Research Article

Retail gasoline pricing: A Bayesian

hierarchical approach to modeling
the effect of brand on elasticity
Received (in revised form): 28th August 2011

David McCaffrey, Tom Liptrot and Barbara Jenkins

KSS Fuels, Manchester, UK

David McCaffrey is Head of Pricing Science at KSS Fuels. He has previously worked as a consultant
statistician for Shell. He has a PhD in Mathematics from the University of Leeds and a BSc in Mathematics
from the University of Glasgow.

Tom Liptrot is Statistician at KSS Fuels. He has previously worked as an analyst at the UK National Audit
Office. He has an MSc in Applied Statistics and Operational Research from the University of Salford and
undergraduate degrees in Economics from Birkbeck College, London and Philosophy from the University of
Edinburgh.

Barbara Jenkins is a Research Analyst at KSS Fuels. She was previously a researcher at the British Antarctic
Survey. She has a PhD in Applied Mathematics from the University of Central Lancashire and a BSc in Physics
from the University of Durham.

Correspondence: Tom Liptrot, KSS Fuels, St. James’s Buildings, 79 Oxford Street,
Manchester M1 6SS, UK
E-mail: [email protected]

ABSTRACT Retail gasoline demand can be modelled using proprietary site level sales data for one
operator or aggregate market share data covering all brands within a region. We show how to incorporate
standard demand models for these situations into Bayesian hierarchies in which the effect of brand on
price elasticity can be estimated. For site level data, we show how to estimate brand effect on the distri-
bution of competitor cross elasticities and, in the case of multi-branded operators, on the distribution of
direct elasticities. For market share data, we show how to estimate regional average direct elasticities by
brand.
Journal of Revenue and Pricing Management (2011) 10, 514–527. doi:10.1057/rpm.2011.30
Keywords: Bayesian hierarchical model; elasticity; gasoline pricing

INTRODUCTION fuels from major oil companies usually contain

Gasoline and diesel motor fuels are sold to retail chemical additive packs with physically differ-
customers at filling station sites. Most such entiating features, such as engine cleaning pro-
sites display a brand on the pole and pumps, perties. However, in general, motor fuel is
often that of a major oil company, a national a homogenised set of products manufactured
superstore chain or a national convenience store to tightly defined technical specifications. The
(c-store) chain. Other sites display regional main value of a brand to the motor fuel retailer
brands. Others still are unbranded. Branded is therefore to communicate to the passing

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527
www.palgrave-journals.com/rpm/
 Retail gasoline pricing

motorist such things as the range of non-fuel inferences as to relative differences in cross
offerings, value for money, cleanliness, security or elasticity between brands, and thus to support
quality of service they can expect to find on-site. decisions as to which competitor brands should
Some fuel retailers, for example, oil com- be used in defining strategic price targets.
panies, own the brand displayed on their sites. Next, consider the site branding decision.
Other retailers franchise rather than own the Many factors in this decision are outside the
brands they display. Both types of retailers have scope of the current discussion. We consider
to determine price strategies for their sites only its impact on the direct elasticity of the site,
which will be consistent with the branding of that is on the sensitivity of sales to changes in
those sites. The latter class of retailers also have own price. The idea we propose is that the
to decide which brand to display on a site. chosen brand should influence the direct elasti-
Consider first the determination of pricing city in a manner which compliments, rather
strategy for a site where the branding decision than counter-acts, the influences of the site
is already fixed. Fuel pricing strategy is gene- location on direct elasticity.
rally expressed as target price points or ranges The biggest factor influencing the elasticity
relative to specified competitor brands. It is of a site is its location. This determines the
therefore useful to understand the sensitivity of volume and type of traffic passing the site,
sales to price changes by each competitor brand, the density of nearby competition and thus the
so that the targets are defined relative to the willingness of passing drivers to purchase fuel at
most important competitors. In other words, a given price. Other factors such as the facilities
one needs to understand cross price elasticity to or branding also influence elasticity, but these
each competitor brand. are generally weaker.
Cross elasticity estimates derived from site Pricing strategies of different retailers take
level demand models tend to have low levels of account of elasticity to differing extents. A
statistical confidence due to the high degree of simplification of a general class of strategies
multicollinearity in the price data. The standard might be stated: price lower for volume at
solution is to fit a hierarchical demand model in high elasticity sites and vice versa. This type of
which sites have separate elasticity estimates, but strategy seeks to differentiate pricing across a
those estimates are assumed to come from some network of sites, gaining volume at sites where
common distribution. This retains the useful it costs less in margin terms to do so, and vice
information present at site level, but improves versa. Within this type of strategy, one should
the statistical confidence in the elasticities by choose the site branding to compliment the
shrinking low confidence estimates towards the effect of location on elasticity, thus ensuring
mean of the common distribution. that both factors act coherently to maximise the
In this article, we show how to build a benefits of the strategy. So it is useful to under-
Bayesian hierarchical model, based on a stan- stand how direct elasticity is influenced by site
dard site level demand model form, in which branding. We propose two ways to do this.
there is a common prior distribution of site Firstly, for retailers who operate more than
level cross elasticities for each competitor brand one brand, we can extend the above site level
present. We fit this model on some (ano- Bayesian hierarchical model to include co-
nymised) site level sales and price data using variate information on own site branding with-
Monte-Carlo Markov chain techniques and in the common prior distribution of site level
examine the resulting posterior estimates for direct elasticities. Again, the resulting estimates
the mean of the common distribution of are tight enough to support inferences as to the
cross elasticities for each competitor brand. We relative impact of brand on direct elasticity.
show that the probability intervals on these Secondly, to assess the impact of brands not
mean estimates are sufficiently tight to support currently represented in the retailer’s proprietary

