Date post: | 06-Apr-2018 |
Category: |
Documents |
Upload: | deepak-manohar |
View: | 218 times |
Download: | 0 times |
of 38
8/3/2019 SSRN-id331541
1/38
The Role of Retail Competition, Demographics and Account Retail Strategy asDrivers of Promotional Sensitivity
by
Peter BoatwrightCarnegie Mellon University
Graduate School of Industrial Administration
Sanjay Dhar and Peter RossiUniversity of Chicago
Graduate School of Business1101 E. 58th StreetChicago, IL 60637
July, 2000
Revised, July 2002
.
AcknowledgementsPeter Boatwright is Assistant Professor of Marketing, Sanjay Dhar is Professor of Marketing, andPeter Rossi is Joseph T. Lewis Professor of Marketing and Statistics. Support from the Kilts Centerfor Marketing, Graduate School of Business, University of Chicago is gratefully acknowledged.
8/3/2019 SSRN-id331541
2/38
- 2 -
The Role of Retail Competition, Demographics and Account Retail Strategy asDrivers of Promotional Sensitivity
July, 2000revised, July 2002
AbstractWe study the determinants of sensitivity to the promotional activities of temporary price reductions,displays, and feature advertisements. Both the theoretical and empirical literatures on pricepromotions suggest that retailer competition and the demographic composition of the shoppingpopulation should be linked to response to temporary price cuts. However, datasets that spandifferent market areas have not been used to study the role of retail competition in determining pricesensitivity. Moreover, little is known about the determinants of display and feature response. Verylittle attention has been focused on retailer strategic decisions such as price format (EDLP vs. Hi-Lo)or size of stores. We assemble a unique dataset with all U.S. markets and all major retail grocerychains represented in order to investigate the role of retail competition, account retail strategy, anddemographics in determining promotional response. Previous work has not simultaneously modeledresponse to price, display, and feature promotions, which we do in a Bayesian Hierarchical model.
We also allow for retailers in the same market to have correlated sales response equations through avariance component specification. Our results indicate that retail strategic variables such as priceformat are the most important determinants of promotional response, followed by demographicvariables. Surprisingly, we find that variables measuring the extent of retail competition are notimportant in explaining promotional response.
8/3/2019 SSRN-id331541
3/38
- 3 -
Introduction
Both the theoretical and the empirical literatures on the response to promotional activities
have focused primarily on response to temporary price reductions. The theoretical literature on
price promotions (c.f. Varian (1980), Narasimhan (1988), Lal, Little, and Villas-Boas (1996), Kim and
Staelin (1999)) emphasizes competition between retailers as the fundamental driver of short-term
price changes. On the other hand, in the empirical literature (e.g., Bolton (1989), Hoch et al (1995),
Shankar and Krishnamurthi (1996)) demographic factors have long been thought to be important
determinants of price sensitivity via various search-theoretic explanations in which the value of time
plays an important role in determining the extent of price awareness. No work has been done on the
importance of account retail strategy variables in determining price elasticities.
This study examines the relative impact of the role of retail competition, demographic
factors and account retail strategy variables in determining not only price elasticities but also display
and feature response. This is particularly important given that little is known about the determinants
of display and feature response. Using scanner level tracking data, manufacturers often observe that
there are enormous differences between accounts and across markets in response to promotional
activities. (Boatwright, McCulloch, and Rossi (1999) advance a new methodology for measuring these
differences.) Due to insufficient data, many studies (such as Hoch et al (1995)) omit display and
feature measurements and focus only on price response. Since price reductions are often correlated
with display and feature response, this may produce an omitted variable problem. Ainslie and Rossi
(1998) document the relationship between demographic variables and display and feature but only
for one retailer in one market area.
Furthermore, the extant empirical literature on price sensitivity has focused on only a limited
number of market areas (at most two as in the Bolton study). In one or two market areas, there can
be no variation in the basic structure of competition as indicated by the number and relative size of
retailers. Hoch et al use store-specific measures of retail competition, but the variation in the
8/3/2019 SSRN-id331541
4/38
- 4 -
competitive climate within one or two market areas may be quite limited.1 Wittink (1977) studied
data from multiple markets, considering how price sensitivity is affected by different advertising
content.
In order to sort out the importance of retail competition, account retail strategy, and
demographic information, we assemble a unique dataset which spans all major U.S. market areas and
all major key accounts (a key account is a retailer-market combination, e.g., Safeway Denver). We
have complete display and feature ad information along with measures of retail competition, account
format, and demographics. We use a comprehensive sales response or demand model that
incorporates the effect of own price, display, feature, competing brand prices, and price at competing
retailers. This model is calibrated to weekly sales data at the key account level. Account aggregation
biases are reduced to the smallest possible extent using a method suggested by Christen et al (1997).
In addition, we provide a formal argument for why aggregation biases cannot affect the relative
magnitudes of our explanatory variable effects.
Our goal of estimating promotion sensitivities and the relationship of these sensitivities to
account level competition in a market, account retail strategy, and demographic variables presents a
formidable estimation problem. There are about 100 key accounts and five basic promotion
variables, yielding some 500 or so response coefficients. A nave strategy of estimating each account
model separately and then using the estimated coefficients in a second stage model does not properly
account for estimation error. If there is substantial estimation error in the account level response
coefficients, this sort of two-step approach may provide misleading views about the importance of
the explanatory variables. That is, the R-squared of the second stage regression can be low simply
due to estimation error rather than that the true coefficients are unrelated to these variables. In
addition, a joint estimation method will provide gains in statistical efficiency.
We use a hierarchical approach in which each account level regression coefficient vector is
viewed as a draw from a random effects distribution that incorporates the account and market level
1Bolton (1987) and Hoch et.al (1995) use data from multiple categories.
