1 POLYGAMOUS STORE LOYALTIES: AN EMPIRICAL INVESTIGATION Qin Zhang † Manish Gangwar P.B. Seetharaman August 31, 2017 Forthcoming in Journal of Retailing † Qin Zhang is Assistant Professor at School of Business, Pacific Lutheran University. Manish Gangwar is Assistant Professor of Marketing at Indian School of Business, Hyderabad, India. P. B. Seetharaman is W. Patrick McGinnis Professor of Marketing at Olin Business School, Washington University in St. Louis. Corresponding author: Qin Zhang, e-mail: [email protected], Ph: 253-535-7253.
49
Embed
POLYGAMOUS STORE LOYALTIES: AN EMPIRICAL INVESTIGATION Category Loyalty.pdf · 6 Our analysis, which uncovers intrinsic store-category attractiveness, confirms our conjecture that
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
POLYGAMOUS STORE LOYALTIES:
AN EMPIRICAL INVESTIGATION
Qin Zhang†
Manish Gangwar
P.B. Seetharaman
August 31, 2017
Forthcoming in Journal of Retailing
†Qin Zhang is Assistant Professor at School of Business, Pacific Lutheran University.
Manish Gangwar is Assistant Professor of Marketing at Indian School of Business, Hyderabad, India.
P. B. Seetharaman is W. Patrick McGinnis Professor of Marketing at Olin Business School, Washington University in
Safeway, Trader Joe’s, and Wild Oats Market, two supercenters ─ Super Kmart and Wal-mart
Supercenter, and three warehouse club chains ─ Costco, Sam’s Club, and Smart & Final.
4. THE MODEL FREE EVIDENCE FOR POLYGAMOUS STORE LOYALTIES
First, we demonstrate the presence of overall store loyalties in our dataset. In Figure 1, we
report the histogram of a number of different stores at which all 1321 households shop.
Throughout the 53-week study period, we see that only 12 out of the 1321 households shop at a
12
single store, the modal value is six, and there are three households that shop at as many as 13
different stores.2 Next, for each household we identify its favorite store, i.e., the store at which the
household makes the largest number of shopping trips over the study period. Subsequently, we
calculate the proportion of shopping trips made by each household at its favorite store over its total
number of shopping trips, and report the probability mass histogram for this proportion across all
1321 households in the dataset in Figure 2. We observe that 50.2% of the households do not visit
their favorite store on 50% or more of their shopping trips. The two figures indicate that
households typically divide their grocery shopping among many different stores and that there
appears to be little overall store loyalty for these households based on the traditional view of store
loyalty.
[INSERT FIGURE 1 and 2 HERE]
Next, we add the category dimension into the analysis, and something interesting emerges.
We observe that each of the 1321 households, including those who shop at multiple stores, makes
all of its category purchases exclusively at the same store for at least one category. Figure 3
displays the frequency distribution of households across the number of categories in which a
household is observed to make all its purchases from a single store. The figure shows that many
households seem to purchase a large number of categories exclusively from one store.3 Since
different households may purchase a different number of categories, we also display the frequency
distribution of households in terms of the percentage of categories that each of these households
buys exclusively from one store in Figure 4. We find that, on average, the percentage of categories
that a household buys exclusively from one store is 38%. We also conduct similar calculations
2 For expositional convenience, we use “store” and “retail chain” interchangeably. 3 We also draw a figure similar to Figure 3 but only consider household-category combinations where at least 10
category purchases are made by corresponding households. As in Figure 3, the figure also shows that many
households seem to purchase a large number of categories exclusively from one store. However, low frequency
categories are shown to be more likely to be purchased exclusively at a single store. This figure is available from the
authors upon request.
13
from the category perspective. In Figure 5 we plot the frequency distribution of categories in terms
of the percentage of households that buy these categories exclusively at one store. We find that, on
average, among the households that make purchases in a category, 49% of them buy the category
exclusively from one store.
[INSERT FIGURE 3, 4 and 5 HERE]
One may argue that the findings in Figures 3, 4 and 5 could be attributed to the fact that
many households make most of their single-store category purchases exclusively at their favorite
stores while the rest of their purchases are scattered across the other stores. To evaluate whether
this is the case, for each of the households that makes single-store category purchases in at least
one category (in this case, all 1321 households), we first count the number of stores at which the
household makes single-store category purchases across all categories. We then plot a probability
mass of this count across all households in Figure 6. We observe that only about 10.2% of the
households make all of their single-store category purchases exclusively at one store. This
provides convincing evidence that many households do not make all of their single-store category
purchases exclusively at their favorite stores. Instead, households make single-store category
purchases at many different stores; in other words, households seem loyal to different stores for
different product categories.
