium. Note that none of the properties of the estimates require uniqueness of equilibrium, although
without uniqueness it is not clear how to use our estimates to examine the effects of policy and environmental changes.
Pierson et al Waiting Time in Fast Food
17
Finally, when multiple outlets are owned by the same franchisee, and the total number of independent
franchisees is F < J, the rst order conditions (8) can also be written in the form (11), however with a
non-diagonal matrix . See Thomadsen (2005b) for details.
3.4. The Two-Stage Competition Model
In Section 6.2, we investigate how the competing rms in the industry select their service levels, i.e. their
waiting time standards. As explained at the beginning of this section, the waiting time standards are deter-
mined at the chain level in advance of the outlets price choices. We therefore assume that the industrys
waiting time standards represent the equilibrium in a two-stage game, with the above price competition
model as the second stage, and a rst stage game in which the chains act as the players.
In the fast food industry, the focus of our study, it has been widely reported (see, for example, Thomadsen
(2005a) and Lafontaine and Shaw (1999)) that the franchisee pays a xed, periodic, franchise fee along
with a percentage of the revenues. However, it is unclear how these two parameters are determined as a
function of the desired waiting time standard. We therefore assume that each chain attempts to maximize
aggregate prots in its supply chain, i.e. its own prots as well as those of its franchisees. (The aggregate
prot measure is independent of the transfer payment scheme.)
It remains an open question whether a pure strategy Nash equilibrium exists in the two-stage game and
whether it is unique. However, when searching for an equilibrium we employ the tat onnemont scheme,
described in subsection 3.3. To evaluate the prot consequences of any vector of waiting times standard, we
solve the associated price competition model, which has a unique equilibrium, see Theorem 3.1(a).
4. Data
We have studied the fast-food industry in Cook County, Illinois. We have chosen this industry both because
of the availability of data and because this is an industry that has historically placed a premiumon competing
via its service levels. The QSR Magazine 2007 Drive-Thru Time Study notes that in 2007 all quick-service
chains made major efforts to improve speed-of-service in their drive-thrus, see Nuckolls (2007). Examples
of new technology adopted to improve speed-of-service include timer systems that allow in-store managers
as well as regional and national ofces to monitor waiting times at outlets, as well as the outsourcing of
Pierson et al Waiting Time in Fast Food
18
drive-thru order taking. There is a plethora of anecdotal evidence that the industry is reacting to consumer
expectations. The same 2007 QSR Magazine Drive-Thru study reported that 70% of surveyed customers
said speed is an important factor in the drive-thru experience. The 2008 study reports that this trend is
continuing, with the fastest chain, Wendys, shaving off an additional 7 seconds from the average waiting
time in the previous year.
We use as our data set, all fast-food outlets belonging to chains selling hamburgers in Cook County Illi-
nois. We consider only outlets with drive-thru windows because outlets without drive-thru windows tend to
be located in places such as malls and airports where consumers are facing a different set of considerations.
We consider only chains with a presence of more than 5 outlets in the county. This results in a total 388
outlets belonging to McDonalds (173), Burger King (92), Wendys (62), White Castle (42), Dairy Queen
(10), and Steak n Shake (9) chains.
For each outlet we gathered prices for the franchises signature burger, a small fries order, and a small
soft-drink by calling the location in 2006. The type of burger selected was standardized by weight and in
the case of White Castle, which sells small burgers, we use the price for four sliders. These prices have
been aggregated to compute the price of a complete standard meal at that outlet. It is worth noting that we
tried to call the chains themselves for price information or for any pricing guidelines that they may give to
their franchisees. We were told that even the practice of suggesting prices to the outlets is illegal (see the
discussion at the beginning of 3). Thus, waiting times are selected centrally by the chains but prices are
chosen by the individual outlets. Indeed, we have noticed very signicant price differences among outlets
belonging to the same chain, with the most expensive McDonalds or Burger King outlet being about 50%
more expensive than the cheapest outlet in the county, see Table 3. This supports the two stage competition
model we have postulated in the previous section.
As mentioned, all chains select and strive for a common waiting time standard among all of their outlets.
In addition, customers often frequent more than a single outlet of a chain and expect to experience a similar
service level, irrespective of the specic chain outlet they go to. We have selected the average steady-state
waiting time, dened as the time spent in the drive-thru queue plus the service time, as the waiting time stan-
dard used in the consumer choice model of subsection 3.1. To arrive at the average waiting time standards for
Pierson et al Waiting Time in Fast Food
19
the different chains, we have employed the QSR Magazines 2005 Drive-Thru Time Study Database, which
we purchased from QSR. The database contains, for a national sample of outlets, two random observations
at lunch and at dinner time. We obtained each chains average waiting time by averaging the recorded obser-
vations over all outlets that belong to the relevant chains, nationwide. These national average waiting times
vary signicantly across chains, with the worst performer being close to twice as slow as the best performer,
see Table 1 below. The chain-wide waiting time standards of the six chains in our study have a mean of
225.92 seconds, a standard deviation of 38.21, and a range of [173.34,269.45]. We have performed a two-
Table 1 Average Waiting Time as Determined from 2005 QSR Drive-Thru Study
Rank Chain Mean Wait (sec)
2 WENDYS 173.34
4 BURGER KING 192.29
14 MCDONALDS 224.27
15 DAIRY QUEEN 231.85
21 STEAKN SHAKE 264.3
24 WHITE CASTLE 269.45
sample t-Test assuming unequal variances on all the national waiting time observations for the six chains
with a major presence in Cook County. We have veried that waiting time observations for different chains
were indeed drawn from different distributions, (with the exception of the McDonalds and Dairy Queen,
