Journal of Econometrics 89 (1999) 269—291

Long-run effects of price promotions in scanner markets

Marnik G. Dekimpe *, Dominique M. Hanssens
,
Jorge M. Silva-Risso 
Department of Applied Economics, Catholic University of Leuven, 3000 Leuven, Belgium

The Anderson Graduate School of Management, University of California, Los Angeles,
CA 90095-1481, USA
 J.D. Power and Associates, Agoura Hills, CA, USA
 The Anderson Graduate School of Management, University of California, Los Angeles, CA, USA

Abstract

Good marketing decisions require managers’ understanding of the response function

relating performance measures to variations in the marketing mix. We use unit-root
techniques to address market response in evolving markets, with a focus on their
response to price promotions. We distinguish between evolution at the primary-demand
vs. selective-demand level, and examine four consumer product categories for which
high-quality scanner records are available. We find category and brand sales to be
predominantly stationary, with differences in promotional impact between national and
private-label brands. Even in the rare occurrence of performance evolution, the long-term
effects of price promotions are not necessarily positive.  1999 Elsevier Science S.A. All
rights reserved.

JEL classification: C22; C32; M31

Keywords: Evolution; Long-term effectiveness; Price promotions

1. Introduction

Marketing managers are principally concerned with the allocation of scarce

marketing resources such as sales force, advertising and promotion, for the

* Corresponding author. E-mail: marnik.dekimpe@econ.kuleuven.ac.be

0304-4076/99/$ — see front matter  1999 Elsevier Science S.A. All rights reserved.
PII: S 0 3 0 4 - 4 0 7 6 ( 9 8 ) 0 0 0 6 4 - 5
270 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

purpose of improving the market and profit performance of their products or

brands. The quality of their decisions greatly depends on their understanding of
the way in which customers and competitors will respond to these efforts, in the
short run as well as the long run. Econometric methods have been used to
enhance this understanding, resulting in a vast body of empirically tested
knowledge on the relationship between market performance and marketing
investments (see, e.g. Hanssens et al., 1990).
The main focus of this response modeling, however, has been on ‘short-run
forecasting and optimization procedures, while assuming an essentially stable
environment’ (Wind and Robertson, 1983, p.13). Still, the environments in which
marketing decisions are made may well be evolving, due to changes in techno-
logy, consumer preferences and/or competition (Dekimpe and Hanssens, 1995b).
Consequently, many of the resource-allocation problems faced by marketers
contain an important long-run dimension. For example, advertising expendi-
tures may affect sales levels immediately, but they can also contribute to the
establishment of a long-run brand image. Customers often react negatively to
a sudden price change, but may adapt to higher prices over time. These and
other long-term marketing behaviors should be conceptually understood and
empirically quantified if we are to improve the practice of effective marketing
resource allocation.
This paper addresses both the concepts and the empirics of market response
in evolving markets. We argue that long-term time-series techniques are ideally
suited to measure market evolution and relate it to marketing decisions. In
doing so, we distinguish between market response at the primary-demand
(product-category) and selective-demand (brand-sales) level, and identify four
potential scenarios depending on their stable/evolving character. We illustrate
these principles on four consumer product categories (catsup, liquid detergent,
soup and yogurt) for which high-quality scanner records on sales and the
marketing mix are available, and focus on the over-time impact of the brands’
price promotions.
Over the last decade, a growing body of literature has documented that
temporary retail price reductions substantially increase sales, that customers
adjust their reference price because of frequent price promotions, and that
cross-promotional effects tend to be asymmetric, with higher-quality brands
impacting the weaker brands disproportionately (see Blattberg et al. (1995) or
Bronnenberg and Wathieu (1996)). The debate is still open, however, on whether
(1) price promotions have any long-run effects, (2) whether they only induce
brand switching, or also result in market-expansion effects, and (3) whether
asymmetric effects are reflected in differing impact durations or only in a distinct
instantaneous effect. The first issue is critical to the effective use of price
promotions, especially in light of the concern that they may actually be detri-
mental to the long-term health of the brand. The second issue, whether or not
part of the observed sales increase during (and shortly after) a price promotion is
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 271

gained at the expense of the other players in the market, affects the strength of
competitive reaction, and ultimately, the profitability of a brand’s promotional
activity. Finally, insights in the total cumulative cross-effect of price promotions
initiated by, respectively, national and private-label brands will contribute to the
ongoing debate on the asymmetric drawing power of national brands.
The remainder of the paper is organized as follows. In Section 2, we make
a case for distinguishing between stable and evolving performance patterns,
both at the primary- and selective-demand level, and introduce four potential
scenarios that emerge when jointly considering both dimensions. In Section 3,
we indicate how long-run time-series techniques can be used to empirically
distinguish between these scenarios, and to model short- and long-run respon-
siveness in each. We briefly review previous marketing applications of these
techniques, and show how they can contribute new knowledge in the promo-
tion-effectiveness literature. Empirical findings are presented in Section 4, and
we conclude in Section 5 with some empirical and analytical generalizations,
and some areas for future research.

2. Marketing in stable and evolving environments

The statistical distinction between stationary (stable) and non-stationary

(evolving) sales or demand behavior has important ramifications for marketers.
If sales are mean reverting without level shifts, marketing actions produce at
most temporary deviations from the brand’s average performance level. If sales
are evolving, on the other hand, there is no such reversion tendency, and there is
a potential for long-term marketing effectiveness. Subsequent multivariate ana-
lyses (cointegration and/or persistence models, see Sections 3 and 4) should then
establish whether or not the brand’s marketing actions actually affect its ob-
served sales evolution (Dekimpe and Hanssens, 1995a,b). Obviously, if discrete
marketing actions (e.g. a single promotion) cause a structural break in the
data-generating process, such as a level-shift in the mean of otherwise stable
sales, that too is strong evidence of long-run marketing effectiveness, as illus-
trated in Leone (1987) and Hanssens et al. (1990, p. 148). As discussed in
Section 4, no evidence of such promotion-induced breaks was found in our data,
so we will quantify the over-time impact of price promotions — operationalized
as unexpected price shocks — on the assumption that the parameters of the
process do not change as a result of these shocks (see Pesaran and Samiei (1991)
for a similar assumption).

