You are on page 1of 14

Intern. J.

of Research in Marketing 20 (2003) 273 – 286


www.elsevier.com/locate/ijresmar

The impact of retailer stockouts on whether, how much,


and what to buy
Katia Campo a,*, Els Gijsbrechts b, Patricia Nisol a
a
University of Antwerp, Prinsstraat 13, 2000 Antwerp, Belgium
b
University of Tilburg, Warandelaan 2, PO Box 90153, 5000 LE Tilburg, The Netherlands
Received 12 July 2001; received in revised form 12 December 2002; accepted 17 December 2002

Abstract

This paper investigates the impact of retailer stockouts on whether, how much, and what consumers buy in a category
during the out-of-stock period. It adds to previous literature by investigating the stockouts’ impact on purchase quantities, by
uncovering the pattern of within-category shifts triggered by them, and by analyzing dynamic effects on incidence, quantity,
and choice decisions. Although essential for an accurate assessment of stockout implications, these aspects of out-of-stock
(OOS) behavior have received little attention so far. Moreover, while most of the previous literature on stockouts relies on
surveys measuring reported or intended behavior, this research examines revealed stockout response in a natural store
environment. To this end, traditional purchase incidence, quantity and choice models are adapted to account for various
stockout effects, and estimated based on scanner panel data for two product categories. While tractable, the models capture
various aspects of stockout reactions within the store, allowing to trace the implications of stockouts for specific SKUs and
household (segment)s. The estimation results demonstrate that out-of-stocks can affect all three purchase decisions: they may
reduce the probability of purchase incidence, lead to the purchase of smaller quantities, and induce asymmetric, non-IIA
choice shifts. Simulation of aggregate implications further suggests that, on the whole, retailer losses from stockouts remain
limited, while manufacturer losses can be quite substantial, depending on the stockout item and the store’s assortment
composition.
D 2003 Elsevier B.V. All rights reserved.

Keywords: Stockouts; Quantity reduction; Asymmetric switching; Dynamic effects; Scanner panel analysis

1. Relationship to the literature and research the potential losses they represent for manufacturer
questions and retailer, was advocated by practitioners and mar-
keting scholars (see, e.g., Peckham, 1963; Schary &
Already in the 60s and 70s, the importance of out- Christopher, 1979; Walter & Grabner, 1975). Recent
of-stock (OOS) occurrences in retail outlets, as well as evolutions in the area of ECR, EDI, and Category
Management have led to a renewed interest in the
occurrence and consequences of stockouts (Campo,
* Corresponding author. Tel. +32-3-275-5045. Gijsbrechts, & Nisol, 2000; Corstjens & Corstjens,
E-mail address: katia.campo@ua.ac.be (K. Campo). 1999).

0167-8116/$ - see front matter D 2003 Elsevier B.V. All rights reserved.
doi:10.1016/S0167-8116(03)00037-5
274 K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286

Previous studies of reactions to retailer stockouts purchase decisions during the OOS period, they may
have indicated that the switch item response is predo- also influence whether, how much and what consum-
minant, and hence, that stockouts may entail sub- ers buy in post-OOS periods. Based on the inventory
stantial sales losses for manufacturers. At the same effects studied in the sales promotions literature (see,
time, a non-negligible portion of consumers seem to e.g., Neslin & Schneider-Stone, 1996), we expect that
switch to another store or to cancel the purchase unusually low home stocks—resulting from the deci-
altogether, implying that retailers, too, may experi- sion to buy nothing or less in response to an OOS—
ence important sales decreases due to stockouts. may lead to purchase acceleration and increased pur-
Stockout response was also found to vary substantially chase quantities in periods succeeding the stockout. In
among consumers, products and situations. An over- addition to these ‘rational’, inventory based motiva-
view of the factors explaining differences in OOS tions, dynamics in OOS reaction may be caused by
reaction is given in Corstjens and Corstjens (1999) psychological reactions. Situational influences that
and Campo et al. (2000). disrupt habitual purchasing can—in combination with
While generating valuable insights, OOS studies to imperfect information on the home inventory level—
date leave a number of issues uncovered, three of result in risk-reducing purchase strategies (Neslin &
which deserve special attention. First, existing papers Schneider-Stone, 1996). In the case of stockouts, this
on OOS reactions have typically concentrated on could lead consumers to buy sooner or more from the
purchase incidence and choice decisions. The pur- category than what is strictly needed to restore
chase quantity issue has been ignored, the implicit normal inventory levels. Also, having temporarily
assumption being that consumers either buy nothing been denied one’s favorite product may reinforce item
or purchase the regular quantity. Research in other preference, enhancing the item’s post-OOS choice
areas suggests that this assumption may be unrealistic probability.
and that consumers may decide to buy smaller quan- Insights into these immediate and delayed OOS
tities (smaller sizes or less units) to reduce the risk of a effects are of crucial importance (i) for an accurate
poor choice, or limit the loss in intrinsic utility assessment of potential OOS losses and (ii) to identify
associated with the purchase of a less preferred alter- appropriate measures for preventing stockouts or
native (Campo et al., 2000; Chiang, 1995). Second, reducing their negative consequences.
while studies to date do analyze whether consumers
switch to another item in case of a stockout, they shed
little light on which replacement item will be selected. 2. Methodology
Yet, as suggested by Anupindi, Dada, and Gupta
(1998), switching in response to an OOS may deviate To shed more light on these issues, traditional
from regular choice and switching behavior. When all purchase incidence, quantity, and choice models are
items are available, item switching predominantly adjusted to account for stockout effects, and estimated
results from positive incentives, such as promotions based on scanner panel data of a large European
or a desire for variety. In the case of stockouts, supermarket chain. This procedure has several advan-
however, the decision to switch to another item is tages over the traditional, survey based research ap-
not purely voluntary and its underlying motivation is proach applied in most previous OOS studies. Using
essentially negative in nature (e.g., to avoid an extra behavioral data may reduce some of the validity threats
visit to a different store). It follows that the group of associated with the use of reported purchase intentions
switchers may comprise consumers who do not nor- and behavior. The fact that surveys somewhat artifi-
mally change items and who may attach a higher cially draw attention to product unavailability may, for
importance to specific product attributes (e.g., brand instance, entail a risk of overestimating OOS reaction
name, as a risk-reducing indicator of product quality). strength. Also, individual buying behavior models
Stockouts may for these reasons produce dispropor- allow for a tractable yet flexible approach to OOS
tionate choice shifts towards some items, compared to reaction analysis, making it possible to account for
the regular preference and choice distribution. Third, mixed response strategies (e.g., combinations of item
stockouts may not only have an immediate impact on switching and purchase quantity reduction).
K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286 275