and the White Castle and Steak n Shake pairs; note from Table 1, that the mean waiting times are virtually
identical within each pair as well). This conrms that different chains offer systematically different waiting
time experiences to the consumer. The results of this analysis can be seen in Table 2. The critical values for
each test, with an alpha of 0.05, consistently rounded to 1.96. The t Statistic is reported in the right-hand
section of the table.
Table 2 Two-Sample t-Test on National Chain-Wide Wait Time Observations
Chain Num. Mean Wait Std. Dev. Mc- Burger Wendys White Dairy Steak
Obs. (sec) Donalds King Castle Queen n Shake
McDonalds 598 224.27 151.38 4.09 6.66 -3.98 -0.70 -3.86
Burger King 600 192.28 116.69 -4.09 2.89 -7.25 -5.15 -7.70
Wendys 596 173.34 109.81 -6.66 -2.89 -9.13 -7.92 -9.93
White Castle 334 269.45 173.83 3.98 7.25 9.13 3.57 0.55
Dairy Queen 528 230.10 128.18 0.70 5.15 7.92 -3.57 -3.40
Steak n Shake 328 262.69 141.22 3.86 7.70 9.93 -0.55 3.40
Pierson et al Waiting Time in Fast Food
20
Demographic and geographic information was gathered with a very ne granularity, the so-called tract
level. Tracts are geographic areas dened by the U.S. Census Bureau to contain 2,500 to 8,000 people.
Cook County consists of 1343 tracts with an average area of only 1.2 square miles. In urban areas a tract
corresponds with a few city blocks. The next smallest geographic area recognized by the U.S. Census, the
so-called block groups, are so small that some demographic data, such as race, cannot be reported without
revealing the exact household being discussed and hence are not available to the public. We have considered
the following age brackets: 0-9, 10-19, 20-39, 40-59, and 60+. We considered black and white consumers
only because these are the racial groups for which we had the necessary data for employing the macro
moments discussed in Section 6. As mentioned in Section 3, consumers are also differentiated based on
whether they are at work or home. As far as the residents in a tract are concerned, we collected the number
of people of each age bracket, race, and gender combination from the 2000 U.S. Census data. As to the
population working in each of the tracts, The Bureau of Transportation Statistics reports the number of
people commuting between every tract pair broken down by age group, race, and gender. We aggregated the
ow of workers into each tract in Cook County from any originating tract (whether or not the originating
tract was within Cook County). The Bureau of Transportation Statistics data are not broken down by age,
gender and race combination at the tract level in Cook County, so, we estimated the population numbers of
each combination by assuming the three social attributes are independent. This means that if a person lives
and works in Cook County they are counted as two consumers. We do this because such consumers have
the potential to consume one meal, i.e. lunch, while at work, and another meal, i.e. dinner, while at home.
Distinguishing between commuters and residents, two genders and two racial groups, as well as among 5
age brackets, we have thus divided the population into 40 different socio-economic groups. The distance
from the consumer to each outlet is calculated as the distance from the tract centroid in which the consumer
is located. To compute the distances between the various tract centroids and the restaurant locations, we
have employed the ArcView Geographic Information System modeling and mapping software.
In addition to the independent variables, we collected data for the so-called instruments used in the esti-
mation method. These are outlet specic variables that are correlated with one or more of the independent
variables but not with the noise terms {
j
: j =1, ..., J} in the cost rates. Following the recommendation in
Pierson et al Waiting Time in Fast Food
21
Thomadsen (2005a), we have selected the following instrumental variables: V
1j
= the distance from outlet
j to the nearest outlet, V
2j
= the number of outlets within two miles of outlet j, V
3j
= the population density
in the tract to which outlet j belongs, and V
4j
= the worker density in this tract. Table 3 shows summary
statistics for these as well as the price variables.
Table 3 Summary Statistics for Collected Data
Variable Name Mean Standard Deviation Min Max
McDonalds Price ($) 4.96 0.25 6.09 4.20
Burger King Price ($) 4.85 0.28 5.39 3.63
Wendys Price ($) 4.75 0.20 5.24 4.27
White Castle Price ($) 4.46 0.09 4.78 4.23
Dairy Queen Price ($) 5.66 0.26 6.07 5.07
Steak n Shake Price ($) 4.99 0.36 5.84 4.67
Distance to Nearest Outlet (mi) 0.55 0.48 0.00 2.52
No. Outlets within 2 mi 5.93 2.54 1 14
Population Density (100K/sq mi) 0.09 0.09 1.71E-04 0.80
Worker Density (100K/sq mi) 0.04 0.05 1.33E-03 0.36
5. Estimation
As mentioned in the introduction, the major hurdle to our estimation of the parameters of the demand func-
tions, and the rms cost structure, is the lack of available demand data. As explained, this challenge is not
unique to the fast food industry, but presents itself in almost all service industries. The unavailability of
sales data prevents the use of standard regression-type estimation techniques. Instead, we employ a tech-
nique that estimates the parameters on the basis of the equilibrium conditions, specied in subsection 3.3,
which characterize the unique Nash equilibrium in the second-stage price competition among the outlets,
see Theorem 3.1. In the context of industrial organization studies, this so-called Generalized Method of
Moments (GMM) technique for circumventing a lack of demand data, was rst introduced by Berry et al.
(1995). See Nevo (2000) for a clear exposition.
The equilibrium conditions (11) represent a system of equations which involve only the observed price
vector P, waiting time standards W, outlet attribute matrix X, and distances {D
j,b
j = 1, . . . , J, b =
1, . . . , B}, as well as the unknown parameter string. (In particular, the system of equations does not involve
the unobservable sales volumes.) Because of the endogeneity of the price vector P, these variables are cor-
related with the cost rate noise terms {
j
: j = 1, . . . , J}. This prevents the usage of standard maximum
Pierson et al Waiting Time in Fast Food
22
likelihood estimation techniques. In addition, the latter requires us to postulate specic distributions for the
noise terms .
The (GMM) technique overcomes both difculties. It employs a vector of so-called instrument vari-
ables Z
j
{Z
1j
, . . . , Z
rj
} which are correlated with (some of) the explanatory variables {P, X, W, D}, but
uncorrelated with the cost rate noise terms , i.e.
E[Z
lj