 In what follows, the mean-stationary model is used as alternative hypothesis, since in most
marketing settings, this corresponds to a more realistic scenario than a trend-stationary model
(Dekimpe, 1992).
272 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

Evolution in sales can exist at the primary-demand (industry sales) and/or the
selective-demand (brand sales) level. While most marketing practitioners and
researchers focus on selective demand, the primary-demand component is
important, even critical, for several reasons:

E Evolution in primary demand allows for persistent market-expansive effects

of marketing investments. For example, Compaq’s introduction of a new line
of aggressively-priced personal computers in the early nineties opened up
a large new market segment (home computing). Likewise, advertising cam-
paigns that attract new buyers to a product category can have a continuing
effect when some of them become regular users.
E In the presence of market expansion, marketing actions are less likely to
cause permanent damage to competition, and retaliatory behavior is expected
to be less severe. Conversely, with stable primary demand, the rules of
zero-sum competitive equilibria are more likely to apply. Absent market
expansive effects of marketing actions, the only way a company can build
a long-run advantage over the competition is by inducing consumers to
switch loyalty.

In sum, the presence or absence of evolution in primary demand may influence

marketing effectiveness, competitive marketing behavior and, ultimately, the
market structure of an industry. Schultz and Wittink (1976) derive a set of
analytical conditions for the presence of primary-demand vs. selective-demand
effects of the marketing mix. The conditions are stated as a set of first-order
derivatives of various performance measures (brand sales, industry sales and
market shares) with respect to the marketing mix, which can be estimated
econometrically. Several authors have made estimates of these derivatives, and
a review of their results may be found in Hanssens et al. (1990). However, none of
these studies have examined whether these estimates have a temporary or
a permanent character to them. Therefore, there is little, if any, empirical
evidence for the existence of permanent primary demand effects of, for example,
advertising or pricing strategies.
From a marketing-strategic perspective, evaluating stable vs. evolving
conditions at the level of industry vs. brand sales gives rise to four possible
scenarios:

E Stable brand sales in a stable category: all sales gains and losses are of
a temporary nature, and brand marketing is tactical in nature. In such
environments, the brand’s relative position or market share is also stable, and
all marketing effects are either intrinsically short-lived, or self-canceling in the
long run (cf. infra). While management decisions may still have strong short-
run share or profit implications, they reflect tactical moves that are unrelated
to the strategic or long-run direction of the brand.
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 273

E Stable brand sales in an evolving category: implies a lack of long-run marketing

effectiveness, as the brand is unable to establish permanent gains in spite of
operating in an evolving category. While marketing activities can have
long-run primary demand effects in such markets, the additional sales do not
accrue to the brand, but rather benefit its competitors.
E Evolving brand sales in a stable category: this scenario implies that the brand is
locked into a strategic battle for long-run position. Moreover, as the category
is not moving away from its historical mean, firms are involved in a zero-sum
game in which the long-run sales gain for one player always occurs at the
expense of a long-run loss for at least one of the other players. This scenario
may result in an unprofitable escalation of the competitors’ marketing expen-
ditures.
E Evolving brand sales in an evolving category: depending on the relative import-
ance of the long-run components in brand and category sales, firms may be
able to improve not only their absolute long-run performance, but also their
relative position. Moreover, if different brands’ performance levels are coin-
tegrated, brands can be seen as riding long-run market waves that could be
driven by their marketing spending.

We will not only diagnose which scenario applies to thirteen different brands
in four product categories, but will also assess the influence of one important
marketing-mix variable, price promotions, on the evolution of brand and cat-
egory sales. Indeed, in their recent review, Blattberg et al. (1995) call the question
whether there are any long-run effects of promotions ‘the most debated issue in
the promotional literature’ and ‘one for which the jury is still out’ (Blattberg et
al. 1995, p. G127).
Previous literature has tried to disentangle the various sources of promo-
tional volume. Gupta (1988) found that the majority of the promotional volume
was due to brand switching, and Bemmaor and Mouchoux (1991) expect the
product-class sales impact of price promotions to be low, if at all existent.
Chintagunta (1993), on the other hand, provided evidence of a substantial
market-expansion effect. As such, there is still ambiguity about the short-run
market-expansion potential of price promotions, and to the best of our know-
ledge, no study has yet considered their potential long-run primary-demand
implications. Indeed, previous studies have implicitly assumed or imposed
a stable market environment, and have therefore precluded the detection and
quantification of any permanent price promotion effects. Unit-root econo-
metrics do not impose this restriction, and will allow us to better quantify the
extent of price promotions’ long-run effectiveness.

 See Blattberg et al. (1995) for a more extensive review.

274 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

3. Modeling long-term marketing effects through unit-root econometrics

Unit-root econometrics is well suited to study long-run marketing effec-

tiveness. First, unit-root tests identify the presence/absence of a long-run
(stochastic-trend) component in the series’ data-generating process, and hence
distinguish stable from evolving variables. When applied to primary and selec-
tive demand, unit-root test results will enable us to classify the different brands
into one of the aforementioned strategic scenarios. To avoid the spurious
identification of a series characterized by one or more level shifts as evolving,
structural-break unit root tests are recommended whenever there is evidence of
discrete events such as new product introductions or changes in advertising
theme. Next, impulse-response functions can be derived from »AR (vector-
autoregressive) or »ECM (vector error-correction) models to study the
over-time impact of price promotions on both performance levels, and the
corresponding multivariate persistence estimates will quantify the long-run
impact of various price shocks. As indicated before, these estimates assume that
the shocks do not alter the structure and/or parameters of the data-generating
process.
Extensive technical reviews of these techniques are given in Harris (1995) and
Mills (1994), and will not be repeated here. Instead, we briefly review existing
marketing applications, and will indicate in Section 4 the actual implementation
adopted in our empirical study. Within a marketing context, unit-root tests have
been used as a first step in studies on the long-run effectiveness of advertising
expenditures (e.g. Baghestani, 1991; Dekimpe and Hanssens, 1995a), and the
absence of a unit root in most published market-share series has been inter-
preted as empirical evidence that many markets are in long-run equilibrium
where the relative position of the players is only temporarily affected by their
marketing activities (Dekimpe and Hanssens, 1995b). In a different context,
Dekimpe et al. (1997) found mean reversion in the brand loyalty of multiple
brands, which lead them to reject the contention that brand loyalty has been
systematically eroded. Structural-break unit root tests (e.g. Perron, 1990) were
used in the latter study to control for the potentially confounding effect of
new-product introductions.
Once a long-run component has been identified in the series of interest,
cointegration analysis can be used to determine whether a long-run equilibrium
relationship exists among the variables. In the marketing literature, cointegra-
tion tests have been used to study whether a brand’s sales and advertising are
moving together over time (e.g. Baghestani, 1991; Zanias, 1994), whether a prod-
uct category’s long-run evolution is linked to the evolution in some macro-
economic variables (Franses, 1994), and whether aggregate advertising spending
is related to macro-economic fluctuations (Chowdhury, 1994). In all instances,
error-correction models were estimated to capture the short-run dynamics to-
wards the identified long-run equilibrium.
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 275