2.1. Purchase incidence and quantity models variable equal to 1 when item i is out-of-stock in t and
equal to 0 otherwise.
Purchase incidence—or the probability that a house- The definition in Eq. (3) is checked against two
hold h will engage in a category purchase on a given alternative definitions, which are also embedded in the
purchase occasion t (PIth)—is modeled using a logistic literature. One is based on the notion that propensity to
function (see, e.g., Bucklin, Gupta, & Siddarth, 1998a; purchase from the category declines when there is a
Bucklin & Lattin, 1991). The probability that—given substantial difference in utility between the OOS and
incidence—a quantity qth will be purchased (PQth) is second preference item(s) (see, e.g., Corstjens &
modeled as a truncated Poisson process (see, e.g., Corstjens, 1999),1 the other that especially consumers
Dillon & Gupta 1996). Purchase utility in the incidence who are uncertain about their preference for the
model (Wth) and purchase rate in the Poisson model (kth) remaining items, will be less inclined to purchase from
are linked to traditional household characteristics and the category (Dhar, 1997).2 Note that the category
h
marketing mix variables (Zk,t ), further specified below. stockout definitions integrate information from the
A special feature of our models is that they account for choice model into the incidence model, making pur-
the effect of stockouts in the category (COOS), in the chase incidence probabilities dependent on expected
current as well as the previous period: choice results.3 The coefficients of the COOS variable

expðWth Þ 1
PIht ¼ In this definition, the (importance of the) category stockout is
1 þ expðWth Þ measured as the difference in systematic utility between (i) the item
 X  with the highest systematic utility when all items are available
h
exp h0 þ hk Zk;t þ f0 COOSht þ f1 COOSht1 h
(U max, t /AV) and (ii) the item with the highest systematic utility
h
k when one or more items are OOS (U max, t /OOS):
¼  X 
ðIÞh h
1 þ exp h0 þ h h h
hk Zk;t þ f0 COOSt þ f1 COOSt1 COOS t ¼ U max;t =AV  U hmax;t =OOS:
k 2
Using entropy as an indicator of preference uncertainty, this
ð1Þ leads to the following alternative definition for the category
stockout variable:
h
expðkht Þ*ðkht Þqt " #
X  
PQht¼ ðIIÞh
COOS t ¼  pchit *ln pchit when there is at least
ð1  expðkht ÞÞ*factorialðqht Þ
 X
i available at t

h h h
kt ¼ expðVt Þ ¼ exp b0 þ bk Zk;t þ g0 COOSht one item out-of -stock in the category,
k ¼ 0 otherwise:

3
þ g1 COOSht1 ð2Þ Interdependencies between incidence and choice decisions are
incorporated in a similar way, through definition of the category
marketing mix variables as weighted sums of item marketing
Given that reactions to item unavailability strongly variables (see below). A similar approach was followed by Dillon
depend on the consumer’s preference for the missing and Gupta (1996), Bucklin, Gupta and Siddarth (1998b), and Jedidi
et al. (1999). Compared to the alternative inclusive value approach
item (see, e.g., Broniarczyk, Hoyer, & McAlister, (see, e.g., Bucklin & Lattin, 1991; Bucklin et al., 1998a), this
1998; Campo et al., 2000; Fitzsimons, 2000), we procedure has the advantage of being less restrictive and of better
specify the category stockout variable (COOS) as a isolating interdependencies due to stockouts. Indeed, instead of one
weighted sum of item stockouts. As weights, we use the inclusive value coefficient, separate incidence parameters are
item’s relative attractiveness for the household, i.e., its estimated for the category price, promotion, and OOS variables—
where especially the latter are of interest to us. To further compare
probability of being selected by the household if all both approaches, we re-estimated (single segment) incidence and
items would have been available: quantity models (1) replacing the aggregate price, promotion, and
X stockout variables with the inclusive value variable; and (2) adding
COOSht ¼ pchit *OOSit ð3Þ the inclusive value variable to the aggregate (category level)
i variable. Only for incidence and cereals did the model with inclusive
value provide a slightly better fit (CAIC, BIC), in all other cases the
where pcith=the
choice probability of item i for house- models with aggregate variables only were found to have superior
hold h in t when all items are available; OOSit=dummy descriptive value.
276 K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286

in expressions (1) and (2) (f0 and g0) are expected to levels relevant for attribute A; DA,l,i=dummy variable
be negative, pointing to a reduced likelihood of a equal to 1 if item i is characterized by level l on attribute
h
category purchase in case of a (more prominent) A, and to 0 otherwise; Zk,it =household and marketing
stockout. The delayed effect coefficients, f1 and g1, mix variables, further specified below; OOSA,it=stock-
on the contrary, are expected to be positive, the out asymmetry variable: number of items similar to i
tendency (quantity) to buy from the category being on attribute A that are OOS in t; OOSi,t1=dummy
reinforced by previous stockouts.4,5 variable equal to 1 if item i was out of stock in t1 and
to 0 otherwise.
2.2. Choice model Recognizing that choices for FMCG are typically
attribute-based (see, e.g., Boatwright & Nunes, 2001;
To model choice in the potential presence of stock- Hoch, Bradlow, & Wansink, 1999), the first term
outs, the traditional multinomial logit (MNL) model expresses item utility explicitly as a function of its
expression is modified in two ways. First, when an item relevant product features. Like Fader and Hardie
is not on the shelves, obviously, it cannot be selected: (1996), we consider discrete attributes (e.g., ‘brand’,
the variable OOSit takes it out of the choice set: ‘size’), each with a finite number of attribute levels
(e.g., different brand names or package sizes). For
expðUith Þ*ð1  OOSit Þ each attribute-level, a separate constant is introduced
pcith ¼ X ð4Þ
½expðUkth Þ*ð1  OOSkt Þ in the item utility expression. These attribute con-
k stants aA,l typically reflect the consumers’ base pref-
erences when all items are available. When OOS
where Uith=utility of item i for household h in t and
switching patterns do not deviate from ‘normal’
OOSit=1 when item i is out-of-stock in t and 0
switching behavior, the attribute-based choice model
elsewhere.
will suffice to capture stockout-induced choice shifts.
Second, for any item i in the category, the utility is
Yet, following Section 1, switching in response to
adjusted to allow for non-IIA switching patterns in
stockouts may not correspond to the regular prefer-
OOS situations and for carry-over effects of previous
ence distribution, and cause disproportionate shifts
stockouts:
towards items with specific attributes. To allow for
XX X X these NON-IIA patterns, we also include attribute-
Uith ¼ aA;l *DA;l;i þ h
dk Zk;it þ qA OOSA;it specific ‘OOS asymmetry variables’ (OOSA,it) in the
A laLA k A
model, which are set equal to the number of items
þ u OOSi;t1 ð5Þ that are OOS in t and similar to the considered item
on attribute A. For example, if one item of brand B is
where A=set of attributes relevant to the product cate- unavailable on occasion t, the brand asymmetry
gory (e.g., brand, flavor, and size); LA=index set of variable, OOSBrand,it, is set equal to 1 for all other
4 items i of brand B in the assortment. A positive
Positive post-stockout effects already occur through reduced
inventory levels (see Section 1). The lagged stockout coefficient (negative) coefficient associated with the OOSA,it
measures the potential presence of post-stockout peaks over and variable would point to asymmetric switching to-
above what is explained by (rational) inventory effects. wards (away from) alternatives that are similar to
5
Note that our models are ‘reduced form’ models, i.e., they do the missing item on attribute A (e.g., items of the
not follow from a dynamic utility framework with forward-looking
same brand).6 Apart from immediate effects, stock-
consumers. As shown by Erdem, Imai, and Keane (2002), this may
result in biased choice estimates, even if (besides choice) incidence outs may impact on choice probabilities in subse-
decisions are explicitly modeled in a nested logit model framework. quent periods through preference reinforcement for
Empirical results demonstrate that this is indeed the case for price and
promotions effects, as a result of consumers integrating future price
6
and promotion expectations in current purchase incidence decisions Note that our model bears some resemblance to the ‘cluster
(Erdem et al., 2002; Sun, Neslin, & Srinivasan, 2002). Whether this asymmetry’ model of Vanden Abeele, Gijsbrechts, and Vanhuele
also holds for stockout events—the core issue of this paper—is not (1990)—which also considers attribute-based asymmetric competi-
clear and should be tested in future research. We thank an anonymous tion—but in the context of aggregate market share models with no
reviewer for pointing this out. stockouts.
K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286 277

