Content uploaded by Harald van Heerde
Author content
All content in this area was uploaded by Harald van Heerde
Content may be subject to copyright.
Vol. 27, No. 6, November–December 2008, pp. 1036–1054
issn 0732-2399 eissn 1526-548X 08 2706 1036
informs®
doi 10.1287/mksc.1080.0358
© 2008 INFORMS
Building Brands
M. Berk Ataman
Rotterdam School of Management, Erasmus University, 3000 DR Rotterdam, The Netherlands, bataman@rsm.nl
Carl F. Mela
Fuqua Business School, Duke University, Durham, North Carolina 27708, mela@duke.edu
Harald J. van Heerde
Waikato Management School, University of Waikato, Hamilton 3240, New Zealand,
and CentER, Tilburg University, 5000 LE Tilburg, The Netherlands,
heerde@waikato.ac.nz
Which marketing strategies are most effective for introducing new brands? This paper sheds light on this
question by ascribing growth performance to firms’ postlaunch marketing choices. We decompose the
success of a new brand into its ultimate market potential and the rate at which it achieves this potential. To
achieve this aim we formulate a Bayesian dynamic linear model (DLM) of repeat purchase diffusion wherein
growth and market potential are directly linked to the new brand’s long-term advertising, promotion, distribu-
tion, and product strategy. We perform the analysis on 225 new-brand introductions across 22 repeat-purchase
product categories over five years to develop generalized findings about the correlates of new-brand success.
We find that access to distribution breadth plays the greatest role in the success of a new brand, and that
investments in distribution and product innovation lead to greater marginal increases in sales for new brands
than either discounting, feature/display, or advertising. Moreover, distribution interacts with other strategies
to enhance their effectiveness. These findings underscore the utility of extending marketing mix models of
new-brand performance to include product and distribution decisions.
Key words: diffusion; new products; marketing mix; dynamic linear model; empirical generalization
History : This paper was received January 16, 2007, and was with the authors 6 months for 1 revision;
processed by Gary Lilien. Published online in Articles in Advance May 15, 2008.
1. Introduction
Markets are often characterized by extensive new-
brand activity, and the pace of innovation is accel-
erating. For example, 1,521 new consumer packaged
goods (CPG) brands were introduced to the United
States in 2004, double the number of brands intro-
duced in 1997 (Figure 1). Manufacturers use new
brands to drive growth in otherwise stable environ-
ments because innovation is often envisioned as piv-
otal to the success of firms. However, the performance
of new brands varies markedly across their roll-outs.
In CPG markets, only 20% of new brands earn more
than $7.5 million in first year sales, and less than 1%
enjoy revenues in excess of $100 million (Information
Resources, Incorporated (IRI) 2005). Although essen-
tial to firms’ overall performance, few new brands
reach the status of an established brand; the major-
ity eventually fail. The IRI survey shows that failure
rates have reached 55%. The tension arising between
the need to innovate and the low success rate coupled
with innovation begs the question of how to facilitate
the success of new brands.
Perhaps as a result, the growth of new brands
has received substantial interest in the marketing
literature (Hauser et al. 2006). Recent research on
new-brand diffusion has advanced our understand-
ing of how external factors such as economic con-
ditions (Van den Bulte 2000), consumer differences
(Van den Bulte and Joshi 2007), competitive setting
(Steenkamp and Gielens 2003), and product and coun-
try characteristics (Tellis et al. 2003) affect diffusion
of new products across space and time. Moreover, a
number of new-product diffusion studies have incor-
porated internal, manageable factors in the diffusion
process. Specifically, these studies have led to impor-
tant insights into how marketing affects the growth
and/or market potential of durable goods (see Bass
et al. 2000 for a review).
In spite of these advances, prior research has
focused on aspects of the marketing mix in isola-
tion (promotion, product, price, and place), often
used durable goods brands, and typically considered
only one or a few products per study. When vari-
ous aspects of marketing strategy (e.g., advertising
and distribution) are coincidental, considering strate-
gies in isolation can give a misleading picture of
which tools are most conducive to a successful launch
because the effect of one coincidental strategy can
1036
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1037
Figure 1 Number of New CPG Brand Introductions, 1997–2004
736
939 884
1,054 991 973
1,361
1,521
1997 1998 1999 2000 2001 2002 2003 2004
Year
Notes. The figures include entirely new brands or new-brand extensions but
exclude stock-keeping unit (SKU)-level variety introductions. All food, drug
and mass merchandising categories in the U.S. market are included.
Source. Information Resources, Inc. (2005). 2004 New Product Pacesetters.
be mistakenly attributed to another. Accordingly, lit-
tle information exists on the drivers of diffusion for
nondurable goods. In this paper, we shed light on
diffusion in repeat-purchase contexts by offering an
integrated view across the entire marketing mix, and
we afford insights into introduction strategies that
enhance the potential for successful roll-outs.
We advance the literature on new-brand diffusion
in two ways: by conducting an empirical generaliza-
tion pertaining to the efficacy of marketing strategies
in the context of new-brand launch, and by develop-
ing a methodology to achieve these aims.
• First, we explore the effect of various marketing
strategies (advertising spending, feature and display
activity, regular price, discount depth, product line
length, distribution breadth and distribution depth in
unison) on new-brand growth across 225 CPG brands.
Although some diffusion studies link certain elements
of the marketing mix to growth and/or market poten-
tial of a new brand (see Table 1), most previous work
focuses almost exclusively on the role of price and
Table 1 Selected Studies from Diffusion Literature Incorporating Marketing Mix Instruments
Growth Market potential
Price Eliashberg and Jeuland (1986), Parker (1992), Parker and
Gatignon (1994)a, Mesak and Berg (1995), Mesak (1996)
Kalish (1983, 1985), Kalish and Lilien (1986), Kamakura
and Balasubramanian (1988), Horsky (1990), Jain and
Rao (1990), Bass et al. (1994), Mesak and Berg (1995),
Mesak (1996)
This paper This paper
Promotion Lilien et al. (1981)a, Horsky and Simon (1983), Kalish (1985),
Simon and Sebastian (1987), Rao and Yamada (1988)a, Hahn
et al. (1994)a, Parker and Gatignon (1994)a, Mesak (1996)
Dodson and Muller (1978), Mesak (1996)
This paper This paper
Place Mesak (1996) Jones and Ritz (1991), Mesak (1996)
This paper This paper
Product This paper This paper
Notes. The studies listed in the table consider diffusion of durable goods unless marked by an “a” for frequently purchased consumer product
categories. Promotion includes advertising expenditure.
advertising. Much less emphasis has been placed on
distribution and product line, due in part to a paucity
of data. As noted by Muller et al. (2007, p. 72). “Of
the four P’s of the marketing mix, diffusion research
so far has created a sound body of knowledge con-
cerning the effects of price and promotion, yet little
has been done concerning the other two elements:
product and place.” By considering launch strategies
in their entirety, we control for potential correlations
across various marketing instruments, and we can
gauge their relative effect to assess which are most
efficacious.
• Second, we develop a diffusion model for fre-
quently purchased CPG brands that simultaneously
(a) considers the effect of repeat purchases, (b) accom-
modates a variety of potential diffusion trajectories,
(c) separates short-term fluctuations in sales from
long-term changes in brand performance arising from
various marketing strategies (e.g., Mela et al. 1997),
and (d) controls for endogeneity in the marketing mix
and models the role of past performance on mar-
keting spend. We do this by formulating a Bayesian
DLM of repeat purchase diffusion. In this approach,
we model long-term effects by considering the growth
process underpinning a brand’s baseline sales. We
posit that growth in baseline sales follows a diffu-
sion process that is affected by changes in long-term
marketing strategies. These strategies (e.g., distribu-
tion penetration or advertising stock) are linked to
both the rate of growth and the market potential. We
further accommodate short-term perturbations about
this growth process that arise from short-term mar-
keting activity (e.g., weekly discounts).
We find that distribution and product play a greater
role than discounting, feature/display, and advertis-
ing in the sales performance of new brands in spite
of a focus on these factors in the preceding litera-
ture. Overall, we find that access to distribution plays
the greatest role in the success of a new brand. Our
Ataman, Mela, and van Heerde: Building Brands
1038 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
results also show that advertising plays a greater role
in accelerating brand growth than increasing market
potential, and that discounting has a positive effect on
the time to maturity but a negative effect on long-term
market potential. We consider the marginal profits
associated with various marketing launch strategies
and find that distribution has the highest payoff; if the
marginal cost of additional distribution is less than
23% of marginal retail revenue, then it is profitable to
expand distribution. In contrast, on average, adver-
tising is profitable only when its marginal costs are
less than 0.8% of marginal retail revenue. Increasing
product line length is profitable when the marginal
cost of doing so are less than 5% of the marginal retail
revenue.
The rest of the paper is organized as follows. First,
we review the extant literature on repeat-purchase
diffusion models. Next, we outline our modeling
approach and provide a brief overview of the esti-
mation process. After discussing the data, we provide
variable operationalizations and develop expectations
about the role of marketing strategy on new-brand
performance. The results are given next followed by
managerial implications drawn from several simula-
tions. We conclude with some overall thoughts on
this paper, its limitations, and possibilities for future
research.
2. Modeling New-Brand Diffusion in
CPG Categories
Though ubiquitous in marketing, the preponder-
ance of diffusion models has been developed for
durable goods categories. Modeling new-brand diffu-
sion in frequently purchased nondurable goods cate-
gories requires a somewhat different approach given
the existence of repeat purchases, flexibility of diffu-
sion patterns, and separating short-term fluctuations
from long-term performance. We address these issues
subsequently.
First, sales arising from repeat purchases are espe-
cially relevant when considering the diffusion of fre-
quently purchased new CPG brands. In contrast,
traditional models of diffusion consider only the
first purchases of the consumers and use aggre-
gate category- or brand-level adoption sales data.
Parameter estimates of traditional diffusion mod-
els are biased when replacement purchases are not
separated from first-time purchases (Kamakura and
Balasubramanian 1987). To prevent such biases and
provide improved sales forecasts, several diffusion
model alternatives with replacement purchases have
been developed for durable goods (see Ratchford
et al. 2000 for a review) as well as non-durable
goods (Lilien et al. 1981, Rao and Yamada 1988, Hahn
et al. 1994). Given our research context, we follow
this stream of repeat-purchase modeling and extend
the earlier work by addressing the second challenge
(flexible diffusion patterns) and the third challenge
(short- versus long-term fluctuations), as we discuss
next.
