MICHAEL LEWIS*

Despite the proliferation of loyalty programs in a wide range of cate-

gories, there is little empirical research that focuses on the measurement
of such programs. The key to measuring the influence of loyalty pro-
grams is that they operate as dynamic incentive schemes by providing
benefits based on cumulative purchasing over time. As such, loyalty pro-
grams encourage consumers to shift from myopic or single-period deci-
sion making to dynamic or multiple-period decision making. In this study,
the author models customers’ response to a loyalty program under the
assumption that purchases represent the sequential choices of cus-
tomers who are solving a dynamic optimization problem. The author esti-
mates the theoretical model using a discrete-choice dynamic program-
ming formulation. The author evaluates a specific loyalty program with
data from an online merchant that specializes in grocery and drugstore
items. Through simulation and policy experiments, it is possible to evalu-
ate and compare the long-term effects of the loyalty program and other
marketing instruments (e.g., e-mail coupons, fulfillment rates, shipping
fees) on customer retention. Empirical results and policy experiments
suggest that the loyalty program under study is successful in increasing
annual purchasing for a substantial proportion of customers.

The Influence of Loyalty Programs and

Short-Term Promotions on Customer
Retention

Loyalty programs have long been an important element Loyalty programs that base rewards on cumulative pur-
of customer relationship management for firms in travel- chasing are an explicit attempt to enhance retention. Such
related industries such as airlines, hotels, and rental cars. programs encourage repeat buying and thereby improve
Information technology that enables firms to practice retention rates by providing incentives for customers to pur-
individual-level marketing has facilitated the spread of loy- chase more frequently and in larger volumes. However,
alty programs into such diverse industries as gaming, finan- dynamically oriented promotions, such as loyalty programs,
cial services, and retailing (Deighton 2000). Accordingly, represent just one possible technique for increasing cus-
academic researchers have begun to study loyalty programs. tomer retention. Repeat buying may also be encouraged
Behaviorally oriented researchers, such as Soman (1998) through various means such as short-term discounts on mer-
and Kivetz and Simonson (2002), have studied the effect of chandise or reduced shipping charges. Therefore, it is
delayed incentives on consumer decisions. Zhang, Krishna, important to develop models that can simultaneously esti-
and Dhar (1999), Kim, Shi, and Srinivasan (2001), and mate the influence of dynamic and current factors on long-
Kopalle and Neslin (2003) have proposed analytical models term customer behavior. In this article, I report on a
to study the impact of loyalty programs in categories with methodology for assessing the relative impact of loyalty-
different structures. This study contributes to the literature based promotions, short-term promotions, and individual-
that is focused on empirically measuring response to loyalty level factors on customer purchasing over time.
programs (Drèze and Hoch 1998; Sharp and Sharp 1997). There is only limited and contradictory published empir-
ical work on the value of loyalty programs. A relevant study
by Sharp and Sharp (1997) analyzes individual-level data
*Michael Lewis is Assistant Professor of Marketing, Warrington College by using a one-period switching model to measure the abil-
of Business, University of Florida (e-mail: [email protected]). The ity of a loyalty program to alter normal repeat-purchase
author thanks the three anonymous JMR reviewers for their helpful rates. Unfortunately, the study’s results are inconclusive. In
comments.
contrast, Drèze and Hoch (1998) report on a category-