the OOS item. This is captured by the lagged OOS tively. Data from 626 (499) households for cereals
variable, OOS i,t1 , the coefficient of which is (margarine) were used for model estimation, providing
expected to be positive. information on 31 599 (19 999) store visits and 6069
(5563) purchases for cereals (margarine). Table 1 gives
2.3. Estimation an overview of the variables included in the models.
Besides attribute constants and OOS asymmetry vari-
To accommodate household heterogeneity in OOS ables (see Section 2), the choice model includes SKU
(and other) reactions, we use latent class estimation price, promotion, and dynamic loyalty variables (see
for each of the three buying behavior models. Like Ailawadi, Neslin, & Gedenk, 1999; Bucklin et al.,
Dillon and Gupta (1996) and Jedidi, Mela, and Gupta
(1999), we start out with a sequential estimation
approach. In a first step, we obtain estimates for Table 1
the incidence, quantity, and choice models separately. Variables in the choice, incidence, and quantity models
These estimates are then used as inputs for one Variable Description Model
iteration of joint likelihood estimation (see Jedidi et OOSit item stockout variable, equal to 1 if choice
al., 1999). This procedure substantially improves item i is out of stock in t, 0 otherwise
tractability, while preserving the possibility to allow OOSA,it stockout asymmetry variable for attribute choice
A, equal to the number of items similar
for correlated errors among the three decision
to i on attribute A that are OOS in t
stages.7 PRICEit price of item i in period t. Because regular choice
prices hardly varied during the analysis
period, this price variable mainly captures
3. Empirical analysis the effect of promotional price cuts.
PROMOit dummy variable equal to 1 if there is an choice
extra amount promotion for item i in
3.1. Data and variables period t and equal to 0 otherwise
LOYGLhit a dynamic loyalty variable (see choice
Scanner panel data are available for one store of a Guadagni & Little, 1983), equal to:
h
large European retail chain. These data cover 1 year of a*LOYGLit1 +(1a)*LPith ,
with LPith=last-purchase dummy variable,
store visits, SKU purchases and in-store promotions.
equal to 1 when item i was last
The data set is split up into a 15-week initialization purchased, and 0 otherwise, a=
period (for inventory and loyalty variables), a 28-week smoothing constant between 0 and 1
estimation period, and a 7-week validation period. Our COOSth category OOS variable, weighted sum of incidence,
application concentrates on two product categories, item stockouts with choice probabilities quantity
as weights: Sipchit *OOSit
cereals and margarine, with 24 and 16 SKUs, respec-
CRh the household’s category consumption incidence,
rate, measured as the ‘daily purchased’ quantity
quantity in the initialisation period
7 INVMCth the (mean-centered) household’s incidence,
A sequential approach was preferred over a joint estimation
inventory level in the category at time quantity
approach for reasons of tractability. Especially in our model,
t=INVthINVMh, where INVM h=mean
tractability is an issue because (i) the category stockout variables
of INVth for household h; INVth=INVt1 h
+
depend in a complex fashion on the choice model parameters, such h
qt1 CRh*INTt,t1, with qt1 h
=quantity
that the joint likelihood is an intricate function of these parameters,
purchased during the last visit, and
and (ii) unlike other studies—which include a reduced subset of items
INTt,t1=number of days between the
from the category only—we leave in a maximum of available items to
two consecutive visits t1 and t.
get a truthful picture of stockout induced choice shifts. This, however,
Cpriceth category price variable, weighted average incidence,
comes at the expense of computational tractability of the joint
of item prices with choice probabilities quantity
estimation procedure. Sequential estimation might entail a bias in the
as weights: Sipchit*PRICEit
parameters. However, as indicated by Jedidi et al. (1999), performing
Cpromoth an index of promotional activity in the incidence,
one last ‘joint’ iteration on the sequentially obtained estimates allows
category, equal to the weighted sum of quantity
to assess the importance of this bias and, eventually, remove it. In our
individual item dummy promotion
case, we found the results of the joint iteration to remain very close to
variables: Sipchit*PROMOit
the sequential parameters.
278 K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286