Second, the sales trajectory of repeat-purchase
goods can follow a litany of diffusion patterns. Ear-
lier applications of repeat-purchase diffusion mod-
els link growth to marketing activity, allowing for
some degree of flexibility, but assume a constant
market potential. The assumption of constant market
potential implies a relatively quick increase in sales
followed by flatness once the brand’s market poten-
tial is reached. However, when actual sales follow a
diffusion pattern with slow take-off, perhaps due to
limited initial availability, repeat-purchase diffusion
models with constant market potential are ill-suited to
capture this phenomenon. Moreover, constant market
potential precludes sales declines following the initial
success of a new brand. Such declines can arise from
cuts in marketing support. A flexible market potential
definition, such as the one proposed in this research,
overcomes these concerns.
Third, short-term fluctuations in sales might mask
the true long-term performance of the new brand (Mela
et al. 1997). Previous applications of repeat-purchase
diffusion models for nondurable goods calibrate the
diffusion model using monthly or quarterly data for
products with relatively smooth sales patterns, such
as therapeutic drugs (e.g., Rao and Yamada 1988,
Hahn et al. 1994). Such sales data do not often exhibit
short-term fluctuations given that these may be aggre-
gated out over the data interval, particularly as short-
term marketing activity is uncommon and seasonal
patterns are not strong. However, for frequently pur-
chased CPG brands, data sampling rate is typically
high, short-term oriented marketing activity is com-
mon, and seasonality assumes greater importance.
Therefore, the series are far from being smooth. Ear-
lier work in the area recommends that the data be
smoothed prior to estimation to eliminate short-term
fluctuations (Lilien et al. 1981). Such smoothing pro-
cedures will bias the parameters, especially when the
variables that build market potential are correlated
with the variables that create the short-term fluctu-
ations in sales. We propose a model that separates
short-term fluctuations from long-term performance
during estimation.
3. Modeling Approach
3.1. General Approach
Consistent with the foregoing discussion, we seek
to determine (1) the rate of new-brand growth (and
the attendant implications for the time to reach peak
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1039
sales) and (2) the new brand’s ultimate market poten-
tial. Accordingly, we predicate our model formula-
tion on the marketing literature on diffusion (Mahajan
et al. 1990). Given our emphasis on repeat-purchase
goods, our modeling approach closely parallels that
of Lilien et al. (1981), Hahn et al. (1994), and Rao
and Yamada (1988) but with several key extensions:
(1) our model is cast in a dynamic Bayesian setting
to accommodate greater modeling flexibility and sta-
tistical efficiency; (2) we link both growth and market
potential to marketing strategy, given the central aims
of our paper; (3) we incorporate performance feed-
back to control for the role of past sales on future
marketing spending; (4) we consider potential com-
petitive effects; and (5) we control for endogeneity
of price and the other marketing instruments. Like
Lilien, Rao, and Kalish (henceforth LRK), we assume
two market segments drive the base demand for a
new brand—those generated from new purchases and
those from retention.
To formalize this notion, we begin by positing a
linear model of brand sales, given by
Salest=t+X
t+t(1)
where Xtis a matrix of regressors containing short-
term-oriented marketing activity that capture short-
term changes in sales around the brand’s growth
trajectory and a control for seasonality. tis a parame-
ter that captures the long-term growth in brand sales,
which is governed by the diffusion process noted
above. Because the Xtinclude weekly discounts, fea-
ture and display, and seasonality, tcan be inter-
preted as baseline sales (which we again presume to
evolve following a diffusion process). The distinction
between long-term and short-term marketing effects
follows Jedidi et al. (1999), as short-term effects are
captured by the effect of a given week’s marketing
activity, Xt, such as a promotion, and long-term effects
are captured by the effect of repeated exposures to
marketing, Zt, on the time-varying parameter t(to
be discussed later). We assume t∼N0V.
We seek to capture nonlinearity in the baseline sales
over time using a diffusion-modeling approach. Fol-
lowing LRK we assume1
t=t−1+−t−1+t(2)
1The diffusion model as developed by LRK applies to pharmaceu-
tical detailing and can be expressed as follows:
t=t−1+−t−1+t−1−t−2 −t−1−t−1+t
where is the innovation parameter, is the imitation parame-
ter, and is the effect of competition. We modify this model in
two key respects to make it suitable to the packaged goods con-
text we consider. First, we specify word-of-mouth effects to be neg-
ligible (≈0). This specification is consistent with the findings
of Hardie et al. (1998), who find no word-of-mouth effects across 19
different CPG data sets. Given (1) high variability in weekly sales
arising from weekly promotions, (2) the fact that most products
where tindicates the base sales for the brand at
time tand is the base market potential. The first
term captures retention effects, as a certain fraction, ,
of the past period’s base t−1will continue to buy on
the subsequent purchase occasions. The second term
captures the attraction of the remaining potential cus-
tomers, as a certain fraction, , of the remaining mar-
ket (given by the deviation between the total market
potential and past base sales t−1) will buy on the
subsequent purchase occasion. The second term there-
fore represents the diffusion process governing the
long-term evolution of baseline sales potential. The
parameters and have an additional interpreta-
tion, as is reflective of the time of adjustment to the
market potential while reflects that potential. We
assume t∼N0W.
Figure 2 depicts growth trajectories in baseline sales
for various parameterizations of Equation (2). As sug-
gested by this figure, baseline sales can grow quickly
early in a brands’ life cycle and then asymptote as the
brand diffuses through the population. This asymp-
tote is given by ·/1−+ if 0 <1−+ < 1.
As increases, sales reach their asymptote faster. As
increases, the sales potential grows. All else being
equal, faster growth and greater potential lead to
higher total sales.
Following LRK we allow the growth parameter to
vary over time →tand specify this parameter as
a function of the long-term marketing strategy used
by the firm that introduces the brand, t≡Z
t. For
example, advertising stock might lead to increased
awareness, thus accelerating trial rates. Like Xie et al.
(1997), who consider the durables context, we allow
the market potential to change over time →t.
In our application, though, we posit market potential
to be a function of the long-term marketing strategy
of a brand, t≡Z
t. For example, advertising stock
might attract new buyers to the brand by changing
their preferences. After substituting the new growth
and market potential definitions in Equation (2) we
obtain
t=t−1+Z
tZ
t−t−1+t(3)
are not consumed the same week of purchase (e.g., detergent has
an eight-week purchase cycle), and (3) limited occasion for social
interactions within a week, word-of-mouth effects are likely min-
imal. In contrast, we note that the LRK model applied directly to
CPG implies that incremental weekly sales drive word of mouth
and that these effects last one week—which are strong assump-
tions in our context. We tested the assumption of no word-of-mouth
effects using a classical approach and find that the fit of the model
with word-of-mouth effects is not significantly better than that
of the model without word-of-mouth effects (likelihood ratio test
statistic =789, p=0444). Taken together, these arguments indi-
cate the lack of word-of-mouth effects in frequently purchased CPG
markets. Second, we capture the effect of competition via the
baseline repeat parameter, =1−; that is,
t=t−1+−t−1−t−1+t≡t−1+−t−1+t
Ataman, Mela, and van Heerde: Building Brands
1040 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
Figure 2 Growth Trajectory Illustrations
(a)
(b)
0 5 10 15 20
0
0.5
1.0
1.5
2.0
2.5
3.0
Week
Sales
0
0.5
1.0
1.5
2.0
2.5
3.0
Sales
μ0 = 3.75 γ = 0.2
μ0 = 3.12 γ = 0.4
μ0 = 2.92 γ = 0.6
μ0 = 2.81 γ = 0.8
0 5 10 15 20
Week
μ0 = 3 γ = 0.2
μ0 = 3 γ = 0.4
μ0 = 3 γ = 0.6
μ0 = 3 γ = 0.8
Notes. Figures assume =09. Note that it is also possible to accommodate
sigmoidal sales trajectories when market potential varies over time t.
where is the repeat-purchase rate, which we esti-
mate without imposing any restrictions. is the
vector of growth parameters, and is the vector
of market potential parameters associated with each
marketing variable.
The Ztin Equation (3) play a long-term role in the
trajectory of brand growth as a result of the carry-
over implied by the lagged tin Equation (3). Condi-
tioned on Ztconstant at Zt=Z, Equation (3) becomes
a Koyck model whose carryover is given by −Z
(to see this, note (3) is equivalent to t=t−1+
Z·Z+twhere =−Z). The model in
Equation (3) therefore implies the rate of innovation
growth is affected by and Z, with lower values
of −Zimplying faster adjustment to the long-
term sales level, given by Z·Z/1−+Z)if0<
1−+Z < 1. The steady-state sales equation fur-
ther implies that an increase in Zyields an increase
in the long-term sales level of a brand when is
positive.
The interaction of the growth and market potential
parameters admit innumerable paths for brand sales.
For example, as −Zapproaches 0 and Z·Z
becomes sufficiently large, sales will adjust immedi-
ately to a high mean but also fall again quickly when
marketing support is withdrawn. Conversely, a low
value for and a low value for imply that the brand
will neither generate large sales nor increase sales
quickly (see Figure 2). The model further allows dif-
fusion speed and market potential to move in oppo-
site directions, as parameters and are estimated
freely: Marketing activities that increase the market
potential might slow the speed to reach that higher
market potential. In sum, Equation (3) provides a flex-
ible model of baseline sales growth, which can change
in response to the marketing mix.
The model defined in Equations (1) and (3) belongs
to a family of Bayesian time-series models known as
the DLMs (West and Harrison 1997). In the next sec-
tion, we discuss model specification and provide a
brief overview of the estimation procedure.
