JIE ZHANG and ARADHNA KRISHNA*

When retailers make product assortment changes by eliminating

certain stockkeeping units (SKUs), how does this affect sales of
individual brands? This is the main question the authors address in this
article. Using data from an online retailer that implemented a permanent
systemwide SKU reduction (SR) program, the authors investigate how
the program affected various components of purchase behavior for
individual brands. They find substantial variations in the SR effects
across brands, categories, and consumers. They explore possible drivers
for these differences and find that higher-market-share, higher-priced,
and more frequently promoted brands tend to gain share and that
reduction in the number of sizes, reduction in the number of SKUs, and
change in SKU share in the category are important in affecting change in
a brand’s purchase share after the SR. They also find that SRs lead to an
increase in category purchase incidence and quantity for highly state-
dependent consumers and frequent buyers but a decrease in category
purchase and quantity for mildly state-dependent consumers and
infrequent buyers. In addition, SRs tend to cause more changes in brand
choice probabilities among consumers of lower state dependence and
higher price and promotion sensitivity. These findings are of importance
both to retailers wanting to make product assortment changes and to
manufacturers affected by them.

Brand-Level Effects of Stockkeeping Unit

Reductions

In the past two decades, supermarkets have experienced a many retailers to experiment with SKU reduction (here-
stockkeeping unit (SKU) explosion (Drèze, Hoch, and Purk inafter, SR) programs. Another reason for reducing the
1994; Kurt Salmon Associates 1993); manufacturers view number of SKUs is the recognition by manufacturers and
SKU proliferation as a way to increase their presence and retailers that carrying too many items could cause clutter in
market share, and retailers fear that eliminating items could the store and increase consumers’ confusion (Broniarczyk,
lower consumer assortment perceptions and decrease store Hoyer, and McAlister 1998). As a result, some retailers and
visits. More recently, however, the higher costs of maintain- manufacturers (e.g., Pier 1 Imports, Sunbeam) have adopted
ing a large number of SKUs and pressure from lower-cost efficient assortment policies by eliminating low-selling
competitors, such as Wal-Mart and Costco, have driven items (Business Wire 1998; Home Textile Today 2005). Mar-
keting academics also have cast doubt on the value of SKU
proliferation.
*Jie Zhang is Assistant Professor of Marketing, Robert H. Smith School Some studies have shown that retailers can eliminate a
of Business, University of Maryland (e-mail: [email protected]). substantial number of SKUs without negatively affecting
Aradhna Krishna is Isadore and Leon Winkelman Professor of Marketing,
Stephen M. Ross School of Business, University of Michigan (e-mail: consumers’ assortment perceptions, store visits, or category
[email protected]). The authors thank an anonymous online retailer for sales (Arnold, Oum, and Tigert 1983; Broniarczyk, Hoyer,
providing the data used in this study. They are grateful to Michel Wedel and McAlister 1998; Boatwright and Nunes 2001, 2004;
and the three anonymous JMR reviewers for their many valuable com- Iyenger and Lepper 2000), and other studies have shown
ments and suggestions. They also thank seminar participants at the Mar-
keting Science Conference, Dartmouth College, University of Delaware,
that SR can decrease store-level shopping frequency and
University of Iowa, University of Maryland, University of Michigan, Uni- purchase quantity (Borle et al. 2005). An important unan-
versity of Wisconsin, Indiana University, and University of Southern Cali- swered question is how elimination of certain SKUs in the
fornia for their input. product assortment affects sales of individual brands. This
is the main question we address in this research. Using data
To read and contribute to reader and author dialogue on JMR, visit from an online retailer that implemented a permanent sys-
http://www.marketingpower.com/jmrblog.
temwide SR program, we examine how consumers reallo-

© 2007, American Marketing Association Journal of Marketing Research

ISSN: 0022-2437 (print), 1547-7193 (electronic) 545 Vol. XLIV (November 2007), 545–559
546 JOURNAL OF MARKETING RESEARCH, NOVEMBER 2007

cate their purchases among the remaining brands after cer- LITERATURE REVIEW
tain SKUs are eliminated in the category and store. To lead Prior research on product assortment changes has mainly
up to this analysis, we conduct an in-depth examination of considered their impact on consumers’ assortment percep-
the effects of SR on category purchase incidence, brand tions, category-level purchase probability and sales, store
choice, and purchase quantity. choice, and store-level shopping frequency and sales. We
Prior research on SR has focused on its impact on the describe five major studies on these topics here.
category or store but not on individual brands. Although a Broniarczyk, Hoyer, and McAlister (1998) focus on how
category- or store-level analysis is useful for retailers, a changes in product assortment affect consumers’ percep-
brand-level analysis provides unique insights valuable not tions of the assortment size, which in turn is shown to influ-
only for retailers but also for manufacturers. For a manufac- ence store choice. In their field study, two treatment stores
turer, the most relevant issue related to SR is what happens had 54% of low-selling SKUs eliminated in the top five
to its brands after an SR program is implemented by a categories (candy, beer, soft drinks, salty snacks, and ciga-
retailer. A related question is what the manufacturer can do rettes). These stores experienced sales increases (2% and
to emphasize its brands’ strengths and minimize negative 8%) in the five categories compared with the control stores.
consequences from the SR. For retailers, the brand-level In addition, shoppers reported finding it easier to shop in
analysis provides information about what may happen to the test stores than in the control stores. The SKU count,
their store brands as a result of SR. It also reveals how mar- availability of favorites, and category space affected store
ket share may shift between different brands (which could choice through assortment perception, and availability of
yield different profits for retailers). This perspective is not favorites also had a direct link to store choice.
possible with a store- or category-level analysis. Iyenger and Lepper (2000) compare consumer reaction to
Our analyses control for changes in the marketing mix small (6 SKUs) versus large (30 SKUs) assortments for
before and after SR and consumers’ responses to these jams and chocolates. They find that shoppers were initially
changes. As we show, such an analysis prevents drawing attracted to retail shelves that offered large assortments.
spurious conclusions. We explore plausible drivers of differ- However, when the shoppers were at the shelves, they were
ences in SR effects among brands and consumers. Specifi- more likely to make a purchase from a small than a large
cally, we examine two groups of brand-specific factors: assortment.
brand characteristics (e.g., market share, price level, promo- In a series of field experiments, Drèze, Hoch, and Purk
tion frequency, store versus national brand) and the nature (1994) measure the effectiveness of two shelf management
of the SR for a brand (e.g., number of SKUs eliminated, techniques: “space-to-movement,” in which shelf sets were
number of sizes eliminated, share of brand sales elimi- customized on the basis of store-specific movement pat-
nated). We further examine the moderating effects of con- terns, and “product reorganization,” in which product place-
sumer characteristics (e.g., the degree of state dependence, ment was manipulated to facilitate cross-category merchan-
purchase frequency, the nature of the SR for a consumer) on dising or ease of shopping. They also examine the impact of
the impact of SR. The effects of each of these factors are shelf positioning and facing allocations on sales of individ-
important to understand for both manufacturers and retail- ual items. In their experiments, they find that category sales
ers. This research helps answer the following questions: increased by approximately 4% when there was an increase
in shelf facings of the high-selling items as a result of the
•Are certain types of brands more likely to gain market share
after an SR? For example, do larger-share brands become even deletion of low-selling SKUs. In contrast to Drèze, Hoch,
larger, or are they more likely to lose share to smaller-share and Purk’s study, we focus on the effects of eliminating cer-
brands? Are high-priced brands likely to gain or lose share? Do tain SKUs for each brand and control for the influence of
brands with frequent promotions obtain a greater share of pur- shelf positioning and facing allocation.
chases in the post-SR market share reallocation period? Boatwright and Nunes (2001, 2004) examine purchase
•How does the nature of SR for a brand (e.g., reduction in num- data collected from an SR field experiment for a large num-
ber of sizes, reduction in number of SKUs) change the brand’s ber of categories and conclude that there were no significant
purchase share? changes in the overall category sales due to the SR. Borle
•How does SR affect the purchase share and quantity of private and colleagues (2005) examine the effects of SR on store
label brands?
and category purchase frequency and dollar sales but find
•Is it valid to view change in market share as a proxy for the
effect of SR on a brand’s choice probability, or are changes in negative results; both shopping frequency and purchase
marketing-mix variables likely to be confounded with SR spending on each shopping trip declined as a result of SR.
changes, thus resulting in spurious conclusions? At the store level, they find that SR led to an average
•Are there any systematic differences in the reaction to SR increase of 23% in expected interpurchase time and an aver-
across consumers? age decrease of 4% in expected purchase spending per
shopping trip. At the category level, for a majority of cate-
The online store environment we use for the analyses gories, reduction in favorite items caused no change in cate-
provides a unique opportunity to study the impact of assort- gory purchase incidence probability or in the category’s
ment changes without confounding it with the effects of share of basket. The assortment reduction had a greater
product display, shelf space allocation, or location on the effect on store visit frequency than on purchase spending
shelf (see Boatwright and Nunes 2001). We organize the per visit. When Borle and colleagues compare the results of
rest of the article as follows: In the next section, we discuss their study with those of Boatwright and Nunes, they find
prior research findings on SR. Then, we describe our that the differences were mainly caused by the different set
model. Following that, we present the data analysis results of categories examined in each study.
and conclude with managerial implications and suggestions The divergent findings in these studies suggest that more
for further research. research is needed on the impact of SR. Moreover, none of
Brand-Level Effects of SKU Reductions 547