1998a,b; Gupta, 1988 for a similar approach).8 In the Table 2


incidence and quantity models, the category stockout Frequency of OOS occurrences
variables are supplemented with category-level price Distribution of observed OOS events over different brand typesa
and promotion variables, household consumption rate, Brand type Cereals: Margarine:
and inventory (see also Bucklin et al., 1998a,b; Jedidi et %OOS events %OOS events
al., 1999). National brand 31.4 58.82
Distributor brand 68.6 0.0
Generic 0.0 41.18
3.2. OOS detection procedure
Distribution of OOS events over low and high selling itemsb
An important problem in using behavioral data for
Market share Cereals: Margarine:
OOS research is that stockout occurrences are seldom %OOS events %OOS events
systematically recorded, and that available OOS data (market share (market share
are thus unreliable and incomplete. Identification of range) range)
stockouts is further complicated by the fact that (i) Low 42.86 (0.00 – 1.49) 23.53 (0 – 3)
zero-sales during a given period do not necessarily Below median 0.00 (1.50 – 2.74) 32.35 (3 – 7)
imply that an OOS occurred and, on the other hand, (ii) Above median 28.57 (2.75 – 6.00) 23.53 (7 – 10)
High 28.57 (>6.00) 20.59 (>10)
that there may have been a stockout in periods with
non-zero sales (for instance, when the product ran out Number of stockouts 35 34
of stock halfway during the observation period). To be during the 28-week
able to identify OOS events based on historical sales estimation period
data, we develop a two-stage tracking procedure. In a a
The cereals category comprises only one generic item, the
first step, potential (‘candidate’) stockout periods are margarine category only one private label product, neither of which
identified. In step two, these periods are further is very succesful. OOS occurrence for national brands should
therefore mainly be compared with the stockout frequency of
screened using a sales model and an iterative outlier
distributor brands in the cereals category, and with generic stockouts
analysis procedure, to distinguish ‘normal’ low sales in the margarine category.
from true stockouts. A detailed description of the b
Items were divided in quartiles based on market share, to
procedure is given in Appendix A. Application of check whether stockouts occur more or less frequently for high/low
the procedure to the cereals and margarine data led selling items.
to the detection of several stockouts, in both categories
and for a wide variety of items (see Table 2). Feedback nitions and is therefore retained. Models are estimated
from managers confirmed that the outcomes have face for a varying number of classes and re-estimated using
validity. In addition, several formal validity checks different sets of starting values.9 Based on the BIC and
were performed to ensure the robustness of the OOS CAIC values (see, e.g., Leeflang, Wittink, Wedel, &
detection procedure (see Appendix A). Naert, 2000, p. 493) reported in Table 3, we opt for four
purchase incidence classes and three purchase quantity
3.3. Estimation results classes, for cereals as well as margarine. For choice, a
one-class model is retained for margarine and a three-
For both purchase incidence and quantity, the ‘pref- class model for cereals. Stability of the latent classes is
erence based’ category stockout definition (Eq. (3)) validated on a holdout period of 7 weeks. In both
clearly provided superior fit over the alternative defi- categories, hit rates, average probability for hits, and
mean absolute deviations are very similar for the
8
The use of dynamic loyalty variables is based on Ailawadi et
al. (1999), who show that a combination of these variables with
9
finite mixture estimation outperforms other modelling approaches to Like Gönül, Carter, Petrova, and Srinivasan (2001), we also
capturing time dependence and heterogeneity in consumer choices examined different values for the smoothing constant a and retained
(see Andrews & Currim, 2002, for a similar recent approach). Note the a with the best model fit, which in our application amounts to 0.7.
that a choice model with separate (static) loyalty and last purchase This smoothing constant provided superior fit in both the
variables resulted in extreme collinearity and, similar to the results homogeneous model and the final model (three latent classes for
of Ailawadi et al., did not improve model fit. cereals).
K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286 279

Table 3
Goodness-of-fit (BIC, CAIC statistic) for various models
Model Cereals: number of latent classes Margarine: number of latent classes
1 2 3 4 5 1 2 3 4 5
Incidence
BIC 14 899 14 175 14 025 13 980 13 996 11 620 11110 11 006 10 927 10 944
CAIC 29 805 28 365 28 074 27 992 28 031 23 247 22 235 22 036 21 886 21 927

Quantity
BIC 5423 5134 5125 5143 6305 5816 5762 5766
CAIC 10 853 10 284 10 274 10 318 12 617 11 647 11 548 11 564

Choice
BIC 15 336 15 061 15 037 15 073 3098 3163
CAIC 30 690 30 156 30 127 30 219 6212 6359

estimation and the validation period, as are the segment more strongly significant in models with a smaller
sizes.10 number of latent classes. For cereals, no impact is
found of stockouts on purchase quantity. Delayed OOS
3.4. Parameter estimates coefficients, g1, are insignificant in all cases, suggest-
ing that in the period succeeding the stockout, con-
For each of the three buying behavior models, the sumers do not buy more than what is needed to
coefficients of price, promotion, loyalty, inventory and replenish the decreased home inventory.12
consumption rate are significant in the majority of The choice model parameters are summarized in
cases and, when significant, have the expected sign. Table 6. The constants for each attribute (brand, flavor,
Results for the purchase incidence model are reported type, or size) reflect the difference in attractiveness vis-
in Table 4. In line with previous papers, we find that à-vis a reference item. Concentrating on stockout
stockouts can reduce the likelihood of a category effects, we find significant OOS asymmetry parame-
purchase: the immediate effect coefficients, f0, being ters, qA, in all classes, with the exception of class three
significant and negative in two out of four classes for for cereals. This finding confirms that stockout induced
cereals, and in three out of four classes for margarine. choice shifts may deviate from regular switching pat-
In addition, delayed stockout coefficients, f1, are terns.13 Product characteristics that produce dispropor-
significant and positive for two out of the four marga- tional choice shifts in response to OOS are consistent
rine incidence classes, suggesting that ‘extra’ purchase between classes of the same product category, yet differ
acceleration takes place in post-stockout periods.11 between categories. For cereals, consumers tend to
Table 5 reports the results for the purchase quantity switch to another brand (negative qB coefficient for
model. The immediate OOS parameters, g0, indicate class 2) in order to be able to buy the same type
that stockouts lead to a reduction of purchase quantity (positive qT coefficient for classes 1 and 2) and/or
in class three for margarine. While this effect is only flavor of cereals (positive qF coefficient for class 2).
weakly significant in the final model, it consistently For margarine, in contrast, consumers who have to
appeared in all estimated versions of the margarine
model (one to four latent classes), and was found to be
12
Note, again, that this does not rule out positive post-stockout
effects. If stockouts lead to immediate incidence probability and/or
10
For reasons of space, we do not report details here. They can quantity reductions, this will lower next period’s inventory, and hence
be obtained from the authors on request. increase next period’s incidence probability and quantity through the
11
Note that stockouts in period t may lead to reduced inventory effect.
13
inventories in t+1, which then automatically lead to higher incidence Fader and Hardie’s (1996) IIA-test was applied to the final
probabilities. A positive coefficient f1 points to purchase acceleration model to check for any remaining conflicts with the MNL model’s
over and above these ‘regular’ adjustments, see also Section 2. IIA-assumption, and provided satisfactory results.
280 K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286