Finally, Dekimpe and Hanssens (1995a) have quantified the relative import-
ance of the identified long-run component in market performance through
univariate persistence estimates, and have interpreted the short- and long-run
interrelationships between performance and advertising spending through im-
pulse-response and multivariate persistence estimates.
Overall, the number of published marketing applications of these long-run
time-series techniques is limited, in spite of their widespread use in other
disciplines such as economics and finance (see, e.g. Harris, 1995; Mills, 1994),
and in spite of repeated calls for a longer-run focus in marketing planning and
decision making (e.g. Wind and Robertson, 1983). A major barrier to their
application in marketing has been a lack of data. Indeed, it is often easier for
marketing researchers to obtain cross-sectional rather than longitudinal data
sets. In contrast, in both economics and finance, specialized agencies exist that
record in a consistent way the over-time behavior of a great variety of variables,
including macro-economic indicators, stock prices, and exchange rates. We
conjecture that the future of long-run time-series modeling in marketing will be
positively and significantly affected by the advent of new data sources that are
based on the automatic, real-time recording of purchase or consumption trans-
actions, as opposed to the retrieval of old accounting records.
Scanner data have already provided a major impetus to cross-sectional
research in marketing, in particular the study of consumer heterogeneity in
market response (see Chintagunta (1993) for a review). This heterogeneity has
been investigated at the level of brand choice, purchase quantity and purchase
timing. The dominant modeling approach has been the multinomial logit model,
not only in published academic research, but also in commercial applications in
the packaged-goods sector, according to a recent survey by Bucklin and Gupta
(1996). Recently, an interest has emerged in using the same scanner data sources
to make inferences about marketing’s long-run effectiveness (e.g. Mela et al.,
1997; Papatla and Krishnamurthi, 1996). However, these studies still use the
conventional battery of statistical techniques to analyze long-run movements in
longitudinal data. For example, Mela et al. (1997) use the Koyck specification
to measure long-run marketing effects. These methods are appropriate for
the study of multi-period sales response in stable markets, where constant
means and variances in performance and marketing support have already been
established, but as Dekimpe and Hanssens (1995a) argue, they are not well
suited to address the more strategically relevant questions about the long-run
evolution of a brand in evolving markets. The Koyck model, for example,
implies that the performance series will return to their pre-expenditure levels,
and hence precludes the detection of any persistent effects. In Section 4, we
illustrate the potential of unit-root econometrics to offer new insights into the
long-run dimension of many substantive marketing problems through a large-
scale study (13 brands, 4 categories) on the over-time effectiveness of price
promotions.
276 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

4. Measuring the over-time impact of price promotions

4.1. Data description

A.C. Nielsen household scanner panel data on the purchases of liquid laundry
detergent, soup, yogurt and catsup in the Sioux Falls market (South Dakota)
were used to construct time series of weekly sales and primary-demand figures.
These data sets were made available to the academic research community
through the Marketing Science Institute, and have been used extensively in the
recent marketing literature (see Chintagunta (1993) for a review).
As some markets have seen a proliferation of brands and sizes (e.g. each brand
in the detergent market is typically offered in several sizes ranging from 32 to
128 ounces), we expressed sales in number of ounces sold, and aggregated all
different sizes of a particular brand into one figure. For the detergent, soup and
yogurt market, we considered the top three brands, while for the catsup market,
we considered three national brands and a major private label brand. This
resulted in a total of thirteen brand-level series. We further constructed a price-
per-ounce variable, and operationalized price promotion as a temporary price
discount (cf. infra). As features and displays have been shown to strongly affect
sales (Blattberg et al., 1995), we control for their presence or absence with
dummy variables.
Each time series consists of 113 weekly observations, from the first week of
1986 until the 9th week of 1988. From a statistical point of view, longer time
spans may be preferred (e.g. Perron, 1989). However, it is of interest to determine
whether long-run inferences can be made from the data that are publicly
available to the marketing community. More importantly, longer time series
lose their managerial relevance, as ever changing market conditions make
managers reluctant to make inferences based on old data (e.g. Glazer and Weiss,
1993).
Our database aggregates consumer choices and marketing conditions to the
weekly market level, which causes some loss of information on possible underly-
ing consumer heterogeneity. Thus the dynamic relationships we uncover may be
sensitive to aggregation bias (Pesaran and Smith, 1995). However, as Allenby
and Rossi (1991) and Leeflang and Wittink (1992) point out, retailers typically
use such aggregated data when setting their price and promotion strategies.
Allenby and Rossi (1991) further prove that applying aggregate logit models on
store-level scanner data is not subject to aggregation bias when three conditions
are met: all consumers are exposed to the same marketing-mix variables, the

 We therefore focus on long-term promotion effectiveness at the brand level (typically a senior
marketing management responsibility), as opposed to the SKU level (typically a junior marketing
management responsibility). See Abraham and Lodish (1993) for a more elaborate discussion.
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 277

Fig. 1. Primary demand in four scanner product categories.

brands are close substitutes, and the price distribution is not concentrated at an
extreme value. We verified that our empirical setting meets these conditions:
price distributions are bell shaped with a few outliers (cf. infra), the product
categories and competitors are tightly defined, and marketing occurs at the
point-of-purchase, which exposes all buyers of the category.