these studies focused on brand-level effects of SR. It is also We define SRt = 0 if the time is before the SR, and SRt =
worth noting another distinction between our study and that 1 if it is after the SR. The utility of brand k at time t for
of Borle and colleagues (2005). Whereas they measure pur- household i is given by
chase quantities in dollar amounts, which could confound
(1) U ikt = δ B
ki SR t + Vikt + ε ikt
changes in purchase volume (in units) with changes in
price, we model changes in purchase volume (in units) = δB
ki SR t + α ki + X ktβ i + γ i LBikt + ε ikt ,
B
directly and then examine the impact of SR on sales reve-
nue in terms of volume and price. This provides a clearer k = 1, …, K ,
picture of the impact of SR on each element of the purchase
decision. where αki, k = 1, …, K – 1, are brand-specific constants;
Xkt is a vector of marketing-mix variables, including regular
MODEL FORMULATION price and price cut; LBikt = 1 if brand k was chosen by
household i on the previous purchase occasion; β iB are coef-
We investigate the impact of SR on three components of
ficients of marketing-mix variables in the brand utility func-
purchase behavior for individual households: category pur-
tions (they are allowed to be different in the pre- and post-
chase incidence, brand choice, and purchase quantity. Previ-
SR periods to capture the notion that price and promotion
ous research has shown that it is important to account for
changes may cause shifts in consumers’ price and promo-
the interdependence in these decisions (e.g., Chiang 1991).
tion sensitivities); γi measures a household’s state depend-
We model these three purchase components jointly using an
ence and is usually interpreted as an indicator of inertia
approach similar to that of Hanemann (1984), Chiang
(γi > 0) or variety seeking (γi < 0) (e.g., Gupta, Chintagunta,
(1991), Bell, Chiang, and Padmanabhan (1999), and Zhang
and Wittink 1997; Seetharaman, Ainslie, and Chintagunta
and Krishnamurthi (2004).
1999); δ Bki, k = 1, …, K – 1, capture the effect of the SR on
To assess the impact of SR, it is necessary to control for
each brand’s utility for household i; and αKi and
the effects of other marketing-mix variables because they
δ Ki
B are fixed to be 0 for identification purposes.
may change over the period used for examining the impact
We model purchase incidence by assuming that house-
of the SR. In addition, the model should accommodate pos-
hold i makes a category purchase at t if and only if at least
sible changes in consumers’ responses to the other
one brand’s utility in the category exceeds a threshold. We
marketing-mix variables in the post-SR period. Our model
specify the category threshold as
controls for the effects of these variables, in particular the
two most important ones, price and promotion. A closer (2) U i 0 t = δ iISR t + Vi 0 t + ε i 0 t
examination of the data reveals that price and promotion
experienced nontrivial changes during the period under = δ iISR t + α 0 i + Yitβ iI
investigation, and there was an increase in the overall cate-
w + ψ iI log( t ) + ε i 0 t ,
gory price and promotion levels in general. Further investi-
gation indicates that these changes were exogenous to the where α0i is the constant; Yit is a vector of covariates,
SR program but coincided with its timing (for details, see including a household’s average purchase frequency in the
the “Data Analysis” section). We assume that changes in initialization period (FREQi) and its mean-centered previ-
consumers’ responses to price and promotion, if any, were ous purchase quantity (LQit); and β iI are coefficients of the
due to changes in price and promotion but not to the SR covariates. The previous purchase quantity variable refers to
program. Our data do not allow us to separate out the possi- the quantity at the previous category purchase occasion and,
ble effects of the SR program on price and promotion coef- in spirit, captures the effect of inventory (Chintagunta and
ficients, because the two types of changes occurred at the Haldar 1998; Jain and Vilcassim 1991). Note that it would
same time. Nonetheless, our empirical analysis indicates not be appropriate to include an inventory variable in our
that there were only minor differences in the price and pro- model, because its computation requires the use of interpur-
motion coefficients in the two periods, and thus, even if the chase duration, which is endogenous to the purchase inci-
SR program may contribute to the changes in the coeffi- dence decision (see Chintagunta and Haldar 1998). In addi-
cients, these effects are negligible in magnitude and thus tion, an inventory variable might not have captured the
unlikely to alter the findings of the study.1 We also control entire inventory, because we have only household purchase
for the seasonality effect by matching the months of the data from the online store. The variable t indicates week,
year of the data with and without the SR. We provide the and the parameter ψ iI captures the possible trend of category
details in the next section. purchase frequency over time at the store.2 Finally, δ iI
Because we are particularly interested in brand-level measures the effect of the SR on the threshold. Note that a
effects of SR, such as what type of brands consumers tend positive ψ iI or δ iI indicates a negative effect on the purchase
to switch to after others are eliminated, we estimate the incidence probability, and vice versa.
model using only brands that remained in the store after the To model purchase quantities, let Q*ikt be a latent variable
SR. In such a model, the change in the conditional brand of household i’s purchase quantity of brand k in week t, and
choice probability gives a direct indication of whether the let Qikt be household i’s actual purchase quantity of brand k
brand has gained or lost market share as a result of the SR, in week t; in addition, we define Iit = 1 if household i makes
which is not confounded by the potential share increase for a category purchase in week t and 0 if otherwise, and we
each remaining brand merely due to the elimination of other define Bikt = 1 if household i purchases brand k in week t
brands.

2The log-transformation of t provides better fit to the data than the linear
1We thank an anonymous reviewer for pointing out this issue. form in our empirical analyses.
548 JOURNAL OF MARKETING RESEARCH, NOVEMBER 2007