8/3/2019 SSRN-id331541
5/38
- 5 -
competition, strategy, and demographic variables. We explicitly incorporate geographical markets in
our model by allowing retail accounts in a given geographical market to have correlated demand
shocks. We are the first to measure the common variance component in the demand shocks across
accounts. Contrary to previous speculation, we find these common components to be small.
Our relatively small set of explanatory variables measuring competition, account retail
strategy, and demographics accounts for around 30 per cent of the variation in response.
Surprisingly, we find that the retail competition variables are the least important determinants of the
responsiveness to promotional activities. Account retail strategy variables such as price format and
chain size are important determinants of both price and display/ feature responsiveness.
8/3/2019 SSRN-id331541
6/38
- 6 -
A Framework for Selection of Explanatory Variables
Our goal in this paper is to relate promotion sensitivity in an account (retailer-market
combination) to the salient characteristics of the retail environment. In order to impose some
discipline on the process of selection of variables, it is useful to develop a framework for thinking
about response to promotions. Fundamentally, the promotions observed in retail scanner data are
various forms of temporary price reductions accompanied, in some cases, by in-store advertising
(displays) or out-of-store print ads (features). This advertising is designed to provide price
information and remind the consumer of the existence of the product. This is not image or
attribute-focused advertising. In our view, price sensitivity is influenced by consumer budget
constraints and wealth, by availability of competitive alternatives, and by consumer search activity.
The variables included in our analysis therefore reflect an understanding what sorts of consumers and
retail environment conditions will promote price sensitivity.
Conceptually, it is possible to organize our discussion by categorizing the variables that may
affect consumer response to temporary shelf-price reductions, display and feature activity into three
groups: consumer characteristics, account retail strategy, and retail competition.
Consumer Characteristics
Income or Wealth: Becker (1965) proposed that systematic differences in price sensitivity
could arise from the opportunity costs of time associated with demographic characteristics (Blattberg
et al. 1978). Higher income consumers (who also tend to live in higher home value neighborhoods)
have higher opportunity costs of time and are likely to be less price sensitive. Consequently, they are
likely to have a lower response to temporary shelf price reductions. However, these very consumers
are time constrained and consequently use feature advertisements and in-store displays to reduce
their search costs and are therefore more likely to respond to non-price retail promotional activity.
Age: Older consumers are likely to have lower opportunity costs of time and therefore more
likely to engage in more search for lower prices and therefore more likely to respond to temporary
8/3/2019 SSRN-id331541
7/38
8/3/2019 SSRN-id331541
8/38
- 8 -
Previous research has shown significant cross-store effects at the brand level for a few product
categories (e.g., Kumar and Leone 1988; Walters 1991). Similarly, in competitive markets, feature
advertisements may be used to compete for customers (as in Lal and Matutes, 1994), and for this
reason features may be more effective in stimulating sales in the more competitive markets. Displays
facilitate in-store decision making of what to purchase and hence may benefit retail chains in more
competitive retail markets.
For ease of reference, we summarize the effects of the different variables in Table 1 below.
Note that price sensitivity is negative, so a '+' in the table indicates increased price sensitivity, which
will result in a negative impact on the price elasticity coefficient.
8/3/2019 SSRN-id331541
9/38
- 9 -
Model and Datasets
In order to study the relationship between account retail strategy, competition, and
demographic variables and promotion sensitivities, we need sales and promotional data across
different retailers and different market areas. We have to combine scanner data on the sales volume
and price with demographic data and data on market and account (retailer-market combination)
characteristics. There are no academic or industry sources of data panels which are comprehensive
enough to include sufficient variation in market and chain characteristics (note: the national panels of
households maintained by IRI and Nielsen do not include audit information on store displays or
feature ads it is impossible to accurately measure what display and feature activity was faced by the
household on any given purchase occasion). There are national samples of stores maintained both
by IRI and Nielsen (INFO SCAN and SCANTRACK). These samples also include audit information
on displays and feature ads. It is the policy of both IRI and Nielsen not to release store level
information (note: the MSI/ Nielsen data sets do not include retailer identifiers and only include a
small number of markets). For this reason, we are limited to what Nielsen and IRI term account
level data. An account is a particular market-retailer combination, e.g. Safeway in Denver. The
information from the Nielsen sample of stores is aggregated to the account level. As discussed
below, we use a method suggested by Christen et al to reduce potential aggregation bias.
Through the intercession of a major Nielsen client, we were able to obtain account level data
for 97 major US retail accounts across 35 Nielsen SCANTRACK markets for the ground coffee
product. For example, this means that SafewayDenver is distinct from SafewaySalt Lake City. We
focus on the two major national brands, Folgers and Maxwell House, which account for 59% of the
category total sales. Other secondary national brands have very spotty distribution, so that it would
be difficult to use all 97 accounts. Our SCANTRACK data contain weekly sales, price, display, and
feature variables for each of 120 weeks.
Our 97 accounts are located in 35 different geographic markets. It has often been
speculated that there are market-wide shocks that affect each of the accounts in a given market.
8/3/2019 SSRN-id331541
10/38
- 10 -
Television advertising and manufacturer coupon drops are often cited as possible examples. If there
were large and regular market-wide shocks, then the accounts within a market would have correlated
sales response equation errors terms. To measure the size of these market-wide shocks, we employ a
variance component model. We postulate a sales response model of the form:
( ) = + = = =K K K'amt amt a amt mln y x a 1, , A m 1, ,35 t 1, ,120 (1)
Equation (1) reflects the structure of our data in which each of the 97 accounts are located in 35
different geographic markets (e.g. Winn-Dixie in Raleigh-Durham or Publix in Miami). Each market
(indexed by m) has up to Amaccounts in it. The error term, , has a variance component structure
( )
( )
( )
= +
=
amt mt amt
mt amt
2
amt am
2
mt m
w v
cov w ,v 0
v ~ N 0,
w ~ N 0,
This variance component structure allows for a positive covariance between accounts in the same
market as they all share the same market-wide error term.