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527 515
 McCaffrey et al

operating data, we can use data on regional For a general introduction to hierarchical
market share and price by brand. We define Bayesian modelling, see for instance Rossi et al
a suitable hierarchical model in which direct (2005), or Raudenbush and Bryk (2002).
elasticity can be related to brand, again based on The article is organised as follows. The next
a standard model form. We use some prior section sets out the technical framework of
economic assumptions to reduce the number Bayesian hierarchical modelling. The subse-
of independent variables in this model, fit it on quent section applies this to analyse the impact
some real data and examine to what extent of brand on price sensitivity at site level. The
inferences can be drawn on the relative impact latter section applies it to analyse the impact of
of brand on elasticity. Owing to the averaging brand on price sensitivity at the US state level.
present in regional data, these inferences are
not as strong as for models based on site level BAYESIAN HIERARCHICAL
data. However, they are strong enough to LINEAR MODEL
support some interesting conclusions, in parti- Our aim is to build a linear model of sales as
cular that operators of different brands appear to a function of price. Sales will be defined for
differentiate their pricing strategies in line with an entity – either a site or all sites of the same
relative differences in elasticity. brand in a state. Each entity has its own model
Academic interest in fuel pricing in the last instance. Entities are similar in structure but
few years has focused on the public availability heterogeneous in their characteristics. We want
of large databases of site level daily pricing to examine how the parameters of the model
(but not sales) data, from which inferences are vary with the attributes of the entity, in parti-
drawn regarding market dynamics, competitive cular its brand. Bayesian hierarchical modelling
practices and strategic pricing objectives – see provides an effective structure for this analysis.
Slade (1992), Eckert and West (2004), Noel At the bottom level of the hierarchy we have
(2007a, b) and Faber (2009). Earlier references, i ¼ 1, y, m regression equations, one for each
such as Clayclamp (1966), Bennavail et al entity i, of the form
(1990), Singh and Bennavail (1993) and Krasteva
et al (1994), looked at modelling site level fuel yi ¼ Xi bi þ ei
demand and competitor cross elasticity using where yi is logged sales volume per time period
proprietary sales data. Authors such as Cohen at entity i, Xi is a k-dimensional data vector
(1999) and Ning and Haining (2003) have of logged prices at entity i and relevant compe-
studied brand effect on pricing. ting entities per time period and bi is a
Various authors have looked at relationships k-dimensional parameter vector. Each regre-
between brand and elasticity in other consu- ssion equation has a separate prior on the error
mer markets. For instance, Erdem et al (2002)
consider food, clothing, toiletries and electro- ei  iidN ð0; s2i Þ
nic goods, while van Nierop et al (2002) con- The independent regression equations are
sider supermarket items. Other authors have tied together by assuming that the {bi} have
incorporated prior economic information into a common prior distribution in the form of
hierarchical Bayesian models to reduce dimen- a multivariate regression model
sionality – see for instance Montgomery and 2 03 2 0 3
Rossi (1999) and Montgomery (2002). Our b1 z1
6 . 7 6 . 7
approach follows that of these latter two refer- B ¼ZD þ V ; B ¼ 6 7 6 7
4 .. 5; Z ¼ 4 .. 5;
ences. As far as we are aware, this is the first 0 0
time that hierarchical Bayesian techniques have bm zm
h 0 i
been applied to estimate brand effects for retail 0 0
D ¼ d1    dk ; ni  N ð0; Vb Þ
fuel markets.

516 & 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527
 Retail gasoline pricing