and 0 if otherwise. The observed purchase quantity Qikt = where φ ∈ (0, 1) is a parameter measuring the similarity
Q*ikt if Iit = 1 and Bikt = 1, and Qikt = 0 otherwise. We spec- among the brands. If Equation 4 is reparameterized into
ify Q*ikt as follows:
1−φ
⎡ ⎤
∑
K
Q*ikt = δ Q ⎢ exp(Vijt ) ⎥
(3) i SR t + Wikt + ξ ikt ⎣ j=1 ⎦
(6) Pr{I it = 1} = 1−φ
,
⎡ ⎤
= δQ
∑
Q Q Q
i SR t + α ki + Z iktβ i + ψ i log( t ) + ξ ikt ,
K
exp(δ iI*SR t + Vi 0 t ) + ⎢ exp(Vijt ) ⎥
⎣ j=1 ⎦
k = 1, …, K ,
the new parameter δ iI* will directly reflect the effect of SR
where α Q ki is a constant for brand k and Zikt is a vector of on the purchase incidence probability Pr{Iit = 1}, where
covariates, including marketing-mix variables of brand k at δ iI* > 0 indicates a decrease in Pr{Iit = 1} and δ iI* < 0 indi-
time t (regular price and price cut) and household i’s aver- cates an increase in Pr{Iit = 1}. It can be shown that
age purchase quantity in the initialization period (AQi) as a
control variable. We allow β Q i , the coefficient vector for Zikt, (7) δ iI* = δ iI
to be different for the pre- and post-SR period. The parame-
⎡ ⎤
∑ ∑
Q K K
ter ψ i captures possible trend of purchase quantity over + (1 − φ) log ⎢ exp(Vijt ) exp(δ Bji + Vijt ) ⎥ .
time. The effect of the SR on purchase quantity is measured ⎣ j=1 j=1 ⎦
by δ Q
i .
3
To accommodate the interdependence of the three pur- We estimate δ iI* for its ease of interpretation. Note that
chase components, we extended a formulation that Zhang Equation 5 is modified accordingly.
and Krishnamurthi (2004) developed. Our model allows for Thus far, the model has been constructed at the individual
a more flexible distribution of the error terms than what was household level. We employ a latent-class formulation to
assumed in their model, which leads to a nested logit for- capture unobserved consumer heterogeneity (see Kamakura
mulation of the purchase incidence and choice components. and Russell 1989), in which parameters are segment
By assuming a flexible bivariate distribution of the error specific, denoted by subscript g = 1, …, G. The discrete
term in the quantity equation (ξikt) and a transformation of latent-class specification has been shown to be empirically
the error terms in the brand utility and category threshold equivalent to continuous approaches for representing
equations (εikt, k = 0, 1, …, K), we were able to derive a heterogeneity, such as the hierarchical Bayesian formula-
closed-form expression of the joint probability of purchase tions (Andrews, Ainslie, and Currim 2002). The log-
incidence, choice, and quantity and thus use standard maxi- likelihood function is given by
mum likelihood estimation procedure to estimate the ⎧ G Ti
N
⎪ ⎡
model.4
In our model, the category purchase incidence probability
(8) LL = ∑ log ⎨ ∑ ∏
qg ⎢ Prg ( I it = 0)1 − I it
⎪⎩g = 1 t = 1 ⎢⎣
i =1
is
K ⎤ ⎫⎪
(4) Pr{I it = 1}
∏ Pr (I
k =1
g it = 1, Bikt = 1, Q ikt = q ikt ) I it × Bikt ⎥⎬,
⎥⎦ ⎪⎭
1−φ
⎡ ⎤
∑
K
⎢ ki SR t + Vijt ) ⎥
exp(δ B
= ⎣ j=1 ⎦ ,
where qg is the probability of belonging to segment g, Ti is
1−φ the number of observations for household i, and other terms
⎡ ⎤
∑
K
exp(δ iISR t + Vi 0 t ) + ⎢ exp(δ B
ki SR t + Vijt ) ⎥ are as shown previously. The number of latent segments G
⎣ j=1 ⎦ is determined empirically by comparing the Bayesian infor-
mation criterion (BIC) of models with different G, and the
and the joint probability of purchase incidence and brand one that yields the lowest BIC is selected. To summarize,
choice is the parameters that we are particularly interested in are δ gI*,
the effect of the SR on category purchase incidence; δ B
kg, the
(5) Pr{I it = 1, Bikt = 1}
effect of the SR on brand k’s utility, k = 1, …, K; and δ Q g,
−φ the effect of the SR on purchase quantity.
⎡ ⎤
∑
K
exp(δ B
ki SR t + Vikt ) ⎢ exp(δ Bji SR t + Vijt ) ⎥
⎣ j=1 ⎦ DATA ANALYSIS
= 1−φ
,
⎡ ⎤ Data Description
∑
K
exp(δ iISR t + Vi 0 t ) + ⎢ exp(δ Bji SR t + Vijt ) ⎥
⎣ j=1 ⎦ Our data are provided by an online grocery retailer that
operates in several metropolitan markets in the United
k = 1, …, K,
States. The retailer implemented a systemwide SR program
on virtually all product categories in January 1999. Our data
3It Q
is possible to make δ i brand specific. In our empirical analyses, the set includes detailed household purchase information on
model with brand-specific δ Q i does not provide significant improvement
three product categories (liquid laundry detergent, mar-
over the one with a common parameter for all brands, for all categories we garine, and spaghetti sauce) collected from a midwestern
analyzed. Therefore, we present the current version in the model market during the January 1, 1997–August 15, 1999,
formulation.
4Details of the model, the likelihood function, and its derivation appear period. As part of the SR program, most brands had some of
in the Web Appendix at http://www.marketingpower.com/content84060. their SKUs eliminated, and a few brands were eliminated
php. altogether. Panels A and B in Table 1 provide a description
Brand-Level Effects of SKU Reductions 549

Table 1
DESCRIPTIVE STATISTICS OF THE SR PROGRAM

A: Overall Category-Level Assortment Changes

Number of Brands Number of SKUs Number of Sizes
Market Share
Category Before After Before After Before After Eliminated (%)
Liquid detergent 14 11 74 50 7 6 9.0
Margarine 12 10 55 45 5 5 3.0
Spaghetti sauce 17 11 127 89 17 12 6.2

B: Description of the Eliminated Brands

Category Brand Number of SKUs Market Share (%)
Liquid detergent Ivory 1 .03
Ultra Yes 1 .83
Value Wise 1 .21

Margarine Move Over Butter 1 .02

Nucoa 1 .08

Spaghetti sauce Alessi 1 .15

Buitoni 1 .13
Del Monte 2 .17
Giannotti 1 .02
Value Wise 1 .25
Weight Watchers 1 .06

of the category-level assortment changes and the brands periods are matched in terms of months to minimize the
eliminated by the SR program. impact of seasonality effects. For concerns of seasonality,
The two panels in Table 1 show that the number of we do not use data from August 16 to December 31, 1998,
brands dropped from 14 to 11 for liquid detergent, from 12 because we do not have data beyond August 15, 1999, in
to 10 for margarine, and from 17 to 11 for spaghetti sauce. the post-SR period. For each category, we choose house-
At the category level, the number of SKUs decreased by holds that made at least two purchases of any brand in the
32% for liquid detergent, 18% for margarine, and 30% for initialization period (for computing the household average
spaghetti sauce, and the cumulative market shares for the purchase frequency and quantity) and at least one purchase
eliminated SKUs were much smaller (9%, 3%, and 6%, of the brands retained for the study (nine in each category)
respectively). The number of sizes dropped from 7 to 6 for in the pre-SR period. This results in 191 households and
liquid laundry detergent, remained at 5 for margarine, and 12,606 observations for liquid detergent, 244 households
dropped from 17 to 12 for spaghetti sauce. The brands that and 16,104 observations for margarine, and 234 households
were completely eliminated had few SKUs and accounted and 15,444 observations for spaghetti sauce in the estima-
for a small market share. Most of the SRs occurred in the tion data. Although these households did not need to make a
brands that remained after the reduction. purchase in the post-SR period to be selected, all made at
As we explained previously, we focus on the brands least one category purchase after the SR. This is consistent
remaining after the SR to investigate how consumers’ pur- with Borle and colleagues’ (2005) finding that there was lit-
chase decisions in the store may have changed because of tle attrition from the store after an assortment reduction
the assortment reduction, including how they may have experiment.
reallocated purchases among the remaining brands. We also Our data show that for most brands, both regular prices
needed to delete a few small brands with too few purchases and price discounts were higher in the post-SR period. As a
for reliable model estimation. Thus, the final data set for result, the average shelf price was 6% higher for liquid
analysis includes 9 of the remaining 11 brands for liquid detergent, 17% higher for margarine, and 8% higher for
detergent (which accounted for 99% of total purchases for spaghetti sauce in the post-SR period. However, note that a
the 11 brands), 9 of the remaining 10 brands for margarine few brands experienced a decrease in shelf price (e.g., the
(99% of total purchases for the 10 brands), and 9 of the private label brands for the liquid detergent and spaghetti
remaining 11 brands for spaghetti sauce (98% of total pur- sauce categories).5 These variations highlight the impor-
chases for the 11 brands). For ease of exposition, here- tance of adopting a model-based approach to adjust for
inafter, we refer to the 9 brands under investigation for each changes in the key marketing-mix variables and responses
product as the “category.”
For each product category, we use January 1–August 15,
1997 (33 weeks), as the initialization period for the house- 5In the Web Appendix (see Additional Table 1; http://www.marketing

hold average purchase frequency and quantity variables. power.com/content84060.php), we present the average regular price, price
discount, shelf price, and market share of each brand before and after the
The estimation data cover January 1–August 15, 1998 (33 SR. Note that the market shares represent each brand’s share of sales
weeks), as the period before the SR, and January 1–August among the brands studied (i.e., not including those eliminated by the SR or
15, 1999 (33 weeks), as the period after the SR. The two taken out because of insufficient purchase observations).
550 JOURNAL OF MARKETING RESEARCH, NOVEMBER 2007