( ) ( ) = + + = 2amt a'mt mt amt mt a'mt mcov , cov w v ,w v
Since consumer promotions are planned on the level of the account (Boatwright, McCulloch, and
Rossi 1999), we assume correlation across accounts within a chain to be negligible, although this and
other types of correlations would be feasible to model.
The explanatory variables in the xamt vector include:
1. intercept
2. ln(price of brand)
3. display without feature
4. feature without display
5. feature and display
8/3/2019 SSRN-id331541
11/38
- 11 -
6. ln(price of competing brand at the retail account)
7. ln(min(price at competing accounts in the market area))
8. ln(lag(sales))
Thus, our sales response model allows for account level sensitivity to all of the marketing mix as well
as competitive brand and competitive account prices. Our model includes lagged sales to allow for
autocorrelated errors. In addition, we investigated other model specifications which included base
(regular) price, multiple variables to capture seasonal effects, alternative specifications of prices of
competing brands in the same account, and alternative specifications of prices at competing
accounts. Our final model contains those variables that were consistently important in the many
specifications that we investigated. Also our final model contains one own price variable instead of a
regular price and discount measure. Base price and actual price are highly correlated (r=0.94),
preventing us from estimating separate short term and long term price effects. In order to investigate
the relationship between the account-level promotion sensitivities (a) and retail competition,
account retail strategy, and demographics variables, we employ a hierarchical model structure with a
second level that allows the promotion sensitivities to be related to both our observable variables and
a random component,
= +am am amz u (2)
where z is the vector of retail and demographic variables and is a matrix of regression coefficients.
u represents the unobservable component of promotion sensitivity.
In previous work (Hoch et al (1995) and Shankar and Krishnamurthi (1996)), the sales
response model in (1) is estimated first and the estimated coefficients are then regressed on
explanatory variables in a second regression. It is well known that this is an inefficient procedure.
Furthermore, the goodness of fit in the second regression is not an unbiased estimate of the extent to
which the explanatory variables can explain variation in promotion sensitivity. Since there is
8/3/2019 SSRN-id331541
12/38
- 12 -
measurement error in the estimates of account level promotion sensitivity, the R-squared from the
second step equation will underestimate the goodness of fit. For these reasons, we do not use a two-
step estimation procedure, but, instead, we estimate the system (1-2) simultaneously in a
Bayesian hierarchical model approach. That is, we regard (2) as a random coefficient model and
complete the model with priors on the hyperparameters of the model.
We first specify the distribution of the error terms in (1) and (2)
( )
( )
2at a
a
~ N 0,
u ~ N 0,V
Note that we allow for correlation between the components of the vector by allowing a non-
diagonal matrix V . Priors on the key hyperparameters complete the model.
( )
( ) ( )
2a ~ IG ,
vec ~ N , V
In any hierarchical model, the prior on V plays a critical role. Adaptive shrinkage requires a
relatively diffuse prior in order for the data to dictate the amount of shrinkage or partial pooling
that takes place across accounts. It is well known that the standard inverted-Wishart prior has a
number of defects. It can be difficult to assess a relatively diffuse prior, and the inverted-Wishart
prior cannot be used in cases where one wants diagonal elements of V to be relatively large or small
for different elements of the regression vector. For these reasons, we adopt the approach of
Barnard, McCulloch, and Meng (2000). In the Barnard et al approach, V is decomposed into the
correlation matrix, R, and the diagonal matrix of standard deviations.
( ) ( )
=
i i i
correl
V diag(s)Rdiag(s)s ~ logn( , )
p R 1 I R
That is, on the correlations we use a flat or uniform prior over the space of positive definite matrices,
and we use log-normal priors on the standard deviations.
8/3/2019 SSRN-id331541
13/38
- 13 -
E x planatory V ariables: A ccount R etail S trategy, Competitive Characteristics, and D emographics
The second level of the hierarchical model (eqn (2) above) relates each of the seven response
coefficients to account retail strategy, retail competition and consumer demographic variables
identified earlier. These data were obtained from several sources. A syndicated data service provided
detailed demographic distributional information on demographic variables for each account. Two
published secondary data sources, Market Scope, and Marketing Guidebook, provided additional
information on the characteristics of the retail chains and the markets served. Both publications are
well-known sources published annually by Trade Dimensions, a unit of the Progressive Grocer Data
Center. We also surveyed each of the retail accounts in the database for additional information.
The variables constructed to measure account retail strategy include 1). Chain size the
number of stores each chain operates in a particular market; 2).EDLP/ Hi-Lo (EDLP) from a
survey of accounts (see Dhar and Hoch (1997)); 3)Floor Space per Capita (FLOOR.SP) average
square footage / population size for an account. This serves as a proxy for the breadth and depth of
assortment carried.
Retail competition is measure by a Herfindahl (HRFNDL) index,
2
aas , where s is the
market share of account a. The Herfindahl index is lower, the higher number of accounts in a
market and the more equal-sized the share of those accounts are. Thus, lower Herfindahl values are
associated with higher levels of retail competition.