Table 4
Estimation results for the purchase incidence model
Variable Cereals Margarine
Class 1 Class 2 Class 3 Class 4 Class 1 Class 2 Class 3 Class 4
a a a a a a
Constant 0.065 2.837 1.341 3.026 2.066 0.756 1.930 0.098
(0.10) (0.07) (0.04) (0.07) (0.07) (0.07) (0.14) (0.16)
Inventory 0.101a 0.058b 0.090a 1.173a 0.364a 0.271a 2.597a 0.370a
(0.02) (0.03) (0.01) (0.07) (0.05) (0.04) (0.21) (0.08)
Cons. rate 11.110a 12.913a 10.227a 39.774a 5.721a 6.019a 28.779a 22.211a
(2.48) (0.89) (0.61) (1.65) (1.18) (1.48) (4.80) (4.67)
Cprice 0.007 0.012 0.005 0.014c 0.037b 0.003 0.002 0.059
(0.02) (0.01) (0.01) (0.01) (0.02) (0.02) (0.03) (0.04)
Cpromo 0.339 0.253 0.220 0.166 0.233b 0.045 0.498a 0.427
(0.30) (0.26) (0.17) (0.25) (0.14) (0.13) (0.20) (0.31)
COOS(f0) 2.720a 1.999 1.524a 0.641 0.813b 1.165a 1.569a 0.566
(1.00) (1.73) (0.54) (0.82) (0.49) (0.33) (0.44) (0.53)
COOSt-1(f1) 0.011 0.978 0.472 0.827 0.525c 0.369 0.989b 0.103
(1.05) (1.09) (0.41) (0.73) (0.42) (0.31) (0.43) (0.56)
Class size (%) 11 13 42 34 22 33 31 14
a
Significant at the 1% level.
b
Significant at the 5% level.
c
Significant at the 10% level (one-tailed significance test).

switch to another item because of an OOS tend to stay levels. To get an overall picture, we group the inci-
with the same brand (positive qB coefficient), and buy dence, quantity, and choice classes into segments with
another size of margarine instead (negative qS coeffi- homogeneous OOS reactions. Table 7 provides an
cient). The coefficient of the delayed stockout variable, overview of observed OOS reaction patterns and the
u, is positive and significant for margarine—pointing fraction of households for which they occur. While
to a preference reinforcement effect for the OOS some segments hold on to a single stockout reaction
item—but insignificant for cereals. (segments 1, 2, and 3), others follow a mixed strategy.
Consumers in segment 4 appear to select one out of
3.5. Heterogeneity in OOS reactions two ‘favorite’ options: switch to items with specific
attributes (asymmetric switching), or give up the
The results above show that consumers differ in planned category purchase. Segment 5 consumers
their stockout reactions and this at each of the decision not only buy less when confronted with an OOS, they

Table 5
Estimation results for the purchase quantity model
Variable Cereals Margarine
Class 1 Class 2 Class 3 Class 1 Class 2 Class 3
Constant 0.522a (0.11) 1.747a (0.13) 0.488a (0.08) 2.115a (0.08) 0.801a (0.09) 0.064b (0.05)
Inventory 0.019 (0.02) 0.106b (0.05) 0.038a (0.02) 0.084c (0.06) 0.090b (0.04) 0.041c (0.03)
Cons. rate 2.510 (2.46) 10.675a (1.52) 9.393a (1.04) 30.647a (1.36) 7.766a (1.89) 9.172a (0.53)
Cprice 0.005 (0.01) 0.002 (0.01) 0.019a (0.01) 0.012 (0.03) 0.009 (0.04) 0.003 (0.01)
Cpromo 0.713a (0.20) 0.955a (0.26) 0.473a (0.14) 0.994a (0.15) 0.100 (0.20) 0.238a (0.07)
COOS(g0) 0.557 (0.90) 2.642 (2.24) 0.174 (0.93) 0.008 (0.62) 0.197 (0.30) 0.483c (0.39)
COOSt-1(g1) 0.706 (0.81) 1.094 (1.39) 0.423 (0.87) 0.144 (0.59) 0.139 (0.43) 0.087 (0.37)
Class size (%) 7 48 45 38 9 52
a
Significant at the 1% level.
b
Significant at the 5% level.
c
Significant at the 10% level (one-tailed significance test).
K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286 281

Table 6
Estimation results for the choice model
Variable Cereals Margarine
Class 1 Class 2 Class 3 Variable Class 1
Brand constant Brand constant
NBc1 3.034a (0.25) 1.231a (0.15) 0.046 (0.10) NBm1 1.315a (0.19)
Distributor 4.510a (1.01) 0.557a (0.16) 1.740a (0.31) NBm2 0.014 (0.15)
Generic 6.482 (51.4) 1.088a (0.14) 5.612 (70.9) NBm3 1.454a (0.22)
Flavor constant NBm4 1.879a (0.52)
Nature 1.382a (0.14) 0.489a (0.12) 0.535a (0.17) NBm5 1.382a (0.79)
Chocolate 1.003a (0.10) 0.970a (0.10) 0.060 (0.12) Distributor 4.738a (0.23)
Honey 0.804a (0.10) 0.269a (0.10) 0.155 (0.13) Generic 7.332a (0.24)
Type constant Type constant
Rice 0.358a (0.08) 0.248a (0.09) 0.042 (0.11) Regular 0.091 (0.12)
Wheat 0.040 (0.10) 0.069 (0.09) 0.034 (0.13) Size constant
2 cereals 1.146a (0.14) 0.387a (0.12) 0.146 (0.10) Large 0.374a (0.13)
Multicereals 0.130 (0.15) 0.432a (0.11) 0.520a (0.16) Loyalty 5.998a (0.07)
Loyalty 4.044a (0.10) 5.196a (0.10) 4.609a (0.13) Price 0.312a (0.03)
Price 0.026a (0.00) 0.014a (0.00) 0.003 (0.01) Promotion 0.321a (0.14)
Promotion 0.340a (0.11) 0.387a (0.09) 0.170 (0.16) OOS variables
OOS variables Brand(qB) 0.410a (0.14)
Brand(qB) 0.779 (0.99) 0.489a (0.16) 0.186 (0.30) Size(qS) 0.230b (0.12)
Type(qT) 0.196b (0.08) 0.172b (0.08) 0.151 (0.11) Type(qT) 0.011 (0.10)
Flavor(qF) 0.048 (0.08) 0.179b (0.08) 0.167 (0.15) OOSt1(u) 0.640a (0.22)
OOSt1(u) 0.219 (0.69) 0.064 (0.26) 0.027 (0.66)
Class size (%) 47 34 19 100
a
Significant at the 1% level.
b
Significant at the 5% level.
c
Significant at the 10% level (1-tailed significance test, with exception of OOS variables, for which a 2-tailed test was applied).

also have a tendency to switch to items of the same with specific attributes, possibly in smaller quantities.
brand—both reactions pointing to risk-reducing be- While the segment characterization is somewhat spec-
havior. Segment 6 combines the previous strategies by ulative, it does point to the need to account for mixed
either dropping the purchase or buying an alternative reaction patterns.