j
] =0 for all cost rates l =1, . . . , r and all outlets j =1, . . . , J (12)
Our instruments are based on the four instrumental variables V
1
j,V
2
j,V
3
j,V
4
j dened in Section 4. In
order to account for asymmetries in the way that different chains are affected by these instrumental variables,
we interact these variables with the chain indicator vectors, i.e. the columns of the matrix X, to arrive at a
total of 24 instruments: for all j =1, . . . , J, Z
j
{Z
l,k,j
=V
lj
X
kj
, l =1, . . . , 4, k =1, . . . , K}. Intuitively,
these instruments affect demand by altering the strength of competition and the size of the potential market.
Moreover, they appear to be uncorrelated with any cost rate differentiated outlets j is experiencing vis-a-vis
the chain norm. In other words, it is reasonable to assume that the moment conditions (12) apply. In view
of the population moment conditions (12) we must have that for the proper parameter vector , the sample
average of the vector of random variables {Z
t
j

j
, j =1, . . . , J}
G() =
1
J
J

j=1
Z
t
j

j
() (13)
is as close to zero as possible. The (GMM) estimator thus determines a parameter vector

which minimizes
a quadratic function of this sample average. More specically the (GMM) estimate is the vector

which
optimizes
min

G()

AG() (14)
where A is a weighting matrix for the moments.
The optimal weighting matrix for the GMM estimator has been shown to be the inverse of the asymptotic
variance-covariance matrix of the moment conditions. However, as this matrix is not available, a-priori, we
follow the commonly used two-step estimation procedure where the rst step is used to estimate the asymp-
totic variance-covariance matrix. In the rst step, we use the GMM with the pre-specied weighting matrix
Pierson et al Waiting Time in Fast Food
23
A
1
= I to get a consistent initial estimator

1
. We use

1
to estimate the asymptotic variance-covariance
matrix of the moment conditions, A
2
= (E[G(