Consumer characteristics are measured by three variables. Income/ wealth is proxied by
home value (HOMEVAL) the fraction of the total number of households for a retail accounts
customer base owning homes with a value higher than $250,000. This variable is highly correlated
with income level. Age is measured by the fraction of an accounts customer base that is older than
55 years (ELD). Finally, we use Private Label Share (PVT.LBL as a proxy for the size of the price
sensitive consumer segement (see Dhar and Hoch (1997))
8/3/2019 SSRN-id331541
14/38
- 14 -
All of the above measures are computed at the account level. There are seven explanatory
variables as well as an intercept, so z a will be dimension 1 x 8. The matrix is therefore 8 x 7.
Table 2 provides summary statistics for the distribution of these explanatory variables. The wide
variability of each of these measures across accounts and markets is striking. For instance, the range
of the Herfindahl index, 0.04 to 0.36, indicates large variation in the extent of retail competition
across the markets. Similarly, the percent of elderly in the shopping populations for these accounts
ranges from around 12% to around 34%. As for EDLP (not shown in Table 2), 34% of the 97
accounts consider themselves to follow an EDLP pricing strategy. It should be emphasized that
even in the empirical I/ O literature it is rare to see studies that consider such a wide range of
competitive conditions.
A ggregation Issues and A djustments
Our account level data are at a lower level of aggregation than typically used in the empirical
I/ O literature (where data are usually aggregated to the market or higher (see Berry, Levinsohn and
Pakes (1995) or Nevo (2001)). However, the nonlinear nature of the basic sales response model in
(1) creates the possibility of aggregation biases. As Allenby and Rossi (1991) point out, aggregation
biases occur under the condition that consumers in the aggregation unit are not uniformly exposed to
the same marketing action. In our data, the display variable is the major problem. Display activity is
a store-specific phenomenon. Thus, without any correction, the display coefficients can be biased
upwards as Christen et al (1997) document. The possible solutions to the aggregation problem
include obtaining store level data or making some of the adjustments that are advocated by Christen
et al. Both IRI and Nielsen have a policy of not releasing store-level data across many markets with
account identifiers. For example, the MSI/ Nielsen store level datasets do not include retail chain
identifiers so that we could not merge our market and retail competition and strategy dataset with the
sales data.
8/3/2019 SSRN-id331541
15/38
- 15 -
Given the lack of store level data, we have chosen to adjust our account level data in a
manner suggested by Christen et al. While the account level data do not offer the detail of store level
data, we are able to obtain a decomposition of unit sales under various merchandizing conditions.
For example, we have not only total unit sales but also unit sales for five groupings of stores: stores
with 1. displays without feature, 2. feature without display, 3. display and feature, 4. price discount
and 5. no promotion. We follow the Christen et al procedure of expanding each account-week
observation into five disaggregate observations corresponding to the five groups above. Christen et
al demonstrate that this method removes a great deal of the aggregation bias.
Since the stores accounting for the sales volume vary over time within an account, i.e.
different collections of stores are grouped over time, the average sales of the groups varies over time.
In order to make the groupings comparable, and thus calculate response statistics across partitions
and time, we normalize sales of a store group and partition lagged sales of the account by "baseline
dollars" of the brand for the group. Calculated from store level data by ACN, baseline dollars
measures the non-promoted sales of the partition, i.e. the typical level of sales of that particular
collection of stores in periods when those stores do not offer any promotions. Our dependent
measure of sales yagt, then, is the log of the ratio of coffee unit sales (pounds) and baseline dollars
sales for account a=1 A , group g=1 5, in week t=1 T and our independent variables are the
corresponding price and merchandizing measures for each of the accounts and each of the five
promotion groups.
It should also be pointed out that we are primarily after the relationship between promotion
sensitivity and various account/ consumer variables. Level shifts due to aggregation bias will not
distort these relationships. That is, the relative importance of variables in explaining variation in
promotion sensitivity is not affected by aggregation bias. For our analyses, aggregation biases can
only cause a serious problem if the magnitude of aggregation bias varies by account andif the extent
of the aggregation bias is correlated to the variables used in our analysis.
8/3/2019 SSRN-id331541
16/38
- 16 -
Results
Figure 1 shows the distributions of each of the promotion response coefficients across the
97 accounts. Price elasticities are negative; most of the mass of the distributions is in the range
between -1.06 and -1.82. For ease of interpretation, we have exponentiated the promotion sensitivity
coefficients so that they report percentage increase in sales due to the promotion (sometimes referred
to as lift). On average, features increase sales by 56%; most of the mass of the effect of features
ranges between an 18% increase and a 112% increase. The effect of a display ranges from 10% to
65%; on average it increases sales by 34%. On average, simultaneous features and displays increase
sales by 100%. The average sensitivity to price at other accounts is 0.10, and the average sensitivity to
price of competing brand is 0.36.2
We now turn to a discussion of how these promotion sensitivities relate to account retail
strategy, retail competition, and consumer demographic variables. We summarize these relationships
by considering the posterior distributions of the elements of , which are regression coefficients
from (2). Each row of the matrix contains a vector of regression coefficients.
=
'intercept
'price
'display
'feature
'display&feature
'comp account price
'compbrandprice
For example, the second row of (2) is the price elasticity regression.
= +'price,a price a price,az u
8/3/2019 SSRN-id331541
17/38
- 17 -
Price Response
The boxplots in figure 2 show the posterior distributions of the elements of that relate to
the own price elasticity.3 Each of the continuous explanatory variables has been standardized with a
mean of zero and standard deviation of 1 so as to facilitate comparisons between coefficients on the
same scale. Figure 2 clearly shows that the EDLP format and private label shares are the most
important determinants of own price elasticity. Because EDLP is a discrete variable, the magnitude
of its coefficient cannot be directly compared with the remaining coefficients. Even so, the mass of
the posterior is positive for both brands. Consumers who shop at accounts with an EDLP pricing
strategy are less sensitive to short term price changes than consumers at non-EDLP accounts. This
is entirely consistent with an interpretation of EDLP as a signal of lower mean price and lower
variance (see Hoch, Drze, and Purk (1994)). Consumers at an EDLP account can be assured of
lower average prices and do not have as much incentive to track deals and switch stores as
consumers in a non-EDLP, Hi-Lo account. Bell and Lattin (1998) report the finding that large
market basket shoppers are less responsive to prices in individual categories and tend to shop at
EDLP chains.