necessary to specify an inverse chi-squared prior

on each error variance s2i in the form
vi s20i
s2i  :
w2vi
The conditional independence structure
allows the model to be estimated via a Gibbs
Figure 1: Hierarchical linear model. sampler. In the first half of each iteration, this
draws bi and s2i for each entity from a set of
univariate normal posteriors conditioned on
At this upper level of the hierarchy, B is the draws of D and Vb from the previous
m  k, Z is m  d and D is d  k. Data on the iteration. In the second half of an iteration, it
d attributes of each entity can be entered in the draws D and Vb from a multivariate normal
z variables. Each column of D has coefficients regression posterior conditioned on bi. This is
which describe how the mean of the corre- repeated as necessary to generate a Monte Carlo
sponding component of b varies as a function of simulation of the posterior distribution of each
the data in the z variables. In the special case parameter.
where d ¼ 1 and Z is a column of 1’s, then D We implemented the Gibbs sampler using
is row vector representing the means of the the bayesm package (Rossi et al, 2005) within
components of b. The term Vb is the k  k the R statistical environment (www.r-project
covariance matrix of b. .org). It can also be implemented using the
Our aim is to use the hierarchical model to WinBUGS package (www.mrc-bsu.cam.ac.uk/
simultaneously estimate the regression coeffi- bugs/winbugs/contents.shtml). We chose diffuse
cients in each individual entity model at the prior settings
lower level together with the global mean
and variance of the distributions of those D ¼0; A ¼ 0:01; n ¼ k þ 3; V ¼ n0:1Ik ;
coefficients. vi ¼3; s20i ¼ varðyi Þ:
The model is estimated using Monte-Carlo
Markov chain techniques. This involves speci- These are acceptable given the large data
fying priors in the form of the sequence of volumes available in both applications below.
conditional distributions shown on the left-
hand side of Figure 1. These represent an SITE LEVEL MODEL
assumption that the parameters D and Vb are We apply the hierarchical framework to analyse
conditionally independent of the sales and price impact of brand on price sensitivity at site level.
data yi, Xi given the bi regression coefficients, This uses data available to an operator of a retail
as shown in the directed graph on the right- site network, namely daily sales volumes and
hand side of Figure 1 – see for instance Rossi prices for each site.
et al (2005). The lower level of the hierarchy consists of
The prior distributions of D and Vb are i ¼ 1, y, m models, one for each own site
specified by normal and inverse Wishart dis- entity, of the form
tributions
X
k
logðVi Þ ¼ai þ b0i logðp0i Þ þ bji logðpji Þ þ ei ;
D  NðD; Vb  A1 Þ Vb  IW ðn; V Þ: j¼1

ei  iidNð0; s2i Þ
The prior of the (m  k)-dimensional joint
distribution of b is thus fully specified by choice where Vi is daily total sales volume at site i for
of the hyper-parameters D, A, n, V. It is also a given grade of fuel, p0i is the average daily

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527 517
 McCaffrey et al

price at site i and pji is the average daily price The first case of interest is how the cross
at j ¼ 1, y, k other sites deemed to compete elasticities bji, j ¼ 1, y, k vary with competitor
with site i. The b coefficients are the elasticities. brand. So set Z to be a column of 1’s. Then D is
The first one b0i acting on the own price is a row vector containing the means of (ai, b0i,
the direct elasticity. The others bji acting on the b1i, y, bki). In particular, element d0 of D is the
competitor prices are the cross elasticities. This mean of the distribution of direct elasticities b0i
is a standard form of demand model and makes and element dj ; j ¼ 1; y; k is the mean of the
the usual linear regression assumptions regard- distribution of cross elasticities bji for competi-
ing the logged data, in particular that there is no tor brand j. The point is that D contains the
serial correlation and no multicollinearity. We mean of bji , which is denoted dj . It doesn’t
highlight these two because they are usually contain bji .
violated by real data. The serial correlation issue The second case of interest is how the direct
can be dealt with by including appropriate time elasticity b0i varies with brand. So here set zi0
series terms in the model as described below. to be a d-dimensional vector of data attributes
The multicollinearity is dealt with quite for the ith site, with a 1 in the first component
effectively via the shrinkage imposed by the and with brand and other relevant site attributes
Bayesian hierarchical model. At those sites represented in the remaining d1 components.
with a high degree of correlation between own D is then a matrix of regression coefficients.
and competitor prices, the elasticity estimates are The elements of column d0 of D indicate how
shrunk towards the mean of the common prior the mean of the distribution of direct elasticities
defined by the upper level of the hierarchy, b0i varies with these own site attributes, in
whereas at sites with more variation in relative particular the brand.
price differences between competitors, the local The models are estimated using a Gibbs
information dominates the elasticity estimate. sampler as outlined above. Time series effects
We assume that each own site has the same can be accounted for via autoregressive volume
total number k of competitor sites, that each of terms and/or seasonal variables. For the data
these competitor sites belongs to a different described below, it was sufficient to include a
competitor brand and that a given choice of six-level factor variable representing day of
jA{1, y, k} represents the same competitor week, with the a intercept term (dummy)
brand at each own site. These assumptions encoding the seventh day of week level.
ensure that each lower level entity has the same To model the variation of cross elasticity with
demand model form. They can be satisfied by competitor brand we used one year of data
arranging the data in a number of ways. For from a single branded US fuel retailer with
instance, for a given own site i we can take pji to a network of around 400 sites. Initial data ana-
be the price at the nearest brand j site. lysis indicated typical characteristic of site level
The upper level of the hierarchy consists of price and sales data. Daily prices are highly
a multivariate regression of the form collinear and driven mainly by changes in
wholesale fuel costs. Volume varies across the
0 3
2 0 3
2
b1 z1 400 sites, but for a given site, average weekly
6 . 7 6 . 7 volume is fairly stable. Day-of-week seasonality
B ¼ZD þ V ; B¼6 7
4 .. 5; Z¼6 7
4 .. 5; is the main influence on variation in site volume,
0
bm zm
0 related to variation in traffic volume. Price
0 explains a relatively small proportion of vari-
D ¼½da ; d0 ; d1 ; . . . ; dk ; ni  Nð0; Vb Þ ation, in part because competing retailers main-
tain reasonably stable relative prices and so the
where bi0 ¼ (ai, b0i, b1i, y, bki) is the parameter data generally contains fairly limited variation in
vector appearing in the ith lower level equation. relative prices between competitors.