to them. For example, if we were to compare the market assessment of the impact of SR. That is precisely what we
share of Tide detergent in the two periods, it would not be do when we examine segment-level SR effects subsequently.
clear whether the drop in market share was due to the SR or We now examine parameters that capture the impact of
to a rise in its price level. Similarly, the market share SR on purchase incidence, brand utility, and purchase quan-
increase for the private label liquid detergent can be tity, denoted by δI*, δ B k , and δ , respectively. As we
Q
explained both by the SR and by the post-SR shelf price explained previously, the signs of δI* and δQ indicate the
decrease. direction of the effect on purchase incidence and quantity
Using this data set, we first estimate the model described per purchase occasion, respectively. The parameter δ B k
in the previous section and then conduct a series of follow- reflects how a brand’s utility is affected by the SR, but it
up analyses based on the model estimation results to inves- does not directly indicate how the brand’s conditional
tigate various aspects of the SR effects. We present three choice probability is affected, because the choice probabil-
sets of results: (1) model estimation results; (2) SR effects ity also depends on the magnitude of changes in other
at the brand level, after we control for changes in other brands’ utilities. We focus on δI* and δQ first and on δ Bk sub-
marketing-mix variables and responses to them; and (3) an sequently. In each product category, these effects seem to be
analysis that attempts to identify drivers of the differences associated with the level of state dependence of the
in the SR effects among consumers and brands. segments.
We expect that the higher the level of state dependence,
Results the more positive the effects of SR would be, that is, if the
Model estimation results. A three-segment model appears consumer’s favorite SKU has not been eliminated. By defi-
to fit the data best for all three categories based on BIC. The nition, highly state-dependent (i.e., inertia-prone) con-
BIC for models with one, two, three, and four segments is sumers are more likely to purchase brands that they like
11,690.8, 11,193.8, 11,026.5, and 11,187.0, respectively, repeatedly and thus should favor a shopping environment
for liquid detergent; 22,318.6, 19,780.9, 18,627.5, and with narrower choice and less clutter. With a less cluttered
18,645.2, respectively, for margarine; and 10,442.4, online store environment, we expect that they will make
9,991.9, 9,984.7, and 10,134.3, respectively, for spaghetti more product category purchases and also purchase larger
sauce. Parameter estimates for the three categories appear in quantities. (An increase in both category purchase fre-
Tables 2–4. The ρ2 and adjusted-ρ2 terms indicate that our quency and quantity is possible because of store switching
model performs well for all three categories. Nonetheless, and consumption expansion.) Our results are consistent
the model could have been further improved if we had data with this expectation in all three product categories.
on other variables that affected purchase behavior. We sum- SR effects at the brand level. In this section, we assess the
marize the effects of the marketing-mix variables and magnitude of the SR impact on purchase incidence proba-
household-specific control variables first, and then we bility, conditional brand choice probabilities, quantity per
report results for parameters that capture the effects of the purchase occasion, total purchase quantity, and total sales
SR. revenue. The basis for this analysis is the posterior values of
In all three product categories, the effects of the these measures for each household, which we obtained by
marketing-mix variables and the household-specific control using the model estimation results and purchase history
variables have the expected direction for all the significant data of each household. The posterior value of a particular
parameter estimates. Specifically, regular price negatively measure for a household is the weighted average of the
affects a brand’s choice probability and the category pur- segment-specific values, weighted by a household’s poste-
chase incidence probability, whereas the effects of price dis- rior segment membership probabilities.
count are the opposite. A higher household purchase fre- To control for changes in the marketing-mix variables
quency in the initialization period is associated with a lower and responses to them, we conduct a “would-be” analysis
category incidence threshold and, thus, a higher purchase using data in the pre-SR period. Specifically, we estimate
incidence probability; the quantity bought on the previous two sets of posterior values for the outcome measures using
purchase occasion increases the category purchase inci- the same data: one set with all pre-SR period parameters in
dence threshold and therefore reduces the purchase inci- the model (no SR effect parameters) and the other set with
dence probability. In addition, purchase quantity decreases the same pre-SR period parameters plus the SR effects
with a brand’s regular price and increases with its price dis- parameters. Note that the pre-SR period values of the inde-
count, and a household’s average purchase quantity in the pendent variables are used in the computation of both sets.
initialization period is positively associated with the quan- The first set represents the expected values of the outcome
tity purchased on a given occasion. measures in the actual pre-SR period. The second set repre-
The parameter estimates also reveal strong consumer sents the expected values of the outcome measures had
heterogeneity. The three segments for each category exhibit there been the SR, all else being equal to the pre-SR period
different brand preferences, marketing-mix effects, and data. Therefore, the differences between the two sets give
degree of state dependence. We also find a significant the effects of the SR on the outcome measures. This
decreasing trend in the category incidence probability over approach is based on the assumption that the price and pro-
time for most segments in all three categories (the parame- motion changes in the data were due to exogenous reasons
ter estimate for log[t] is positive and significant or margin- other than the SR program. Our information about the
ally significant for seven of the nine category segments). It online retailer’s operations and a check of external market-
is critical to control for these overall trends in consumers’ place price and promotion data indicate that the changes
purchase behaviors in the online store to obtain an accurate were indeed exogenous. Alternatively, we could compute
Brand-Level Effects of SKU Reductions 551

Table 2
PARAMETER ESTIMATES FOR LIQUID DETERGENT

Variables/Parameters Segment 1 Segment 2 Segment 3

Brand Utility (baseline: private label)
αk: Wisk 2.867** –.472 1.078*
All .519 .984 .802
Tide 2.568** 1.236** 1.842***
Cheer 1.788* .472 1.574**
Arm & Hammer –18.469*** –11.018*** .988
Era 2.441** .447 .661
Dreft 4.979*** –2.103** 2.981***
Surf 1.588* .573 .340
βB: Regular price (before SR) –.833** .211 –.470**
Price cut (before SR) .427* .424 .519**
Regular price (after SR) –.960** –.523* –.243
Price cut (after SR) .609* .423 .456
State dependence (γ) 5.098*** 3.262*** 2.510***
δBk (SR): Wisk .482 2.515*** .321
All –.126 .250 .399
Tide 2.121** 1.736** .402
Cheer .591 .630 –.635
Arm & Hammer 10.247*** –9.200*** –.201
Era –.694 1.038* –.369
Dreft –.088 4.053*** –1.186*
Surf –.008 –.007 .329

Category Threshold
Constant 2.495** 2.404*** 2.702***
Purchase frequency –5.721*** –5.754*** –7.057***
Previous purchase volume –.003 .025** .032***
log(t) .249*** .176** .144*
δI* (SR) –.362* –.123 .180

Purchase Quantity
αQk: Wisk 4.367*** 9.860*** 8.574***
All 3.446*** 19.546*** 8.960***
Tide 4.915*** 19.567*** 8.886***
Cheer 3.264*** 9.959*** 8.927***
Arm & Hammer 6.727*** 12.091*** 8.943***
Era 3.368*** 10.065*** 9.042***
Dreft 2.447*** 18.896*** 3.856**
Surf 9.915*** 9.813*** 8.678***
Private label 4.581*** 51.466*** 6.119***
βQ: Regular price (before SR) .291 –1.046*** .372
Price cut (before SR) 3.630*** .811** 1.520*
Regular price (after SR) –.151 –.942** .367
Price cut (after SR) 3.698*** .753** 1.514**
Average purchase volume .485*** .055 .048
log(t) –.111 –.105 –.051
δQ (SR) 2.792*** –.541 .164

φ (brand similarity) .109 .902 .989

θ (interdependence of incidence, choice, quantity) –.692
Segment size 18.5% 27.0% 54.5%

–Log-likelihood [–L(β)] 10,403.3 ρ2 .787

Number of parameters (M) 132 Adjusted ρ2 .784
*p < .10.
**p < .05.
***p < .01.
Notes: ρ2 = 1 – L(β)/L(0), and adjusted ρ2 = 1 – [L(β) – M]/L(0), where L(0) = log-likelihood when all parameters are zero.

the SR effects using the post-SR period values of the inde- endogenous to an SR program, in which case the pre-SR
pendent variables and the post-SR period price and promo- period price and promotion data and coefficients should be
tion coefficients, which is also consistent with the observa- used in the first set and the post-SR period price and pro-
tion that changes in prices and promotions were exogenous. motion data and coefficients should be used in the second
We compared the two methods and found the results to be set of outcome measures.
similar. Note that our approach can be easily modified for We begin by summarizing the overall impact of SR at the
the situation in which price and promotion changes are category level in terms of purchase incidence probability,
552 JOURNAL OF MARKETING RESEARCH, NOVEMBER 2007