Our results are somewhat different than those of Shankar and Krishnamurthi (1996) who
found that shoppers in EDLP chains have higher regular or long-run price sensitivities, using a much
more limited sample of only 2 accounts. There is not necessarily a contradiction between our results
on short-term promotional price response and the results of Shankar and Krishnamurthi. As noted
above, we do not include regular price terms in the sales response model due to insufficient variation
in regular prices. It could well be that EDLP shoppers do not respond to deals but are sensitive to
longer-run price changes.
2The averages given here are medians, while the ranges given are the 10% and 90% quantiles. The statistics
and boxplots for thecoefficients give results only for Folgers.3In the boxplots, the box indicates the range from the 10% to the 90% quantiles, while the whiskers show the
full range of the posterior.
8/3/2019 SSRN-id331541
18/38
8/3/2019 SSRN-id331541
19/38
- 19 -
response. This is consistent with a search cost view of display and feature advertising. Displays
reduce in-store search costs while features reduce out of store search costs. The wealthy have a high
opportunity cost of time and, thus, the reduction in search costs is more important in the sense that
more search behavior can result from a given reduction in search costs. Similarly, the elderly
consumers are more display responsive and feature responsive, which again is consistent with the
search cost view of promotions, since the elderly may have greater cognitive and/ or physical
constraints.
Our findings resolve a number of puzzles in the literature on price sensitivity (at least in the
context of the coffee category). Hoch et al find that the elderly are often more price sensitive. Since
display activity was not measured in the Hoch et al data and displays and price cuts are correlated, the
findings of Hoch et al can be explained by an omitted variable bias. The proper interpretation of the
data is that the Elderly are more promotion rather than price sensitive. Since promotions and price
cuts are often correlated, the Elderly will appear more price sensitive in models without these
variables. Dhar and Hoch (1997) hypothesized that if the elderly were more price sensitive, they
would be more likely to purchase private labels. However, Dhar and Hoch did not find the elderly to
be more likely to purchase private labels. If in fact elderly are promotion sensitive rather than price
sensitive, as our results suggest, we would expect the elderly to appear price sensitive with respect to
the heavily promoted national brands but not particularly prone to purchase the less promoted and
lower priced private labels.
Finally, we find that private label share is positively correlated with display sensitivity and
with feature sensitivity, suggesting that private label buyers are more promotion sensitive than other
buyers. The coffee category is unusual in that private labels, at a national level, command 9.3%
market share, which is 3rd place, ahead of all brands except for Folgers and Maxwell House. Those
accounts with stronger private label coffees may be category leaders in their market, i.e. the strength
of the private label share could in part reflect strategic decisions of a retail account in the coffee
category
8/3/2019 SSRN-id331541
20/38
- 20 -
Figure 5 provides the results for feature sensitivity and display and feature sensitivity. These
results are similar to display activity and to feature activity, with the elderly showing a greater
response to promotions and response is lower in the EDLP format account.
Also, the feature sensitivities are negatively related to chain size. This supports our earlier
discussion that the higher the number of stores that an account has in a market, the closer they are
located to consumers. In this situation, displays are less likely to be effective, as consumers are
already familiar with their neighborhood stores layout and hence do not reduce in-search costs.
Furthermore, the proximity that the accounts stores have to consumers in the market also reduces
the potential number of consumers that can switch leading to feature advertising being less effective
in reducing out-of-store search costs.
R esponse to Competing A ccount Price
In order to measure the effect of prices at other competing accounts in an accounts market
area, we included a measure of the price of the item at competing accounts (defined as the minimum
of price in that week at other accounts in the market). There are large variations in both the degree
of response to this variable (Figure 1) as well as in the competitive conditions across market areas
(Table 2). This is the only sensitivity for which we find an effect of variation in retail competition.
As might be expected, in market areas with greater competition as defined by the Herfindahl index,
we find greater sensitivity to competing account price.
R esponse to Competing Brand Price
In order to capture brand-switching effects within the same account, we also included the
price of the major competing brand (Folgers in the Maxwell sales model and Maxwell in the Folgers
model). The EDLP format variable is important in explaining account variation in this cross-
elasticity. As might be expected from the discussion of own price elasticities, customers at EDLP
format stores are less sensitive to competing brand prices. The chain size variable also explains some
of the account variation in this elasticity, indicating that consumers at accounts with more stores per
8/3/2019 SSRN-id331541
21/38
- 21 -
capita are less sensitive to price changes of competing brands. This result is similar to the findings of
Ho, Tang, and Bell (1998), where in this case the lower travel costs associated with their trip to a
close store may more than compensate for the higher prices that they may pay.
Relative Contributions of E x planatory V ariable T ypes
We have related three groups of variables to promotion sensitivity: i. Retail competition, ii.
Account retail strategy, and iii. Consumer demographics. In order to assess the relative importance of
each group, we compute an R-squared like measure of the contribution of each group to the
explanation of the variability of the promotion sensitivities. To do this, we partition the matrix Z
into Z1, Z2, and Z3 that correspond to the three groupings of variables, fitting the models where Z is
replaced by Z1 and by [Z1 Z2]. Z1 contains the intercept, so all models were fit with an intercept. We
calculate the R-squared measure for each model using the formula (Model SS)/ (Total SS), where
Total SS refers to the sums of squares of the estimates of the sensitivities, and Model SS refers to the
sums of squares of the predicted sensitivities, which is Z for the full model. We model the two
brands, and we report price, display, feature, feature & display, store competitor price, and brand
competitor price results for each brand, giving 12 R-squared measures for each model. Also, we can
calculate R-squared for each draw, meaning that we get 12 distributions of R-squared measures for
each model. We report the medians of these distributions in Table 3.