518 & 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527
 Retail gasoline pricing

As mentioned above, the Bayesian hierarch- M 2.38

H 1.91
ical approach deals quite effectively with the B 1.18
multicollinearity of prices, better than other J 0.9
E 0.51
shrinkage techniques such as ridge regression. L 0.41
In a given lower level model, the direct elasti- C 0.28
D 0.07
city and at least some of the cross elasticities F −0.06
will generally be significant, in the Bayesian UNB −0.06
K −0.18
sense that the 95 per cent probability intervals G −0.2
for their posterior distributions will not con- Other −0.26
I −0.76
tain zero, and the means of these distributions A −2.15
will have the expected signs and reasonable OWN −4.04

magnitudes. The insignificant cross elastici- −4 −2 0 2 4

Mean Elasticity
ties will usually have mean estimates close to
zero and so have little impact on prediction Figure 2: Level 2 posterior elasticities.
accuracy. The parameter estimates are reason-
ably stable when fitted on different subsets of
the data.
Figure 2 shows the 95 per cent probabi- K Own sales are equally sensitive to the pricing
lity intervals for the posterior estimates of of E, L and C – E is a regional brand.
the parameters of the upper level of the hier- K Own sales are insensitive to the pricing of
archy, namely the mean direct elasticity (labe- brands D, F, UNB, K and G
lled OWN) and mean cross elasticity for
each (anonymised) competitor brand. The The cross elasticities for brands Other, I, A are
probability intervals on these estimates are positive. This could be due to a high degree of
tight enough to support inferences of the correlation between own prices and prices of
following types, where again an estimate is these brands, or to specific dynamics of the
significant if its probability interval does not market where this retailer operates.
contain zero, and two estimates are significantly To model the variation of direct elasticity with
different if their respective intervals do not own brand we used one year of data from a US
overlap: fuel retailer with a network of around 500
sites displaying 10 different brands. The general
K Own sales are most sensitive to pricing of comments made above on the structure and
competitor brands M and H – the former is sources of variation in site level data apply here
a national superstore chain selling discounted also. The data matrix Z included the following
gasoline, the latter is a national brand with a own site attribute data: brand, US state, location
mid-market average price point and strong type and restaurant type. Dummy encoding
regional presence. was used to represent the attribute data, so that
K Own sales are about half as sensitive to in the vector d0 of regression coefficients for the
pricing of M and H as they are to own mean of the distribution of direct elasticities, the
pricing. first component (that is, the intercept term)
K Own sales are more sensitive to the pricing corresponds to one particular choice of brand,
of M and H than to B and J – these latter two state, location type and restaurant type (labelled
are national brands again with mid-market respectively I, N, Highway and B). The additive
average price points. effect of the other factor levels is captured in the
K Own sales are more sensitive to the pricing of remaining components of d0. Figure 3 shows the
B and J than to L and C – L and C are national 95 per cent probability intervals for the posterior
brands with higher average price points. estimates of these regression coefficients. These

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527 519
 McCaffrey et al

State (Reference N)
O
Instead we apply the framework of the
T
Restaurant (Ref B) section ‘Bayesian Hierarchical Linear Model’
No Restaurant
C
A
to analyse the impact of brand on elasticity
Own Brand (Ref I)
G using market share and average price data
A
B
F
for all brands within a given region. Availabi-
H
C lity and granularity of such data varies from
D
Unbranded
E
country to country. In the United States it
City Type (Ref Highway)
Urban
is publically available at the state level from
Rural
Tourist sources such as Oil Price Information Service
(Intercept) (www.opisnet.com), which reports (for a fee)
−6 −4 −2 0 2 4 6 on gasoline market share by brand, US state
Figure 3: Effect of site attributes on own elasticity. and week. The data are estimated from brand
share of volume purchased on a widely accep-
ted fuel card.
So for a given US state, define the following
are tight enough to support the following types
notation
of inference:
j¼ brand (1, y, m)
K Mean direct elasticity associated with brand t¼ week
I, state N, highway location and restaurant Njt ¼ number of sites for brand j week t
type B is about 2 and is significant. Wt ¼ total number of sites over all brands
K Brands A, B, C, D, F, G, H or Unbranded week t
have the same impact on direct elasticity as Mjt ¼ market share for brand j week t (as
brand I (the additive effect of any one of estimated by the above data source)
these is insignificant). Pjt ¼ average retail price for brand j week t
K Brand E approximately doubles mean direct (that is, average over all sites belonging
elasticity as compared to any of the other to
Pmthat brand)
brands (the additive effect is about 2 and is P̄t ¼ j ¼ 1Pjt/m ¼ (unweighted) average re-
significant). tail price over all brands week t
P^jt ¼ Pjt/P̄t ¼ relative price for brand j week t
In addition, we can infer that being located
in state O has the same impact on direct We build a model of m simultaneous equations,
elasticity as being in state N, but being in state one for the market share of each brand. To start
T adds 1 to the direct elasticity. Restaurant with we assume:
type overall has no impact on direct elasticity.
Being located in an urban, rural or tourist type
K there is a single global parameter a0 which
location each make a site more elastic than
scales the outlet share into sales share;
being located in a highway location.
K there is a single global direct elasticity b0
which represents (for all brands) the com-
mon rate at which market share of any brand
MARKET SHARE MODEL changes as a function of the relative price of
The third case of interest involves assessing that brand;
the impact on direct elasticity of displaying K there are m global cross elasticities gi for
a brand which is not currently represented in i ¼ 1, y, m. The ith cross elasticity repre-
the own site portfolio. Models estimated on sents the common rate at which the market
proprietary site level sales data provide no shares of all brands other than i change as a
information here. function of the relative price of brand i.