[INSERT FIGURE 6 HERE]
Based on Figures 1-6, we can conclude that, despite the lack of appearance of overall store
loyalty for their grocery shopping, households do exhibit polygamous store loyalties, that is, the
consumers are attracted to different stores for different categories. Next, we further explore this
phenomenon to understand the key influencers of store-category loyalty from a long-term
perspective, particularly those that relate to the strategic merchandising programs of retailers as
opposed to tactical promotions. To achieve this goal, we propose a model that decomposes store-
14
category attractiveness, which determines store-category loyalty, into store-specific, store-
category-specific and store-household-specific effects.
5. EMPIRICAL ANALYSIS
5.1 The Proposed Model
In this study, we focus on a household’s long-term relationship with a store at the category
level. We conceptualize that store-category loyalty represents a household’s long-term propensity
to choose a store in a category. Specifically, we consider a market where H households
(h=1,2,…,H) make purchases in C categories (c=1,2,…,C) among S stores (s=1,2,…S). The
observed category purchases of household h in category c across S stores is represented by
1 2( , ,..., )h h h h
c c c ScN n n n , where h
scn denotes the total number of purchases made by household h in
category c at store s during a period. We assume that each observed store choice outcome vector in
a category, h
cN , is generated by a multinomial process determined by the underlying latent store-
category loyalty vector h
cSCL of a household h in category c across S stores, where
1 2, ,..., ,h h h h
c c c ScSCL scl scl scl i.e., ~ ( )h h
c cN Multinomial SCL , and h
cSCL is normalized to one such
that 1
1S
h
sc
s
scl
.4
We assume that h
cSCL follows the axioms of Bell, Keeney and Little (1975), which provide
a theoretical foundation for the attraction models. Following their conceptual framework, we also
assume that the household h’s loyalty for store s in category c, h
scscl , is proportional to the
attractiveness of store s to household h in category c , h
scA ; i.e. h h
sc scscl A . Cooper and Nakanishi
4 Alternatively, one can conceptualize a household’s SCL based on the household’s category expenditures at the store.
Since the correlation between shares of category expenditure and shares of category purchase incidences across stores
in our data is 0.977, we do not expect meaningful differences in results to emerge from using the alternative
conceptualization of SCL in our model.
15
(1988) propose a multinomial logit (MNL) specification for the attraction models for its logical
consistency. Accordingly, we also choose this widely accepted MNL specification for our
proposed model, which is written as follows:
1
exp
exp
h
sch
sc Sh
rcr
Ascl
A
(1)
Next, we discuss the factors that influence store-category attractiveness, h
scA , which in turn,
affects latent store-category loyalty, h
scscl .
5.2 Store-Category Attractiveness
The attractiveness of a store in a given category to a household can arise from multiple
factors that can be attributed to store characteristics, category characteristics, household
characteristics, and various combinations of these three types of characteristics. Thus, we
decompose store s’s attractiveness in category c to household h, h
scA , into various corresponding
components as follows:
,h h h h
sc s sc s c sc scA X (2)
where s represents the mean intrinsic attractiveness of store s; sc represents the intrinsic
attractiveness of store s in category c; s and sc together s sc represent the intrinsic store-
category attractiveness; h
s denotes household heterogeneity in store attractiveness.5 The term
h
c scX represents the household store-category attractiveness attributed to store s’s merchandising
effort in category c; scX is a vector of K variables representing store s’ merchandising strategies in
5 We also control for observed heterogeneity in store category attractiveness both in category and household
dimensions, but for ease of exposition we discuss it separately in section 6.2.3. See equation (5) for full specification
of the model.
16
category c (such as product assortments and pricing);h
c denotes household h’s response specific
to category c, which we call merchandising effectiveness (we will discuss h
c and scX in detail in
the next two subsections); and finally, the last term,h
sc , accounts for the household’s store-
category level idiosyncrasies in store category attractiveness (Cooper and Nakanishi 1988).
5.3 Merchandising Effectiveness
A household’s response to merchandising strategies may differ across categories. To better
understand these differences, we decompose merchandising effectiveness, h
c , into the following
components: (1) the mean effect common across categories and households, denoted by ; (2) the
effect specific to category but common to all households, c ; and (3) the effect unique to
household h but common across all categories, denoted by h (Ainslie and Rossi 1998,
Seetharaman; Ainslie and Chintagunta 1999; Singh, Hansen and Gupta 2004; and Prasad, Strijnev,
and Zhang 2008). Mathematically, the decomposition can be written as follows: 6
,h h h
c c c (3)
where h
c captures the residual unobserved deviation that is specific to both household h and
category c. Next, we discuss each merchandising variable in the vector of scX that represents store s’
merchandising strategies in category c.
5.4 Key Merchandising Variables
We are particularly interested in how merchandising strategies, which are under the control
of retailers, influence a store’s attractiveness in a category to households. The typical
merchandising variables can be constructed to represent a store’s product assortment, pricing and
6 Similar to intrinsic store attractiveness, we also control for observed heterogeneity in merchandising effectiveness in
both category and household dimensions. Again for ease of exposition, we discuss it separately in section 6.2.3. See
equation (6) for full specification.