The explanatory factors account for around 30% of the variation in almost all of the
consumer sensitivities. For instance, the explanatory factors explain 31% and 26% of the variance in
price sensitivity for Maxwell House and Folgers, respectively. Retail competition accounts for very
little of the variation in price sensitivities (3% and 4%). Account retail strategy accounts for more of
the variation in price sensitivities (11% and 12%), and demographics explain close to the same
amount of variation as account retail strategy (17% and 10%).
As for response to display and feature activity, demographics and account retail strategy each
explain roughly equal portions of the variance (around 12%), and retail competition explains about
8/3/2019 SSRN-id331541
22/38
- 22 -
6% of the variance in consumer response. The explanatory factors also account for a large percent
of the variation in the cross-price sensitivities, close to 40% for Maxwell House for both the brand
competition and the account competition.
M ark et and A ccount-specific V ariance Components
Our sales response model includes a variance components error specification. This
specification breaks the regression error into a market wide component, which is common to all
accounts in the same geographic market, as well as an account-specific component. In some recent
work in marketing (c.f. Villas-Boas and Winer (1999)), this common component has been identified
as a possible source of endogeneity (if retailers can predict this common component and change
prices as a function of the component). We can provide some direct evidence on the size of this
component by exploring the relative size of the variances of the market-wide and account variances.
Figure 8 provides a histogram of the posterior median of ( ) + 2 2 2m m a for each of our 97
accounts for the Maxwell House regression errors. For the bulk of the accounts, the common
market-wide error component is much smaller than the account component.
Specification Searches
We approached the choice of explanatory variables using the framework outlined above.
Only variables that could be justified on a priori grounds were included as measures of consumer
demographics, retail competition or account retail strategy. In some cases, variables were collected
but found to have insufficient independent variation to be included in the model. In this section, we
briefly discuss some variables that were not included and explain why.
In our model, we used home value as a proxy for household wealth. Other possible proxies
include income and education. Income is often found to be a poor proxy for wealth, as it does not
include flows from physical, financial, and human capital assets. In addition, census measures of
8/3/2019 SSRN-id331541
23/38
- 23 -
income have been found to be subject to large measurement error. Education is highly but not
perfectly correlated with home value.
Some have included ethnicity as an explanatory variable. We have elected not to include
ethnicity in our results reported above as there is no real theory that can account for effects of
ethnicity once wealth and household composition is controlled for. In specifications not reported
here, we found that Hispanic and black consumers are less price responsive.
We employ the Herfindahl index as a measure of degree of retail competition. It captures the
effect of both the number of chains in a market and the distribution of shares across these chains.
We also tried to include the retail market share that an account has as a measure of clout they have
in a market. However, this measure was correlated with the Herfindahl index and could not be used.
Finally, the presence/ absence of private label coffee was not included as a account retail
strategy variable, as only one of the 97 accounts in our sample did not have a private label coffee
brand.
8/3/2019 SSRN-id331541
24/38
- 24 -
Conclusion
From an empirical point of view, little is known about the determinants of differential
response to promotion activities. Earlier work has focused only on price elasticities in a limited
number of market areas. The use of one or two market areas means that account retail strategy and
competition variables cannot be investigated. With the exception of Ainslie and Rossi (1998),
virtually no work has been done on the determinants of response to display and feature advertising.
We assemble a unique data set that spans all major US market areas and includes account
retail strategy, retail competition, and demographic variables. We also have access to display and
feature advertising information that allows computation of display/ feature sensitivities for each
major retailer-market combination. We conduct our analysis at the key account level. This is the
relevant unit of analysis for retail competition and account retail strategy. We adjust our analysis for
possible aggregation bias, following the suggestions of Christen et al. Two stage estimation methods
used in past studies are less efficient than our simultaneous Bayesian hierarchical method. This
means we bring maximum statistical power to bear on the problem. 4
Collectively, account retail strategy, retail competition, and demographic variables can
explain about 30 per cent of the variation in promotion sensitivity. Most surprising is our finding
that retail competition variables are the least important in explaining response to promotional
activities. The only detectible relationship is between retail competition and the cross-elasticity of
prices across accounts in the same market area. On the other hand, account retail strategy variables
such as price format and store size are quite important in explaining both price elasticity as well as
display/ feature advertising response. Since our analysis includes advertising variables, we are able to
better understand some of the puzzling results of earlier work relating demographics to price
4In order to address potential endogeneity concerns and to see if instrumental variables approaches would
affect our results, we projected prices on market-level wholesale price data (specifically collected for this studyfrom Leemis) and used the fitted values from these regressions in our hierarchical model (the regressions alsoincluded account dummies and interactions of account dummies with the market-level wholesale prices).Unfortunately, we find that with instruments in our hierarchical model removes all account level time seriesinformation in the data and there is insufficient information to measure the relationship between the
8/3/2019 SSRN-id331541
25/38
- 25 -
response. In fact, it turns out that certain demographic groups (such as the Elderly) are not more
price elasticper sebut are more advertising responsive. Taken as a whole, our results call attention to
the importance of account retail strategy variables and to the non-trivial interaction between price
and other promotional variables.
promotional sensitivities and our explanatory variables.Finally, we want to argue that endogeneity biases, likeaggregation biases, are unlikely to affect the relative importance of the explanatory variables.