Table 3
PARAMETER ESTIMATES FOR MARGARINE

Variables/Parameters Segment 1 Segment 2 Segment 3

Brand Utility (baseline: Shedd’s Country Crock)
αk: Brummel & Brown .903* .626 –2.626***
Fleischmann’s –.609 .572 –2.462***
I Can’t Believe It’s Not Butter –.487 .272 –2.358***
Imperial –.305 .225 –.792**
Land O’ Lakes –1.562*** .267 –1.793***
Parkay –.166 .762 –1.288***
Promise .626* –.265 –3.129***
Private label .559 –.096 –.405
βB: Regular price (before SR) .088 –.055 .132
Price cut (before SR) .019 .020 –.043
Regular price (after SR) .073 –.014 .118
Price cut (after SR) –.048 .107 .004
State dependence (γ) 4.292*** 3.929*** 3.048***
δBk (SR): Brummel & Brown .686 –1.204* –.127
Fleischmann’s –.003 –1.170* –.699
I Can’t Believe It’s Not Butter .912 –.326 .096
Imperial .266 –.934 –.294
Land O’ Lakes 1.233* –.747 –.493
Parkay .477 –1.701** –.495
Promise –.425 –.601 .438
Private label –1.074 –.423 –.010

Category Threshold
Constant 7.584*** 7.181*** 6.704***
Purchase frequency –6.097*** –9.177*** –5.302***
Previous purchase volume .004 .004 –.000
log(t) .214** .186** .071
δI* (SR) –.234* .487* .635**

Purchase Quantity
αQk: Brummel & Brown 5.089*** 17.806*** 8.485***
Fleischmann’s 1.342 10.130*** 10.062***
I Can’t Believe It’s Not Butter 2.201** 17.194*** 8.561***
Imperial .649 17.065*** 8.498***
Land O’ Lakes 2.836** 17.141*** 8.896***
Parkay 1.376 33.675*** 16.271***
Promise 2.574** 9.357*** 9.121***
Shedd’s Country Crock 2.309** 47.942*** 15.424***
Private label 4.649*** 16.644*** 15.355***
βQ: Regular price (before SR) .375 –1.025** –.519
Price cut (before SR) 2.663*** 1.268** .492
Regular price (after SR) .334 –1.025** –.594
Price cut (after SR) 2.637*** 1.253** .593
Average purchase volume .854*** .017 .046*
log(t) –.502*** .036 .017
δQ (SR) .533 .140 .899

φ (brand similarity) .042 .053 .035

θ (interdependence of incidence, choice, quantity) .082
Segment size 36.0% 33.6% 30.4%

–Log-likelihood [–L(β)] 17,988.2 ρ2 .766

Number of parameters (M) 132 Adjusted ρ2 .768
*p < .10.
**p < .05.
***p < .01.
Notes: ρ2 = 1 – L(β)/L(0), and adjusted ρ2 = 1 – [L(β) – M]/L(0), where L(0) = log-likelihood when all parameters are zero.

average quantity per purchase occasion, total purchase tity and revenue (dropped by 17.8% and 15.7%, respec-
quantity, and revenue. The three categories exhibit vast dif- tively).6 It appears that the average purchase incidence
ferences. The spaghetti sauce category enjoyed the greatest probability across households decreased substantially for
increase in total purchase quantity and revenue (12.4% and margarine (21.2%) and a little for detergent (1.4%) but
15.4%, respectively) due to the SR. The liquid detergent
category also experienced an increase in purchase quantity 6Note that our category-level findings may be due to the SR program
and revenue (7.1% and 10.2%, respectively). In contrast, being storewide. These results are likely to be modified if an SR is imple-
the overall reaction to the SR in the margarine category was mented in a single category. We thank an anonymous reviewer for this
negative, with a sharp decline in both total purchase quan- insight.
Brand-Level Effects of SKU Reductions 553

Table 4
PARAMETER ESTIMATES FOR SPAGHETTI SAUCE

Variables/Parameters Segment 1 Segment 2 Segment 3

Brand Utility (baseline: Ragu)
αk: Barilla .670 .788** –3.540***
Classico .871 .829** –3.382**
Five Brothers .855 .566 –4.654***
Healthy Choice –.865 –1.757** –2.442***
Hunt’s –4.535*** –3.772*** –1.039
Newman’s Own 1.030 .299 –3.881***
Prego –.171 –.027 –.147
Private label –.875 –1.554** –2.114***
βB: Regular price (before SR) –.451*** –.472** –.003
Price cut (before SR) .138 .438*** .126
Regular price (after SR) –.356* –.401* –.760***
Price cut (after SR) .130 .386 .607*
State dependence (γ) 6.277*** 3.033*** 2.458***
δBk (SR): Barilla 1.298 .112 2.436**
Classico .222 .004 2.735**
Five Brothers .139 –.049 3.407**
Healthy Choice 1.982** .674 1.758**
Hunt’s –10.732*** .576 –.296
Newman’s Own –.405 –.643 1.781*
Prego 2.578*** 1.275 .723
Private label –1.422* –17.332 1.236*

Category Threshold
Constant 5.059*** 2.357*** 1.808***
Purchase frequency –6.797*** –7.084*** –2.533***
Previous purchase volume .079*** –.007 .028*
log(t) .039 .252** .341*
δI* (SR) –.397 –.042 .278

Purchase Quantity
αQk: Barilla 2.351*** 8.515*** 3.001***
Classico 1.300** 9.247*** 2.940***
Five Brothers .880 8.401*** 2.413***
Healthy Choice –1.430** 9.540*** 2.503***
Hunt’s .643 5.925*** 2.082***
Newman’s Own .530 7.646*** 2.602***
Prego .122 7.894*** 2.404***
Ragu –.114 6.656*** 2.458***
Private label –1.110* 6.746*** 2.349***
βQ: Regular price (before SR) –1.819** –6.634*** –1.690***
Price cut (before SR) 4.176*** .125 5.639***
Regular price (after SR) –1.914** –6.720*** –1.613**
Price cut (after SR) 3.828*** .518 5.236***
Average purchase volume 1.173*** .475*** .263***
log(t) .031 –.322** .049
δQ (SR) 1.141*** .454 –.660

φ (brand similarity) .182 .290 .951

θ (interdependence of incidence, choice, quantity) –.899
Segment size 15.9% 53.2% 30.9%

–Log-likelihood [–L(β)] 9348.1 ρ2 .781

Number of parameters (M) 132 Adjusted ρ2 .778
*p < .10.
**p < .05.
***p < .01.
Notes: ρ2 = 1 – L(β)/L(0), and adjusted ρ2 = 1 – [L(β) – M]/L(0), where L(0) = log-likelihood when all parameters are zero.

increased slightly for spaghetti sauce (.9%), whereas the chase quantity in the 33 weeks of the pre-SR period. The
average quantity per purchase occasion increased moder- pattern of results for sales revenue is similar to the one for
ately for all three categories (3.5%–5.9%). Regardless of quantity, and therefore we do not discuss it separately.
whether category-level effects were positive or negative, Because our study focuses on the impact of SR on individ-
effects of SR at the brand level were mixed in every cate- ual brands, we average both measures (choice and purchase
gory. We examine this next. quantity) across households. Note that these two measures
Two key measures at the brand level appear in Table 5: do not always move in the same direction, because total
the average conditional brand choice probability given a purchase quantity is affected not only by changes in the
category purchase incidence and the average household pur- choice probability and purchase quantity on each occasion,
554 JOURNAL OF MARKETING RESEARCH, NOVEMBER 2007

Table 5
BRAND-LEVEL EFFECTS OF SR

Average Conditional Brand Choice Probability Average Household Purchase Quantity in 33 Weeks (Ounces)
a Without SR With SR Differencea % Difference Without SR With SR Differencea % Difference
Liquid Detergent
Wisk .097 .116 .019** +20 44.6 54.5 9.9** +22
All .078 .085 .007* +9 34.0 31.9 –2.1 –6
Tide .502 .548 .046** +9 331.9 372.7 40.8** +12
Cheer .085 .055 –.030** –35 32.1 23.3 –8.8** –27
Arm & Hammer .058 .044 –.014** –24 26.1 17.3 –8.8** –34
Era .044 .029 –.015** –34 22.7 18.4 –4.3* –19
Dreft .067 .068 .001 +1 21.5 40.8 19.3** +90
Surf .033 .024 –.009** –27 15.9 11.9 –4.0** –12
Private label .037 .031 –.006** –16 14.2 11.1 –3.1** –22

Margarine
Brummel & Brown .057 .055 –.002 –4 4.9 4.2 –.7* –14
Fleischmann’s .097 .084 –.013** –13 5.1 4.1 –1.0** –20
I Can’t Believe It’s
Not Butter .231 .255 .024** +10 18.7 18.1 –.6 –3
Imperial .127 .121 –.006 –5 9.5 7.4 –2.1** –22
Land O’ Lakes .097 .101 .004 +4 5.6 4.4 –1.2** –21
Parkay .077 .065 –.012** –16 8.4 4.8 –3.6** –43
Promise .081 .078 –.003 –4 6.7 6.9 .2 +3
Shedd’s Country
Crock .177 .190 .013** +7 20.0 15.2 –4.8** –24
Private label .055 .050 –.005* –9 5.0 3.7 –1.3** –26