1
)G(

1
)

])
1
. We then run the GMM procedure a second
time with this new weighting matrix to arrive at our parameter estimate
2
. For a nice discussion of the
two step method and a proof that the inverse of the asymptotic variance -covariance matrix is the optimal
weighting matrix, see Hall (2005).
There are well documented technical difculties associated with the optimization problem (14). Its objec-
tive function has many local optima quite far from the globally optimal solution. In addition, there are large
regions where this function is close to at, creating formidable difculties for standard gradient methods.
To mitigate these difculties, we restrict the feasible region for the parameter vector by imposing several
reasonable constraints. Let
con()
b,m
=
J

j=1
S
j,b,m
(P, W, X|)/h(b, m), b =1, . . . , B; m=1, . . . , M
= the fraction of the population in tract b and socio-economic group m which purchases a fast food meal;
con = an upper bound for the fraction of the population in any geographical area and any socio-economic
group to purchase a fast food meal;
con = a lower bound for the fraction of the population in any geographical area and any socio-economic
group to purchase a fast food meal;
c
k
() = best estimate of chain ks standard cost rate c
k
=
1
J
k

jJ
k
_
P
j
+(P, X, W|)
1
Q(P, X, W|
j
)

, k =1. . . , K, where
J
k
= {j : k(j) =k} denotes the set of outlets belonging to chain k
c
k,min
= 0
c
k,max
= min
jJ
k
P
j
: k =1, . . . , K.
We impose the constraints:
con con()
b,m
con, for all b =1. . . , B and m=1. . . , M (15)
c
k,min
c
k
() c
k,max
, forall k =1. . . , K (16)
Pierson et al Waiting Time in Fast Food
24
Thus, instead of of solving the unconstrained optimization problem (13), we propose solving the con-
strained optimization problem:
(P) min

{(12) s.t. (15) and (16)}.

To solve the constrained optimization problem, we propose replacing the soft constraints (15) and (16) by
penalty functions added to the objective function. We thus obtain:
G