3.2. Model Specification
Our goal is to explain how marketing mix activity
generates growth and builds market potential for a
new brand. We achieve this by estimating the trans-
fer function DLM developed in the previous section
(see Bass et al. 2007; Van Heerde et al. 2007, 2004 for
other DLM applications in marketing). The observa-
tion equation, which separates short-term fluctuations
from long-term sales, is specified as a linear sales
model,
Salesjt =jt +
X
jtj+S
jt(4)
where Salesjt is the vector of sales of brand jat
time t, and
Xjt includes variables that might gener-
ate short-term fluctuations in sales. We standardize
all variables within brands and indicate this with a
superscripted bar. jt is the baseline sales for brand j
and evolves over time, following the repeat-purchase
diffusion process as specified in the following evolu-
tion equation:
jt =jjt−1+
Z
jt
Z
jt−jt−1+0jt(5)
Z
jt is a vector of standardized marketing strategy
variables posited to affect diffusion. The standard-
ization ensures we can pool different units across
categories and control for unobserved time-invariant
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1041
brand effects.2The parameter jcaptures the brand-
specific repeat-purchase rate, whereas and cap-
ture growth and market potential due to marketing
effort, respectively.
The observation equation and the evolution equa-
tion specified in (4) and (5) can be compactly writ-
ten as
Yjt =Fjtjt +Xjtj+jt(6)
jt =Gjtjt−1+hjt +jt(7)
where Yjt is the standardized sales of brand jin week t,
and Fjt =1. Xjt is the matrix of standardized regres-
sors that create short-term fluctuations in sales. We
assume jt ∼N0V, where Vis the observation
equation error variance. The time-varying parameter
vector jt =jt evolves as described in Equation (7).
Rearranging the terms in Equation (5) gives Gjt =j−
Z
jt. Then the second term on the right-hand side of
Equation (7) is hjt =
Z
jt
Z
jt. The stochastic term
jt are distributed N0W, where Wis the evolution
equation error variance.
3.3. Marketing Mix Endogeneity, Performance
Feedback, and Competition
We specify an additional equation for each market-
ing mix instrument to control for endogeneity in the
marketing mix, partial out the role of past perfor-
mance, and control for competitive effects. To address
endogeneity, we follow an approach analogous to
instrumental variables wherein lagged endogenous
variables serve as instruments. Moreover, we allow
for correlation between the demand-side error term
and the supply-side error term to account for com-
mon unobserved shocks in the system.
We control for performance feedback (i.e., sales gains
lead to increased marketing) by including lagged
national sales in each marketing equation. Past suc-
cess in distribution—and hence sales, for example—
might lead to increased distribution in subsequent
periods. Given that firms might also react to changes
in the competitive landscape, we also consider com-
petitive marketing activity in the category.
For each marketing mix instrument the foregoing
discussion results in a time-varying mean DLM,
Zij t =ij t +Z
ij t (8)
ij t =0ij +1ij ij t −1+2ij Salesjt−1
+
K
k=1
k+2ijCZkct−1+ijt (9)
2Standardization does not affect the ease of managerial interpreta-
tion because we can always convert standardized values back to
their original scale.
where Zijt is the ith marketing mix instrument of
brand jin week t, and CZkct−1is the sales-weighted
average of the kth marketing mix instrument for com-
petitors in category c(j∈c in week t−1.3Equa-
tion (8) posits that observed marketing spending is a
manifestation of an underlying latent national strat-
egy (ij t ), and deviations from this strategy arise from
random shocks. Equation (9) defines the evolution of
this latent strategy as a function of its past value, the
past performance of the focal brand, and past mar-
keting activities of competition. The parameter 1ij is
associated with the lagged national strategy and cap-
tures inertia in the marketing spending. Salesjt−1is the
focal brand’s lagged standardized national sales. Thus
the parameter 2ij captures own-performance feed-
back effect for the marketing mix instrument i. Finally,
the parameters k+2ij capture correlations between
the focal brand’s marketing and that of its competi-
tors. The superscripted bar indicates that the variable
is standardized.
3.4. Estimation
We estimate Equations (8) and (9) together with Equa-
tions (4) and (5) and let error terms S
jt and Z
ij t
be correlated to account for common unobserved
shocks in the observation equations.4We place nor-
mal priors on all parameters of the observation equa-
tion, the evolution equation, and the marketing mix
equations. The evolution equation error covariance
matrix is assumed to be diagonal, and we place
an inverse Gamma prior on its diagonal elements.
As we allow for correlation between the observa-
tion equation error terms and the marketing mix
equation error terms, the associated error covari-
ance matrix is full. Therefore, we place an inverse
Wishart prior. Given these priors, the estimation is
carried out using DLM updating within a Gibbs sam-
pler. Conditional on ,,V,W,ht, and Gt, the
time-varying intercepts are obtained via the forward-
filtering, backward-sampling procedure (Carter and
Kohn 1994, Frühwirth-Schnatter 1994). The param-
eters of the baseline sales evolution are estimated
using a random walk Metropolis-Hastings algorithm,
because the evolution equation is nonlinear in param-
eters. The details of the sampling chain are provided
in the appendix.
3The kth marketing mix instrument of the composite competitor in
a given category is computed as the weighted average of marketing
mix instruments of top five brands in that product category. We
use average market share in the most recent 13 weeks (t−1t −
2t−13) as rolling weights in the aggregation.
4We estimated an alternative diagonal error correlation. The log
Bayes factor (BF) (West and Harrison 1997) favored the full
matrix specification over the diagonal matrix specification log BF =
18642.
Ataman, Mela, and van Heerde: Building Brands
1042 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
4. Data and Variables
4.1. Data
We calibrate our model on a novel data set pro-
vided by IRI (France). The data cover more than five
years (1/1/1999 to 2/1/2004) of weekly SKU store-
level scanner data for 25 product categories sold in a
national sample of 560 stores operated by 21 differ-
ent chains. We also use matching monthly brand-level
advertising data provided by TNS Media Intelligence
(France).
Data are aggregated from the SKU store level to
national brand level following the procedures out-
lined in Christen et al. (1997) to avoid any biases
due to aggregation. Because the sales model in Equa-
tion (7) is linear, we first aggregated the data from
SKU-store to brand-store level in a linear fashion (dis-
cussed in §4.2). Using lagged all commodity volume
(ACV), we then calculated an ACV-weighted average
of brand-store level independent variables to obtain
national-brand level data.
Between January 1, 1999 and February 1, 2004,
we observe 365 new national brand introductions in
25 product categories. Of these new brands, 55 fail
within the mentioned time window. For a single cat-
egory, the number of new-brand introductions varies
between 5 and 38, with an average of 17 brands. On
average, we observe the first 152 weeks of the new
brand’s life cycle, with a minimum of 15 weeks and a
maximum of 264 weeks. We select brands with at least
two years of data, regardless of whether they succeed
or fail, which leaves us with 225 new-brand introduc-
tions in 22 categories (as opposed to line extensions
within existing brands).5See Table 2 for descriptive
statistics.
4.2. Variables
The selection of variables is linked to our goal of con-
trasting the relative efficacy of the marketing mix in
generating new-brand growth. The variables consid-
ered represent the conjunction of those suggested by
theory and those available in the data. In this section
we detail each variable and its anticipated effect on
the diffusion of new brands. We first discuss the vari-
ables in the observation, or sales, equation and then
consider the variables in the growth equation.
4.2.1. Sales Equation Variables. The dependent
variable in Equation (1), Salesjt, is the sales volume of
a new brand, which is calculated as the sum of sales
across all stores in a given week. We posit the sales
to be affected by a number of short-term variables,
including brand-level discount depth Discjt, feature
5Of these 225 new brands, 8 are brand extensions and 217 are
entirely new brands. The substantive results are robust to the exclu-
sion of these eight brands.
or display support FoDjt, and average weekly tem-
perature (Tempt. Thus, Xjt =DiscjtFoDjtTempt.We
measure the SKU store-level depth of promotion by
one minus the ratio of the actual price to the regu-
lar price. The brand store-level promotion depth vari-
able is chosen as the maximum discount depth across
SKUs (e.g., Mela et al. 1997), and the national brand
level variable is calculated as the store ACV-weighted
average of the brand store-level data. The brand store-
level feature and display variable take the value of
one if at least one SKU from the brand’s product line
is on promotion in a given week. The national brand-
level averages for these variables are calculated across
stores in a linear fashion using lagged store ACV
as weights. We expect discounts and feature/display
intensity to have a positive short-term effect on sales,
while temperature affords a parsimonious control for
seasonality.
4.2.2. Evolution Equation Variables. We now dis-
cuss the operationalization of the marketing mix vari-
ables in Zjt in the evolution Equation (5), along
with our expectations about the role they play in
growth and market potential. Table 3 summarizes our
expectations.
Price. We define the price of a brand as the regular
price in a given store week. Consistent with previous
studies (e.g., Mela et al. 1997), we select the minimum
regular price per 1,000 volume units across SKUs of a
brand. The national brand-level average price is cal-
culated across stores in a linear fashion, using lagged
store ACV as weights.
Previous research provides unequivocal evidence
that regular price reductions influence the growth of
new-brand sales (Parker and Gatignon 1994, Parker
1992). However, there is a lack of consensus on
whether price also affects the market potential. Bass
et al. (1994) and Kamakura and Balasubramanian
(1987, 1988) find no impact from price, whereas
Mesak and Berg (1995) and Kalish and Lilien (1986)
report negative impact. However, like Eliashberg and
Jeuland (1986), we expect that lower prices stimu-
late additional demand as the brand matures. More-
over, the brand can achieve high market-penetration
rate rather quickly because lower initial prices moti-
vate the potential buyers to make the purchase earlier
(Bass and Bultez 1982). In sum, we expect lower prices
to facilitate growth and increase market potential for
a new brand.