Table 7
OOS response patternsa
OOS-reaction Reduced Reduced Asymmetr. Margarine Cereals
incidence quantity (non-IIA)
switching
1. ‘‘Switch to other item’’ No No No 194 (31%)
2. ‘‘Do not buy’’ Yes No No 217 (34.7%)
3. ‘‘Switch to item with No No Yes 38 (7.6%) 98 (15.6%)
specific attributes’’
4. ‘‘Do not buy or switch to Yes No Yes 200 (40.1%) 117 (18.7%)
item with specific attributes’’
5. ‘‘Buy less and switch to item No Yes Yes 33 (6.6%)
with specific attributes’’
6. ‘‘Do not buy, or buy less and Yes Yes Yes 228 (45.7%)
switch to item with specific attributes’’
a
These figures were obtained after assignment of households to latent classes for the incidence, quantity, and choice models, based on a
posteriori class membership probabilities.
282 K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286

4. Implications for manufacturers and retailers can be computed in the same fashion and also have
face value.15
From the estimation results presented above, in- Similarly, based on the quantity model Eq. (2), a
sights can be obtained into the consequences of stock- stockout for item j reduces the household’s purchase
outs for retailers and manufacturers. The retailer is rate with the fraction:
primarily interested in the implications of stockouts
on total category sales within the store. The manufac- kht j AV  kht j OOSjt
RQhjt ¼ ¼ 1  expðg0 pchjt Þ ð7Þ
turer, in contrast, is preoccupied with shifts within the kht j AV
category to competitors’ items.
where (kht jOOSjt) is the household purchase rate when
4.1. Loss in category sales due to purchase incidence j is out-of-stock and (kht jAV) the purchase rate when
and quantity adjustments all items are available. Like for incidence, the reduc-
tion in purchase quantity increases with the absolute
As indicated in the literature review, category value of the stockout parameter (g0) and the house-
losses implied by a stockout vary across households hold’s preference for the stockout item (pcjth). For
and items. The incidence and quantity models allow to instance, for households in class 3 of margarine
predict stockout implications for particular (groups of) (where g0=0.483), their purchase rates would de-
households and products. Based on Eq. (1), a stockout crease by 9% for stockouts of items with a 20%
for item j reduces the purchase incidence probability choice probability and by as much as 32% for items
for household h in the stockout period with the with a 80% choice probability.
following fraction: To obtain insights into overall retailer losses, we
use the actual market environment in the calibration
ðpiht j AVÞ  ðpiht j OOSjt Þ period as a basis for simulation. On average, unavail-
RIhjt ¼
ðpiht j AVÞ ability of one item from the assortment leads to an
1 immediate reduction of category sales of 2.0% for
¼1 h
ðpit j AVÞ þ expðf0 pchjt Þ*ð1  ðpiht j AVÞÞ margarine and of 1.83% for cereals. While immediate
ð6Þ OOS effects are similar for both product categories, a
much larger part of the sales losses can be recovered in
where (PIthjAV) is the purchase incidence probability the post-OOS period for margarine (1.66%) than for
when all items are available and (PIthjOOSjt) is the cereals (0.02%), implying much smaller net losses for
purchase incidence probability with a stockout for margarine (0.34%) than for cereals (1.81%).
item j. To illustrate, consider a household in class 1
of margarine with an average purchase probability of 4.2. Loss in brand sales due to competitive switching
0.3. If this household encounters a stockout for an
item for which it has a 0.8 choice probability, its From a manufacturer’s perspective, choice shifts in
propensity to purchase from the category will reduce response to stockouts are not problematic as long as the
by 39%14 (drop from 0.3 to 0.18). Following expres- consumer switches to other items of his own product
sion (6), this reduction in purchase incidence proba- line. Our choice models allow to compute how the
bility is larger (i) for segments with more negative purchase probability of a brand is affected by a stockout
stockout parameters f0 and (ii) for highly preferred of one of its SKUs, taking regular switching patterns as
OOS items (high pcjth). Incidence effects for cereals
15
For instance, with a stockout parameter of 1.52 (class 3), a
stockout for a household’s favorite item—for which a ‘typical’
14
Note that the latter result is surprisingly close to the finding choice probability in the cereals category is 50%—would reduce
of Campo, Gijsbrechts, and Nisol (2000), that 34% of consumers this household’s purchase propensity from say, 0.2 to 0.104. This
decide not to buy from the category in the store (that is: to switch 48% reduction is largely comparable to that reported by Campo et
stores, defer or cancel the purchase) when their favorite margarine is al. (2000), who find that 56% of consumers do not purchase any
not on the shelves. cereals in the store when their favorite product is missing.
K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286 283

well as OOS induced asymmetric choice shifts into how much to purchase from the category, and what to
account. Based on expressions (4) and (5), a stockout choose. At the same time, quantity effects—contrary to
for the i-th SKU reduces brand B’s choice share (equal incidence and choice effects—only occur in one of the
to SjaBrand Bpcjth) with the following fraction: two categories. No such effects are found for cereals,
where variety seeking is typically high (Campo, 1997;
RChB;t Givon, 1984) and consumers—who regularly switch
! among items—may not need to decrease purchase
X X
ðpchjt Þ*exp qA OOSA;jt quantity to limit the risk or utility loss of item substi-
jaB;jpi A tution. For margarine, the quantity effect is significant
¼1 ! but small, which may be due to the fact that most
X X X
ðpchjt Þ* ðpchlt Þ*exp qA OOSA;lt consumers purchase only one package per occasion.
jaB lpi A Second, stockout induced choice shifts are found to
ð8Þ deviate from regular switching patterns. Disproportion-
ate shifts are observed towards items of the same type
The loss in brand sales becomes less dramatic as (i) or flavor for cereals, and of the same brand for
more SKUs of the brand are available, (ii) the remain- margarine. These differences in non-IIA switching
ing SKU’s are relatively more attractive compared to patterns appear to be in line with product category
competitive items (incorporated in the denominator of characteristics and category differences in OOS re-
Eq. (8)), and (iii) the brand’s other SKUs share impor- sponse reported in a recent stockout study (Sloot,
tant attributes with the stockout item that produce Verhoef, & Frances, 2002). Cereals’ brand lines typi-
positive asymmetric choice shifts. Simulation results cally contain similar varieties—the private label line,
show, for instance, that asymmetric choice shifts to for instance, containing a copy of most of the market
items of the same type in class 1 of the cereals category leaders’ product variants. Also, consumers—being
(qT=0.196), could lead to twice as large choice share familiar with more product alternatives—rely less on
losses for brand NBC2 when the brand assortment the brand as a quality cue. Margarine manufacturers
would not comprise any other item of the same type rather try to differentiate themselves from competitors
(share loss of 43% instead of 23% in case an alternative through distinctive positioning (e.g., ‘low cholesterol’
of the same type would be available). To assess the vs. ‘tasty’ brands). They also offer different package
implications of this phenomenon at the market level, sizes, making it easier for consumers to stay with their
we, again, use our data set as a simulation environment. familiar product and brand. On the whole, our findings
For the margarine national brand NBm5, which has indicate that traditional choice models—which do not
only one item in its line, a stockout automatically account for non-IIA (stockout) choice shifts—may lead
reduces brand sales to zero (100% loss). For the other, to biased estimates of manufacturer OOS losses and
multiple item brands, sales losses range between 7% that insights into the category-specific characteristics
and 58% of the original choice share. This large driving stockout-based switches may be crucial to
variation in brand sales losses is not at all surprising, counteract their negative effects. Third, our results
given that the availability of substitute items in the show that stockouts may positively affect purchase
brands’ assortments, as well as the OOS items’ attrac- incidence and choice decisions in post-OOS periods,
tiveness (choice share), varies considerably. allowing retailers and manufacturers to recover some of
the immediate losses. Interestingly, dynamic effects
over and above those needed to restore inventories
5. Discussion and limitations are only observed for margarine—a more ‘necessary’
and ‘storable’ category than cereals.
This research contributes to the literature in two From a methodological viewpoint, our approach
ways. At the substantive level, the paper improves our has the advantage that it allows to assess stockout
understanding of consumer response to OOS in retail implications on the basis of scanner panel data. The
outlets. First, our results reveal that stockouts may alter OOS tracking procedure permits to trace the occur-
each of the three consumer decisions: whether to buy, rence of stockouts when no or incomplete OOS
284 K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286