()AG() +

m
{log[concon()
b,m
] +log[con()
b,m
con] (17)
+

k
log[c
k,max
c
k
()] +log[ c
k
() c
k,min
]}
with as the weighting factor, a multiplier applied to the penalty functions.
The fact that the chain indicator variables and the waiting time standard are common to all outlets in a
chain creates a co linearity problem. We therefore normalize one of the chain indicator -parameters (that
of Wendys) to zero. (Recall that all market share formula in Section 3.1 are invariant to a common additive
shift in the utility measures.)
We have developed a special algorithm to solve (P) via the modied objective (17). The algorithm begins
with a large value for , the weight of the penalty functions, relative to the objective function value at the
starting point and involves a quasi-Newton search method. During this search, we restrict movement in the
direction of the barriers imposed by the penalty functions so that any point within the interior of the feasible
region can be reached, but points along the barrier are not approached very quickly, thus preventing the
algorithm from trapping itself in unfavorable points. When a stopping condition is reached, the penalty
weight is halved and the quasi-Newton search re-run. In the rst iteration, when the penalty is large,
this results in the algorithm moving to a point which is quite far from the barriers. The algorithm iterates
until the penalty weight is small enough to render the penalty terms is insignicant compared to the regular
objective function (14).
To arrive at the reported estimates, we used a process in which, in the rst stage, we took 19 starting
points and ran both the above algorithm as well as the general KNITRO algorithm with default options -
resulting in 38 estimates. We then took the (almost) 20 best local optima found, for the objective function
Pierson et al Waiting Time in Fast Food
25
(13), generated weighting matrices for each of these points and, in the second stage, ran our algorithm as
well as the KNITRO algorithm from each of these 20 points, generating 40 nal estimates.
In order to validate the statistical signicance of these estimates, we constructed condence intervals
using a bootstrapping procedure. This procedure is used when attempting to measure the variance of an
estimator when there is no sample data available beyond those used to obtain the estimate. The idea is to
use subsets of the sample and calculate the value of the estimator on each subset in order to estimate the
variance. To that end, we selected 20 random subsets of the tracts, and ran the algorithm on each subset
for each of the 20 (second stage) starting points of the two algorithms, resulting in a total of 800 parameter
values. Each subset has 134 tracts (10% of total number of tracts). We estimated the coefcients for each
sub-model and used the empirical distribution to construct the condence intervals for each parameter.
6. Results
In this section, we report the results of the estimation process. We focus on the key parameters of interest,
emphasizing those that turned out to be statically signicant. Table 4 reports the estimated value of each
of the main demand coefcients: price, waiting time and distance sensitivity (, , ). We also report the
estimated marginal cost for each of the top three chains: McDonalds, Burger Kind and Wendys. The table
reports, in addition, the 90% percent condence intervals for each of the reported estimates.
Table 4 Estimates of Price, Waiting Time and Distance Coefcients
Variable Name Coefcient Est. 90% Condence Interval
Price Sensitivity ($) 0.4364 [0.3349, 0.6088]
Waiting Time Sensitivity (sec) 0.0259 [0.0070, 0.0541]
Distance Sensitivity (mi) 0.00350 [0.0001, 0.0143]
Marginal Cost McD 2.563 [1.972, 3.313 ]
Marginal Cost BK 2.460 [1.868, 3.209]
Marginal Cost WN 2.354 [1.746, 3.105]
First note that both the price sensitivity and waiting time sensitivity parameters and , see (1), are
signicantly positive, at the 90% condence level. (The two are, in fact, signicant at the 99% condence
level). We thus conclude that both price and waiting time parameters have a signicant impact on the
consumers decision at which outlet to purchase a fast-food meal, if at all. This result conrms our initial
Pierson et al Waiting Time in Fast Food
26
conjecture, as well as the belief expressed by industry experts, that in the fast-food drive-thru industry
customers trade off price and waiting time. In addition, our estimates in Table 4, indicate that consumers
attribute a very high cost to the time they spend waiting. In particular, to overcome an additional second
of waiting time, an outlet will need to compensate an average customer by as much as $0.06 in a meal
whose typical price ranges from $2.25 to $6. This corresponds to an hourly cost rate of more than ten
times the (pre-tax) average wage of $18/hour. Even when considering the 90% condence intervals for
the estimates, comparing (opposite) extreme values of these condence intervals, reveals that the average
consumer assigns a cost to waiting, which corresponds with an hourly rate of at least $40, more than twice
the average pre-tax wage in the US. Since price differences in this, as in many other industries, are rather
modest, this implies that in the drive-thru market waiting time plays a more signicant role than pricing
in explaining sales volumes. Moreover, these results seem to justify the continuing substantial trend of
investments chains make to improve their waiting time performance.
We also observe that the sensitivity to the distance measure is not statistically signicant. The contrast
with the waiting time sensitivity is, at rst, striking. After all, the consumers need to traverse a given
distance translates into an amount of time expended in traveling as well as the waiting process in the drive-
thru queue. The fact that this loss of time is valued so differently from the time spent waiting in the drive-
thru queue itself is consistent with nding in the behavioral economics literature, see e.g. Kahneman and
Tversky (1984) and Larson (1987) which reveal that individuals value time very differently, depending on
the context and the degree to which the time is spent is pleasurable or not: most people mind time spent
driving far less than time waiting idly; many even enjoy the ride. In addition, the distance between the
consumers residence and the outlet is not the best possible measure for the additional effort and time she
needs to expand to travel to the outlet. After all, many consumers stop in a drive-thru on the way from one
point to the other, and thus there is no real disutility to distance. A last explanation is the fact that our study
was carried out in Cook County, which is fairly populated and dense in fast food outlets.
An outlet derives a shift of its utility value if it belongs to either the McDonalds and Burger King chains:
the shift is positive, yet not statistically signicant. On the other hand a Dairy Queen franchise has a negative
Pierson et al Waiting Time in Fast Food
27
shift in the outlets mean utility value, which is again insignicant. Due to the lack of statistical signicance,
these -parameters are not reported in the paper, see (1).
As stated above, one of the advantages of the estimation method is that it also allows us to estimate
the marginal costs. The average marginal costs for the three largest chains: McDonalds, Burger King and
Wendys are fairly close. These, in conjunction with the average prices can be used to infer market power
of each of the chains. Given that the average prices for these chains are $4.96, $4.86 and $4.75, we get
that the markups are $2.414, $2.407 and $2.413. One can infer that the market power of all chains is quite
signicant, and note that the differences between them are insignicant.
6.1. 1 Stage Estimation With Macro Moments
In order to test the robustness of our estimates and also to attempt to improve the efciency of our esti-
mates, we supplemented the twenty-four micro-moments, introduced in 5, with additional so-called macro-
moments. Imbens and Lancaster (1994) suggest supplementing micro-moments with macro-moments to
increase the efciency of the estimates, and this approach has been used in industrial organization studies
by Petrin (2002) and Davis (2006).
We have added macro-moments that are based on 3 demographic features: age, race, and gender. We use
the study by Paeratakul et al. (2003), which reports the proportion of people in various demographic groups
that consume fast food over a two day period. The following twelve macro-moments were added to the
micro-moments (10 based on comparisons between age brackets, one between genders and one between
races).
G
09,1016
() =
1
J
J