Journal of Marketing Research

281 Vol. XLI (August 2004), 281–292
282 JOURNAL OF MARKETING RESEARCH, AUGUST 2004

specific loyalty program that results in increases for both ment strategy in two respects. First, it is a means for assess-
the specific category and total store traffic. The contrasting ing the influence of dynamic promotions such as loyalty
findings are consistent with the lack of consensus on the programs that operate over an extended period. Second, by
ability of loyalty programs to increase customer retention controlling for the dynamic elements of a customer’s deci-
(see Dowling and Uncles 1997). It should be emphasized sion problem, the model removes a possible source of bias
that studies that question the value of loyalty programs and may provide better estimates of the effects of short-
(e.g., Dowling and Uncles 1997; Sharp and Sharp 1997) are term marketing tactics.
largely based on research that uses single-period switching The analysis employs a research methodology known as
models. Additional research with models that more fully discrete-choice dynamic programming (Eckstein and
replicate the dynamics of consumer response is needed to Wolpin 1989). The approach is based on the idea that a cus-
judge the effectiveness of dynamically oriented loyalty tomer’s observed sequence of decisions may be interpreted
programs. as the solution to a dynamic optimization problem. These
For a frequency program to be effective in increasing loy- methods are established in economics (Eckstein and Wolpin
alty, it must have a structure that motivates customers to 1989; Rust 1994) and are becoming more common in mar-
view purchases as a sequence of related decisions rather keting (Erdem, Imai, and Keane 2003; Erdem and Keane
than as independent transactions. That is, the structure must 1996; Gönül and Shi 1998; Gönül and Srinivasan 1996;
give customers an incentive to adopt a dynamic perspective. Sun, Neslin, and Srinivasan 2004). Dynamic programming
O’Brien and Jones (1995) suggest that the major factors methods are ideal for analyzing individual choices that are
that customers consider when evaluating programs are the based on both current and future expected benefits. A loy-
relative value of awards and the likelihood of achieving a alty program that bases awards on the level of purchasing
reward. Furthermore, the likelihood of achieving a reward is over a specified period is a prime example of such a deci-
a function of cumulative buying thresholds and time con- sion problem. A further benefit of dynamic programming
straints. These design elements (e.g., thresholds, rewards, methods is that the estimated coefficients can be used to
time constraints) combine with individual-level require- conduct simulations that replicate the consumer’s dynamic
ments and preferences to determine the customer’s expected decision process.
benefits of participating in a loyalty program. The primary contribution of this research is a framework
A special characteristic of loyalty programs is that their for modeling the influence of a reward program and other
attractiveness may change dynamically with a customer’s marketing instruments on customer retention. Firms have
decisions. As purchases are made, both the customer’s multiple options for their promotional budgets, so models
investment in the program and the customer’s likelihood of that can quantify the long-term effects of loyalty programs
earning a reward increase. Conversely, when a customer and other options (e.g., pricing, coupons, shipping fees) can
decides not to purchase in a given period, the likelihood of help the firm justify its choices. Although most database
earning a reward decreases, because the customer moves no marketing applications focus on tasks such as customer
closer to the reward threshold, and the time left to earn scoring that are designed to maximize the profitability of
rewards shrinks. The assessment of a program’s attractive- single-period mailing efforts (Bult and Wansbeek 1995), the
ness is further complicated because customers usually have current research focuses on customers’ response to a range
imperfect knowledge of their future requirements and of the of marketing instruments over an extended period. The
marketing policies of the firm. These dynamic factors are a model provides the means to support multicampaign direct
challenge in the modeling of customer response to loyalty marketing in environments in which customers have a
programs. dynamic orientation.
This study empirically estimates the impact of a reward In terms of substantive findings, the results suggest that
program and other elements of the marketing mix on cus- the loyalty program under examination is successful in
tomer buying behavior over time by developing a model changing customer behavior and in motivating customers to
that replicates dynamic consumer response to a loyalty pro- increase purchasing. In addition to a statistically significant
gram. In contrast to previous models, the current model estimate for the loyalty reward parameters, formulations
considers the impact of previous purchasing activity and that assume that customers are dynamically oriented fit bet-
customer expectations. The underlying behavioral assump- ter than do models that do not include a dynamic structure.
tion is that a reward program can motivate customers to The simulation experiments also provide a means for esti-
base their purchasing decisions both on the current environ- mating the magnitude of response to the program.
ment and on a long-term goal of achieving a frequent buyer The remainder of the article is organized as follows: In
reward. In other words, an effective reward program can the next section, I develop a dynamic programming model
encourage customers to make decisions that maximize of customer response to a loyalty program. Next, I discuss
expected utility over an extended time horizon rather than at model estimation issues and present empirical results. I then
each purchase occasion. This assumption is consistent with report the results of policy experiments that are designed to
previous findings in the literature that expectations of the assess the impact of the loyalty program and other promo-
future can affect consumers’ current-period decisions (e.g., tions on customer retention. I subsequently address limita-
Boulding et al. 1993; Lemon, White, and Winer 2002). tions and future research opportunities and conclude with a
The empirical section of this article uses individual-level discussion of key findings and managerial applicability.
customer data from an Internet grocer to develop a dynamic
model of customer retention. The model identifies the key MODELING THE DYNAMICALLY ORIENTED
factors that influence customers to make repeat purchases CUSTOMER
over time. A model based on an appropriate dynamic behav- The model is intended to replicate the decision-making
ioral specification can assist an overall customer manage- process of a customer when dynamic incentives exist. As
Loyalty Programs and Customer Retention 283