Spaghetti Sauce
Barilla .081 .096 .015** +19 9.6 12.6 3.0** +31
Classico .163 .234 .071** +44 19.5 26.4 6.9** +35
Five Brothers .057 .070 .013** +23 4.9 6.2 1.3** +27
Healthy Choice .028 .035 .007** +25 3.1 3.7 .6** +19
Hunt’s .022 .009 –.013** –59 2.7 .8 –1.9** –70
Newman’s Own .080 .058 –.022** –28 8.7 9.6 .9 +10
Prego .293 .311 .018** +6 47.8 58.8 11.0** +23
Ragu .243 .176 –.067** –28 40.3 37.5 –2.8 –7
Private label .033 .012 –.021** –64 3.2 1.5 –1.7** –53
*p < .05.
**p < .01.
aBased on t-test of the difference.

which are brand specific, but also by changes in the pur- effects of marketing-mix variables. We find that merely
chase incidence probability, which are category specific comparing a brand’s market share before and after the SR
and, therefore, the same for all brands in the category. does not provide an accurate assessment of the effects of the
Table 5 shows a high degree of variation among brands SR on brand choice, because it does not account for other
on the two measures. It appears that SR not only changed changes in the marketing-mix variables. For example, in the
the category purchase incidence but also altered consumers’ liquid detergent category, a simple comparison of market
choice among the remaining brands, so that some brands share before and after the SR (see Additional Table 1 in the
gained market share and others lost market share. As a Web Appendix at http://www.marketingpower.com/content
result, in every category, some brands gained in total sales 84060.php) would lead to the conclusion that the SR caused
quantity, and others suffered substantial sales loss due to the Tide to lose market share by 3.2% and caused the private
SR, regardless of the direction of the category-level effect. label brand to gain market share by 2.2%, whereas our
For example, although the total category purchase quantity analysis in Table 5 indicates the exact opposite; the pur-
for spaghetti sauce increased by 12.4% with the SR, the chase share increased by 4.6 percentage points for Tide and
Hunt’s brand and private label both suffered a severe loss in decreased by .6 percentage points for the private label, both
sales (by 70% and 53%, respectively). Conversely, despite a of which are statistically significant.
loss in total category sales quantity of 17.8% for margarine, Drivers of the differences in SR effects among consumers
there was no significant change in sales for the I Can’t and brands. To obtain a comprehensive picture of how the
Believe It’s Not Butter brand and even a slight increase in SR affected purchase behavior across consumers and brands
sales for the Promise brand. Note that in all three cate- differently, we conduct regression analyses based on data
gories, SR seems to have caused a substantial drop in pooled across categories on each of the three purchase deci-
choice probability and total purchase quantity for the pri- sion components: category purchase incidence, conditional
vate label brand, which should be alarming to the retailer. brand choice probability, and quantity given a purchase
We discuss the possible reasons for this subsequently. occasion. We analyze the first and third components at the
The results for choice probabilities demonstrate the household level because the effects do not vary across
importance of a would-be analysis that controls for the brands (with 669 observations in each analysis), and we
Brand-Level Effects of SKU Reductions 555

analyze the second component at the household × brand the table shows, state dependence and purchase frequency
level (6021 observations resulted from having nine brands have positive, significant effects on the change in purchase
in each category and 669 households in the data of three incidence probability, and the former also has a positive,
categories combined). significant effect on the change in quantity given a purchase
We first describe the regression analyses on the SR occasion, which indicates that the higher a household’s state
effects on households’ average purchase incidence probabil- dependence and purchase frequency, the more positive its
ity and purchase quantity, given a category purchase occa- reaction was to the SR. A closer examination of the data
sion. Both dependent variables are standardized within each reveals that even for the margarine category, which experi-
category because the categories differ in the magnitude of enced substantial sales reduction due to the SR, the highly
purchase incidence probabilities and purchase quantities. state-dependent and frequent buyers of this category still
We use the following five household-specific factors as increased their purchase incidence probability and/or aver-
explanatory variables: age quantity for purchase occasion. This is particularly
encouraging news for retailers because consumers of high
1. Level of state dependence, state dependence and purchase frequency are arguably their
2. Purchase frequency, most valuable customers. We did not find significant effects
3. Whether a household’s favorite brand was eliminated by the of the other three factors, possibly because of low variations
SR, in these variables across households and strong collinearity
4. Whether a household’s favorite SKU was eliminated by the among them in the data.
SR, and
5. Eliminated SKUs’ share of purchase quantity for a
We now describe the regression analysis for the SR
household. effects on the conditional brand choice probabilities. The
dependent variable is the difference in a conditional choice
We standardize the level of state dependence and pur- probability (with and without SR) obtained from the would-
chase frequency within each category to control for the be analysis. We examine two groups of brand-specific fac-
magnitude differences across categories. A favorite brand/ tors: brand characteristics and brand-level SR variables.
SKU is defined as the most frequently purchased brand/ The brand characteristics are as follows:
SKU by a household in the pre-SR period. We did not •Market share: This refers to the market share before the SR.
include a variable for whether a household’s favorite size •Price level: Because prices are not directly comparable across
was eliminated, because no household in our data experi- categories, we use the standardized average shelf price for a
enced it. The results appear in Table 6 (Models 1 and 2). As brand within each category.

Table 6
DRIVERS OF THE DIFFERENCES IN THE SR EFFECTS ACROSS CONSUMERS AND BRANDS

Dependent Variable Explanatory Variable Parameter Estimate p-Value

Model 1: Intercept –.0147 .5686
Standardized change in category Standardized state dependence .8042 <.0001
purchase incidence Standardized purchase frequency .0469 .0112
Favorite brand being eliminated –.2059 .6320
Favorite SKU being eliminated .1730 .2578
Eliminated SKUs’ share of household purchase .0954 .6587

Model 2: Intercept .0090 .8221

Standardized change in quantity Standardized state dependence .3563 <.0001
given a purchase occasion Standardized purchase frequency .0516 .1623
Favorite brand being eliminated –.9686 .1475
Favorite SKU being eliminated .1972 .4065
Eliminated SKUs’ share of household purchase –.2339 .4861

Model 3: Intercept –.0278 <.0001

Change in conditional choice Market share .1259 <.0001
probability Standardized price level .0095 <.0001
log(promotion frequency) .0355 <.0001
Number of sizes eliminated (DSIZE) .0028 .1238
DSIZE × share of purchase –.0557 <.0001
Number of SKUs eliminated (DSKU) .0064 <.0001
DSKU2 –.0013 <.0001
Change in SKU share .1119 <.0001
Eliminated SKUs’ share of brand sales .0210 .0978

Model 4: Intercept .0260 <.0001

Sum of squares of changes in brand Standardized state dependence –.0211 <.0001
choice probabilities Standardized purchase frequency .0006 .7826
Favorite brand being eliminated –.0253 .5160
Favorite SKU being eliminated .0059 .6680
Eliminated SKUs’ share of household purchase .0333 .0887
Standardized regular price sensitivity –.0232 <.0001
Standardized price promotion sensitivity .0364 <.0001
556 JOURNAL OF MARKETING RESEARCH, NOVEMBER 2007