8/3/2019 SSRN-id331541
26/38
- 26 -
References
Ainslie, Andrew and Peter E. Rossi (1998) Similarities in Choice Behavior Across ProductCategories. M arketing Science17, 2, 91-106.
Allenby, Greg M. and Peter E. Rossi (1991) There is No Aggregation Bias: Why Macro LogitModels Work, Journal of Business and E conomic Statistics, 9, 1 (Jan), 2-14.
Berry, S., J. Levinsohn, and A. Pakes (1995), Automobile Prices In Equilibrium, E conometrica 63,841-890.
Barnard, J., McCulloch, R., and Meng, X. (2000), "Modeling Covariance Matrices in Terms ofStandard Deviations and Correlations, with Application to Shrinkage," Statistica Sinica, 10, 4(Oct), 1281-1311.
Becker, Gary (1965), A Theory of the Allocation of Time, E conomic Journal, 75, 493-517.
Bell, David R. and James M. Lattin (1998) "Shopping Behavior and Consumer Preference for StoreFormat: Why 'Large Basket' Shoppers Prefer EDLP," M arketing Science17, 1, 66-88.
Blattberg, Robert C., Gary D. Eppen, and Joshua Lieberman (1978), A Theoretical and EmpiricalEvaluation of Price Deals for Consumer Nondurables, Journal of Mark eting, 45, 1, 116-129.
Boatwright, Peter, Robert McCulloch, and Peter E. Rossi (1999) "Account Level Modeling for TradePromotion: An Application of a Constrained Parameter Hierarchical Model," Journal of the A merican Statistical A ssociation, 94, 448, 1063-1073.
Bolton, Ruth (1989) "The Relationship Between Market Characteristics and Promotional Price
Elasticities,"M arketing Science, 8, 2, 153-169.
Chevalier, Judith A., Anil K. Kashyap, and Peter E. Rossi (2001), Why Don't Prices Rise DuringPeriods of Peak Demand? Evidence from Scanner Data, University of Chicago GraduateSchool of Business Working Paper.
Christen, Markus, Sachin Gupta, John C. Porter, Richard Staelin, and Dick R. Wittink (1997), UsingMarket-Level Data to Understand Promotional Effects in a Nonlinear Model, Journal ofM ark eting R esearch, 34, 3 (August), 322-334.
Dhar, Sanjay K. and Stephen J. Hoch (1997) "Why Store Brand Penetration Varies by Retailer,"M arketing Science16,3, 208-227.
Hausman, J. A. (1996), Valuations of New Goods Under Perfect and Imperfect Competition, in T.Bresnahan and R. Gordon, Eds, T he E conomics of N ew Goods, Chicago: University of ChicagoPress.
Ho, Teck-Hua, Christopher S. Tang, David R. Bell (1998) "Rational Shopping Behavior and theOption Value of Variable Pricing,"M anagement Science44, 12, S145-S160.
8/3/2019 SSRN-id331541
27/38
- 27 -
Hoch, Stephen J., Xavier Drze and Mary E. Purk (1994) "EDLP, Hi-Lo, and Margin Arithmetic,"Journal of Mark eting, 58, 4 (Oct.). 16-27.
Hoch, Stephen J., Byung-Do Kim, Alan L. Montgomery, and Peter E. Rossi (1995) "Determinants ofStore-Level Price Elasticity," Journal of Mark eting Research32,1 (Feb) 17-29.
Kim, S.Y., and R. Staelin (1999) "Manufacturer Allowances and Retailer Pass-Through Rates in aCompetitive Environment," M arketing Science18, 1, 59-76.
Kumar, V. and Robert P. Leone (1988), Measuring the Effect of Retail Store Promotions on Brandand Store Substitution, Journal of Mark eting Research, 25 (May), 178-185.
Lal, Rajiv, John D.C. Little, and J. Miguel Villas-Boas (1996) A Theory of Forward Buying,Merchandising, and Trade Deals, M ark eting Science, 15, 1, 21-37.
Lal, Rajiv; and Carmen Matutes (1994) Retail Pricing and Advertising Strategies, T he Journal ofBusiness; 67, 3, 345-370.
Lal, Rajiv, and Ram Rao (1997) Supermarket Competition: The Case of Every Day Low Pricing,M arketing Science, 16, 1, 60-80.
Narasimhan, C. (1988), Competitive Promotional Strategies, Journal of Business, 61, 4 (October),427-449.
Nevo, A. (2001), Measuring Market Power in the Ready-to-Eat Cereal Industry, E conometrica,forthcoming.
Raju, J. S. (1992), The Effects of Promotions on Variability in Category Sales, M ark eting Science 11,207-220.
Shankar, Venkatesh and Lakshman Krishnamurthi (1996) "Relating Price Sensitivity to RetailerPromotional Variables and Pricing Policy: An Empirical Analysis," Journal of R etailing, 72, 3,249-272.
Tanner, M. A. (1993), T ools for Statistical Inference: Methods for the E x ploration of Posterior D istributions andL ik elihood Functions (2nd ed.), New York: Springer-Verlag.
Varian, Hal R. (1980) "A Model of Sales," The A merican E conomic Review70, 4 (Sept.), 651-659.
Villas-Boas, M. and R. Winer (1999), "Endogeneity in Brand Choice Models," M anagement Science, 45,
1324-1338.
Walters, Rockney G. (1991), Assessing the Impact of Retail Price Promotions on ProductSubstitution, Complementary Purchase, and Interstore Sales Displacement, Journal ofMark eting, 55 (April), 17-28.
Wittink, Dick R. (1977) "Exploring Territorial Differences in the Relationship Between MarketingVariables," Journal of Mark eting Research14, 2 (May) 145-55.