Discounts. Discounts encourage trial purchases for
the first-time buyers. They reduce search costs for
the consumer, generate awareness, and increase the
likelihood of adoption (Kalish 1985). Anderson and
Simester (2004) find that deep discounts also increase
repeat rates of first time buyers; thus, discounts accel-
erate growth. However, the effect of discounting on
market potential is not clear. Discounting can build
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1043
Table 2 Descriptive Statistics
Sales volume Sales volume
No. of new No. of brands new brand (×1000) Sales value Advertising Price Distribution Product Distribution Discount Feature/display
Category brands in category (×1000) category mean (×1000) (×1000) (per 1,000) breadth (%) line length depth (%) depth (%) /100
Bath products
M 25 326 500 639362975013323270831184
SD 119225639108211287748190424129
Beer
M 36 961 1030422639 24932753945231022183
SD 2330226826544024782298110514120
Butter
M 12 382 2183620299197974095097302027211
SD 2182566099209466982019920
1024196
Cereals
M 7 118 647124540 401827614639232562
SD 119401229085856—305243144454
Chips
M 5 86 414306144114562— 2399524228135
SD 8011816199530412—0919184241186
Coffee
M 16 306 93214445 1212— 9019401570340
SD 1432730901756—29285608108315
Feminine needs
M 3 65 490 1808 7215— 71718726170947
SD 30142666613—27714209020649
Frozen pizza
M 3 72 20029186111514332457118686020136
SD 10131449219988—352978250930
Ice cream
M 19 211 1046324651 8052235082601323115
SD 152117615911469—26111
76071995
Mayonnaise
M 9 234 2486 8799 250363910783221916162
SD 572847731499388445179191409121
Mineral water
M 3 143 74199 651982353103283411228
SD 1116565539—69125110811333
Paper towel
M 2 66 252311 9917 — 2690261014215185
SD 667183093—13802002510130
Pasta
M 16 334 656625628931174324621732172
SD 167511730031248133430520826106
Shampoo
M 9 172 1211911430201378909822861151070
SD 1950228952313699584726666130649
Shaving cream
M 4 51 1384 5294 2138— 1047719390982
SD 1307126523061—86621115
0663
Soup
M 21 333 1643522443 58483123389661918124
SD 37141181350123585072016273101496
Tea
M 8 178 138 1092 17914364054442419140
SD 110481315996726963191120134
Toothpaste
M 1 84 03 5220538 — 877267101805100
SD — 17209——— ———— —
Water
M 14 189 587884293869363422035260799
SD 54234213101455646204170603126
Window cleaner
M 1 54 988 7520139— 09301012211109
SD — 19903——— ———— —
Ataman, Mela, and van Heerde: Building Brands
1044 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
Table 2 (cont’d.)
Sales volume Sales volume
No. of new No. of brands new brand (×1000) Sales value Advertising Price Distribution Product Distribution Discount Feature/display
Category brands in categor y (×1000) category mean (×1000) (×1000) (per 1,000) breadth (%) line length depth (%) depth (%) /100
Yogurt
M 8 226 5341107620 2797 1320474824091370
SD 84673722913637—176208030963
Yogurt drink
M 3 37 1777150435 7514— 3610721580733
SD 124811410764954—224610190113
Notes. M, mean; SD, standard deviation of average marketing support across all brands. The mean and standard deviation of advertising, discount depth, and
feature/display are calculated using nonzero observations.
customer loyalty through rewards and thus might
help the brand build baseline sales through increased
familiarity and experience, or simply through pur-
chase reinforcement or habit persistence (Ailawadi
et al. 2007, Keane 1997, Pauwels et al. 2002, Slotegraaf
and Pauwels 2006). On the other hand, discounting
can also have a negative long-term impact because it
could erode brand equity (Ataman et al. 2006, Jedidi
et al. 1999).
Feature/Display. We also consider the role of non-
price promotions in the diffusion of a new brand.
Feature promotions, retail displays, and other in-store
communication tools are manufacturer-retailer joint
advertising efforts. Such nonprice promotions make
the new brand salient and promote it to the shop-
per traffic (Gatignon and Anderson 2002). In a sense,
they work in the same way that advertising does.
Therefore, we expect features and displays to facili-
tate growth and increase market potential at the same
time.
Advertising. We construct the weekly advertising
support variable from the available monthly advertis-
ing expenditure data by dividing the monthly figures
by the number of days in a month, and then totaling
across days for the corresponding weeks (Jedidi et al.
1999). This enables us to model marketing response
at the highest frequency available in the data, avoid-
ing the potential for data aggregation bias (Tellis and
Franses 2006).
A number of studies have investigated the role of
advertising in new-product diffusion (e.g., Dodson
and Muller 1978, Horsky and Simon 1983, Kalish
1985, Simon and Sebastian 1987). National brand-
oriented advertising, which serves information and
Table 3 Summary of Expectations
Growth Market potential
Advertising ++
Regular price −−
Discounting +0
Feature and display ++
Distribution breadth ++ ++
Distribution depth ++ ++
Line length ++ ++
persuasion functions simultaneously in the context
of new brands, produces high awareness levels, dif-
ferentiates brands, and builds brand equity (Aaker
1996). Thus, it helps build market potential. Elberse
and Eliashberg (2003) find that advertising is crucial
for new-brand performance, especially in the early
stages of introduction. Moreover Lodish et al. (1995)
find that advertising works better when brands are
new, implying a positive growth effect.
Distribution breadth. We use ACV-weighted dis-
tribution as a measure of distribution breadth
(Bronnenberg et al. 2000). ACV weights a brand’s dis-
tribution by the total dollar volume sold through a
particular store, giving more distribution credit to a
large dollar volume store than it does to a small dollar
volume store.
Early work on new-product diffusion tended to
overlook the role distribution plays in building new
brands. These studies typically explain the success of
a new brand by factors such as advertising or price,
and they assume that the brand is always available
to the consumers. A notable exception is the study by
Jones and Ritz (1991), where the authors note that a
new brand cannot build sales if the consumers can-
not find a store in which they can purchase it. Recent
research on new products devotes more attention to
distribution decisions and explains realized demand
conditional on product availability. Such an approach
is appropriate especially in competitive environments
where customers visit the retail stores and decide
what to buy based on which brands are available
(Krider et al. 2005). Taking this view Bronnenberg
et al. (2000) show that in new repeat-purchase prod-
uct categories market shares are strongly influenced
by retailer distribution decisions. Other studies con-
firm that distribution is a critical factor influenc-
ing new-product performance (Elberse and Eliashberg
2003, Gatignon and Anderson 2002, Neelamegham
and Chintagunta 1999). In light of these findings,
we expect distribution to be an important element
in explaining the new brand’s growth and market
potential.
Distribution depth. We measure distribution depth
as the number of SKUs a brand offers in the category
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1045
in a given store relative to the total number of SKUs in
that category in that store. This measure reflects how
many different SKUs of a particular brand are carried,
on average, at each point of ACV distribution. We cal-
culate the distribution variables at the store level and
then calculate national averages.
Any marketing activity that spreads information in
proportion to the number of products in the market,
such as self-advertising by just being on a supermar-
ket shelf, could generate awareness for a new brand
(Eliashberg and Jeuland 1986). Therefore, we expect
distribution depth to facilitate growth and build mar-
ket potential.6
Line length. We measure the brand line length by
the number of SKUs a brand offers in a given week.
Our discussion about the role that brand line length
plays in the diffusion process of a new brand is
rather tentative because theoretical and empirical evi-
dence on this issue is virtually nonexistent. We argue
that, holding all else constant, more SKUs provide
assortment and increase the probability of trying an
item from the new brand’s line. Also, having more
alternatives might serve more segments. Therefore,
we expect line length to increase market potential
and facilitate growth. Because the marginal change in
baseline arising from the addition of new SKUs might
decrease, we specify log-transformed line length in
the model.7
Relative effects. As indicated in Table 1, thus far no
research has incorporated all marketing mix instru-
ments into a single diffusion framework, let alone
into a repeat-purchase diffusion framework for CPG
categories. Therefore, the relative importance of mar-
keting instruments in building new CPG brands is
undocumented. However, relative effects sizes are of
central interest to managers because they point out
areas in which it might be more desirable to allocate
marketing funds. We argue that line length and distri-
bution (breadth and depth) should assume the great-
est importance (denoted by ++ in Table 3) because
(1) a consumer, given her reluctance to shop across
stores or markets, will not adopt a brand if it is not
available in the stores she visits (Bronnenberg and
6Distribution depth and line length capture different aspects of
brand strategy, as reflected in a relatively modest correlation of
0.35. The correlation is modest because line length is the number
of products a brand offers at the national level, while distribution
depth is the fraction of a store’s category assortment that belong
to the brand, which is thus measured at the store level (and next
averaged across stores). As a consequence, line length is a decision
variable that is under direct control for a manufacturer, whereas
distribution depth depends to a large extent on the willingness of
the retailer to carry the brand’s products. Hence, a high line length
does not necessarily coincide with a deep distribution.
7The log BF (927.5) favored the model with decreasing returns over
the model with constant return.
Mela 2004, Jones and Ritz 1991), and (2) said con-
sumer will also be unlikely to purchase goods if there
are not variants or items that match her needs. Yet
availability and alternative options require awareness,
hence advertising and feature/display should be in
the second tier of critical elements of the diffusion
process.
5. Results
We estimate the DLM specified above using a Gibbs
sampler and run the sampling chain for 30,000 iter-
ations (15,000 for burn-in and 15,000 for sampling
with a thinning of 10). The repeat-purchase diffu-
sion model with flexible growth and market poten-
tial specification, coupled with the ability of the DLM
methodology to accommodate potential nonstationar-
ity in brand launch, provides excellent fit to the data
(see Figure 3). Across 225 brands we analyze in the
paper, the median correlation between actual and pre-
dicted sales is 0.88.
For all 225 brands we consider three sets of param-
eters: (1) the short-term marketing effects () on sales
model specified in Equation (4); (2) the long-term
marketing strategy effects on growth () and mar-
ket potential (), as well as the repeat-purchase rate
Figure 3 Actual vs. Predicted Sales of Two Selected Brands
0 20 40 60 80 100 120 140 160 180 200
–2.5
–2.0
–1.5
–1.0
–0.5
0
0.5
1.0
1.5
2.0
Week
0 20 40 60 80 100 120 140 160 180 200
Week
Brand A
Brand B
Sales (standardized)
–2.5
–3.5
–2.0
–3.0
–1.5
–1.0
–0.5
0
0.5
1.0
1.5
Sales (standardized)
Actual
Predicted
Confidence interval
Ataman, Mela, and van Heerde: Building Brands
1046 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
parameter () in the baseline sales evolution model
as shown in Equation (5); and (3) the marketing mix
inertia, performance feedback, and cross-marketing
mix parameters () in the marketing mix endogeneity
model specified in Equation (9). We discuss each set
of parameters in sequence.
5.1. The Sales Model
Table 4 shows the inverse-variance weighted aver-
age (to afford more weight to more reliable esti-
mates) of discounting, feature/display, and average
weekly temperature estimates at the category level.