4.2. Diagnosing category evolution

Total category sales in the four product categories are depicted in Fig. 1. To
formally assess the stable or evolving character of these series, we applied
several versions of the Augmented Dickey Fuller (ADF) test. Table 1 first
lists the standard test statistics, in which we varied the number of lagged
difference terms (p) between 0 and 4, and subsequently used the Schwarz
Bayesian Criterion (SBC) to select the order of the test equation within
that range. We found evidence of I(1) or evolution in only one category,
soup, suggesting its potential for long-run market-expansion effects. The deter-
gent, catsup and yogurt markets, on the other hand, were all found to be
stationary.
The graphs in Fig. 1 also suggest the presence of occasional outliers, some of
which are due to price promotions. As shown in Franses and Haldrup (1994),
not accounting for these outliers might produce spurious stationarity. We
278 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

Table 1
Unit-root tests

A. Primary demand
ADF p ADFO p ADF w/entries p
Detergent !2.97 2 !3.26 1 !8.31 0
Yogurt !4.29 0 !4.52 0 !5.21 0
Catsup !3.46 4 !11.95 0 —
Soup !2.14 2 !2.15 2 —

B. Detergent sales B. Detergent prices

Tide !3.58 2 !6.13 0 Tide !2.21 2 !4.36 0
Wisk !7.35 0 !5.59 1 Wisk !2.79 1 !3.28 1
Era !6.76 0 !8.56 0 Era !4.1 0 !4.51 0

C. ½ogurt sales C. ½ogurt prices

ADF p ADFO p ADF p ADFO p
Dannon !3.04 2 !4.28 1 Dannon !8.12 2 !6.33 0
Yoplait !4.44 0 !3.94 0 Yoplait !2.82 2 !3.01 2
Private label !2.84 3 !5.84 0 Private label !3.22 2 !2.93 2

D. Catsup sales D. Catsup prices

ADF p ADFO p ADF p ADFO p
Hunts !6.55 2 !6.65 1 Hunts !5.3 0 !5.68 0
Delmonte !5.44 1 !7.27 1 Del Monte !3.42 3 !6.72 0
Heinz !8.88 0 !9.4 0 Heinz !3.5 2 !3.17 2
Private label !7.91 0 !8.69 0 Private label !3.66 0 !4.07 0

E. Soup sales E. Soup prices

ADF p ADFO p ADF p ADFO p
Swanson !5.3 0 !3.67 0 Swanson !5.3 0 !3.32 1
Campbell !2.27 2 !2.26 2 Campbell !3.61 1 !5.32 0
Private label !7.65 0 N.A. 0 Private label !2.93 2 N.A. 0

Critical values (5%): !2.89 (ADF and ADFO), !3.33 (ADF w/ entries).

assessed in a stepwise fashion whether or not these outliers affected our

unit-root test results. First, we controlled for price promotions — characterized
by simultaneously high sales and low prices — with a single dummy variable.
Next, we added separate dummy variables for additional outliers that were
identified as isolated extreme points in a variable’s histogram. As shown in
Table 1, ADFO column, our substantive conclusions were not affected.
Two of our categories (detergent and yogurt) experienced two new-product
entries in the considered time span, which could have caused a structural break

 See Abraham and Lodish (1993) for a conceptually similar two-step procedure.
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 279

in the mean level of primary demand. Even though not accounting for such
events has been shown to bias the test results toward the unit-root conclusion
(Perron, 1990), there was still no evidence of evolution in those series. When
applying the testing procedure in Perron and Vogelsang (1992b) to explicitly
account for these a priori determined potential break points, the evidence in
favor of stability became even stronger (Table 1, ADF w/ entries-column).
Finally, we applied the testing procedure in Perron and Vogelsang (1992a)
where the potential break-point is not set a priori but rather determined
empirically, to the evolving soup series. Again, our results were found to be
robust, in that the unit-root null hypothesis was not rejected with this testing
procedure either (min t "!2.64, with t "!4.25).
?  
The combined findings of the different unit-root tests indicate that the
potential for long-run market expansion effects of price promotions is restricted
to the soup market. In the three other instances, we found that, in spite of
numerous marketing interventions by the incumbents over the considered time span,
category demand for detergents, catsup and yogurt behaved as a series of fluctu-
ations around a constant mean.

4.3. Diagnosing the incumbents+ selective demand and pricing behavior

The unit-root test results for the brands’ sales series are also presented in
Table 1. While the ADF test occasionally indicates a unit root, the outlier-
controlled ADF tests reject the unit-root hypothesis in all but one case. Again,
we conclude that in spite of numerous marketing interventions, the demand
fluctuations of all incumbent catsup, detergent, yogurt and soup brands are
stable. The sole exception is the mild evolutionary behavior in the sales of
Campbell which, given Campbell’s dominant share in the soup category, is also

 We also applied formal outlier-detection methods, in particular Chen and Liu (1993), but were
limited by the many potential outlier observations due to repeated promotional activity, relative to
the time sample. Extensive sensitivity analyses were therefore performed to ensure that the results
were insensitive to the choice of outliers. For example, when adding or subtracting one data point
identified as an outlier relative to the number used in the ADFO column of Table 1, the substantive
findings were unaffected.
 For the stable series (i.e. where the unit-root null hypothesis was rejected), the question
remains whether there is any evidence of a structural change in the parameter values, and if so,
whether this change could be attributed to a marketing (in casu, promotional) activity.
Successive-window tests, as implemented in the E-views 2.0 software, were performed on the
parameter estimate for the lagged dependent variable in the unit-root test equations.
A visual inspection revealed very stable estimates, and hence, no evidence of such a break (detailed
results are available from the authors upon request). Subsequent analyses on the impact of price
promotions over time will therefore assume that these shocks will not cause such a structural change
either.
280 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

found at the primary-demand level. In terms of our strategic scenarios, the

following classification results:

Stable brand sales Evolving brand sales

Stable category sales Catsup brands (4)

Detergent brands (3)
Yogurt brands (3)

Evolving category sales Swanson soup Campbell soup

Private Label Soup

Most brands (10) fall in the stable-brands/stable-industry category, which im-

plies that their relative positions (market shares) will only be temporarily
affected by marketing activities as well. Marketing activities for these brands
may still have market-expansive and/or selective demand effects, and these may
differ in magnitude and duration across brands (cf. Section 4.4). However, the
effects are all bound to be temporary in nature.
The finding of predominantly stable sales is corroborated by the unit-root test
results on prices, since all thirteen ADF0 tests reject the null hypothesis of a unit
root in prices (Table 1). The test statistics, however, were generally lower in
absolute value for prices than for brand sales, indicating that prices will, on
average, take a longer time to return to their pre-shock mean than consumer
sales. This observation will be validated in Section 4.4 through multivariate
impulse-response calculations.
In sum, while the statistical power of each individual unit-root test could be
improved (Perron, 1989), the combined evidence of finding stationarity in 28 out
of 30 different series is very strong.

4.4. Diagnosing the effectiveness of price promotions

Unit-root tests are only indicative of the potential for long-run marketing
effectiveness. To capture the potential long-term impact of price promotions,
multivariate analyses are needed linking price promotions to the longitudinal
behavior in performance. We therefore derived impulse-response functions from
VAR models specified in the levels of the stable variables, and in the first

 Also in this case, an application of the Perron and Vogelsang (1992a) procedure confirmed the
presence of a unit root (min t "!2.71 for Campbell sales).
?
 Similar findings were reported by Lal and Padmanabhan (1995) who conducted deterministic-
trend regressions on the market shares of multiple frequently purchased consumer goods, and found
few significant slope coefficients.
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 281

difference for the evolving variables (i.e. soup primary demand and Campbell
selective demand). To avoid potential degrees-of-freedom problems when esti-
mating extended VAR models (e.g. when several competitors’ performance and
marketing-mix variables are included simultaneously as endogenous variables),
we estimated separate models for category and brand sales. Furthermore, we
estimated a separate VAR model for each brand’s selective demand. Apart from
the performance measure, we included the prices of all major competitors in
each VAR model, and controlled for the presence of features and display
through exogenous dummy variables. In the detergent and yogurt analyses,
two additional step dummy variables were added to control for the new-brand
entries. As an example for the detergent market, the following VAR model
was used to derive the over-time impact on the primary demand for liquid
detergents (S ) of price shocks to the major brands Tide (¹I), Wisk (¼I) and

Era (E):

    
s C G G G G S
 R       R\
G G G G
2' R " 2' # '
P C P
    ; 2' R\
P C G G G G P
5' R 5' G     5' R\G
P C G G G G P
#R #     # R\G

  
g$ g$ g$
   F
g$ g$ g$ 2' R
#    ; F
g$ g$ g$ 5' R
   F
g$ g$ g$ #R
  

  
g" g" g"
   D
g" g" g" 2' R
#    ; D
g" g" g" 5' R
   D
g" g" g" #R
  

    
dd e
  R
d d entry e
#   ;  # 2' R
d d entry e
   5' R
d d e
  #R
where D indicates whether or not Tide was on display in any given week
2'
(similar definitions apply for the other feature and display variables), entry
G
(i"1,2) is a step dummy variable, and I is the order of the VAR model
determined on the basis of the SBC criterion.
282 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

As with any econometric model, restrictions on the information set may affect
one’s conclusions. We therefore tested our findings for consistency across
primary and selective demand specifications (cf. infra) and, in unreported ana-
lyses, we verified the robustness of the results across alternative specifications of
the final VAR model. Impulse-response functions derived from VAR models
have been criticized because of the ambiguity that arises when the equation
error terms are correlated. Following Dekimpe and Hanssens (1995a), we
imposed a temporal ordering on the endogenous variables, in that we always
assigned causal priority to the price variable that was shocked, and always
ordered the performance variable last, so that it could be influenced instan-
taneously by all prices. Putting sales last in the temporal sequence makes
intuitive sense when working with weekly data, as firms cannot immediately
adjust retail prices to incoming market information. For the ordering among the
price variables, numerous robustness checks were conducted, and our results
were not sensitive to the imposed causal ordering. This robustness is due to the
fact that all instantaneous cross-price effects, as reflected in the residual correla-
tion matrix, were very small. This corroborates Leeflang and Wittink’s (1992)
finding that it takes firms and retailers some time to react to competitive price
promotions.
Price promotions are temporary (Blattberg et al., 1995) price reductions
offered to the consumer, which can be operationalized as one-time shocks to the
VAR system. To enhance the comparability of the findings across brands and
product categories, we trace the over-time impact of a one-standard-error shock.
This approach captures price promotions in relative rather than in absolute
dollar terms (Blattberg et al., 1995), and reflects that the perceived depth of
a monetary discount depends on the unpredictability of the action (Leeflang and
Wittink, 1996).

4.4.1. Price dynamics in the stable catsup market

The market-expansive effects of the various brands in the catsup market are
depicted in Fig. 2A, which shows that both the magnitude and duration of their
temporary impact on category sales differ considerably. We observe a larger
instantaneous effect for the bigger brands (especially Heinz and Hunts), while
the private-label brand has a more prolonged impact. Table 2 compares the
different brands in terms of their total cumulative primary-demand effect. In
computing this total, we summed all impulse-response weights with a t-statistic
greater than one in absolute value (see Dekimpe and Hanssens (1995a), Pesaran
et al. (1993), or Van de Gucht et al. (1996) for a conceptually similar procedure).
This cutoff point is generous and translates into fairly wide confidence intervals,
causing our results to be indicative, rather than precise estimates of the resulting
cumulative effectiveness. When interpreting our findings, we therefore focus on
(1) the overall form of the response functions, and (2) consistent patterns across
product categories.
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 283

Fig. 2. (a) Market expansion effects in the stable catsup market. (b) Market expansion effects in the
evolving soup market.