8/3/2019 SSRN-id331541
28/38
- 28 -
Table 1
Anticipated Signs for Explanatory Variables Effects
Price Sensitivity Display Sensitivity Feature Sensitivity
Income/ Wealth - + +Age + + +Private Label Share + + +EDLP Format - - -
Number of Stores/ Assortment - - -Retail Competition + + +
8/3/2019 SSRN-id331541
29/38
- 29 -
Table 2
Account Retail Strategy, Retail Competition and Consumer Characteristics:
Descriptive Statistics
Median St Dev Min Max
Elderly 0.1964 0.0375 0.1139 0.3430House Value 0.1731 0.2741 0.0551 0.8754Private Label Share 0.0639 0.0578 0.0000 0.2560Floor Space/ Capita 4.3664 1.4889 2.1198 9.3488Herfindahl 0.1182 0.0628 0.0402 0.3554Chain Size 60 41.5805 15 192
8/3/2019 SSRN-id331541
30/38
- 30 -
Table 3
Relative Importance of Account Retail Strategy, Retail Competition and Demographics in Explaining
Promotion Sensitivity
Note: Figures in parentheses are standard deviations of posterior distributions of the R2 measure
Model Price Display FeatureFeat &Disp
AccountPriceComp
Brand PriceComp
Full Model Maxwell 31% (10%) 33% (11%) 24% (9%) 23% (9%) 41% (14%) 43% (13%)
Folgers 26% (10%) 22% (9%) 28% (10%) 31% (10%) 35% (14%) 24% (11%)
Retail Strategy & Maxwell 14% (8%) 17% (9%) 17% (8%) 15% (8%) 11% (8%) 18% (9%)
Competition Folgers 16% (8%) 8% (6%) 15% (7%) 14% (7%) 29% (14%) 15% (9%)
Retail Maxwell 3% (3%) 7% (6%) 6% (5%) 9% (7%) 4% (5%) 3% (4%)
Competition Folgers 4% (4%) 3% (4%) 3% (3%) 6% (5%) 26% (14%) 4% (5%)
8/3/2019 SSRN-id331541
31/38
- 31 -
PriceResponse
-3.0
-2.0
-1.0
All Accts Selected Accounts
Price
DisplayLift
0
1
2
3
All Accts Selected Accounts
Display
FeatureLift
0
2
4
6
8
All Accts Selected Accounts
Feature
F&D
Lift
0
5
10
15
20
All Accts Selected Accounts
Feature and Display
ResponsetoCompet(Account)
-0.6
0.0
0.6
All Accts Selected Accounts
Price at Competing Account
ResponsetoCom
pet(Brand)
-1.0
0.0
1.0
All Accts Selected Accounts
Price of Competing Brand
Figure 1
Distribution of Promotion Sensitivities
8/3/2019 SSRN-id331541
32/38
- 32 -
Figure 2
Relationship between Price Response and Account Retail Strategy, Competition and
Consumer Demographic Variables
-0.2
0.0
0.2
0.4
Price Response
Eld
Homeval
Pvt.Lbl
EDLP
ChainSize
Floor.Sp
Hrfndl
M F M F M F M F M F M F M F
M MaxwellF Folgers
8/3/2019 SSRN-id331541
33/38
- 33 -
Figure 3
Relationship between Display Response and Account Retail Strategy, Competition and Consumer
Demographic Variables
-0.4
-0.3
-0.2
-0.1
0
.0
0.1
0.2
Display Response
Eld
Homeval
Pvt.Lbl
EDLP
ChainSize
Floor.Sp
Hrfndl
M F M F M F M F M F M F M F
M MaxwellF Folgers
8/3/2019 SSRN-id331541
34/38
- 34 -
Figure 4
Relationship between Feature Response and Account Retail Strategy, Competition and Consumer
Demographic Variables
-0.6
-0.4
-0.2
0.0
0.2
Feature Response
Eld
Homeval
Pvt.Lbl
EDLP
ChainSize
Floor.Sp
Hrfndl
M F M F M F M F M F M F M F
M MaxwellF Folgers
8/3/2019 SSRN-id331541
35/38
- 35 -
Figure 5
Relationship between Feature and Display Response and Account Retail Strategy, Competition and
Consumer Demographic Variables
-0.6
-0.4
-0.2
0.0
0.2
Feature & Display Response
Eld
Homeval
Pvt.Lbl
EDLP
ChainSize
Floor.Sp
Hrfndl
M F M F M F M F M F M F M F
M MaxwellF Folgers
8/3/2019 SSRN-id331541
36/38
- 36 -
Figure 6
Relationship between Response to Competing Account Price and Retail Strategy, Competition and
Consumer Demographic Variables
-0.2
-0.1
0.0
0.1
0.2
Response to Competing Account Price
Eld
Homeval
Pvt.Lbl
EDLP
ChainSize
Floor.Sp
Hrfndl
M F M F M F M F M F M F M F
M MaxwellF Folgers
8/3/2019 SSRN-id331541
37/38
- 37 -
Figure 7
Relationship between Response to Competing Brand Price and Account Retail Strategy, Competition
and Consumer Demographic Variables
-
0.4
-0.2
0.0
0.2
Response to Competing Brand Price
Eld
Homeval
Pvt.Lbl
EDLP
ChainSize
Floor.Sp
Hrfndl
M F M F M F M F M F M F M F
M MaxwellF Folgers
8/3/2019 SSRN-id331541
38/38
0.1 0.2 0.3 0.4
0
5
10
15
Medians of Mkt Variance/(Mkt Variance + Acct Variance)
Figure 8
Posterior Median of Relative Variances Maxwell House Regression Errors