Both discounting and feature/display parameter esti-
mates exhibit face validity because each stimulates
same-week sales (average estimates across all brands
are 0.03 and 0.05, respectively). The 90% posterior
confidence interval of the average weekly temper-
ature coefficient typically excludes zero for brands
Table 4 Parameter Estimates (Sales Model)
Category Discounting Feature/display Temperature Repeat rate
Observation equation parametersa
All categories 003 005 −000 094
Bath products 004 014 −000 089
Beer 002 009 000 086
Butter 000 002 −001 094
Cereals 009 −001 −001 098
Chips 006 000 −002 092
Coffee 004 004 −000 085
Feminine needs 005 002 000 098
Frozen pizza 022 −008 −001 092
Ice cream 005 010 000 095
Mayonnaise −000 011 −000 091
Mineral water 003 001 000 099
Paper towel 024 005 −000 091
Pasta 000 004 −000 091
Shampoo 004 011 000 094
Shaving cream 003 004 −001 097
Soup 003 006 −001 094
Tea 0 00 008 −001 095
Toothpaste 003 001 −001 102
Water 000 000 −000 100
Window cleaner 003 010 −000 101
Yogurt 003 005 −001 098
Yogurt drink 001 003 −001 097
Growth Market potential
Marketing activity Median 5th and 95th percentile Median 5th and 95th percentile
Growth and market potential parametersb
Constant 00948 00868 01035 01008 00760 01288
Advertising 00077 00003 00153 00275c−00036 00609
Regular price −00110 −00147 −00074 −00793 −01098 −00488
Discounting 00148 00116 00183 −00323 −00515 −00102
Feature and display −00088 −00106 −00071 02971 02517 03517
Distribution breadth 00231 00194 00268 07951 07418 08507
Distribution depth −00003 −00041 00037 01344 01056 01698
Line length (log) 00089 00049 00132 00998 00668 01312
Notes. (a) Variance-weighted average of median estimates across brands. (b) Bold indicates that 90% posterior confidence interval
excludes zero. (c) The market potential effect of advertising crosses zero at 92nd percentile.
from product categories that are expected to exhibit
seasonal patterns (e.g., soup and ice cream), whereas
the coefficient is negligible for others.
5.2. The Baseline Sales Evolution Model
Of central interest to this research are the estimates
on the evolution of baseline sales (t), including
(1) repeat-purchase effects , how marketing mix
instruments correlate to sales growth () for new
brands; and (2) the role these instruments play in
the market potential () for a new brand. Table 4
indicates that increases in advertising support, dis-
tribution breadth, line length, and discount correlate
with faster growth for new brands, whereas increases
in regular prices inhibit the diffusion process. These
findings are in line with the expectations. The effect
of distribution depth on growth is negligible. Sur-
prisingly, we find that feature and display intensity
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1047
slows diffusion of new brands, although the effect is
quite small. When combined with positive short-term
effects and the large effect of feature/display on mar-
ket potential, the net effect is positive (as we show in
the subsequent section).
Table 4 further reveals that feature and display
activity, brand line length, distribution breadth, and
distribution depth correlate positively with market
potential for new brands, whereas the 80% posterior
predictive interval for advertising excludes zero (tan-
tamount to a one-sided p-value of 10% in classical
statistics). As expected, low prices are associated with
higher market potential. Consistent with the literature
on the long-term effect of discounts, the effect of dis-
counting on market potential is negative (Mela et al.
1997). Note the dual role of discounts in leading to
faster growth but lower market potential.
Across the 225 brands, the repeat-purchase param-
eters, , range between 0.83 (25th percentile) and 0.98
(75th percentile), with a median of 0.94. The varia-
tion of repeat purchase parameter estimates across
product categories does not reveal major differences.
This median repeat-purchase rate across all brands
suggests that for most brands, 90% of the long-term
sales effect for new brands materializes within the
first 52 weeks (Leone 1995). To our knowledge, this is
the first study to conduct an empirical generalization
of time to peak sales for new packaged goods brands.
5.3. Relative Effect Sizes
The foregoing discussion reveals that marketing strat-
egy plays a role in the diffusion of new brands but
affords little insight into which strategies explain the
greatest amount of variation in the sales performance
of new brands. Accordingly, we consider the relative
Figure 4 Relative Effects Across Marketing Mix Instruments
31%
20%
15%
12% 12% 10%
0%
54%
2%
5% 7%
20%
2%
9%
Distribution
breadth
Discounting Regular price Line length (log) Feature/display Advertising Distribution depth
Growth
Market potential
Note. The bars represent the size of the instrument’s absolute parameter estimate divided by the sum of the absolute parameter estimates.
effect sizes of the marketing mix variables by comput-
ing the ratio of (1) the standardized coefficient for a
given marketing mix instrument to (2) the sum of all
standardized marketing mix coefficients. In the cal-
culation, we use the absolute values of the standard-
ized coefficients for the growth and market potential
parameters, respectively. Figure 4 presents the relative
effects of the marketing mix instruments.
Figure 4 makes it apparent that distribution breadth
is the single most important marketing mix instru-
ment in generating growth (relative effect of 31%) and
building market potential (relative effect of 54%) for
a new brand. Although the result is not altogether
surprising (a brand cannot have sales without dis-
tribution), the precise effect of size relative to other
variables is less obvious as (1) the effect of distribu-
tion exceeds all other strategies combined in generating
growth; and (2) it is also the case that a brand can-
not have sales without a product line, yet this effect
is not as considerable. Distribution breadth and depth
assume greater importance in building market poten-
tial (jointly 63%) than accelerating growth (jointly
31%). After distribution, discounting has the second-
largest impact on growth (20%). Feature and display
have the second-largest effect on market potential
(20%), which implies that their short-term effect on
weekly sales is supplemented by their ability to build
long-run demand for new brands.
5.4. Marketing Mix Models
Now we briefly summarize the results of the mar-
keting mix instrument equations presented in Equa-
tions (8) and (9). First, we find that past sales typically
has little effect on current marketing. A notable excep-
tion is the effect of past sales on distribution breadth.
Ataman, Mela, and van Heerde: Building Brands
1048 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
In 57% of all cases, we observe that past sales have
a positive effect on future distribution. Second, the
coefficients for the inertial effect of past marketing
range from 0.39 (for feature/display) to 0.80 (for dis-
tribution breadth), indicating the largest effect again
arises from distribution breath. Collectively, these two
results underscore the importance of controlling for
performance feedback and amplify our finding that
distribution plays a major role in brand building.
Finally, we find that competitor effects are pre-
dominantly zero for the marketing mix instruments.
Steenkamp et al. (2005) observe a similar result in the
context of advertising and pricing for mature brands,
and Pauwels (2007) observes the same in discounting
and feature/display. We extend this finding across the
marketing mix.
5.5. Model Comparison
Our analysis presumes that (1) baseline sales fol-
low a dynamic growth process, and (2) this pro-
cess is linked to marketing strategy. To test the
first assumption, we contrast our full model (M0)
to one wherein no dynamics are exhibited in base-
line growth (M1), that is jt =
Z
jt+0jt. To test
the second assumption, we contrast our full model
(M0) with one wherein observed growth is indepen-
dent of long-term marketing strategy (M2), that is
jt =jjt−1+−jt−1+0jt. Table 5 indicates the
full model outperforms these benchmarks on the log
BF for one-step-ahead forecasts.
6. Managerial Implications
We next consider the ramifications of our analysis
for new-brand launch marketing strategies. As a pre-
lude, we note limits inherent in the archival data
analysis that we propose, namely that parameter esti-
mates might not be invariant to our policy simula-
tions. That said, in the context of a dynamic problem
with many agents, states, and controls, the imposi-
tion of assumptions to identify a more structural solu-
tion may induce more problems (dimensionality of
the state-space, restrictive assumptions, etc.) than it
redresses.
6.1. Long-Term Marketing Mix Elasticities for
New Brands
Procedure. Using our model, one can assess how
marketing strategies affect brands’ steady-state sales
Table 5 Predictive Fit of Focal Model and Nested Benchmark Models
Model Time-varying parameters Log BF
M0 Dynamics and marketing —
M1 No dynamics 9870
M2 No marketing 408
Note. Log Bayes factor is relative to M0.
and rate of growth. Our analysis proceeds by using
our model to forecast a brand’s sales with all market-
ing mix variables set to their historical means. Denote
this estimate as S0.S0serves as the basis for a compar-
ison to sales forecasted under an alternative strategy.
In this strategy, we increase the considered market-
ing activity by 10% and calculate a new level of sales,
denoted S1. One can then obtain the percentage of
sales change due to 10% permanent marginal increase
in marketing spending by comparing the sales level
of the new case to the base case S1−S0/S0≡.
In these calculations, we considered only the first
52 weeks after launch because, as noted above, 90%
of the long-term marketing effects materialize within
52 weeks (see also Leone 1995). Table 6 summarizes
the results of our policy simulation.
Findings. The first column in Table 6 reports the
average sales change across 225 brands analyzed in
this study. The large variation in effect sizes across
brands is largely driven by variation in marketing
spending across brands. The table indicates three
strata of effect sizes. The most effective stratum
comprises distribution breadth (a 10% arc elasticity
of 7.6%), regular price (5.1%), and distribution depth
(3.1%). The implied average regular price elasticity
(0.51) is low relative to meta-analytical results for
regular prices and new brands (Bijmolt et al. 2005).
This result might reflect the lower price sensitivity of
consumers who try new brands (Ghosh et al. 1983,
Parker 1992). The next stratum includes line length
and feature/display (1.5%). The least effective group
of strategies for affecting new-brand sales includes
discounting (which actually has a negative marginal
effect) and advertising. This finding is notable, as
Table 1 also suggests these are the most-often consid-
ered instruments in past research.
Marginal Profit Analysis. Table 6 illuminates a mar-
ginal profit approximation. Let C0denote the cost
Table 6 Equilibrium Sales Value Impact of 10% Permanent Increase
in Marketing Support (%)
Standard
Mean deviation 1st Quartile Median 3rd Quartile
Advertising 025 024 008 016 031
spending
Regular price −513 372 −687 −445 −281
Distribution 761 406 527 661 880
breadth
Line length 153 111 076 125 173
Distribution 318 191 198 268 378
depth
Discount depth −024 140 −018 −012 −007
Feature/display 153 109 078 121 199
Note. As a result of a 10% permanent increase in regular prices, sales
reaches a 5.1% lower equilibrium level than it would have reached had the
price been kept constant at its mean.