520 & 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527
 Retail gasoline pricing

We show first that gi depends on b0. The initial global model (in log-log form) consists of
system of j ¼ 1, y, m simultaneous equations j ¼ 1, y, m equations
for market share for brand j in week t is 
 a0 Njt
Njt Y
m logðMjt Þ ¼ a0 log
Mjt ¼
b
Pbjt 0
g
Pbiti Wt
Wt !
i¼1 X Mi
þ b0 logðPbjt Þ  logðPbit Þ
i 6¼ j 1  Mi
i6¼j

Market shares sum to one, that is, with two independent coefficients a0 and b0.
X
m Now suppose that there is a brand com-
1¼ Mit ð1Þ ponent to each direct elasticity, that is, each
i¼1 brand j has direct elasticity of the form b0 þ bj
where b0 is the global direct elasticity and bj is
It follows that, for each j ¼ 1, y, m, a brand specific adjustment. Suppose, with a
view to constructing a hierarchical model, that
q1 Xm
qMit qMjt Xm
qMit these bj for j ¼ 1, y, m are random variables
0¼ ¼ ¼ þ
qPbjt b
i¼1 qPjt qPbjt qPbjt with mean zero. Then b0 þ bj can then be
i ¼1 thought of as being drawn from a global prior
i 6¼ j distribution with mean b0.
Mjt X
m
Mit There is then also a brand specific adjustment
¼b0 þ gj to the global cross elasticity of brand i of
Pbjt Pbjt
i¼1 the form gi þ gij for jai. This adjustment gij is
i 6¼ j the extent to which the global cross elasticity
for brand i is modified when it appears in the
or equation for brand j market share. We suppose
that these cross elasticity adjustments are defi-
X
m ned as follows, for jai
0 ¼ b0 Mjt þ gj Mit ð2Þ
i¼1 j Mi
gi þ gi ¼ ðb0 þ bj Þ
i 6¼ j 1  Mi

This only gives m1 relations involving gi.

Since this is true for all t, we can average Mjt
As it stands, the mean of the random variables gji
with respect to t to get
for jai is not zero. To obtain an extra rela-
Mj tion consider the market share model incorpor-
gj ¼ b0 P ating the brand adjustments to direct and cross
i6¼j M i elasticity

Time averaged market shares must also sum  a0 Y

m
Njt b þb g þg
j

to 1, so Mjt ¼ Pbjt 0 j Pbiti i

Wt
i¼1
Mj i 6¼ j
gj ¼ b0
1  Mj
Substituting the relevant form of this expre-
that is, the m global cross elasticity parameters ssion, for each i ¼ 1, y, m, into equation (1)
are dependent on the direct elasticity. So the and differentiating with respect to P^jt gives the

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527 521
 McCaffrey et al

following analogue of equation (2) the form

 
X
m Njt
0 ¼ ðb0 þ bj ÞM j þ ðgj þ gij ÞM i ð3Þ logðMjt Þ ¼ aj log
Wt
i¼1 !
X Mi
i 6¼ j þ bj logðPbjt Þ  logðPbit Þ þ ejt
i6¼j
1  Mi
This says that the (market share weighted) ejt  iidN ð0; s2j Þ
direct elasticity for brand j balances the sum of
the (market share weighted) cross elasticities and an upper level multivariate regression for
for brand j appearing in the m1 equations for the common prior of the aj and bj of the form
the other brands iaj, that is, the change in 2 3
a1 b1 2 3
market share of brand j due to a price change by 1
brand j balances the sum of the change in market 6 . 7 6 .. 7
B ¼ZD þ V ; B ¼ 6 7
4 .. 5; Z ¼ 4 . 5;
share of the other brands due to this same price
change. am bm 1
Now define an adjustment term gjj as D ¼½a0 b0 ;
0
nj  N ð0; Vb Þ