Primary-demand effects are due to a variety of factors, such as the attraction

of new customers to the product category, increased consumption by current
buyers, purchase acceleration and stockpiling. The latter two factors have an
inherent temporary character. However, the absence of persistent effects sug-
gests that the first two factors have, at most, a short-lived effect as well. As for
284 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

Table 2
Short- and long-run market-expansion effects of price promotions

Product category Brand Cumulative effect Persistence

Catsup Heinz 16,198 0

Hunts 4717 0
Del Monte !4901 0
Private label 10,332 0

Liquid detergent Tide 0 0

Wisk 15,675 0
Era 16,634 0

Yogurt Yoplait 11,134 0

Dannon 1638 0
Private label 10,999 0

Soup Campbell R !8703

Swanson R !6754
Private label R 8749

Effects resulting from a one standard error price promotion.

Table 3
Selective-demand effects of price promotions by the market leader

Market leader Cumulative effect on own Persistent effect on own

performance performance

Heniz 21,879 0
Tide 25,776 0
Yoplait 6418 0
Campbell !R !9159

the selective-demand effects, we report in Tables 3 and 4 the cumulative own-

demand effects of price shocks to the market leader (Heinz) and the private-label
brand. When comparing these figures with their market-expansion effects, we
see that the competitive implications differ vastly. For Heinz, its own sales gains
(21,900) exceed its market-expansion effect (16,200), implying that a substantial
fraction of its gains come at the expense of the competition. Price promotions for
the private-label brand, on the other hand, benefit some competitors, as its
market-expansion effect (10,300) is greater than its own gains (1400). This
supports the scenario that price promotions by the cheaper private label brand
temporarily attract new buyers to the product category, who also try out some
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 285

Table 4
Selective-demand effects of price promotions by private-label brands

Product category Cumulative effect on own Persistent effect on own

performance performance

Catsup 1402 0
Yogurt 5592 0
Soup 6845 0

of the national brands in their repeat purchases. Part of this expansion effect of
the private label brand, however, is also due to price-matching behavior of the
two leading national brands, Heinz and Hunts.
Finally, we observe that, consistent with the univariate results in Section 4.3,
prices exhibit a substantial degree of ‘stickiness’: they revert to the mean at
a slower rate (i.e. after approx. 8 weeks) than their corresponding sales level. This
price stickiness has a negative effect on the profitability of promotional plans
because customers can take advantage of a still lower than average price even
though their purchase quantities have already returned to mean levels.

4.4.2. Price dynamics in the stable yogurt and detergent markets

The results in yogurt and detergents follow similar patterns: (1) all primary-
and selective- demand effects of price promotions are temporary, (2) price
promotions by the private-label brand cause a market-expansion effect that
partially accrues to the national brands, and (3) the price fluctuations of
a majority of the brands are stickier than their customers’ responsiveness. In the
detergent market, price promotions by the market leader (Tide) are competitive
only, in the terminology of Schultz and Wittink (1976). They cause a substantial
own-demand effect that is completely at the expense of the other brands. Overall,
the dynamics of price promotions and market response are comparable across
these three different product categories.

4.4.3. Price dynamics in the evolving soup market

The primary-demand effects in the soup market are depicted in Fig. 2B. The
incremental impact of price promotions by the three major brands does not
converge to zero as in the stable catsup market (Fig. 2A), but stabilizes at
a non-zero level. For the leading brand, Campbell, we see a large immediate
impact of 59,700. This differential impact then oscillates over time, and eventual-
ly reaches an asymptotic persistence level of !8,703, which is negative but small
relative to the initial positive impact. Similar results are obtained for Campbell’s
selective demand (see Table 3). Thus, even though Campbell’s price promotions
are highly effective in the short run, they set into motion a set of opposite forces
286 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

that cause the eventual long-run impact to be self-canceling or even damaging to

the brand.
Once again, a different picture is obtained for the private-label brand. Its price
promotions have a persistent market-expansion effect, even though they only
temporarily benefit its own performance. Private-label price promotions appear
to be powerful enough to either increase consumption or to attract previous
non-buyers to the category who subsequently become national-brand buyers.
This scenario is validated when analyzing the private-label’s promotional
cross-effect on the sales of Campbell, with a persistent positive impact of 3800.
In conclusion, the observed evolution in the soup category and market leader
sales is somewhat, but not strongly affected by price promotions. Their predomi-
nant effects occur in the short run, and are different for the market leader vs.
private label.

5. Discussion

Our unit-root based econometric analyses of thirteen brands in four product

categories have quantified the short- and long-run effects of price promotions,
both at the primary- and selective-demand level. Even though there is variability
across brands and product categories, some empirical generalizations emerge:

E Category sales, brand sales and brand prices in scanner markets generally
follow a mean-stationary process. In a majority of cases, there is faster mean
reversion for sales than for prices.
E Price promotions have a temporary impact on the brand’s and the market’s
future sales levels. Only in the highly concentrated soup market is there
evidence of a long-run promotion effect, with relatively low persistence.
E Even though price promotions by the market leader tend to have the largest
immediate effect, their cumulative impact is more limited, and mostly com-
petitive in nature.
E Private-label brand promotions, on the other hand, can expand the market
and actually enhance the performance of national brands.

The latter findings establish the existence of another asymmetry between price
promotions initiated by national vs. private-label brands. Previous research had
provided evidence that price promotions by national brands have a higher
instantaneous drawing power (e.g. Blattberg et al., 1995; Bronnenberg and
Wathieu, 1996). This phenomenon was also observed in our impulse-response
functions. However, when considering the total over-time impact on both
selective and primary demand, this conclusion should be qualified. Moreover,
our findings invite national brands to rethink their perception of private-label
brands as detrimental to their long-run viability. Instead, the behaviors we
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 287

observe are more akin to co-opetition than to competition (Brandenburger and

Nalebuff, 1996). While private-label and national brands visibly compete for
market share, they may mutually benefit in other dimensions, for example the
stimulation of primary demand.
Our findings also invite some scrutiny of the question as to why there is so
much observed stationarity in performance and marketing effort. After all,
brand managers are compensated and motivated to improve their market
position and profitability over time, so one would expect to see evolutionary
behavior in sales and marketing spending. Instead, we find that most position
improvements are temporary in our four major consumer product categories.
There are two possible explanations:

(a) the marketing-mix variables of these brands, as well as the cross-effects

from their competitors, intrinsically have only temporary effects. Whether or not
competitive reaction is desirable depends on the trade-off between lower sales
with same marketing costs vs. same sales with higher marketing costs.
(b) they intrinsically have long-run effects, but because of competitive activ-
ities, they cancel each other out in the long run. In this scenario, a brand
manager has no choice but to respond to an aggressive action of a competitor,
such as a price promotion, lest (s)he wants to risk the permanent loss of sales.