such, the model needs to include the factors that affect the written in terms of alternative specific expected value
attractiveness of a customer’s options at a given time and a functions as V[S(t)] = maxj ∈ JVj[S(t)], where the alterna-
structure that captures dynamic considerations. I begin with tive specific value function Vj[S(t)] is the expected value of
a customer who makes repeated choices from among J the customer selecting option j at time t and then selecting
mutually exclusive alternatives in each of T discrete peri- optimal actions subsequently. Furthermore, the alternative
ods, where the choice set is defined in terms of purchase specific value functions satisfy the Bellman (1957)
quantity (I dropped customer-specific indexes for clarity). equation:
The choice set also includes a no-purchase option. This
(3) Vj[S(t)] = E[Rj(t)|S(t)] + αE{V[S(t + 1)]|S(t), dj(t) = 1},
model allows a customer to choose the extent to which he
or she participates in a program over a sequence of when t = 1, ..., T – 1,
occasions.
The quantity decision is likely to be based on marketing and
factors such as prices and on individual-level factors such Vj[S(T)] = E[Rj(T)|S(t)], at t = T.
as inventories. A myopic decision maker would select from
the options 1, ..., J and consider the relative benefits asso- The form of the alternative specific value functions
ciated with each option, denoted as Rj. The standard implies that decisions are based on both the immediate
assumption is that consumers select the option that yields utility provided by an option and the expected total future
the optimal benefits for the purchase occasion. However, a utility from subsequent purchases and loyalty rewards. A
loyalty program may provide an incentive for customers to key detail is that the expected future benefits from period
view weekly transactions as a sequence of related deci- t + 1 forward may depend on the option selected. In many
sions. If rewards are earned over time, a multiple-period loyalty programs, an important element of the state space,
objective function is appropriate for the customer’s deci- which evolves according to customer actions, is the level
sion problem. The objective function of a dynamically ori- of cumulative spending. In general, the evolution of the
ented customer1 is state space is a function of both the current environment,
S(t), and the choices made by the customer, dj(t), over
 T J  time.
(1) max E 
 ∑ αt −1 ∑ R (t)d (t)|S(t) ,
j j The discount factor, α, defines the degree to which con-
t = 1 j = 1  sumers are forward-looking or dynamically oriented. A pos-
itive α (0 < α < 1) implies that customers consider future
where Rj(t) is the single-period reward associated with benefits when making current decisions. If α is equal to
option j at time t, dj(t) is an indicator variable that is set zero, customers completely discount future benefits. In the
equal to 1 if option j is chosen at time t and is set equal to context of a loyalty program, a positive α means that cus-
zero otherwise, and α is a one-period discount factor. An tomers partially base current purchasing decisions on the
important element of this formulation is that the reward expectation of earning a reward. When α is equal to zero,
functions Rj(t) are conditional on the state of the environ- customers do not consider future benefits.
ment, S(t). The state space, S(t), is a vector of information Customer Choices and State Description
about the environment that is relevant to the customer’s
forward-looking optimization problem. The state space may As I previously stated, the setting for the empirical analy-
consist of marketing-mix elements, such as the pricing envi- sis is an online retailer that specializes in nonperishable
ronment in a given week, and customer-specific informa- grocery and drugstore items. I model customer choices in
tion, such as cumulative purchases. terms of four purchase levels during weekly periods: no
The customer’s decision problem involves selecting the order, small order, medium order, and large order. Small
level of buying in each period to maximize the expected orders are ones in which merchandise costs less than $50.
discounted utility for the remainder of the time horizon.2 Medium orders are ones in which merchandise costs at least
The value function, V, is the maximum value of discounted $50 but less than $75. Large orders cost at least $75. The
expected utility over the horizon. options available in each week are indexed by j, where