j=1
[R
09
Q
j,1019
()
Pop
1019
R
1019
Q
j,09
()
Pop
09
] (18)
. . .
G
4059,60+
() =
1
J
J

j=1
[R
4059
Q
j,60+
()
Pop
60+
R
60+
Q
j,4059
()
Pop
4059
] (19)
G
J
() =
1
J
J

j=1
[R
Male
Q
j,Female
()
Pop
Female
R
Female
Q
j,Male
()
Pop
Male
] (20)
G
J
() =
1
J
J

j=1
[R
Black
Q
j,White
()
Pop
White
R
White
Q
j,Black
()
Pop
Black
] (21)
Pierson et al Waiting Time in Fast Food
28
where R
09
denotes the national fraction of fast-food consumers who belong to the 0-9 age bracket as
estimated by the Paeratakul et al. (2003) study, Pop
09
denotes the Cook County population in this age
bracket, and Q
j,09
() denotes the demand of consumers age 0-9 at outlet j. Similar denitions pertain
to the other R , Pop , and Q numbers. As suggested in Thomadsen (2005a) The macro-moments are
constructed based on the idea that the consumption ratio of related demographic groups in Cook County
should be close to the national consumption ratios. For example, the local ratio of men to women consuming
a fast food meal should match the national ratio. The addition of these moments yielded the results detailed
in Table 5, which reports the estimated coefcients for the main parameters.
Table 5 Estimates of Price, Waiting Time and Distance Coefcients Employing Additional Macro-Moments
Variable Name Coefcient Est. Condence Interval
Price Sensitivity($) 0.4411 [0.267, 0.6098]
Waiting Time Sensitivity (sec) 0.025 [0.0016 0.0490]
Distance Disutility (mi) 0.0455 [5.7E-05 1.409 ]
The results are reported based on a single stage of estimation. We were unable to perform the second
estimation stage since the moment variance-covariance matrices at most candidate optima were close to
singular. The estimates are close to those obtained without the macro-moments, thus showing that our results
are robust. Due to the fact that we could not perform the second stage of the estimation procedure, the new
condence intervals are slightly wider than when estimating without the macro-moments.
6.2. Counter Factuals
How much, then, is it worth to reduce the waiting time standards? We mentioned the industry maxim that a
7 second reduction in waiting times increase a chains market share by 1%. We investigated the impact of
a single chain reducing its waiting time standard by 7 seconds, allowing all outlets to adjust their prices to
the new price equilibrium. The results of this experiment can be seen in Table 6. In the rst two rows, we
give the estimated daily demand and market share of each chain at the current waiting time standards and
prices. The subsequent rows show the resulting changes in the chains market share and demand volume.
The percentage of the total market captured at the current waiting time standards closely matches the results
in Paeratakul et al. (2003), providing further validation of our estimates.
Pierson et al Waiting Time in Fast Food
29
Our results conrm that the industry maxim is, on average, correct. However, the absolute change in
market share ranges from 3% at McDonalds (the market leader) to 0.04% at Dairy Queen, with Wendys,
the fastest chain in 2007 and 2008, experiencing an increase by 1.33%. (The percentage increase in market
share ranges between 4% at McDonalds and 20% at Dairy Queen.) Even more importantly, an unmatched
reduction of McDonalds waiting time standard by 7 seconds results in an increase of its sales volume by
approximately 15%. Note that the increase in demand comes primarily from attracting newcustomers to the
market. The percentage of the potential fast food market captured by all the chains grows by more than 1%
when any of the three large players lower their waiting time. Any chains unilateral waiting time reduction,
Table 6 Change In Market Share Following 7 second Wait Standard Reduction at a Chain
McD BK WN Wh. Castle DQ S n S % of Total Market
Initial Demand 1.15E+06 2.22E+05 1.11E+05 2.27E+04 2.73E+03 4.84E+03 22.92%
Init. Market Share 75.98% 14.68% 7.33% 1.5% 0.18% 0.32%
McD ( Demand) 1.82E+05 -7.44E+03 -3.71E+03 -761.76 -91.38 -162.07 25.50%
( Market Sh.) 3.16% -1.93% -0.96% -0.20% -0.02% -0.04%
BK -7.64E+03 4.24E+04 -737.17 -151.31 -18.15 -32.19 23.43%
-2.16% 2.42% -0.21% -0.04% -0.01% -0.01%
WN -3.83E+03 -739.64 2.16E+04 -75.78 -9.09 -16.12 23.18%
-1.09% -0.21% 1.33% -0.02% 0.00% 0.00%
WC -787.63 -152.22 -75.98 4.51E+03 -1.87 -3.32 22.97%
-0.23% -0.04% -0.02% 0.29% 0.00% 0.00%