j 1 X m
This model is estimated using a Gibbs sam-
gj þ gj ¼ ðgj þ gij ÞM i
1  Mj pler. We note again that time series effects can
i¼1 be taken into account in this model in the form
i 6¼ j of autoregressive terms and seasonal variables.
We omit discussion as it doesn’t affect the
This represents the notional adjustment to qualitative conclusions.
the global cross elasticity for brand j, should it The model was estimated on data for one US
appear in the equation for brand j market share. state for the 12-month period from March 2008
Then, from equation (3), to March 2009. The data covered weekly
market share, share of sites and average price
j Mj by brand within that state. There were 15
gj þ gj ¼ ðb0 þ bj Þ
1  Mj brands covered in the data set. We pick out the
seven largest brands by share in the discussion
which is of the same form as the defining below. In descending order of share, these are
relations for terms gj þ gij for iaj. It follows labelled Multi-National 1, Multi-National 2,
that the mean of the random variable gj þ gij US Mid-Tier, Reg.Indep.1:Brand 1, Reg
over all i, including i ¼ j, is equal to gj. Thus, .Indep.2, Reg.Indep.1:Brand 2 and Unbranded.
for given j, the random variable gij has mean These labels are mostly self-explanatory. Reg
zero. .Indep stands for regional independent. Reg
It is also necessary, for modelling reasons, to .Indep.1:Brand 1 and Reg.Indep.1:Brand 2
introduce a brand specific component into the are two distinct brands operated by the same
scale parameter a ¼ a0 þ aj, so that these are retailer. The former is a high value brand with
thought of as drawn from a global prior dis- presence outside this particular state. The latter
tribution with mean a0. is a convenience brand local to the state.
Now write a0 þ aj simply as aj and b0 þ bj as Figure 4 shows a time series of the mean, 5
bj and let a0 and b0 denote the means respec- and 95 centile of the distribution of weekly
tively of the distributions of aj and bj. We thus price for each brand. The period of high retail
have a hierarchical model of brand market share prices (over $3 per gallon) is the run up to the
with j ¼ 1, y, m equations at the lower level of collapse of Lehman Brothers, when crude

522 & 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527
 Retail gasoline pricing

400 prices in the middle throughout. The more

Price − Cents Per Gallon

expensive brands are slower to reduce their

350 prices than the cheaper brands during the
period of falling prices. It is interesting to note
300
that a differentiated pricing strategy is being
250 applied across the two brands Reg.Indep.1
:Brand 1 and Reg.Indep.1:Brand 2 operated by
200 the same retailer. The relative price of the
former is higher than that of the latter through-
0 10 20 30 40 50 out the whole year and during weeks 26–40 the
Week
price of the former declines more slowly than
Figure 4: Distribution of prices. that of the latter.
Figure 6 shows time series of weekly market
share for the seven largest share brands in the
Highest Priced Brand state. Some of the movement in market share
1.04 in the first 25 weeks is due to changes in the
number and share of sites. The total number of
1.02
sites declines by about 3 per cent during this
period. About 1 percentage point of this
Price

1.00
decline is due to divestments by the largest
brand Multi-National 1. Figure 7 shows weekly
0.98
Lowest Priced Brand
market share, share of sites and the ratio of
0.96
market share to site share for this brand. The
10 20 30 40 50
divestment can be seen in the sharp fall in share
Week of sites up to week 20. Movements in market
Figure 5: Highest and lowest relative price. share in the second half of the year seem to be
related to the divergence in relative price
between the brands which commences in week
prices were high. The Lehman collapse occurs 26. The three relatively highest priced brands
in week 26 of the plot, after which point prices during weeks 26–40, namely Multi-National 1,
decline rapidly to around $1.50 per gallon in Multi-National 2 and Reg.Indep.1:Brand 1 all
week 40. The tightness of the centiles indicates display a loss of market share in this same
that all brands moved their prices broadly in line period. Conversely, the three relatively lowest
with changes in product cost. priced brands, Unbranded, Reg.Indep.2 and
Figure 5 shows the spread of weekly relative Reg.Indep.1:Brand 2, all display gains in market
price over the brands. Relative prices are share.
reasonably stable until around week 26, after The relevant model outputs here are the
which there is a divergence and increased posterior estimates of the parameters of the
volatility. The three most expensive brands lower level model, that is, the direct elasticity
during the period week 26–40 are Multi- parameters bj for each brand j ¼ 1, y, m. These
National 1, Multi-National 2 and Reg.Indep are shown in Figure 8. They are not as well
.1:Brand 1. These generally follow close to differentiated nor as significant as the estimates
the upper line on the plot during this period. obtained in the previous section from site level
The three cheapest brands during the same models. In part this is due to the use of state
period are Unbranded, Reg.Indep.2 and Reg average weekly prices which contain less infor-
.Indep.1:Brand 2. These generally follow close mation. Also it may be due to the particular
to the lower line. The US Mid Tier brand type of fuel card data from which the market