17
promotional strategies (e.g., features, display and coupons, etc.). Empirical researchers have
shown that although temporary promotional activities are effective in achieving short-term goals
such as store traffic and sales, their effects on long-term measures such as market shares (Nijs,
Dekimpe, Steenkamp and Hanssens 2001) are minimal. Since store category attractiveness is
relatively stable over time and governed by consumers’ perceptions, in this study, we focus on
retailers’ product assortment and pricing strategies which are long-term in nature. In our model,
the effects of retailers’ promotional strategies are essentially subsumed in intrinsic store-category
attractiveness.
One of the challenges of modeling purchases in multiple categories in a single united
framework is that we need measures that are comparable across categories, across stores and
across different strategies. To achieve this objective, we follow Briesch, Chintagunta and Fox
(2009) to construct index variables that are normalized by corresponding average values and thus
operationalized to be unit-free and scale-free to ensure appropriate comparison.
5.4.1 Product Assortment Variables
Several studies, based on surveys and lab experiments, have revealed that product
assortments play an important role in consumers’ store evaluations and/or store choice decisions
(Meyer and Eagle 1982; Arnold, Oum and Tigert 1983; Craig, Ghosh and McLafferty 1984; and
Louviere and Gaeth 1987). It has further been shown that consumers’ perceptions of product
assortments are multi-dimensional (Broniarczyk, Hoyer and McAlister 1998; and Chernev and
Hamilton 2009). Therefore, we construct product assortment variables that capture the two most
important dimensions, namely, breadth and exclusiveness, of the category assortments.
First, we measure the assortment breadth in category c at the store s from two aspects –
brand breadth and SKU breadth – within a brand. We explain the two variables below:
Number of Brands in the Category at the Store (BRANDsc).
18
This variable is defined as
1
scsc S
rc
r
BRANDSBRAND
BRANDS S
, where BRANDS sc stands for
the total number of brands in category c available at store s.
Average Number of SKUs per Brand in the Category at the Store (SKUsc).
This variable is defined as
1
/
/
sc scsc S
rc rc
r
SKUS BRANDSSKU
SKUS BRANDS S
, where SKUSsc stands
for the total number of SKUs in category c available at store s.
Next, we construct variables to measure assortment exclusiveness. To the extent that a
private label is exclusive to the store, this measure can serve as a proxy for the exclusiveness of
the store’s product assortments.7 We construct the following assortment exclusiveness variable:
Number of Private Labels in the Category at the Store (PVTLABELsc).
This variable is defined as
1
scsc S
rc
r
PVTLABELSPVTLABEL
PVTLABELS S
, where, PVTLABELSsc
stands for the number of private label SKUs in category c available at store s.
5.4.2 Price Variables
One of the robust findings in research on store choice decisions is that the perception of
low prices is an important factor in driving positive consumer evaluations of stores (Baumol and
Ide 1956; Brown 1978; Meyer and Eagle 1982; Arnold, Oum and Tigert 1983; Bell and Lattin
1998; and Bell, Ho and Tang 1998). We construct the following variable to measure the
attractiveness of the store’s pricing in the category:
Price Index of the Store in the Category (PRICEsc).
7 We do recognize that retailers’ decisions of providing shoppers different private label options go beyond the purpose of just being
exclusive; for example, retailers offer private label products to gain higher profit margins or to have better negotiating leverage with
manufacturers of national brands (Ailawadi, Pauwels and Steenkamp 2008).
19
This store-category level variable is operationalized as the average of normalized SKU
prices (i.e., price of an SKU divided by the average SKU price in the dataset) and defined
as 1
scN
scu
u cu
sc
sc
P
AvgPPRICE
N
where Pscu stands for the average price of SKU u in category c
in store s over time, AvgPcu stands for the average price of SKU u in category c across all
stores over time, and is the total number of SKUs in category c at store s. When
constructing this price variable, we aim at eliminating the effects of non-pricing factors,
such as differences in package sizes (thus, potential quantity discounts) and quality (e.g.,
organic products may be priced higher than non-organic products), on price levels by
normalizing the SKU prices ─ dividing the SKU prices by average prices.
Given the same level of average category prices at two stores, the store with lower price
variability over time may be interpreted as a consistent and dependable provider of good value in
the category from a long-term perspective, which may result in greater store attractiveness to
consumers in the category. Conversely, a store with greater price variability over time may
encourage consumers to shop at that store only when low prices are offered in the category and
drive consumers, during periods of high prices, to search for lower prices at other stores. In other
words, greater price variability in a category at a store may reduce the overall attractiveness of the
store to households in the category from a long-term perspective. For this reason, we include the
following variable to measure the price variability of the store in the category:
scN
20
Price Variability of the Store in the Category (PRICEVARsc).