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1049
of the base marketing strategy, denote the sales
increase in Table 6 arising from a 10% increase in
the marketing mix, R0indicate the revenue of the
base strategy, MM denote the manufacturer gross
margins, and RM denote the retailer gross margins.
Then the manufacturer profits under the base case
are 0=1−RM ∗MM ∗R0−C0. With a 10%
increase in the marketing expenditure, profits become
1=1+ ∗1−RM ∗MM ∗R0−1+010∗C0
(assuming that a percent increase in costs leads to
a percent increase in marketing). The condition that
1>
0therefore implies that it is profitable to
increase marketing spend on the margin when the
resulting increase in marginal revenue, 1−RM ∗
MM∗∗R0is greater than the resulting increase in
marginal cost, 01∗C0. Assuming a retailer gross mar-
gin of RM =25% of retail sales (Agriculture and Food
Canada Report 2005) and a manufacturer gross mar-
gin of MM =40% (Grocery Management Association
2006), this condition reduces to C0/R0<3∗.
Stated differently, the marginal profits of market-
ing investment become positive when costs as a
percent of retail revenue are less than 3 ∗. For
distribution (=0076), this implies it is profitable
on the margin to invest in distribution when dis-
tribution costs are less than 23% of retail revenue.
On the other end of the spectrum, it is profitable to
advertise (=00025) only when the marginal cost
of advertising is less than 0.8% of revenue. Given
most firms budget about 5% of manufacturer sales
for advertising (or 3.75% of retail sales), this sug-
gests that further increases in advertising are, on aver-
age, unwarranted (though variation across categories
imply different strategies dominate in different cate-
gories). The thresholds for line length and distribution
depth are 5% and 10%, respectively, while the thresh-
old for feature display is 5%.
6.2. Strategic Launch Options
Prior research has speculated on the relative merits of
various marketing strategies in the context of product
diffusion (skimming versus penetration pricing, con-
stant versus decreasing advertising, national versus
phased brand roll-out, and simultaneous versus
sequential product line entry). Our empirical general-
izations not only afford empirical insights into these
strategies, but also extend prior work to consider their
interactions. In particular, we contend that the effi-
cacy of marketing strategies is amplified by strong
distribution.
Simulation Design. To explore these interactions
we generate a 2 (skimming/penetration pricing) ×2
(constant/decreasing advertising) ×2 (national dis-
tribution/phased roll-out) ×2 (simultaneous/phased
brand entry) design and ensure the strategies are
well within the observed range of the marketing mix
instruments. We use a 52-week duration because most
brands reach their maximum sales by this time. Ini-
tializing new-brand sales at zero, we forecast demand
for all 225 brands over the 52 weeks after launch using
the parameters estimated in our model.
Price Skimming vs. Penetration Pricing. Penetration
pricing is considered optimal for new durable goods
(e.g., Horsky 1990, Kalish 1985, Mesak and Berg
1995). Our skimming/penetration condition contrasts
(1) a strategy wherein the launch price is one standard
deviation above the historical mean price at launch
and one standard deviation below the historical mean
price at 52 weeks (price skimming) to (2) a strategy
wherein the regular price is held constant one stan-
dard deviation below the mean (penetration).
Constant vs. Monotonically Decreasing Advertising
Spending. Prior literature argues that decreasing
returns to scale in advertising favors a monotoni-
cally decreasing advertising strategy (Dockner and
Jørgensen 1988, Horsky and Mate 1988, Horsky and
Simon 1983, Kalish 1985). Our constant/decreasing
advertising manipulation contrasts (1) advertising
held at one standard deviation above its histori-
cal mean (constant) to (2) a case where advertis-
ing decreases from one standard deviation above the
mean to one standard deviation below (decreasing).
National Launch vs. Phased Roll-Out. Despite the
pivotal role distribution plays in new-brand diffu-
sion, little academic research exists on distribution
strategies over time in the context of new-brand dif-
fusion (Bronnenberg and Mela 2004, Jones and Ritz
1991). Our national launch/regional condition manip-
ulation contrasts (1) holding distribution at one stan-
dard deviation above its historical mean (national
launch) with (2) increasing distribution from one stan-
dard deviation below the mean to one standard devi-
ation above the mean (phased roll-out).
Simultaneous vs. Phased Brand Entry. Moorthy and
Png (1992) and Wilson and Norton (1989) argue that it
is effective to release all variants early in the brand life
cycle except when cannibalization is present. In the
simultaneous/phased entry manipulation we com-
pare (1) an increase from one standard deviation
below the mean to one standard deviation above the
mean (phased) to (2) a constant level of brand line
length held at one standard deviation above the his-
torical mean observed in the data (simultaneous).
Table 7 reports the sales and growth effects of the
strategic launch options. The sales impact is expressed
as percentage gains relative to a base case wherein
marketing activity is held fixed at historical mean
levels over the 52-week duration (see Panel A). In
this case, sales peak at week 41, with 90% of growth
within 14 weeks. We express the growth impact as
the difference between the time it takes a brand to
reach 90% of maximum sales in the base case and the
Ataman, Mela, and van Heerde: Building Brands
1050 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
Table 7 Sales and Growth Impact of Strategic Trade-Offs
Marketing mix instruments Sales Growth
Pricing Advertising Distribution Brand line M SD M SD
Panel A: Base case
At mean At mean At mean At mean 401×107105 ×10814 —
Marketing mix instruments Sales impact Growth impact
Pricing Advertising Distribution Brand line M SD M SD
Panel B: Interaction effects (relative to base case)
Penetration Decreasing National Simultaneous 611333−4902
Penetration Constant National Simultaneous 636348−4000
Skimming Decreasing National Simultaneous 529293−1308
Penetration Decreasing National Phased 523288−1010
Skimming Constant National Simultaneous 5553070316
Penetration Constant National Phased 5493030821
Skimming Decreasing National Phased 4422496039
Skimming Constant National Phased 4682648449
Penetration Decreasing Phased Simultaneous 894727341
Penetration Constant Phased Simultaneous 1136127738
Skimming Decreasing Phased Simultaneous 302428536
Penetration Decreasing Phased Phased 221728636
Skimming Constant Phased Simultaneous 543728834
Penetration Constant Phased Phased 463029034
Skimming Decreasing Phased Phased −310129632
Skimming Constant Phased Phased −071229930
Notes. M (mean) and SD (standard deviation) are computed across 225 brands. For example, with penetration pricing, decreasing
advertising, national launch, and simultaneous line entry, an average brand enjoys 61.1% more sales in the first year and reaches the
90% mark 4.9 weeks earlier than it does in the base case.
time to reach 90% of maximum sales under an alter-
native strategic option. Of note, distribution effects
are considerably larger than the effect of all other
strategies. Moreover, national launch interacts with
(1) low price and broader lines to enhance market
potential and growth and (2) advertising to facilitate
growth.8Taken together, these interactions suggest
broad access to distribution is a necessary condition
for effective marketing.
7. Conclusions
Although new brands are central to the success of
organizations, large numbers of these brands fail
each year. For example, Hitsch (2006) reports that
75% of new-product introductions fail in the ready-
to-eat breakfast cereal category. It is therefore a
long-standing and central question in marketing to
explain why some brands fail and some succeed.
This research seeks to be a step in that direction
by linking the sales outcomes for 225 new brands
across 22 product categories over a five-year period
to ascertain which marketing strategies discriminate
8We tested for these interactions using a classical analysis of vari-
ance of the sales and growth columns in Table 7 on the design
variables in the rows of Table 7.
successful brands in terms of sales and time to pen-
etrate the market. In contrast to prior research per-
taining to the effects of marketing strategy on the
sales of new brands, we generalize our analysis across
many categories and incorporate an array of mar-
keting strategies that span the entire marketing mix.
Moreover, we use statistical controls for marketing
mix endogeneity and performance feedback in our
analysis. We contend an empirical generalization that
assesses the relative efficacy of launch strategies has
remained heretofore unaddressed in the marketing
literature.
To achieve this aim, we formulate a Bayesian
DLM of repeat-purchase diffusion. The methodology
extends the literature on repeat-purchase diffusion
models (e.g., Lilien et al. 1981) to incorporate dynam-
ics in the growth process over time and the endogene-
ity of marketing spend. Our state-space formulation
of the repeat-purchase model enables us to achieve
these goals. This innovation also enables a multitude
of additional potential specifications given its inherent
flexibility in estimation. Using this approach, we find:
• The relative effect sizes of the various strategies
(standardized to sum to one) on market potential are
as follows: distribution breadth 54%, feature/display
20%, distribution depth 9%, line length 7%, regular
price 5%, advertising 2%, and discounting 2%. Thus,
over the range of our data, the effect of distribution
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1051
exceeds the combined effect of all other marketing
effects. This underscores the importance of obtaining
distribution for new brands.9This finding supple-
ments that of Ataman et al. (2007), who find that dis-
tribution plays a central role in explaining differences
in sales across geographic regions in France. The
result further underscores the desirability of ascer-
taining the antecedents of distribution including, for
example, the use of slotting allowances (Sudhir and
Rao 2006) and suggests the study of penetration into
distribution is an substantially under-researched area
in marketing (we suspect this might be due in part
to a lack of good data). The relative effect sizes of
the various strategies on the growth parameter (stan-
dardized to sum to one) are as follows: distribution
breadth 31%, discounting 20%, regular price 15%, line
length 12%, feature/display 12%, advertising 10%,
and distribution depth 0.1%.
• With the exception of discounting, all strategies
have a positive total effect on sales. Discounts quicken
diffusion but have a negative effect on long-term mar-
ket potential.
• Not only does distribution have the largest direct
impact on sales, but it also interacts with other strate-
gies to enhance their efficacy.
• Using a simulation predicated on our data, we
find the breakeven thresholds to be lowest for distri-
bution breadth and depth and highest for advertising
and discounting.
Our findings have a number of managerial impli-
cations. First, the results of our analysis can be
informative to firms seeking to allocate funds across
the mix in a means consistent with their growth
objectives. Given that discounting accelerates growth
at half the rate of distribution breadth, firms can
trade off the cost of a two-standard-unit increase in
discounting with a one-standard-unit increase in dis-
tribution breadth. Second, like all diffusion models,
the model developed herein can be used to fore-
cast the sales growth of new brands; however, in
this instance the model can be used under various
marketing scenarios for repeat-purchase goods. Given
the empirical generalization, firms can choose ana-
log products to engage these forecasts even with lit-
tle data and then update them as new data become
available; the Bayesian nature of our model allows the
modeler to readily update the parameter estimates.