•Promotion frequency: This is defined as the percentage of their minds more easily. The effects of market share, price
weeks in which a brand was on price promotion. level, and promotion frequency may also explain why the
private label brand suffered sales loss in each category;
Because the retailer or manufacturers could potentially specifically, private labels tend to have small market share,
change the prices and promotions as a response to the SR, low price levels, and less frequent promotions (as is the case
which could create an endogeneity problem, we use price in our data).
level and promotion frequency before the SR as explanatory Furthermore, we find that the nature of the assortment
variables. change also affects how consumers reallocate their pur-
The brand-level SR variables are as follows: chases. Brands with a larger reduction in the number of
sizes tend to lose share. Notably, this effect appears to
•Number of SKUs eliminated for the brand (DSKU): We include increase with a household’s share of purchase of a brand. In
the quadratic terms of DSKU to capture possible nonlinear other words, the more loyal a household is to a brand, the
effect of this variable. Boatwright and Nunes (2001) find that more severe the negative effect is for the household when
the number of SKUs eliminated for a category has a nonlinear
effect on the category sales; a moderate cut increases sales, and
the brand had a size reduction. Another notable finding is
a deep cut decreases sales. We investigate whether such an that the nonlinear effect of the number of SKUs eliminated,
effect also exists at the brand level. as first documented by Boatwright and Nunes (2001) for
•Number of sizes eliminated (DSIZE): We include both a main category-level sales, exists at the brand level. All else being
effect of DSIZE and its interaction with a brand’s share of pur- equal, a brand’s choice probability first increases with the
chase for a household. As in many previous studies (e.g., number of SKUs eliminated and then decreases after a cer-
Bucklin, Gupta, and Siddarth 1998; Tellis and Zufryden 1995), tain point. This implies that a brand could benefit from the
we use the share-of-purchase variable as a measure for a elimination of certain SKUs to reduce clutter and thus ease
household’s loyalty to a given brand. The interaction term is consumers’ purchase decisions, but a deep reduction in its
intended to capture the possible moderating effect of a house- number of SKUs would decrease its appeal to consumers.
hold’s loyalty to a brand on the impact of size reduction for the
brand.
Finally, we find that an increase in a brand’s share of SKUs
•Change in the share of SKUs (ΔSKUSH): We define a brand’s in the category tends to increase its share of purchases; all
share of SKUs (SKUSHR) as its number of SKUs divided by else being equal, if share of SKUs increases by 1 percent-
the total number of SKUs in the category, and ΔSKUSHR = age point, on average, the market share would increase by
SKUSHRafter – SKUSHRbefore. .11 percentage points.
•Eliminated SKUs’ share of brand sales: This is the proportion To examine how households differed in the SR effects on
of a brand’s sales in the pre-SR period contributed by the their brand choice probabilities, we first computed the sum
eliminated SKUs. of squares of changes in the conditional brand choice
probabilities for each household and then ran a regression
Parameter estimates from the regression appear in Table analysis of this variable on household characteristics.9 This
6 (Model 3).7 For the SR effects on brand choice, we iden- sum-of-squares measure captures a household’s total
tified six significant drivers: (1) market share, (2) price amount of changes in brand choice probabilities due to SR.
level, (3) logarithm of the promotion frequency,8 (4) num- In addition to the five household characteristics listed previ-
ber of sizes eliminated and its interaction with a brand’s ously, we also included standardized household-specific
share of purchase for a household, (5) the quadratic terms posterior regular price and price cut coefficients for the pre-
of number of SKUs eliminated, and (6) change in the share SR period as explanatory variables in the regression model.
of SKUs. Notably, we did not find a significant effect of the (We tested the regular price and price cut coefficients in
eliminated SKUs’ share of brand sales. both the choice and the quantity parts and found the choice
The effects of market share, price level, logarithm of the coefficients to be nonsignificant. Therefore, we kept only
promotion frequency, and change in the share of SKUs are the quantity price and promotion sensitivity in the final
positive. It appears that after the SR, all else being equal, model.) The results appear in Table 6 (Model 4). Model 4
market shares tend to shift toward larger brands, higher- indicates that the SR program caused more changes in
priced brands, and brands with more frequent promotions. brand choice probabilities among consumers of lower state
These effects imply that when consumers are faced with a dependence and higher price and promotion sensitivity.
reduced product assortment and reallocate their purchases When we combine findings from Models 1, 2, and 4, it
among the remaining brands and product options, they are appears that the SR program affected consumers of high
likely to switch to familiar brands or premium brands; state dependence and purchase frequency mainly through
higher-priced brands tend to be premium brands, larger its effects on purchase incidence and quantity decisions,
share brands have higher market exposure, and more fre- whereas its impact on brand choice behavior is most pro-
quent promotions also help bring consumers’ attention to a found for consumers of low state dependence and high price
brand. A possible reason behind the switch is that the SR and promotion sensitivity.
program may have reinforced consumers’ need for ease of The significant drivers identified in Models 1–3 all con-
shopping, and familiar or premium brands tend to come to tribute to the differences across brands and households in
the SR effects on a household’s total brand purchase quan-
7We report the correlation coefficients of the variables in Additional
tity because this purchase quantity results from the three
Table 2 (see the Web Appendix at http://www.marketingpower.com/content
individual purchase components. This is confirmed by our
84060.php). regression analysis for the SR effects on total brand pur-
8Our empirical analysis shows that the logarithm of promotion fre- chase quantity during the 33 weeks in the pre-SR period.
quency provides better fit to the data than the linear form of promotion
frequency. 9We thank an anonymous reviewer for suggesting this analysis.
Brand-Level Effects of SKU Reductions 557

DISCUSSION detergent also had an increase in its total sales volume

As we stated previously, prior research on product assort- (7.1%) and revenue (10.2%). In contrast, margarine suf-
ment reductions has mainly focused on their impact at the fered a substantial decrease in both measures (–17.8% and
store and category levels, which is important for retailers. –15.7%, respectively). The pattern seems related to the total
By focusing on brand-level effects, our study provides addi- number of SKUs in each category. It appears that reducing
tional insights for both manufactures and retailers. The assortment in a category with a large number of SKUs helps
most relevant issue for a manufacturer regarding SR is the cut down its clutter and thus has a beneficial effect, whereas
impact on its brands. The manufacturer wants to know what eliminating SKUs in a category with an already low number
it can do to emphasize its brands’ strengths and minimize of SKUs can elicit strong negative assortment perceptions
among consumers (Broniarczyk, Hoyer, and McAlister
negative repercussions. Retailers benefit from a brand-level
1998). This finding implies that retailers should take cau-
analysis by obtaining a better understanding of how market
tion in selecting categories for implementing SR, a point of
share may shift between brands with different profitability
view we share with Borle and colleagues (2005).
and what may happen to their private label brands; this type
We find that attributing differences in market share to the
of insights is not possible with a store- or category-level
effect of SKU change may be misleading because there can
approach.
be confounding changes in other marketing-mix variables.
To summarize the key findings of our analyses, we found
For example, in the liquid detergent category, a simple com-
a large variation in the impact of SR across consumers and
parison of market share before and after the SR suggests
individual brands and categories. Consumers with higher
that the SR decreased the choice for Tide and increased the
state dependence and more frequent purchases welcomed choice for the private label brand; however, our analysis
the change by increasing category purchase frequency and indicates that the opposite occurred. Our results also indi-
purchase quantity, whereas consumers with lower state cate that eliminated SKUs’ share of brand sales does not
dependence and less frequent purchases reduced their pur- predict changes in a brand’s choice probability and pur-
chases as a result of SR. In addition, the SR program caused chase quantity after the SR. Retailers have mainly focused
more changes in brand choice probabilities among con- on sales measures in efficient assortment decisions. How-
sumers of lower state dependence and higher price and pro- ever, our findings imply that they should instead focus
motion sensitivity. attention on other factors that may play a more important
At the brand level, consumers appeared to reallocate their role in contributing to the differences in SR effects among
purchases among brands substantially after the SR. Both brands, such as those we identified herein.
brand characteristics and the nature of the SR influenced For manufacturers of large-share brands or premium
how they chose among these brands. As a result of the brands, our results suggest that even if an SR program
effects at the category and brand levels, total sales quantity decreases the overall category sales in a store, they need not
and revenue for some brands were not affected much by the worry as much as manufacturers of small-share brands or
SR, whereas other brands experienced drastic reductions (or lower-priced brands. In addition, it would help a brand gain
increases) in brand choice and purchase quantity. We identi- market share if the brand were made more prominent in
fied six significant drivers for the differential effects on consumers’ minds, such as by increasing its promotion fre-
brand choice probabilities: market share, price level, pro- quency. Moreover, our results imply that though a manufac-
motion frequency, number of sizes eliminated, number of turer may not have control over retailers’ SR initiatives, it
SKUs eliminated, and change in the share of SKUs in the can mitigate the potential negative consequence on its
category. Brands with higher market shares, higher price brands. For example, if a manufacturer must eliminate a
levels, and more frequent promotions tended to gain share. certain number of SKUs, it should negotiate with the
A moderate reduction in the number of SKUs increased a retailer to minimize reduction in the number of sizes of its
brand’s choice probability, whereas a deep cut hurt its brands.
chances of being chosen. In addition, an increase in a Our study also offers valuable implications for retailers.
brand’s share of SKUs was likely to translate into higher The finding that consumers with high state dependence and
purchase share. Finally, brands that experienced a smaller purchase frequency reacted favorably to the SR is encourag-
cut in the number of sizes gained share from those that had ing news to retailers that are faced with the dilemma of
greater size reduction; this effect was accentuated by a whether to implement “efficient assortment” policies. The
household’s share of purchase of the brand. The significant finding that large-share brands tend to gain shares after an
effect of the number of sizes is consistent with the findings SR is also good news for retailers from the perspective of
of Guadagni and Little (1983), who show that consumers trade relations. In addition, because the number of sizes
exhibit high loyalty to size. In summary, our findings eliminated, the number of SKUs eliminated, and the change
demonstrate that how a brand is affected by SR is not only in the share of SKUs all play a role in how a brand’s choice
driven by the assortment change but also influenced by the probability is affected by an assortment reduction, a retailer
characteristics of the brand and its consumers. should use discretion in determining which SKUs to elimi-
Although our study did not set out to investigate how and nate. The results from our first-stage analysis suggest that
why categories differ in their SR effect on sales (for in- private label brands are likely to suffer more unfavorable
depth analyses of this topic, see Boatwright and Nunes consequences of SR than national brands. This cautions the
2001; Borle et al. 2005), we observe that the effects varied retailer to minimize the negative effects of SR on its private
substantially across the three categories examined here. label brands; a retailer can take some preventive measures,
Spaghetti sauce experienced the largest increase in total such as maintaining the number of private label SKUs (and
sales volume (12.8%) and revenue (14.8%), and liquid thus increasing its share of SKUs in the category) and
558 JOURNAL OF MARKETING RESEARCH, NOVEMBER 2007