This variable is defined as
1
,scsc S
rc
r
CVPRICEVAR
CV S
where CVsc stands for the
coefficient of (temporal) variation over time of category prices in category c at store s at
time t, 1
Cat_Price
scN
scut
u cu
sct
sc
P
AvgP
N
. Notice that Cat_Pricesct is defined similarly to
scPRICE but with a time subscript, and it is also normalized to eliminate the effects of non-
pricing factors on prices as in PRICEsc.
In summary, we include three variables to represent a store’s assortment strategies in a
category, among which BRANDsc and SKUsc measure the relative breadth of the assortment, and
PVTLABELsc measures the relative exclusiveness of the assortment. We also include two variables
to measure two aspects of the pricing strategies of the store in the category PRICEsc measures
the relative price level and PRICEVARsc measures the relative price variability over time. In Table
1, we provide descriptive statistics pertaining to these variables in our dataset, which indicate that
all variables have comparable scales in the data.
[INSERT TABLE 1 HERE]
5.5 Estimation
Given the total number of purchases made by household h in category c across different
stores during the study period, 1 2( , ,..., )h h h h
c c c ScN n n n , we can write the likelihood function for
household h in c and then the total likelihood across all categories and households as follows:
1 1 1
1
exp
exp
h
sc
hH C S
sc
Shh c src
r
n
AL
A
(4)
21
Since our data contains individual household level purchases across multiple stores in
multiple categories, this enables us to separate store-category level effects from overall store level
effects and household heterogeneity. We employ a hierarchical Bayes technique to estimate the
two-way random effect model and make standard parametric assumptions. Specifically, we
assume that sc ,h
s and h
sc follow S-1 dimension independent multivariate normal distribution
with respective zero mean vectors and (S-1) by (S-1) full variance covariance matrix sc ,
hs
and
hsc
.8 For c , h , and
h
c , we assume they follow K-dimension independent multivariate normal
distribution, where K is the number of variables representing the merchandising strategies, with
respective zero mean vectors and K by K full variance covariance matrix c
, h
and .hc
After
augmenting a latent step to estimateh
scA , via a Metropolis Hastings step, the rest of the Gibbs
sampler is fairly straightforward under conjugate priors.
6. ESTIMATION RESULTS AND MANAGERIAL IMPLICATIONS
We estimate the proposed model using the data described in section 3. To attenuate the
concern about the impact of low purchase frequency categories on store-category attractiveness,
we drop categories and households that have fewer than 100 purchase observations from the
original dataset. This results in 925,153 category purchase incidences generated by 1,280
households purchasing across 244 categories. For the proposed model, we estimate 13 mean
intrinsic store attractiveness terms s , 2719 category-specific intrinsic store attractiveness terms
sc , 16,640 household-specific intrinsic store attractiveness terms h
s , five merchandising
effectiveness terms , 244 category-specific merchandising effectiveness terms c , and 1,280
8 For identification purpose, we use one store as the base store and restrict its elements in
h
scA to 1; therefore, the
dimension is S-1.
22
household-specific merchandising effectiveness terms h .9 In addition to the proposed model,
we also estimate three benchmark models:
1) Benchmark Model 1: No category-level variations and no household heterogeneity in intrinsic
store attractiveness and in merchandising effectiveness. Mathematically, the model can be
written as:
;h h h
sc s c sc scA X where h h
c c
2) Benchmark Model 2: No category-level variations but allow for household heterogeneity in
intrinsic store attractiveness and in merchandising effectiveness. The model can be written as:
;h h h h
sc s s c sc scA X where h h h
c c
3) Benchmark Model 3: Allow for category-level variations but no household heterogeneity in
intrinsic store attractiveness and in merchandising effectiveness. The model can be written as:
;h h h
sc s sc c sc scA X where h h
c c c
Table 2 lists the log-likelihood and the deviance information criterion (DIC) of the
proposed model and the three benchmark models. The comparison shows that the proposed model
outperforms all benchmark models, implying that the attractiveness of a store varies substantially
by category as well as by household. Besides addressing household heterogeneity, it is also
important to pay attention to category differences in store attractiveness to households.
Next, we discuss the parameter estimates of the proposed model, which are most pertinent
to our main research questions.
[INSERT TABLE 2 HERE]
9 In addition, to account for observed category heterogeneity in category purchase frequency and budget share and
household heterogeneity in family size and family income, we also estimate 13*4 = 52 terms for intrinsic
attractiveness across 13 stores and 5*4 = 20 terms for 5 merchandising effectiveness variables. For ease of exposition,
we discuss the details in section 6.2.3.