As with any research, the findings summarized
above are subject to several extensions/limitations.
Many limitations are not unique to this study but are
9The finding that feature/display is the second most important
instrument to build market potential coincides with results from
Pauwels and Hanssens (2007), who find that promotional vari-
ables are especially effective in turning brand performance around.
Together, these results imply that “marketing strategy is in the
details.” We thank an anonymous reviewer for this observation.
inherent in empirical models of sales response pred-
icated on secondary data. These extensions/limits
include the following. First, we focus exclusively
on national brand introductions and exclude private
labels; presumably, retailers would be quite inter-
ested in private label brands and the strategies that
ensure their viability. Second, brand extensions are
an important topic in their own right, and compar-
ing marketing strategy efficacy of brand extensions
to that of entirely new brands can further enhance
our understanding of successful roll-out strategies.
Third, traditional models of diffusion in repeat pur-
chase contexts separate growth due to word of mouth
effects from innovation effects. We focus on the latter,
given that word-of-mouth effects are largely absent
in packaged goods (Hardie et al. 1998). Nonethe-
less, it would be desirable to extend this model
for durable goods contexts in which word-of-mouth
plays a greater role. Fourth, the data preclude us from
taking a more nuanced view of innovation and diffu-
sion. For example, we do not consider how personal
and product characteristics or organizational capabil-
ities moderate diffusion rates (Cooper 1998, Gatignon
and Robertson 1986, Rogers 1976). Sixth, our model
does not provide a formal accounting of retailer deci-
sion making. A national roll-out strategy is not only
incumbent on a firm’s choice to distribute nationally,
but also on the willingness of retailers to adopt a new
brand (Bronnenberg and Mela 2004).
Our analysis is a step toward a more complete
view of the role of postlaunch marketing strategy on
the diffusion of frequently purchased CPG brands. In
light of our findings and the foregoing limitations,
we hope this work will stimulate further research on
new-brand launch, especially with regard to the role
distribution plays in the success of new brands.
Acknowledgments
The authors thank Information Research, Incorporated and
TNS Media Intelligence for providing the data; and Jason
Duan, Vithala Rao, Song Yao, and seminar participants at
the 2007 Marketing Science Conference, the 2007 Marketing
Dynamics Conference, the 2007 Yale Center for Customer
Insights Conference, and the 2007 European Marketing
Academy Conference, as well as the editor-in-chief, the area
editor, and four anonymous reviewers for their comments.
The first and third authors thank Netherlands Organization
for Scientific Research for research support.
Appendix. Model Estimation
The observation equation and the evolution equation of the
multivariate DLM for brand j(j=1225) are
Yjt =Fjtjt +Xjtj+jt (10)
jt =Gjtjt−1+hjt +jt(11)
where Yjt is a vector that stacks standardized sales and mar-
keting mix instruments. From now on, we drop the brand
subscript jfor simplicity. Ft=IM+1, where M=7is the
Ataman, Mela, and van Heerde: Building Brands
1052 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
number of marketing mix variables. Xtis the matrix of
regressors that create short-term fluctuations in sales. For
a given brand, we assume t∼N0V and t∼N0W,
where Vand Ware full and diagonal matrices, of size
M +1×M +1, of error variances, respectively. The time-
varying parameter vector,
t=
t
t, evolves as described
in (11).
Step 1. tYt,V,W,,Gt,ht.
For each brand we sample from the conditional distri-
bution of using the forward-filtering, backward-sampling
algorithm proposed by Carter and Kohn (1994) and
Frühwirth-Schnatter (1994). First, for t=1T we for-
ward filter to obtain the moments mtand Ct. Conditional
on
Yt,V,W,,Gt,htand 0D0∼Nm
0C
0, where
Yt=
Yt−X
t:
• The prior at time tis tDt−1∼Na
tR
t, where at=
Gtmt−1+htand Rt=GtCt−1G
t+W.
• One-step-ahead forecast at time tis
YtDt−1∼
Nf
tQ
t, where ft=Ftatand Qt=FtRtF
t+V.
• The posterior distribution at time tis tDt∼
Nm
tC
t, where mt=at+RtF
tQ−1
t
Yt−ft, and Ct=Rt−
RtF
tQ−1
tFtRt.
At t=Twe sample a matrix of evolution parameters
from the distribution Nm
tC
t. Next we sequence back-
ward for t=T−11 sampling from ptt+1rest∼
Nq∗
tQ∗
t, where q∗
t=mt+Btt+1−at+1,Q∗
t=Ct−BtRt+1B
t,
and Bt=CtG
t+1R−1
t+1. We select m0=0 and C0=01asthe
initial values.
Step 2. Vt,Yt,.
For a given brand, we assume that the observation equa-
tion error variance matrix, of size M +1×M +1, is full.
We place an inverse Wishart prior, with (nV0S
V0. Then
the full conditional posterior distribution is also inverse
Wishart, with nV1=nV0+Tand SV1=SV0+T
t=1Yt−X
t−
FttYt−X
t−Ftt. We use a diffuse prior with nV0=
M +1+2 and SV0=0001 ×IM+1.
Step 3. Wt,,,.
We assume that the evolution equation error-variance
matrix, of size M +1×M +1, is diagonal for a given
brand. We place an inverse Gamma prior on the elements
of this matrix, with nW0/2 degrees of freedom and a scale
parameter of SW0/2. The full conditional posterior distribu-
tion is also distributed inverse Gamma with nW1=nW0+
T−1 and SW1=SW0+T
t=1t−Gtt−1−htt−Gtt−1−ht.
We use a diffuse prior with nW0=3 and SW0=0001.
Step 4. t,W,,,t,W,,, and t,W,,.
Conditional on the sampled baseline sales series across
all brands, the evolution equation is nonlinear in parame-
ters and there is no closed-form density for the parameters.
Therefore, we use a random walk Metropolis-Hastings step
within the Gibbs sampler to obtain the parameter estimates.
We discuss only the estimation of the brand-specific repeat
rates. The estimation of t,W,,and t,W,,
follows directly. We generate the candidate repeat purchase
rate draw by m
j=m−1
j+z, where m denotes mth iter-
ation, and zis a random draw from N0I. We select
such that the acceptance rate is between 20%–50% (Chib
and Greenberg 1995). The candidate draw is accepted with
the probability ∗=min1, where
=m
jtW
m−1
jtW(12)
and ·is conditional likelihood of Equation (11) evaluated
at each draw.
Step 5. t,W.
To obtain the conditional posterior distribution of the
brand-specific evolution equation parameters associated
with the ith marketing mix instrument i, we define KiT =
1T−1iT −1SalesjT−1SalesjT−1and WiT =Wi⊗IT−1. We place a
normal prior on the parameters, i∼N
. Then the
full conditional posterior is also normal with i∼N¯
,
where =¯
−1
+KiT W−1
iT iT , and ¯
=−1
+
KiT W−1
iT K
iT −1.
Step 6. t,V.
To obtain the brand-specific conditional posterior distri-
bution of the nontime-varying observation equation param-
eters , we define
Yt=Yt−Fttand VT=V⊗IT.We
place a normal prior on the parameters, ∼N
.
Then the full conditional posterior is also normal with ∼
N¯
, where =¯
−1
+XtV−1
T
Yt, and ¯
=
−1
+XtV−1
TX
t−1.
References
Aaker, D. A. 1996. Measuring brand equity across products and
markets. California Management Rev. 38(3) 102–120.
Agriculture and Food Canada Report. 2005. Overview of the
retail grocery market in the U.S. south central states. http://
www.ats.agr.gc.ca.
Ailawadi, K. L., K. Gedenk, C. Lutzky, S. A. Neslin. 2007. Decom-
position of the sales impact of promotion-induced stockpiling.
J. Marketing Res. 44(3) 450–467.
Anderson, E., D. Simester. 2004. Long run effects of promotion
depth on new versus established customers: Three field stud-
ies. Marketing Sci. 23(1) 4–20.
Ataman, M. B., C. F. Mela, H. J. Van Heerde. 2007. Consumer pack-
aged goods in France: National brands, regional chains, local
branding. J. Marketing Res. 44(1) 14–20.
Ataman, M. B., H. J. Van Heerde, C. F. Mela. 2006. The long-term
effect of marketing strategy on brand performance. Working
paper, RSM Erasmus University, Rotterdam, The Netherlands.
Bass, F. M., A. V. Bultez. 1982. A note on optimal strategic pricing
of technological innovations. Marketing Sci. 1(4) 371–378.
Bass, F. M., D. Jain, T. V. Krishnan. 2000. Modeling the marketing
mix influence in new product diffusion. V. Mahajan, E. Muller,
Y. Wind, eds. New-Product Diffusion Models. Kluwer Academic
Publishers, Dordrecht, The Netherlands, 99–122.
Bass, F. M., T. V. Krishnan, D. Jain. 1994. Why the Bass model fits
without decision variables. Marketing Sci. 13(3) 203–223.
Bass, F. M., N. Bruce, S. Majumdar, B. P. S. Murthi. 2007. Wearout
effects of different advertising themes: A dynamic Bayesian
model of advertising-sales relationship. Marketing Sci. 26(2)
179–195.
Bijmolt, T. H. A., H. J. Van Heerde, R. G. M. Pieters. 2005. New
empirical generalizations on the determinants of price elastic-
ity. J. Marketing Res. 42(2) 141–156.
Bronnenberg, B. J., C. F. Mela. 2004. Market roll-out and retailer
adoption for new brands. Marketing Sci. 23(4) 500–518.
Bronnenberg, B. J., V. Mahajan, W. R. Vanhonacker. 2000. The emer-
gence of market structure in new repeat-purchase categories:
The interplay of market share and retailer distribution. J. Mar-
keting Res. 37(1) 16–31.
Carter, C. K., R. Kohn. 1994. On Gibbs sampling for state space
models. Biometrika 81(3) 541–553.
Ataman, Mela, and van Heerde: Building Brands
Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS 1053
Chib, S., E. Greenberg. 1995. Understanding the Metropolis-
Hastings algorithm. Amer. Statist. 49(4) 327–335.
Christen, M., S. Gupta, J. C. Porter, R. Staelin, D. R. Wittink. 1997.