DQ -94.54 -18.27 -9.12 -1.87 542.63 -0.40 22.93%
-0.03% -0.01% 0.00% 0.00% 0.04% 0.00%
SS -167.66 -32.40 -16.17 -3.32 -0.40b 962.04 22.93%
-0.05% -0.01% 0.00% 0.00% 0.00% 0.06%
is likely to result in waiting time changes by the competing rms. Indeed, between 2005 and 2008, almost
all chain have gradually reduced their waiting time standards, with McDonalds going from 224 to 158
seconds and Wendys reducing its standard from 173 to 131. To invesitgate the chains equilibrium waiting
time choices, we have considered the two-stage game outlined in Section 3.4. In the rst stage, chains
choose the target waiting times and in the second stage the outlets compete in prices, taking the waiting
times as given. As explained in Section 3.4, we assume that each chain maximizes chain wide prots. This
counter factual study requires knowledge of the chains variable cost ratios {d
k
}, see (6). Like the {c
k
}-
parameters these are not directly observable; moreover they cannot be estimated either, since they do not
Pierson et al Waiting Time in Fast Food
30
appear in the equilibrium condition (8). All we know is that 0 d
t
c
k
. We have, therefore, computed
the two-stage equilibrium assuming d
k
=0.5c
k
, k =1, ..., K and a minimum feasible waiting time standard
of 2 minutes, approximately 10% lower than the 2008 best practice of 131 seconds (by Wendys). Indeed,
in equilibrium all chains reduce their waiting time standards to the minimum feasible value, providing an
analytical justication for the industry trend over the past 5-10 years.
In conclusion, reducing waiting time standards pays off handsomely in the fast food industry. Consumers
assign an implicit value to waiting in the drive-thru queue which amounts to at least $40/hr, more than twice
the pre-tax U.S. wage. A 7 second reduction, the magnitude of Wendys improvement from 2007 to 2008,
implies an average increase of a chains market share by approximately one percentage point but for a
large chain like McDonalds, it would result in an increase by more than 3% and an increase in its sales
volume by 15%. The competitive dynamics are such that, to the extend feasible via marginal process and
technological improvements, it is in all chains interests to reduce their waiting times; this occurs to a large
extent because such service improvements result in more potential consumers selecting the fast food option.
References
Allon, G., A. Federgruen. 2006. Service competition with general queueing facilities. Forthcoming in Operations
Research .
Allon, G., A. Federgruen. 2007. Competition in service industries. Operations Research 55(1).
Allon, G., A. Federgruen, M. Pierson. 2009. Price competition under multinomial logit demand functions with random
coefcients. Working Papaer, Kellogg School of Management, Northwestern Univeristy, Evanston, IL .
Armony, M., M. Haviv. 2001. Price and delay competition between two service providers Technical Report, Stern
School, NYU, NY.
Berry, S., J. Levinsohn, A. Pakes. 1995. Automobile prices in market equilibrium. Econometrica 63(4) 841890.
Berry, S.T. 1994. Estimating discrete-choice models of product differentiation. The RAND Journal of Economics
242262.
Bresnahan, T.F. 1987. Competition and collusion in the American automobile industry: The 1955 price war. The
Journal of Industrial Economics 457482.
Cachon, G.P., F. Zhang. 2007. Obtaining fast service in a queueing system via performance-based allocation of
demand. Management Science 53(3) 408.
Cachon, P. C., P. T. Harker. 2002. Competition and outsourcing with scale economies. Management Science 48(10)
13141333.
Pierson et al Waiting Time in Fast Food
31
Caplin, A., B. Nalebuff. 1991. Aggregation and imperfect competition: On the existence of equilibrium. Econometrica:
Journal of the Econometric Society 2559.
Chen, H., M.Z. Frank, O.Q. Wu. 2005. What actually happened to the inventories of American companies between
1981 and 2000? Management Science 51(7) 1015.
Chen, H., M.Z. Frank, O.Q. Wu. 2007. US Retail and Wholesale Inventory Performance from 1981 to 2004. Manu-
facturing & Service Operations Management 9(4) 430.
Chen, H., Y-W. Wan. 2003. Price competition of make-to order rms. IIE Transactions 35(9) 817832.
Davis, P. 2006. Spatial competition in retail markets: movie theaters. RAND Journal of Economics 964982.