increasing the promotion frequency of the private label REFERENCES

brands. If a reduction of private label SKUs is inevitable, Andrews, Rick L., Andrew Ainslie, and Imran S. Currim (2002),
the retailer should try to minimize the reduction in the num- “An Empirical Comparison of Logit Choice Models with Dis-
ber of sizes offered. crete Versus Continuous Representations of Heterogeneity,”
Our analysis is based on data provided by an online Journal of Marketing Research, 39 (November), 479–87.
retailer. As we mentioned previously, the online store envi- Arnold, Stephen J., Te H. Oum, and Douglas J. Tigert (1983),
ronment provides a unique opportunity to study the impact “Determinant Attributes in Retail Patronage: Seasonal, Tempo-
of assortment changes without confounding it with the ral, Regional, and International Comparisons,” Journal of Mar-
effects of product display, shelf space allocation, or location keting Research, 20 (May), 149–57.
Bell, David R., Jeongwen Chiang, and V. Padmanabhan (1999),
on the shelf. Nonetheless, we expect the effects to be some- “The Decomposition of Promotional Response: An Empirical
what different in online and bricks-and-mortar stores Generalization,” Marketing Science, 18 (4), 504–526.
because of differences in consumer purchase behavior Boatwright, Peter and Joseph C. Nunes (2001), “Reducing Assort-
between the two channels. Previous research has shown that ment: An Attribute-Based Approach,” Journal of Marketing, 65
online consumers are more convenience conscious, more (July), 50–63.
state dependent, and less promotion sensitive than their ——— and ——— (2004), “Correction Note for ‘Reducing
offline counterparts (e.g., Degeratu, Rangaswamy, and Wu Assortment: An Attribute-Based Approach,’” Journal of Mar-
2000; Zhang and Wedel 2007). Combining this with the keting, 68 (July), iv.
findings from the current study, we expect that, all else Borle, Sharad, Peter Boatwright, Joseph B. Kadane, Joseph C.
being equal, the SR effects on purchase incidence, quantity Nunes, and Galit Shmueli (2005), “The Effect of Product
Assortment Changes on Customer Retention,” Marketing Sci-
given a purchase occasion, and total sales and revenue will ence, 24 (4), 616–22.
be more negative in bricks-and-mortar stores than the Broniarczyk, Susan M., Wayne D. Hoyer, and Leigh McAlister
results of this study suggest. In addition, we expect that (1998), “Consumers’ Perceptions of the Assortment Offered in a
there will be stronger brand-switching effects (i.e., those in Grocery Category: The Impact of Item Reduction,” Journal of
Model 4) among bricks-and-mortar consumers than among Marketing Research, 35 (May), 166–76.
online consumers. Bucklin, Randolph E., Sunil Gupta, and S. Siddarth (1998),
This research attempts to direct academic focus on brand- “Determining Segmentation in Sales Response Across Con-
level effects of SR. Many worthwhile topics remain to be sumer Purchase Behaviors,” Journal of Marketing Research, 35
explored in this area. In terms of methodology, a limitation (May), 189–97.
of our model is that it does not incorporate serial correlation Business Wire (1998), “Sunbeam Completes Record Year for
Sales, Earnings & Global Expansion,” (January 28), 1.
across observations for the same household. Our model Chiang, Jeongwen (1991), “A Simultaneous Approach to the
accounts for state dependence, consumer heterogeneity, and Whether, What, and How Much to Buy Questions,” Marketing
interdependence in the three purchase components, which Science, 10 (Fall), 297–315.
helps mitigate the impact of serial correlation. Nonetheless, Chintagunta, Pradeep K. and Sudeep Haldar (1998), “Investigating
further research could extend our model by incorporating Purchase Timing Behavior in Two Related Product Categories,”
serial correlation. The model could also be constructed at Journal of Marketing Research, 35 (February), 43–53.
the SKU level and then aggregated to the brand level. It Degeratu, Alexandru, Arvind Rangaswamy, and Jianan Wu (2000),
would be worthwhile to compare the results from SKU- “Consumer Choice Behavior in Online and Traditional Super-
level models with our results from brand-level models. In markets: The Effects of Brand Name, Price, and Other Search
terms of substantive topics, an important issue for further Attributes,” International Journal of Research in Marketing, 17
(1), 55–78.
research would be to incorporate product attributes/features Drèze, Xavier, Stephen J. Hoch, and Mary E. Purk (1994), “Shelf
in the assessment of the impact of assortment changes. Management and Space Elasticity,” Journal of Retailing, 70 (4),
Another important issue would be to conduct direct prof- 301–326.
itability analyses of SR programs, which would require Guadagni, Peter M. and John D.C. Little (1983), “A Logit Model
cost-saving information from both retailers and manufactur- of Brand Choice Calibrated on Scanner Data,” Marketing Sci-
ers. In addition, issues such as the nature of product cate- ence, 2 (3), 203–238.
gory on brand-level effects of SR and the impact of a Gupta, Sachin, Pradeep K. Chintagunta, and Dick R. Wittink
brand’s SR in one category on the same brand sales in (1997), “Household Heterogeneity and State Dependence in a
another category, among many other topics, could be exam- Model of Purchase Strings: Empirical Results and Managerial
ined. Finally, the finding that brands with larger market Implications,” International Journal of Research in Marketing,
14 (4), 341–57.
shares and higher prices tend to gain shares from those with Hanemann, W. Michael (1984), “Discrete Continuous Models of
smaller market shares and lower prices, such as private Consumer Demand,” Econometrica, 52 (3), 541–61.
labels, has important implications for both manufacturers Home Textile Today (2005), “Pier 1 Reduces SKUs to Up Sales,”
and retailers. This study is the first to demonstrate such a (March 14), 4.
pattern. Although these effects are consistent with our con- Iyengar, S. and M. Lepper (2000), “When Choice Is Demotivating:
jecture, we believe that more empirical studies are needed Can One Desire Too Much of a Good Thing?” Journal of Per-
to test their generalizability. This would be another worthy sonality and Social Psychology, 79 (6), 995–1006.
direction for further research.10 Jain, Dipak C. and Naufel J. Vilcassim (1991), “Investigating
Household Purchase Timing Decisions: A Conditional Hazard
Function Approach,” Marketing Science, 10 (1), 1–23.
Kamakura, Wagner A. and Gary J. Russell (1989), “A Probabilistic
10We thank an anonymous reviewer for suggesting several of the ideas Choice Model for Market Segmentation and Elasticity Struc-
discussed here. ture,” Journal of Marketing Research, 26 (November), 379–90.
Brand-Level Effects of SKU Reductions 559

Kurt Salmon Associates (1993), “Efficient Consumer Response: Zhang, Jie and Lakshman Krishnamurthi (2004), “Customizing
Enhancing Consumer Value in the Grocery Industry,” Food Promotions in Online Stores,” Marketing Science, 23 (4),
Marketing Institute Report No. 9-526, Washington, DC. 561–78.
Seetharaman, P.B., Andrew Ainslie, and Pradeep Chintagunta ——— and Michel Wedel (2007), “The Effectiveness of Cus-
(1999), “Investigating Household State Dependence Effects tomized Promotions in Online and Offline Stores,” working
Across Categories,” Journal of Marketing Research, 36 paper, Robert H. Smith School of Business, University of
(November), 488–500. Maryland.
Tellis, Gerard J. and Fred S. Zufryden (1995), “Tackling the
Retailer Decision Maze: Which Brands to Discount, How
Much, When, and Why?” Marketing Science, 14 (3), 271–99.