Using market-level data to understand promotion effects in a
nonlinear model. J. Marketing Res. 34(3) 322–334.
Cooper, R. G. 1998. Product LeadershipCreating and Launching Supe-
rior New Products. Perseus Books, Reading, MA.
Dockner, E., S. Jørgensen. 1988. Optimal advertising policies for
diffusion models of new product innovations in monopolistic
situation. Management Sci. 34(1) 119–130.
Dodson, J., E. Muller. 1978. Models of new products diffusion
through advertising and worth-of-mouth. Management Sci.
24(15) 1568–1578.
Elberse, A., J. Eliashberg. 2003. Demand and supply dynamics for
sequentially released products in international markets: The
case of motion pictures. Marketing Sci. 22(3) 329–354.
Eliashberg, J., A. P. Jeuland. 1986. The impact of competitive entry
in developing market upon dynamic pricing strategies. Mar-
keting Sci. 5(1) 20–36.
Frühwirth-Schnatter, S. 1994. Data augmentation and dynamic lin-
ear models. J. Time Ser. Anal. 15(2) 183–202.
Gatignon, H., E. Anderson. 2002. Inter-channel competition and
new product diffusion: Market making, market taking, and
competitive effects in several European countries. Working
paper, INSEAD, Fontainebleau, France.
Gatignon, H., T. Robertson. 1986. A propositional inventory for new
diffusion research. J. Consumer Res. 11(4) 849–867.
Ghosh, A., S. A. Neslin, R. W. Shoemaker. 1983. Are there asso-
ciations between price elasticity and brand characteristics?
P. E. Murphy, E. R. Laczniak, eds. AMA Educators’ Conf. Proc.,
American Marketing Association, Chicago, 226–230.
Grocery Managament Association. 2006. Insights into the
food, beverage and consumer products industry. http://
www.gmabrands.com.
Hahn, M., S. Park, L. Krishnamurthi, A. A. Zoltners. 1994. Analy-
sis of new-product diffusion using a four-segment trial-repeat
model. Marketing Sci. 13(3) 224–247.
Hardie, B. G. S., P. S. Fader, M. Wisniewski. 1998. An empirical com-
parison of new product trial forecasting models. J. Forecasting
17(3–4) 209–229.
Hauser, J., G. J. Tellis, A. Griffin. 2006. Research on innovation: A
review and agenda for marketing science. Marketing Sci. 25(6)
687–717.
Hitsch, G. J. 2006. An empirical model of optimal dynamic product
launch and exit under demand uncertainty. Marketing Sci. 25(1)
25–50.
Horsky, D. 1990. A diffusion model incorporating product benefits,
price, income and information. Marketing Sci. 9(4) 342–365.
Horsky, D., K. V. Mate. 1988. Dynamic advertising of competing
durables good producers. Marketing Sci. 7(4) 356–367.
Horsky, D., L. Simon. 1983. Advertising and the diffusion of new
products. Marketing Sci. 2(1) 1–17.
Information Resources, Inc. 2005. 2004 New Product Pacesetters.
Chicago.
Jain, D., R. C. Rao. 1990. Effect of price on the demand for durables.
J. Bus. Econom. Statist. 8(2) 163–170.
Jedidi, K., C. F. Mela, S. Gupta. 1999. Managing advertising and
promotion for long-run profitability. Marketing Sci. 18(1) 1–22.
Jones, J. M., C. J. Ritz. 1991. Incorporating distribution into new
product diffusion models. Internat. J. Res. Marketing 891–112.
Kalish, S. 1983. Monopolist pricing with dynamic demand and pro-
duction cost. Marketing Sci. 2(2) 135–160.
Kalish, S. 1985. A new product adoption model with pricing, adver-
tising and uncertainty. Management Sci. 31(12) 1569–1585.
Kalish, S., G. L. Lilien. 1986. A market entry timing model for new
technologies. Management Sci. 32(2) 194–205.
Kamakura, W. A., S. Balasubramanian. 1987. Long-term forecasting
with innovation diffusion models: The impact of replacement
purchases. J. Forecasting 6(1) 1–19.
Kamakura, W. A., S. Balasubramanian. 1988. Long-term view of the
diffusion of durables: A study of the role of price and adoption
influence process via tests of nested models. Internat. J. Res.
Marketing 51–13.
Keane, M. 1997. Modeling heterogeneity and state dependence in
consumer choice behavior. J. Bus. Econom. Statist. 15(3) 310–327.
Krider, R. E., T. Li, Y. Liu, C. B. Weinberg. 2005. The lead-lag puzzle
of demand and distribution: A graphical method applied to
movies. Marketing Sci. 24(4) 635–645.
Leone, R. P. 1995. Generalizing what is known of temporal aggre-
gation and advertising carryover. Marketing Sci. 14(3, Part 2)
G141–G150.
Lilien, G. L., A. Rao, S. Kalish. 1981. Bayesian estimation and con-
trol of detailing effort in a repeat-purchase diffusion environ-
ment. Management Sci. 27(5) 493–506.
Lodish, L. M., M. Abraham, S. Kalmenson, J. Livelsberger,
B. Lubetkin, B. Richardson, M. E. Stevens. 1995. How T.V.
advertising works: A meta-analysis of 389 real world split cable
T.V. advertising experiments. J. Marketing Res. 32(2) 125–139.
Mahajan, V., E. Muller, F. M. Bass. 1990. New product diffusion
models in marketing: A review and directions for research.
J. Marketing 54(1) 1–26.
Mela, C. F., S. Gupta, D. R. Lehmann. 1997. The long-term impact of
promotion and advertising on consumer brand choice. J. Mar-
keting Res. 34(2) 248–261.
Mesak, H. I. 1996. Incorporating price, advertising and distribution
in diffusion models of innovation: Some theoretical and empir-
ical results expectations in diffusion models. Comput. Oper. Res.
23(10) 1007–1023.
Mesak, H. I., W. D. Berg. 1995. Incorporating price and replace-
ment purchases in new products diffusion models for con-
sumer durables. Decision Sci. 26(4) 425–449.
Moorthy, K. S., I. P. L. Png. 1992. Market segmentation, cannibal-
ization, and the timing of product introductions. Management
Sci. 38(3) 345–359.
Muller, E., R. Peres, V. Mahajan. 2007. Innovation diffusion and new
product growth: Beyond a theory of communications. Working
paper, Leonard Stern School of Business, New York University,
New York.
Neelamegham, R., P. Chintagunta. 1999. A Bayesian model to fore-
cast new product performance in domestic and international
markets. Marketing Sci. 18(2) 115–136.
Parker, P. M. 1992. Price elasticity dynamics over the adoption life
cycle. J. Marketing Res. 29(3) 358–367.
Parker, P. M., H. Gatignon. 1994. Competitive effects in diffusion
models. Internat. J. Res. Marketing 11 17–39.
Pauwels, K. 2007. How retailer and competitor decisions drive
the long-term effectivness of manufacturer promotions for fast
moving consumer goods. J. Retailing. Forthcoming.
Pauwels, K., D. M. Hanssens. 2007. Performance regimes and mar-
keting policy shifts. Marketing Sci. 26(3) 293–311.
Pauwels, K., D. M. Hanssens, S. Siddarth. 2002. The long-term
effects of price promotions on category incidence, brand
choice, and purchase quantity. J. Marketing Res. 39(4) 421–439.
Rao, A., M. Yamada. 1988. Forecasting with a repeat purchase dif-
fusion model. Management Sci. 34(6) 734–752.
Ratchford, B. T., S. K. Balasubramanian, W. A. Kamakura. 2000.
Diffusion models with replacement and multiple purchases.
V. Mahajan, E. Muller, Y. Wind, eds. New-Product Diffusion Mod-
els. Kluwer Academic Publishers, Dordrecht, The Netherlands,
123–140.
Rogers, E. M. 1976. New product adoption and diffusion. J. Con-
sumer Res. 2290–301.
Ataman, Mela, and van Heerde: Building Brands
1054 Marketing Science 27(6), pp. 1036–1054, © 2008 INFORMS
Simon, H., K.-H. Sebastian. 1987. Diffusion and advertising: The
German telephone company. Management Sci. 33(4) 451–466.
Slotegraaf, R. J., K. Pauwels. 2008. The impact of brand equity
and innovation on the long-term effectiveness of promotions.
J. Marketing Res. Forthcoming.
Steenkamp, J.-B. E. M., K. Gielens. 2003. Consumer and market
drivers of the trial probability of new consumer packaged
goods. J. Consumer Res. 30 368–384.
Steenkamp, J.-B., V. R. Nijs, D. Hanssens, M. Dekimpe. 2005. Com-
petitive reactions to advertising and promotion attacks. Mar-
keting Sci. 24(1) 35–54.
Sudhir, K., V. R. Rao. 2006. Do slotting allowances enhance effi-
ciency or hinder competition? J. Marketing Res. 43(2) 137–155.
Tellis, G. J., P. H. Franses. 2006. Optimal data interval for estimating
advertising response. Marketing Sci. 25(3) 217–229.
Tellis, G. J., S. Stremersch, E. Yin. 2003. The international takeoff
of new products: The role of economics, culture and country
innovativeness. Marketing Sci. 22(2) 188–208.
Van den Bulte, C. 2000. New product diffusion acceleration: Mea-
surement and analysis. Marketing Sci. 19(4) 366–380.
Van den Bulte, C., Y. V. Joshi. 2007. New product diffu-
sion with influentials and imitators. Marketing Sci. 26(3)
400–421.
Van Heerde, H. J., K. Helsen, M. G. Dekimpe. 2007. The impact of a
product-harm crisis on marketing effectiveness. Marketing Sci.
26(2) 230–245.
Van Heerde, H. J., C. F. Mela, P. Manchanda. 2004. The dynamic
effect of innovation on market structure. J. Marketing Res. 41(2)
166–183.
West, M., J. Harrison. 1997. Bayesian Forecasting and Dynamic Models,
2nd ed. Springer-Verlag, New York.
Wilson, L. O., J. A. Norton. 1989. Optimal entry timing for a product
line extension. Marketing Sci. 8(1) 1–17.
Xie, J., M. Song, M. Sirbu, Q. Wang. 1997. Kalman filter estima-
tion of new product diffusion models. J. Marketing Res. 34(3)
378–393.