records are available. The purchase incidence, quanti- variables problem, constitute interesting areas for fu-
ty, and choice models—adjusted to account for imme- ture research.
diate and delayed stockout effects—can be used to
simulate OOS implications for specific SKUs and/or
consumer segments, and this for the manufacturer as Acknowledgements
well as the retailer. Our reliance on revealed (not
stated) behavior may produce a more realistic view This research was conducted with the financial aid
on the importance of stockout losses. Among consum- of the Fund for Scientific Research Flanders. The
ers for whom the unavailable item is their category authors are much indebted to the managers of the retail
favorite, our model predicts similar stockout losses as chain for their valuable inputs and comments. They
previous survey studies (see, e.g., Campo et al., 2000). also thank the editor, Prof. Dr. Hubert Gatignon, and
Yet, as the retailer’s clientele also comprises many three anonymous reviewers for their comments and
customers for whom the missing item is not their suggestions.
favorite, overall losses predicted for the retailer are
found to be lower than is often believed. In contrast,
manufacturer losses may be quite substantial, and are Appendix A . OOS tracking system
largely governed by the consumers’ preference for the
missing item and the availability of acceptable alter- OOS detection takes place for each SKU in the
natives in the brand’s line, a finding in line with category and proceeds in two steps.16 Obviously, not
Boatwright and Nunes (2000), Broniarczyk et al. all low-sales periods correspond with true stockouts
(1998), Campo et al. (2000), and Fitzsimons (2000). (TSO). The first step consists of a screening procedure
Clearly, our study also has important limitations and to reduce the number of periods that have to be
opens up opportunities for future research. The empir- examined more closely in the second stage. For each
ical application covers only two categories in one store. item, we identify ‘candidate stockout’ (CSO) periods,
Our substantive findings, therefore, need to be inves- in which the item may have been unavailable. House-
tigated in a wider range of settings. Moreover, category hold purchase observations are aggregated into item
losses in our study were formulated from the viewpoint sales with periodicity equal to (maximum) one third of
of a specific retail outlet. Future research may add the the minimum replenishment time (MRT) for the
distinction between store switch, deferment and can- category (which, for this application, amounts to 1
cellation; requiring scanner panel data across stores. On day). This data periodicity implies that, if an item has
the methodological side, it needs to be mentioned that been OOS, there should be at least one observation
our models do not follow from a unified utility frame- with zero sales (i.e., in at least one of the three
work with forward-looking consumers: our paper treats subperiods). This is true because if a stockout oc-
retailer stockouts as exogeneous and unexpected by the
consumer. Future research could formally investigate
the presence of systematic patterns in retailer stockout 16
Note that our tracking procedure does NOT lead to a ‘self-
occurrences, as well as the impact of such patterns on fulfilling prophecy’. In the stockout identification stage, TSOs are
the behavior of forward-looking consumers. The ap- detected using an aggregate model for the stockout item’s sales.
proaches developed by Erdem et al. (2002) and Sun, This is a ‘definition’ issue: if an item is unavailable, it cannot be
sold. For the TSOs thus identified, we later assess whether and how
Neslin, and Srinivasan (in press) could prove useful for
strongly they affect category purchase incidence, category purchase
this purpose, allowing for a more accurate estimation of quantity, and choice shares of other items in the category. Even if
stockout effects as well as a more comprehensive unavailability of an item results in huge sales losses for that item, it
evaluation of the impact of replenishment policy does not automatically imply large incidence or quantity losses at
changes. Moreover, despite the validation checks, our the category level, nor does it condition the nature of the choice
shifts within the category. Note also that the procedure—in contrast
‘inferred’ stockout variables are likely to contain errors,
with models used to fill missing prices of non-promoted brand in
which may somewhat dampen the estimated impact of scanning data—does not entail what Erdem et al. (1999) refer to as
stockouts. Further validation of the stockout detection the ‘self selection bias’, since it does not infer missing out-of-stock
method and/or accounting for the potential errors in observations from available observations on stockouts.
K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286 285