DeVany, A., T. Savings. 1983. The economics of quality. Journal of Political Economy 91(6) 9791000.
Dube, J.P.H., J.T. Fox, C.L. Su. 2008. Improving the Numerical Performance of BLP Static and Dynamic Discrete
Choice Random Coefcients Demand Estimation. Working Paper. University of Chicago, Chicago, IL .
Feenstra, R., J. Levinsohn. 1995. Estimating markups and market conduct with multidimensional product attributes.
Review of Economic Studies 1952.
Gans, N. 2002. Customer loyalty and supplier quality competition. Management Science 207221.
Gans, N., G. Knox, R. Croson. 2007. Simple models of discrete choice and their performance in bandit experiments.
Manufacturing & Service Operations Management 9(4) 383.
Gans, N., G. Koole, A. Mandelbaum. 2003. Telephone call centers: Tutorial, review and research prospects. Manu-
facturing & Service Operations Management 5(2) 79141.
Gaur, V., M.L. Fisher, A. Raman. 2005. An econometric analysis of inventory turnover performance in retail services.
Management Science 51(2) 181194.
Gilbert, S.M., Z.K. Weng. 1998. Incentive effects favor nonconsolidating queues in a service system: The principal-
agent perspective. Management Science 16621669.
Hall, A.R. 2005. Generalized method of moments. Oxford University Press, USA.
Hall, J., E. Porteus. 2000. Customer service competition in capacitated systems. Manufacturing & Service Operations
Management 2(2) 144.
Hasija, S., E.J. Pinker, R.A. Shumsky. 2007. Call center outsourcing contracts under information asymmetry. Man-
agement Science .
Hassin, R., M. Haviv. 2003. To Queue or Not to Queue: EquilibriumBehavior in Queueing Systems. Kluwer Academic
Publishers, Boston, MA.
Hughlett, M. Nov 28, 2008. Drive-throughs done right ring up returns. Chicago Tribune .
Imbens, G.W., T. Lancaster. 1994. Combining micro and macro data in microeconometric models. The Review of
Economic Studies 655680.
Kahneman, D., A. Tversky. 1984. Choices, values, and frames. American Psychologist 39(4) 34150.
Pierson et al Waiting Time in Fast Food
32
Kalai, E., M. Kamien, M. Rubinovitch. 1992. Optimal service speeds in a competitive environment. Management
Science 38(8) 11541163.
Lafontaine, F., K.L. Shaw. 1999. The dynamics of franchise contracting: Evidence frompanel data. Journal of Political
Economy 107(5) 10411080.
Larson, R.C. 1987. Perspectives on queues: social justice and the psychology of queueing. Operations Research
895905.
Lederer, P. J., L. Li. 1997. Pricing, production, scheduling and delivery-time competition. Oper. Res. 45(3) 407420.
Levhari, D., I. Luski. 1978. Duoply pricing and waiting lines. European Economic Review 11 1735.
Li, L., Y. S. Lee. 1994. Pricing and delivery-time performance in a competitive environment. Management Science
40(5) 633646.
Loch, C. 1991. Pricing in markets sensitive to delays. Ph.D. Dissertation, Stanford University, Stanford, CA .
Luski, I. 1976. On partial equilibrium in a queueuing system with two servers. The Review of Economic Studies 43
519525.
Nevo, A. 2000. A practitioners guide to estimation of random coefcients logit models of demand. Economics &
Management Strategy 9(9) 513548.
Nuckolls, M. W. 2007. Building a better drive-thru: Consumer drive-thru preferences. QSR Magazine Drive-Thru
Time Study.
Olivares, M., G. Cachon. 2007. Competing Retailers and Inventory: An Empirical Investigation of US Automobile
Dealerships .
Paeratakul, S., D.P. Ferdinand, C.M. Champagne, D.H. Ryan, G.A. Bray. 2003. Fast-food consumption among US
adults and children: dietary and nutrient intake prole. Journal of the American Dietetic Association 103(10)
13321338.
Petrin, A. 2002. Quantifying the benets of new products: The case of the minivan. Journal of Political Economy
110(4) 705729.
So, K. 2000. Price and time competition for service delivery. Manufacturing Service and Operations Management
2(4) 392409.
Thomadsen, R. 2005a. The effect of ownership structure on prices in geographically differentiated industries. The
Rand Journal of Economics 36(4).
Thomadsen, R. 2005b. Existence of pure strategy equilibria among geographically dispersed rms. Economics Bulletin
12(12) 14.
Topkis, D.M. 1998. Supermodularity and complementarity. Princeton University Press.
Wang, H., T.L. Olsen. 2008. Service Capacity Competition with Peak Arrivals and Delay Sensitive Customers. Work-
ing Paper. Washington University, St Louis, MO .