 T J 
j = no: no purchase in a given week,
(2) V[S( t )] = max E 
d j (t) 
∑ ατ − t ∑ R j (τ)d j (τ)|S(τ)  .

j = sm: purchase of a small basket of goods,
j = med: purchase of a medium basket of goods, and
τ = t j = 1  j = lrg: purchase of a large basket of goods.
The value function is the utility that can be achieved over The order options are defined to be mutually exclusive so
the time horizon if the customer selects the optimal that Σdj(t) = 1. This discretization is not arbitrary but is
sequence of quantities. The value function may also be based on the shipping fees charged by the firm.
I use weekly purchasing decisions to update a state vari-
able that tracks cumulative expenditures Cumit for individ-
1The model is developed for a representative customer. Given that ual i at time t. This variable is maintained by rounding the
response to this type of program is likely to vary across the population, I actual cumulative expenditures to the nearest $33 incre-
estimate a finite mixture model, which is detailed in the “Model Estima-
tion” section.
ment.3 For purposes of the customer’s expectations of the
2The loyalty program under study includes a time constraint. Specifi- future state, the variable is updated according to Equation
cally, the program is based on customers’ annual purchasing, and points
are not carried over into future annual cycles. This is an important element
of the program’s structure because it dictates the formulation of the 3These increments roughly reflect the sizes of the average orders
dynamic program used to model the customer’s optimization problem. observed in each category.
284 JOURNAL OF MARKETING RESEARCH, AUGUST 2004