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527 523
 McCaffrey et al

Multi−National 1 Multi−National 2 US Mid−Tier

0.140 0.110

230 0.108

Market Share
0.138

Market Share
0.106
225
0.136 0.104

220 0.102
0.134
0.100
215
0.132 0.098

0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40
Week Week Week
Reg. Indep 1: Brand 1 Reg. Indep 2 Reg. Indep 1: Brand 2
106 0.104
0.090
105 0.102

Market Share
Market Share
104
0.100
103 0.085
0.098
102
0.096
101 0.080
0.094
100
0.092
099
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40
Week Week Week

Unbranded
0.066

0.064
Market Share

0.062

0.060

0.058

0 10 20 30 40 50
Week

Figure 6: Weekly market share by brand.

Market Share Site Share Market Share / Site Share

1.52
0.155
0.230
1.50
0.154

1.48
Market Share

Site Share

0.225
MS / SS

0.153
1.46

0.152
0.220
1.44

0.151
1.42
0.215

0.150 1.40

0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Week Week Week

Figure 7: Multi-National 1.

524 & 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527
 Retail gasoline pricing

shares are estimated. This card is mainly used Small Brand B

Small Brand E
by business drivers. The purchase behaviour of Small Brand D
this type of driver may be less price sensitive Small Brand C
Reg. Indep 1: Brand
than that of the general population. Small Brand G
That said, we can still infer that of the seven Multi−National 2
Reg. Indep 2
largest brands, Multi-National 2, US Mid Tier
US Mid−Tier
and Reg.Indep.1:Brand 1 do not display any Small Brand F
significant direct elasticity, that is, the relation- Unbranded
Multi−National 1
ship between market share and price for these Reg. Indep 1: Brand
brands is inelastic. This result generally makes Small Brand A

sense. One would expect these brand types to −4 −2 0

Own Price Elasticity
2 4

be less price sensitive. It makes particular sense

in respect of Reg.Indep.1:Brand 1, since the Figure 8: Own elasticity by brand.
brand owner operates it as a differentiated high
value brand and prices it towards the top end of
the range of relative prices. those of other brands. Future work will con-
We can also infer that, among the seven sider a revised model which includes data on
largest brands, Reg.Indep.2, Unbranded and various non-price factors representing the ave-
Reg.Indep.1:Brand 2 all have significant price rage quality of sites belonging to each brand.
elasticity. Again this makes sense. One would The probability intervals on the estimates
expect regionally branded or unbranded sites to in Figure 8 are not tight enough to support
be more price sensitive. It makes particular inferences about differences in magnitude of
sense in respect of Reg.Indep.1:Brand 2, since elasticity between brands. However, we can test
the brand owner operates it as a differentiated the extent to which the operators of these
local convenience brand and prices it towards brands are following pricing strategies consis-
the bottom end of the range of relative prices. tent with their brand elasticity. Figure 9 shows
Lastly note that Multi-National 1 has a signi- the mean elasticity estimate for each brand
ficant elasticity. This is perhaps contrary to plotted on the y-axis against their 52-week
expectations for multi-national brands, which average relative price on the x-axis. The seven
tend to be less price sensitive than regional or largest brands are highlighted in bold type.
independent brands. In part, the elasticity esti- We have already presented arguments that
mate for this brand is being skewed by changes the elasticity estimate for Multi-National 1 is an
to its market and site shares related to divest- outlier. We argue that Small Brand H is also an
ment of sites in the first half of the year. Fitting outlier. It represents a national c-store brand,
the model on data from just the second half but its sites generally display one or other oil
year reduces the elasticity estimate for this brand major brand on their pumps. A small number of
from about 3 to about 2, although it their sites display the c-store brand on the
remains significant. We believe that the estimate pumps. The branding definition rules used by
is also being skewed to some extent by the the data supplier classify sites according to the
omission from the model of data on the average pump brand. The Small Brand H market share
size and quality of the sites belonging to each estimate is probably not reliable since it is based
brand. Multi-National 1 sites, on average, have on only 20 or so sites in this state which display
larger sales volumes than those of other brands. Small Brand H on their pumps, when in fact
Some of this may be explained by non-price they have a larger number of c-store sites with
effects specific to that brand, such as that brand’s major brands displayed on the pumps.
sites having on average more pumps, or larger A regression line through the remaining 13
forecourt areas, or being better located than points, excluding Multi-National 1 and Small

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527 525
 McCaffrey et al

Slope = 61.6 (3, 120 )

impact on direct elasticity of the brand displayed
Small Brand H
2
on the site. The third scenario uses publically
1
Small Brand B
available data on market share and price by brand
in a hierarchical model of market share at the
Own Price Elasticity

Small Brand E

0
Small Brand D
Small Brand C
state level. Estimates obtained from this model
−1 Reg. Indep 1: Brand 1 are not as tight as those obtained using site level
Small Brand G
Reg. Indep 2
US Mid−Tier
Multi−National 2 data, due to the more aggregate nature of the
Small Brand F
−2 Unbranded data. However, they are still strong enough to
−3 Reg. Indep 1: Brand 2
Multi−National 1 support some interesting conclusions, in particu-
Small Brand A lar that the operators of different brands appear
0.98 0.99 1.00 1.01 1.02 to differentiate their pricing strategies in line
Relative Price
with relative differences in elasticity.
Figure 9: Own price elasticity plotted against relative
price.