23
c
6.1 Intrinsic Attractiveness
6.1.1 Intrinsic Overall Store Attractiveness
We first look at the estimates for intrinsic store attractiveness, s , which represents the
overall attractiveness of the stores (such as overall perceptions of store quality, store images,
convenience of store locations, etc.) common to all households and across all categories after
accounting for the impact of retailers’ assortment and pricing strategies. This is reported in Table
3.10 We find that Fry Food Store has the highest intrinsic store attractiveness (10.18), suggesting
that after accounting for the impact of retailers’ assortment and pricing strategies, Fry Food Store
is intrinsically the most attractive store to households. The intrinsic attractiveness of Albertsons,
Bashas’, Safeway and Wal-mart Supercenter is greater than that of the Trader Joe’s while the
intrinsic store attractiveness of Food 4 Less and IGA is slightly less. For the rest of the six stores,
this estimate turns out to be insignificant, suggesting that after accounting for the impact of
retailers’ assortment and pricing strategies, these stores have similar overall attractiveness as
Trader Joe’s. Next, we look at store attractiveness from the category perspective.
[INSERT TABLE 3 HERE]
6.1.2 Intrinsic Store-Category Attractiveness
The extent to which intrinsic store attractiveness varies across categories is measured by
the variance of intrinsic store category attractiveness, sc . Relatively low diagonal values in
sc imply that intrinsic store-category attractiveness is perhaps only a function of overall store
characteristics (e.g., store-level customer services, store-level check-out services and store
location, etc.), while a large value implies that the intrinsic store attractiveness of a category
depends on characteristics that are specific to the category (we will further discuss the issue
10 For identification purposes, we set the intrinsic attractiveness for Trader Joe’s at zero.
24
related to observed category characteristics in section 6.2.3). We observe large variations in store
attractiveness across categories and substantial heterogeneity across households (which is
measured by hs
); both standard deviations are reported in Table 3. Specifically, we find that
53.84% of the intrinsic store-category attractiveness estimates sc are significantly different
from zero, indicating that 53.84% of the categories have their own category specific store
attractiveness that is different from mean store attractiveness s . This underscores our conjecture
that store loyalty is best viewed as a category specific trait.
Next, we discuss how a retailer can use our estimates of intrinsic store-category
attractiveness to help manage categories within the store and compete with other stores.
6.1.3 Competitive Positions of Stores based on Intrinsic Store Category Attractiveness
We provide a unique perspective of stores’ competitive positions based on stores’
intrinsic category specific attractiveness, sc . The correlation of sc between two stores (derived
from sc ) across categories indicates how similar (or dissimilar) two stores are in consumers’
minds in terms of their categories’ standing of intrinsic category attractiveness. For example, a
positive correlation indicates that two stores have a similar rank ordering of categories in terms of
intrinsic category attractiveness and hence compete closely with each other at the category level. A
correlation matrix can help a retailer understand its position against other stores at the category
level in the competitive environment and thus formulate appropriate marketing strategies. In Table
4, we report the correlation matrix of sc across stores. From the table, we see that the correlations
among Albertsons, Bashas’, Fry Food Store and Safeway are more than 0.9, suggesting that these
stores compete closely with each other at the category level.
[INSERT TABLE 4 HERE]
25
We use a factor analysis to analyze the correlation matrix and find that the first two factors
explain 89.37% of cumulative variance. We plot the stores on a map using the two factors. The
perceptual map is reported in Figure 7. The map shows that this market can be best described by
four clusters. The first cluster consists of Albertsons, Bashas’, Fry Food Store and Safeway, which
are supermarkets generally implementing HiLo pricing strategies. They have high correlations (0.9
and above) among themselves on intrinsic store category attractiveness and therefore compete
head to head in consumers’ minds, after accounting for the stores’ assortment and pricing
strategies. Their higher overall store attractiveness also confirms that they are major players in this
market. The second cluster consists of Food 4 Less, Food City and IGA. These stores have limited
assortments and generally implement everyday low price (EDLP) strategies (Gauri, Trivedi and
Grewal 2008). They have high correlations (0.8 and above) among themselves and therefore are
direct competitors. The two club stores Costco and Sam’s Club are in cluster three, which can be
characterized as supercenters with EDLP strategies but only accessible to club members. The two
stores have a positive correlation between themselves (0.6) but have negative correlations with the
rest of the stores. This suggests that the two club stores compete with each other (but not head to
head) and are complementary to most of the other stores. Super Kmart, Smart & Final and Wild
Oats Market can be grouped into one cluster. They exhibit relatively small correlations with other
stores, implying that they are perceived to have niche positions and indirectly compete with other
stores. Finally, Wal-Mart Supercenter can be categorized as a group of its own. It has moderate
positive correlations with the stores in cluster one (e.g., Albertson’s and Safeway, etc.) and Super
Kmart, and relatively small correlations with the rest of the stores. This implies that it competes
(though not head to head) with HiLo stores.