In Appendix A, we address these explanations analytically, and derive

whether or not brand actions and counteractions that intrinsically have long-
run effects can produce a time series of sales that is mean-stationary, in the
absence of structural breaks in the data-generating process. We consider three
competitive scenarios: firms set their budgets independently, a leader/follower
scenario, and both firms set their budget as a function of the other brand’s
decision. We show that case (b), where the apparent stationarity of sales would
mask competing long-run effects cannot occur in the first two competitive-
reaction scenarios, and is unlikely to occur in the third scenario. ¹hus, actual ‘do
or die’ promotional wars are unlikely to exist in stationary markets and many of the
promotional battles one observes in the current market place may actually reflect
unnecessary escalations.
This analytical result is tight and our combined empirical evidence of sales
stationarity in scanner markets is strong. Still, in some instances it is difficult to
empirically designate an individual time series as being stationary or evolving
with a low persistence level. More work is therefore needed to enhance the
power of unit-root tests. Moreover, for any selected size (a) of the empirical
testing procedure, the possibility remains that one erroneously rejects the
unit-root hypothesis. As such, one should not be surprised to observe unnecess-
ary promotion escalations, as risk-averse managers with asymmetric loss func-
tions are likely to assume persistent competitive effects and react accordingly
even if the evidence for their brand and category is weak.
288 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

Acknowledgements

The authors are indebted to Dr. Keiko Powers for excellent research assist-
ance.

Appendix A

We consider whether brand actions and counteractions that intrinsically have

long-run effects can still result in performance series which are mean-stationary.
For the sake of simplicity, we assume that both prices (P and CP) are mean-
reverting, but a similar reasoning applies when the control variables are evolv-
ing. Three scenarios are considered, which have been observed repeatedly in
empirical research (see, e.g. Hanssens et al., 1990):

A.1. Case 1. Independent price setting

The simplest case of independent price setting is given by

S "a #b (¸)P #b (¸)CP #e , (A.1a)
R Q  R  R 1R
P "a #e , (A.1b)
R . .R
CP "a #e , (A.1c)
R !. !. R
where e , e and e are white-noise residuals, and where
1R .R !. R
cov(e , e )"0, ∀i. After appropriate substitutions, we get
. R !. R>G
S "a*#b (¸)e #b (¸)e #e . (A.2)
R  .R  !. R 1R
A temporary price reduction will have a continuing impact if the partial
derivative of S (kPR) with respect to e is non-zero (assuming, as before,
R>I .R
that they do not cause a structural break). Obviously, this can only occur if b (¸)

is an infinite-lag polynomial whose coefficients do not converge to zero. In that
case, however, the variance of the right-hand side of Eq. (A.2) will grow without
bound, while the left hand-side (S ) is a finite-variance variable, which would
R
create an inconsistent model specification (Granger, 1981). Hence, b (¸) cannot

be an infinite-order polynomial, and neither P nor CP can have a continuing
impact.

A.2. Case 2. One firm is the leader, the other the follower

In this second scenario, Eq. (A.1c) is changed to reflect the fact that the
competitor sets his/her prices as a function of our current and/or past prices,
CP "a #i(¸)P #e , (A.3)
R !. R !. R
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 289

with cov(e , e )"0, ∀i. After appropriate substitutions, we get

. R !. R>G
S "a*#[b (¸)#b (¸)i(¸)]e #b (¸)e #e . (A.4)
R   .R  !. R 1R
Using a similar reasoning, P and CP can only have a continuing (and supposed-
ly canceling) impact if b (¸) and b (¸) are infinite-lag polynomials whose
 
coefficients do not converge towards zero. This situation leads to a similar
inconsistency. Even in the unlikely event that the i( ) coefficients cancel ‘an
infinite number of contributions to the total variance’ in the term between
square brackets, the b -terms would still cause an infinite-variance right-hand

side.

A.3. Case 3. Both firms react to each other’s (current and past) prices

Eq. (A.1b) is updated to reflect this new scenario:

P "a #i (¸)CP #e , (A.5a)
R   R .R
CP "a #i (¸)P #e . (A.5b)
R !.  R !. R
After appropriate substitutions, the performance equation becomes
[b (¸)i (¸)#b (¸)] [b (¸)i (¸)#b (¸)]
S "a*#    e #    e #e .
R 1!i (¸)i (¸) !. R 1!i (¸)i (¸) .R 1R
   
(A.6)
In this case, it is possible that, even when b (¸) and b (¸) are infinite-lag
 
polynomials (and thus reflect underlying long-run effects), both the left- and
right-hand side have a finite variance. However, this masking of long-run effects
would only occur if both competitors react in such a way that they completely
cancel out the other firm’s promotional effect for an infinite number of periods to
come. It is only if this very stringent (and therefore unlikely) condition is met,
that one would see a masking of underlying long-run or continuing effects.

References
Abraham, M.M., Lodish, L.M., 1993. An implemented system for improving promotion productivity
using store scanner data. Marketing Science 12, 248—269.
Allenby, G.M., Rossi, P.E., 1991. There is no aggregation bias: Why macro logit models work.
Journal of Business and Economic Statistics 9, 1—14.