curred in a certain period (say, between 8 and 12 a.m. stockouts obtained from the retail manager judging
of the 12-h opening day), it cannot have been replen- item sales data. In both cases, a satisfactory rate of
ished within the same period (between 8 and 12 a.m.) agreement was obtained, providing additional support
nor in the period after that (between 12 and 4), and for the robustness of the tracking procedure.
sales can at best reappear during the third part of the
day, the mid-period necessarily showing zero sales.
All periods with length equal to MRT (in our case, References
daily periods) which comprise at least one zero sales
sub-period (third of a day) are designated as candidate Ailawadi, K., Neslin, S., & Gedenk, K. (1999). Heterogeneity and
stockouts. purchase event feedback in choice models: An empirical analy-
sis with implications for model building. International Journal
Step 2 consists of an outlier analysis to distin- of Research in Marketing, 16, 177 – 198.
guish ‘normal’ low-sales periods from those due to Andrews, R., & Currim, I. (2002). Identifying segments with iden-
item unavailability. This is done by estimating, for tical choice behaviors across product categories: An intercate-
each item, a simple sales model with item sales per gory Logit mixture model. International Journal of Research in
Marketing, 19(1), 65 – 80.
MRT (in our case, daily sales) as dependent, and
Anupindi, R., Dada, M., & Gupta, S. (1998). Estimation of con-
item, time of week, and promotion indicators as sumer demand with stockout based substitution: An applica-
explanatory variables. Next, for each CSO identified tion to vending machine products. Marketing Science, 17(4),
in step 1, we re-estimate the model excluding the 406 – 423.
observation and assess the probability for the ob- Boatwright, P., & Nunes, J. C. (2001, July). Reducing assort-
served sales level to be ‘predicted’ by the estimated ment: An attribute-based approach. Journal of Marketing, 65,
50 – 63.
model. If this probability is too low (below a cut-off Broniarczyk, S., Hoyer, W., & McAlister, L. (1998, May). Consum-
value a, which is typically set at 0.05) and if the ers’ perceptions of the assortment offered in a grocery category:
observed sales level is below the predicted value, the The impact of item reduction. Journal of Marketing Research,
observation is labelled to comprise a TSO. Finally, 35, 166 – 176.
all successive periods with zero sales between two Bucklin, R., Gupta, S., & Siddarth, S. (1998a, May). Determining
segmentation in sales response across consumer purchase be-
subsequent TSO are also labelled as stockout peri- haviors. Journal of Marketing Research, 35, 189 – 197.
ods. Note that, to prevent item sales per period from Bucklin, R., Gupta, S., & Siddarth, S. (1998b). Modelling the effect
being too strongly skewed towards zero, model of purchase quantity on consumer choice of product assortment.
estimation is conducted at a higher level of time Journal of Forecasting, 17, 281 – 301.
Bucklin, R., & Lattin, J. (1991). A two-state model of purchase
aggregation (data periodicity equal to MRT and not
incidence and brand choice. Marketing Science, 10(1), 24 – 39.
one third of MRT) than in step 1. Campo, K., 1997. Variety seeking and the sensitivity to in-store
Validation. The outcomes of the OOS detection promotions. Unpublished Doctoral Dissertation, Antwerp: Uni-
procedure depend on the choice of the cut-off value a. versity of Antwerp-UFSIA.
To shed some light on the stability of the results, a Campo, K., Gijsbrechts, E., & Nisol, P. (2000). Towards under-
sensitivity analysis was conducted: in addition to an a standing consumer response to stock-outs. Journal of Retailing,
76(2), 219 – 242.
of 0.05, TSOs were identified for an a level of 0.025. Chiang, J. (1995). Competing coupon promotions and category
We thus identified two ‘series’ of OOS occurrences, sales. Marketing Science, 14(1), 105 – 122.
one corresponding with each level of a. We then re- Corstjens, J., & Corstjens, M. (1999). Store Wars: The battle for
estimated our incidence, quantity, and choice mod- mindspace and shelfspace. New York: Wiley.
els—in which stockouts intervene as explanatory Dhar, R. (1997, September). Consumer preference for a no-choice
option. Journal of Consumer Research, 24, 215 – 231.
variables. The results confirm the stability of the Dillon, W., & Gupta, S. (1996). A segment-level model of category
findings: even though the stockout parameters slightly volume and brand choice. Marketing Science, 15(1), 38 – 59.
change, the pattern of the coefficients across models, Erdem, T., Keane, M., & Sun, B. (1999). Missing price and coupon
households, and categories remains the same. As an availability in scanner panels: Correcting for the self selection
bias in choice model parameters. Journal of Econometrics, 89,
extra check, we cross-validated the results of the
177 – 196.
tracking procedure against two other data sources: Erdem, T., Imai, S., & Keane M.P. (2002, September). Brand and
(i) available objectively registered (but incomplete) choice dynamics under price uncertainty. Working Paper.
stockout information and (ii) subjectively identified Fader, P. S., & Hardie, B. G. S. (1996, November). Modeling con-
286 K. Campo et al. / Intern. J. of Research in Marketing 20 (2003) 273–286

sumer choice among SKU’s. Journal of Marketing Research, Neslin, S. A., & Schneider-Stone, L. G. (1996). Consumer inven-
33, 442 – 452. tory sensitivity and the postpromotion dip. Marketing Letters,
Fitzsimons, G. (2000). Consumer response to stockouts. Journal of 7(1), 77 – 94.
Consumer Research, 27(2), 249 – 266. Peckham, J. (1963, October). The consumer speaks. Journal of
Givon, M. (1984). Variety seeking through brand switching. Mar- Marketing, 21 – 26.
keting Science, 3(1), 1 – 22. Schary, Ph., & Christopher, M. (1979). The anatomy of a stock-out.
Gönül, F. F., Carter, F. C., Petrova, E., & Srinivasan, K. (2001, Journal of Retailing, 55(2), 59 – 70.
July). Promotion of prescription drugs and its impact on phys- Sloot, L., Verhoef, P. C., & Frances, Ph. H. (2002). The impact of
ician’s choice behavior. Journal of Marketing, 65, 79 – 90. brand and category characteristics on consumer stockout reac-
Guadagni, P., & Little, J. (1983). A Logit model of brand choice tions. Working Paper. Rotterdam: Erasmus Research Institute of
calibrated on scanner data. Marketing Science, 2(3), 203 – 238. Management.
Gupta, S. (1988, November). Impact of sales promotions on when, Sun, B., Neslin, S., & Srinivasan, K. (2002). Measuring the impact
what, and how much to buy. Journal of Marketing Research, 25, of promotions on brand switching when consumers are forward
342 – 355. looking. Working Paper, Carnegie Mellon University.
Hoch, S., Bradlow, E., & Wansink, B. (1999, Winter). The variety Vanden Abeele, P., Gijsbrechts, E., & Vanhuele, M. (1990). Spec-
of an assortment. Marketing Science, 18, 527 – 546. ification and empirical evaluation of a cluster-asymmetry market
Jedidi, K., Mela, C., & Gupta, S. (1999). Managing advertising and share model. International Journal of Research in Marketing, 7,
promotion for long-run profitability. Marketing Science, 18(1), 223 – 247.
1 – 22. Walter, C., & Grabner, J. (1975, July). Stockout cost models:
Leeflang, P., Wittink, D., Wedel, M., & Naert, Ph. (2000). Build- Empirical tests in a retail situation. Journal of Marketing, 39,
ing models for marketing decisions. Boston: Kluwer Academic 56 – 68.
Publishers.

You might also like