[INSERT FIGURE 7 HERE]
26
6.1.4 Rank Ordering Most Intrinsically Attractive Categories
Retailers can rank order categories based on intrinsic store category attractiveness. This
can help them identify top categories to effectively allocate merchandising and marketing
resources across categories and leverage their stores’ competitive advantages to improve overall
store patronage. For each of the 13 stores, we rank the categories based on their intrinsic store-
category attractiveness and list the top 10 categories that have the highest values in Table 5.11
Given our conjecture about store-category attractiveness, it is not surprising to see that different
stores are strong in different categories but exhibit similar intrinsic store-category attractiveness
patterns within a cluster as mentioned in section 6.1.3.
[INSERT TABLE 5 HERE]
Retailers can further investigate categories that have high intrinsic store-category
attractiveness to understand what makes those categories intrinsically attractive in their stores and
cultivate their attractiveness further. For example, in Albertsons, frozen fruits, croutons and frozen
pizza are ranked as the top three categories. Albertsons can investigate what additional factors
unrelated to assortment and pricing strategies, such as salient aisle placement of the categories,
may contribute to their high attractiveness. Furthermore, our analysis can also help a retailer
estimate the attractiveness of a category that is not currently present in the store. Leveraging the
correlation of attractiveness across stores sc , a retailer can derive the attractiveness of the new
category that was not present in the focal store based on information extrapolated from the same
category in other stores. For example, though Food City currently doesn’t carry skin care products,
our model can predict, on average, how attractive skin care products would be to consumers if
introduced in Food City.
11 Note that these categories are ranked only based on intrinsic store-category attractiveness. The overall attractiveness
of a category at a store depends on both merchandising effectiveness of retailers as well as the category’s intrinsic
store attractiveness.
27
6.2 Effects of Merchandising Programs on Store-Category Attractiveness
6.2.1 Effects Common across Categories and Households
Retail managers are interested in understanding the main effects of their merchandising
strategies on store-category attractiveness. We report the estimated mean effects that are common
across categories and households, , for each of the assortment and price variables in the first
column of Table 6. We notice that the estimated mean effect ( ) for the number of brands
(BRANDsc) is positive (1.47), suggesting that offering more brands in an average category
increases the category’s store attractiveness. The estimated mean ( ) for the average number of
SKUs per brand (SKUsc) is positive (0.60), implying that, on average, the number of SKUs within
a category also has a positive effect on its store-category attractiveness. The results also show
substantial variation across categories (discussed further in section 6.2.2). Although intuitively it
would seem that increasing assortment breadth may help store-category attractiveness, the existing
empirical research on the effect of assortment breadth on category revenues has shown mixed
results. For example, Dreze, Hoch and Purk (1994) find that sales go up after assortment
reduction, while Broniarczyk, Hoyer and McAlister (1998) and Boatwright and Nunes (2001) find
no effect. Borle et al (2005) find that assortment reduction has a negative effect on both shopping
frequency and purchase quantity but observe that the impact varies widely by category. Our
research supports the findings of Borle et al (2005), as we find that increasing assortment breadth
increases store-category attractiveness and also that the effect varies greatly by category.
[INSERT TABLE 6 HERE]
Table 6 also shows that the estimated mean is insignificant for the number of private
labels (PVTLABELsc). Corstjens and Lal (2000) analytically demonstrated that only when the
quality of store brands exceeds a threshold level does carrying store brands increase the store’s
attractiveness. Our finding is consistent with theirs, particularly in that, on average, merely
28
increasing the breadth of private label assortments within a category has no impact on store-
category attractiveness.12
The estimated effects of both price variables are consistent with our a priori expectations.
Specifically, the estimated mean for PRICEsc is negative (-0.49), while that of PRICEVARsc is
negative (-0.09). These findings suggest that, on average, a retailer who adopts an EDLP strategy
within a category enjoys higher store-category attractiveness. However, we find that there is
substantial variation across categories, which we discuss in the next section.
6.2.2 Category-Specific Merchandising Effects
The proposed model enables us to identify the merchandising effects of a specific category
on store-category attractiveness to consumers. This can help retailers improve their stores’
attractiveness to consumers through improved category management. The magnitude of the
estimated c represents the degree of deviation in category c from the mean effect of the kth
merchandising (e.g., product assortments or pricing) program, , on store-category attractiveness.
Between two categories, the store-category attractiveness of the category with a higher absolute
value of c is more responsive to changes in the kth merchandising program. Therefore, rank-
ordering the categories based on the magnitudes of the estimated values of c would help retailers
prioritize among categories particularly for merchandising program k; this, in turn, would enable
retailers to appropriately allocate their limited marketing resources across categories. As an
illustration, in Table 7, we list the top 10 categories with positive deviations ( 0c ) that differ
significantly from the mean effect ( ) and top 10 categories with negative deviations ( 0c )
12 As our data does not contain the information on the quality of the private label brands, in this empirical application we cannot
distinguish private label brands based on quality differences.