REFERENCES
Bennavail, J.-C., Harding, T. and Spears, D. (1990) An Experiment
Brand H is shown on Figure 9. The slope of this on the Use of a Knowledge Support System for Dynamic Pricing of
line is positive and significant. This indica- Gasoline Products in France, Proceedings of the IEEE Interna-
tional Conference on Systems, Man and Cybernetics; 4–7
tes that operators of lower elasticity brands are November, Los Angeles. Piscataway, NJ: IEEE, pp. 861–866.
following a strategy of pricing higher than Clayclamp, H. (1966) Dynamic effects of short duration price
operators of higher elasticity brands. This is differentials on retail gasoline sales. Journal of Marketing Research
3(2): 175–178.
consistent with the simple elasticity-based pri-
Cohen, M. (1999) Pricing peculiarities of the UK petrol market.
cing strategy set out in the introduction. Journal of Product and Brand Management 8(2): 153–162.
Eckert, D.S. and West, A. (2004) Retail gasoline price cycles
across spatially dispersed gasoline stations. Journal of Law and
Economics 47(1): 245–273.
CONCLUSION Erdem, T., Swait, J. and Louviere, J. (2002) The impact of brand
In this article, we have shown how to build credibility on consumer price sensitivity. International Journal of
Bayesian hierarchical demand models, based on Research in Marketing 19(1): 1–19.
Faber, R.P. (2009) Asymmetric Price Responses of Gasoline
standard demand model forms, which can sup- Stations: Evidence for Heterogeneity of Retailers. Tinbergen
port inferences on the effect of brand on price Institute Discussion Papers 2009-106/1.
elasticity. We have considered three modelling Krasteva, E., Singh, M., Sotirov, G., Bennavail, J.-C. and
Mincoff, N. (1994) Model Building for Pricing Decision Making
scenarios. The first uses proprietary site level in an Uncertain Environment, Proceedings of the IEEE Interna-
sales data for one operator and incorporates tional Conference on Systems, Man and Cybernetics; 2–5
information on competitor brands into the October, San Antonio. Piscataway, NJ: IEEE, pp. 194–199.
Montgomery, A.L. (2002) Reflecting uncertainty about
specification of the common prior distributions
economic theory when estimating consumer demand. In:
of site level cross elasticities. We have fitted T.B. Formby and R. Carter Hill (eds.) Advances in Econometrics
this model on some real data and shown that 16. Bingley, Yorkshire, England: Emerald Group Publishing,
the resulting estimates are tight enough to allow pp. 257–294.
Montgomery, A.L. and Rossi, P.E. (1999) Estimating price
the retailer to determine the relative impact on elasticities with theory based priors. Journal of Marketing
own sales of pricing by each competitor brand. Research 36(4): 413–423.
The second scenario, which is open to multi- Ning, X. and Haining, R. (2003) Spatial pricing in interdepen-
dent markets: A case study of petrol retailing in Sheffield.
branded operators, uses the same data and Environment and Planning A 35(12): 2131–2159.
incorporates information on the brand of each Noel, M.D. (2007a) Edgeworth price cycles, cost-based pricing
own site in the specification of the common and sticky pricing in retail gasoline markets. Economics &
prior distribution of site level direct elasticity. Statistics 89(2): 324–334.
Noel, M.D. (2007b) Edgeworth price cycles: Evidence from the
Again, the resulting estimates are tight enough Toronto retail gasoline market. Journal of Industrial Economics
to allow the retailer to determine the relative 55(1): 69–92.

526 & 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527
 Retail gasoline pricing

Raudenbush, S. and Bryk, A. (2002) Hierarchical Linear Models. Slade, M. (1992) Vancouver’s gasoline price wars: An empirical
Thousand Oaks, CA: Sage. exercise in uncovering supergame strategies. Review of Economic
Rossi, P.E., Allenby, G.M. and McCulloch, R. (2005) Bayesian Studies 59(2): 257–276.
Statistics and Marketing. Chichester, UK: John Wiley. van Nierop, E., Fok, D. and Franses, P.H. (2002) Sales Models
Singh, M.G. and Bennavail, J.-C. (1993) Experiments in the use for Many Items Using Attribute Data, ERIM Report Series
of a knowledge support system for the pricing of gasoline ERS-2002-65-MKT. Rotterdam, the Netherlands: Erasmus
products. Information & Decision Technologies 18(6): 427–442. Research Institute of Management.

& 2011 Macmillan Publishers Ltd. 1476-6930 Journal of Revenue and Pricing Management Vol. 10, 6, 514–527 527