 To give a specific example, if b (¸) is an infinite-lag polynomial, it is not sufficient that the

competitive reactions result in a net zero effect in period 1, period 2,2, period 20; they must cancel
the firm’s promotional effect in so many periods that there are only a finite number of periods which
offer a contribution to the variance on the right-hand side.
290 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291

Baghestani, H., 1991. Cointegration analysis of the advertising—sales relationship. Journal of

Industrial Economics 34, 671—681.
Bemmaor, A.C., Mouchoux, D., 1991. Measuring the short-term effect of in-store promotions and retail
advertising on brand sales: A factorial experiment. Journal of Marketing Research 28, 202—214.
Brandenburger, A.M., Nalebuff, B.J., 1996. Co-opetition. Doubleday, New York.
Blattberg, R.C., Briesch, R., Fox, E.J., 1995. How promotions work. Marketing Science 14,
G122—G132.
Bronnenberg, B.J., Wathieu, L., 1996. Asymmetric promotion effects and brand positioning. Market-
ing Science 15, 379—394.
Bucklin, R.E., Gupta, S., 1996. Business adoption of advances in the analysis of scanner data. Paper
presented at the Marketing Science Institute Conference on ‘Use and Usability: Business-
Focused Market Research’, Boston, MA.
Chen, C., Liu, L.M., 1993. Joint estimation of model parameters and outlier effects in time series.
Journal of the American Statistical Association 88, 284—297.
Chintagunta, P.K., 1993. Investigating purchase incidence, brand choice and purchase quantity
decisions of households. Marketing Science 12, 184—208.
Chowdhury, A.R., 1994. Advertising expenditures and the macro-economy: Some new evidence.
International Journal of Advertising 13, 1—14.
Dekimpe, M.G., 1992. Long-run modeling in marketing. Unpublished Ph.D. Thesis, UCLA, Los
Angeles.
Dekimpe, M.G., Hanssens, D.M., 1995a. The persistence of marketing effects on sales. Marketing
Science 14, 1—21.
Dekimpe, M.G., Hanssens, D.M., 1995b. Empirical generalizations about market evolution and
stationarity. Marketing Science 14, G109—G121.
Dekimpe, M.G., Steenkamp, J.B.E.M., Mellens, M., Vanolen Abeele, P., 1997. Decline and variabil-
ity in brand loyalty. International Journal of Research in Marketing 14, 405—420.
Franses, P.H., 1994. Modeling new product sales; An application of cointegration analysis. Interna-
tional Journal of Research in Marketing 11, 491—502.
Franses, P.H., Haldrup, N., 1994. The effects of additive outliers on tests for unit roots and
cointegration. Journal of Business and Economic Statistics 12, 471—478.
Glazer, R., Weiss, A.M., 1993. Marketing in turbulent environments: decision processes and the
time-sensitivity of information. Journal of Marketing Research 16, 509—521.
Granger, C.W.J., 1981. Some properties of time series data and their use in econometric model
specification. Journal of Econometrics 16, 121—130.
Gupta, S., 1988. Impact of sales promotions on when, what and how much to buy. Journal of
Marketing Research 25, 342—355.
Hanssens, D.M., Parsons, L.J., Schultz, R.L., 1990. Market Response Models: Econometric and
Time Series Analysis. Kluwer Academic Publishers, Boston.
Harris, R., 1995. Cointegration Analysis in Econometric Modelling. Prentice Hall, Wheatsheaf.
Lal, R., Padmanabhan, V., 1995. Competitive response and equilibria. Marketing Science 14,
G101—G108.
Leeflang, P.S.H., Wittink, D.R., 1992. Diagnosing competitive reactions using (aggregated) scanner
data. International Journal of Research in Marketing 9, 39—57.
Leeflang, P.S.H., Wittink, D.R., 1996. Explaining competitive reaction effects. Working paper,
University of Groningen.
Leone, R.P., 1987. Forecasting the effect of an environmental change on market performance: An
intervention time-series approach. International Journal of Forecasting 3, 463—478.
Mela, C., Gupta, S., Lehmann, D.R., 1997. The long-term impact of promotion and advertising on
consumer brand choice. Journal of Marketing Research 34, 248—261.
Mills, T.C., 1994. The Econometric Modelling of Financial Time Series. Cambridge University
Press, Cambridge.
 M.G. Dekimpe et al. / Journal of Econometrics 89 (1999) 269–291 291

Papatla, P., Krishnamurthi, L., 1996. Measuring the dynamic effects of promotions on brand choice.
Journal of Marketing Research 33, 20—35.
Perron, P., 1989. Testing for a random walk: A simulation experiment of power when the sampling
interval is varied. In: Raj, B. (Ed.), Advances in Econometrics and Modeling. Kluwer Academic
Publishers, Boston, pp. 47—67.
Perron, P., 1990. Testing for a unit root in a time series with a changing mean. Journal of Business
and Economic Statistics 8, 153—162.
Perron, P., Vogelsang, T.J., 1992a. Nonstationarity and level shifts with an application to purchasing
power parity. Journal of Business and Economic Statistics 10, 301—320.
Perron, P., Vogelsang, T.J., 1992b. Testing for a unit root in a time series with a changing mean:
Corrections and extensions. Journal of Business and Economic Statistics 10, 467—470.
Pesaran, M.H., Samiei, H., 1991. Persistence, seasonality and trend in the UK egg production.
Applied Economics 23, 479—484.
Pesaran, M.H., Pierse, R.G., Lee, K.C., 1993. Persistence, cointegration and aggregation: A disag-
gregated analysis of output fluctuations in the U.S. economy. Journal of Econometrics 56, 57—88.
Pesaran, M.H., Smith, R., 1995. Estimating long-run relationships from dynamic heterogeneous
panels. Journal of Econometrics 68, 79—113.
Schultz, R.L., Wittink, D.R., 1976. The measurement of industry advertising effects. Journal of
Marketing Research 13, 71—75.
Van de Gucht, L., Dekimpe, M.G., Kwok, C.C.Y., 1996. Persistence in foreign exchange rates. The
Journal of International Money and Finance 15, 191—220.
Wind, Y., Robertson, T.S., 1983. Marketing strategy: New directions for theory and research.
Journal of Marketing 47, 12—25.
Zanias, G.P., 1994. The long run, causality, and forecasting in the advertising—sales relationship.
Journal of Forecasting 13, 601—610.