29
that differ significantly from the mean effect ( ).13 Depending on the direction of the mean effect
(i.e., the sign of ) of a specific merchandising variable, these categories are respectively the
most responsive and the least responsive to changes in the merchandising program. For example,
since the mean effect of the number of brands within the category is positive BRANDS( 1.47) , the
store-category attractiveness of motor oil is the most responsive and that of dish detergent is the
least responsive to changes in the number of brands within the category. On the other hand, since
the mean effect of price is negative PRICE( 0.49) , the store-category attractiveness of non-fruit
drinks is the most price sensitive while that of frozen seafood is the least price sensitive.
[INSERT TABLE 7 HERE]
Our study allows retailers to make informed decisions on how different types of
merchandising programs (e.g., assortment versus price), or even different levels in a given
merchandising program (e.g., few versus many brands), can be customized for different categories
to improve overall store attractiveness to consumers. Careful analysis of categories with high
intrinsic category attractiveness (Table 5) in a store combined with how these categories respond
to various merchandising efforts (Table 7) can help retailers craft customized strategies to improve
store patronage.
6.2.3 Category Heterogeneity
We uncover substantial variation in the estimated effects of product assortments and
pricing strategies across categories. The variation is captured by the standard deviations of the
estimated category-specific effects, c , across categories. Similarly, the heterogeneity among
households is captured by the standard deviations of the estimated household-specific effects,
13 For ease of exposition, we report the rank-ordering for three merchandising variables only. The results for rest of the
merchandising variables are available from authors upon request.
30
h , across households. We report these two standard deviations in the second and third columns
of Table 6, respectively. A comparison between these two columns shows that except for the
price variable, there are large variations in the effects of retailers’ assortment and pricing
programs across categories. Though the variation across categories in the effect of the number of
private labels is relatively small in terms of absolute value, the variation is directional, i.e., for
some categories, increasing the number of private label SKUs will increase their store-category
attractiveness, while for other categories, it will decrease their store-category attractiveness. These
results suggest that when planning merchandising programs, retailers should pay close
attention to the differences across categories. Our results are also consistent with studies where
opposite effects of various merchandising strategies are found due to choices of different
categories (e.g., Dreze, Hoch and Purk 1994; Broniarczyk, Hoyer and McAlister 1998; Boatwright
and Nunes 2001; and Borle et al 2005, to name a few).
We next investigate how specific observed category characteristics potentially explain the
heterogeneity. First, we explain the estimation procedure before discussing the results. To analyze
the effects of observed category characteristics, we had decomposed the category specific effect in
merchandising effectiveness, c , see equation (6), into two components: (i) cZ , where cZ is a
vector of variables that represent observed category characteristics (i.e., purchase frequency and
budget share) and is the vector of corresponding responses; and (ii) c , the remaining effect
specific to category c that is not explained by the observed category characteristics. To maintain
consistency, we also decompose intrinsic store-category attractiveness, sc , see equation (5), into
two components: (i) s cZ , where s is a vector of corresponding responses to cZ ; and (ii) sc , the
remaining intrinsic store-category attractiveness that is not explained by the observed category
characteristics. Additionally, we also estimate the effects of household characteristics by including
31
a vector of demographic variables (i.e., family size and family income), hD . Similar to category
heterogeneity decomposition, household heterogeneity in merchandising effectiveness, h , is
decomposed into observed heterogeneity,hD , where is a vector of coefficients for the
observed household demographic variables, and unobserved heterogeneity, h . Similarly, we also
decompose household heterogeneity in intrinsic household store attractiveness, ,h
s into observed
household heterogeneity,h
sD , and unobserved household heterogeneity h
s . To ensure consistent
estimates, we incorporate the above decomposition into equation (2) and (3) via a hierarchical
structure and ran a one-step Bayesian hierarchical estimation. Specifically, equation (2) and (3)
were estimated as the following two equations respectively:
;h h hh hsc s s c sc c sc scs s
A Z XD (5)
h hh hc c c cZ D (6)
We report the estimates for , in Table 8a and , in Table 8b, respectively. In Table
8a, we can see that category purchase frequency has positive effects on intrinsic store-category
attractiveness for most stores. Category budget share has positive effects on intrinsic store-
category attractiveness only for the two club stores. For most stores, store attractiveness is higher
for larger families. Costco is more attractive for households with higher family income, which is
consistent with the positioning of Costco. Table 8b shows that only one observed category
characteristic – category purchase frequency – positively affects how the store-category
attractiveness of a category responds to the number of brands in the category. Other observed
category and household characteristics have no impact. With more data on observed category
characteristics, retailers can use our model to comprehend how the store-category attractiveness of
different categories will respond differently to merchandising strategies.
32
[INSERT TABLE 8a and 8b HERE]
7. SUMMARY AND CONCLUSIONS
Most of the marketing literature views store loyalty as a behavioral trait of consumers that
relates to consumers’ store choice decisions, particularly for explaining which stores are the most
frequently visited by those consumers for their overall grocery shopping needs. However, if a
consumer is observed to shop at multiple grocery stores over time, thus appearing to not be store
loyal, she/he may still purchase some product categories consistently from one store, exhibiting
store loyalty at the category level. We call this consumer behavior “store-category loyalty”. Our
empirical findings indicate that capturing pertinent information on category-specific consumer
shopping behavior can shed new light on store loyalty, which would be of interest to the broader
community of researchers as well as retailers.
Using purchase data from 1321 households for 284 grocery categories across 14 retail
chains in a large southwestern city in the US, we demonstrate strong empirical evidence of store
category loyalty in the data, even though the overall store loyalty based on the traditional view is
low. We propose a model to examine the effects of key factors influencing store-category
attractiveness, which determines store category loyalty. By simultaneously studying households’
purchases in multiple categories at multiple stores on a large scale, we are able to decompose such
effects into those that are common across categories and across households and those that are
specific to a particular category or household. Furthermore, we demonstrate how the estimation
results from the proposed model can assist retailers in designing appropriate retail strategies at the
category level with the overall aim of improving overall store patronage.
Our paper augments prior literature on the effect of category attractiveness on store loyalty.
Our integrated approach of incorporating category-specific store attractiveness provides deeper
33
insights into consumers’ store choice behavior. In particular, our approach can provide a different
perspective of the relative positioning of stores in consumers’ minds based on intrinsic store-
category attractiveness after controlling for retailers’ merchandising programs. Our study also
enables retailers to make informed decisions on how to employ merchandising programs of
different types (e.g., assortment versus price), and different levels (e.g., few versus many brands),
which can be customized at the category level to improve overall store patronage. In a nutshell, we
believe that viewing store loyalty as a category-specific trait and understanding category-specific
store attractiveness can help managers boost overall store patronage.
8. Limitations and Directions for Future Research
There are some possible areas for future research. First, our analysis aggregates the time
dimension since we do not have weekly promotional information (e.g. features, displays, coupons,
etc.) in the data. Moreover, the weekly data for product assortments and prices is sparse for a large
number of categories. We are unable to accommodate the time dimension in this study. It would be
useful to analyze a dataset that contains weekly store environment data with comprehensive
information in these areas. That analysis can help parse out the influence of short-term marketing
activities on long-term store-category attractiveness. Second, although we model correlations
among stores and control for household-level preferences that are common across categories, due
to the curse of dimensionality, we are unable to explicitly model cross-category correlations that
may arise due to demand complementarity (e.g., cake mix and cake frosting). Accounting for such
correlations may be useful to uncover how a household’s store-category loyalty may be related
across categories. Third, although we do not expect meaningful differences in results to emerge
from using purchase quantity or expenditure in the model due to the high correlation between
households’ shares of category expenditure and shares of category purchase incidences across
34
stores in our data, it would be of interest for future studies to use purchase quantity or expenditure
to compare alternative approaches for conceptualizing store-category loyalty. Last but not least, it
would be interesting for future research to link households’ store-category loyalty with
households’ overall store loyalty, as understood in the store loyalty literature. Such a study will
assist retailers in better understanding how piecemeal management of category loyalty can
eventually lead to an overall advantageous position for their stores in the market.
We hope that our work prompts future research on modeling and understanding the
influence of categories on the relationship between a consumer and a store and its implications on
retail practice.
35
REFERENCES
Aaker, D. A., & Jones, J. M. (1971). Modeling Store Choice Behavior. Journal of Marketing
Research, 8(1), 38-42.
Ailawadi, K., Pauwels, K., & Steenkamp. J-B. (2008). Private Label Use and Store Loyalty.
Journal of Marketing, 72(6), 19-30.
Ailawadi, K., & Keller, K. L. (2004). Understanding Retail Branding: Conceptual Insights and
Research Priorities. Journal of Retailing, 80, 331-342.
Ainslie, A., & Rossi, P. E. (1998). Similarities in Choice Behavior across Multiple Categories.
Marketing Science, 17(2), 91-106.
Arnold, S. J., Oum, T. H., & Tigert, D. J. (1983). Determinant Attributes in Retail Patronage:
Seasonal, Temporal, Regional, and International Comparisons. Journal of Marketing Research,
20( 2), 149-157.
Baumol, W. J., & Ide, E. A. (1956). Variety in Retailing. Management Science, 3(1), 93-101.
Bell, D. R., Bonfrer, A., & Chintagunta, K. (2005). Recovering Stockkeeping-Unit-Level
Preferences and Response Sensitivities from Market Share Models Estimation on Item
Aggregates. Journal of Marketing Research, XLII, 169-182.
Bell, D. R., Ho, T., & Tang, C. S. (1998). Determining Where to Shop: Fixed and Variable Costs
of Shopping. Journal of Marketing Research, 35(3), 352-369.
Bell, D. R., & Lattin, J. M. (1998). Shopping Behavior and Consumer Preference for Store Price