4.4 For example, if a customer has $500 in cumulative within the previous two weeks, k = 2 indicates between
expenditures at a given time t and makes a small purchase, three and five weeks since the previous purchase, k = 3 indi-
the customer is assumed to behave as if the future value of cates between six and eight weeks since the previous pur-
Cumit + 1 is $533. chase, and k = 4 indicates no purchase in the previous nine
weeks. The level of duration for a representative consumer
4
at time t is denoted by a series of indicator variables,
(4) Cum it + 1 = Cum it ∑ (100/3)( j − 1)d (t).
j Rcn(k)i,t, that are defined as mutually exclusive, such that
j = 1 Σ4k = 1Rcn(k)i,t = 1.5
The Rcn and FM variables are consistent with the
I use the cumulative buying variables to determine when recency, frequency, and monetary (RFM) value variables
a loyalty-based reward is earned. A reward variable, Lit, is that are popular in direct-marketing response models. The
set to one if a purchase causes a reward threshold to be RFM variables reflect important individual-level differ-
crossed and is set to zero otherwise. This occurs when a ences and have been used in previous dynamic program-
customer whose previous level of cumulative purchases is ming models of consumer behavior (Gönül 1999; Gönül
less than a reward threshold makes a purchase that brings and Shi 1998).
cumulative purchasing over the next reward threshold. For An important marketing-mix element for direct retailers
the program under study, the reward thresholds are set at is shipping and handling fees. A unique element of the data
$1,000, $1,500, and $2,000 worth of annual purchases. If a is that the online retailer in this study experimented with
customer with cumulative purchases of $925 makes a pur- various shipping fee structures during data collection. The
chase worth $75 or more, Lit equals one, but Lit equals zero shipping fee structure in place at a given time is important,
if only a $50 purchase is made. Rewards are paid in blocks because it provides incentives and penalties for various
of 500 frequent-flyer miles for the $1,000 and $1,500 order sizes and thus may affect both order incidence and
thresholds and 1000 miles for the $2,000 level. Therefore, distribution of order sizes. Specifically, graduated shipping
when the $2,000 threshold is crossed, the Lit variable equals fee structures impose an element of nonlinear pricing on
two, to reflect that the reward earned is double the value of buying decisions. For example, a shipping fee structure used
the previous rewards. in the data is a schedule that charges $6.97 in fees for orders
I do not treat the cumulative expenditure variables as a less than $75 and provides free shipping for orders greater
direct component of customers’ reward functions. Rather, I than $75. This type of policy provides an incentive for cus-
define variables that account for the potential benefits of tomers to increase order sizes. Customers may find them-
being an experienced customer. My strategy is to use indi- selves in situations in which the total order cost (i.e., the
cator variables to define broad categories of previous activ- sum of merchandise and shipping fees) is less for a basket
ity. I define two indicator variables (FM1 and FM2) that with a higher merchandise total, because the marginal cost
define classes of previous purchasing activity for a given of an incremental item (that causes the merchandise order to
customer as follows: FM1 = 1 if Cumit ≥ $300, and FM1 = 0 cross a threshold) can be negative. Similarly, a schedule that
otherwise, and FM2 = 1 if Cumit ≥ $500, and FM2 = 0 involves fees that increase as order size increases provides
otherwise. The variables are intended to capture dynamic an incentive for customers to purchase smaller baskets of
effects beyond the loyalty program. For example, as cumu- goods. The cost of shipping each order size under each ship-
lative purchasing increases in terms of number and size of ping fee schedule is provided in Table 1.
transactions, it is likely that customers gain a measure of Shipping variables, SH(h)t, indicate the shipping fee
expertise that effectively reduces transaction costs. structure used at a given time, where h = 1 indicates a
Another element of the state space is the amount of time mixed/decreasing fee schedule, h = 2 indicates increasing
since a customer’s previous order. This measure is intended fee schedule (high), h = 3 indicates increasing fee schedule
to control for possible inventory or customer attrition (low), and h = 4 indicates free shipping for all order sizes.
effects and is adjusted each period as follows: Durit = weeks For example, SH(4)t = 1 indicates that the firm used a free
since a purchase was made; if j ∈ {small, medium, large} shipping policy at time t. The shipping and handling fee
(i.e., a purchase is made), then Durit + 1 = 1; and if j = no, structures are mutually exclusive, such that ΣSH(h)t = 1.
then Durit + 1 = Durit + 1. I subsequently define the actual The online retailer under study carries in excess of
implementation of the duration variable, which is capped at 10,000 items. Therefore, rather than use specific prices in
nine weeks through a series of categorical variables. the model, I use a summary measure that describes the pric-
Before the empirical analysis, the role of the duration ing environment. This price variable is defined as the aver-
variable is not well defined. It may capture inventory- age price of the 50 top-selling items during the data collec-
related decision making if the natural buying cycle is tion period in a given week and is calculated as Pt =
greater than the weekly cycle used in the model. However, (Σ50
item = 1Pitem,t)/50. The price variable accounts for the rela-
if the probability of observing a purchase decreases as the tive attractiveness or competitiveness of the firm’s pricing
time since the previous purchase increases, the duration in a given week. An additional price-related factor is e-mail
variable will capture customer attrition. The duration cate-
gories are indexed by k, where k = 1 indicates a purchase
5Recency also could be constructed as a continuous variable. The use of
a series of categorical variables is a simplifying assumption that helps
4For purposes of the loyalty rewards, this variable is updated exactly on maintain computational tractability. In addition, I based the selection of the
the basis of previous levels of buying and the quantity purchased. I use the break points on patterns of incidence observed in the data. For example,
discretization to approximate the customer’s expectations of the next purchase incidence drops from 9.5% when duration is eight weeks to 5%
period state. when it is nine weeks; after nine weeks, the rate stabilizes near 3%.
Loyalty Programs and Customer Retention 285

promotions. The firm’s practice has been to send occasional exp(θLow + θP Pt )

e-mail coupons to existing customers for a 10% discount on (5) Pr (Pt + 1 = PLow ) = ,
[1 + exp(θLow + θP Pt )]
purchases during a given week. The variable Cit is set equal
to one if a 10% discount coupon is sent in week t and is set exp(θHigh + θLow + θP Pt )
equal to zero otherwise. (6) Pr (Pt + 1 = Pave ) =
[1 + exp(θHigh + θLow + θP Pt )]
In addition to the role of current-period prices, expecta-
tions of future prices may also be salient. Recall that in exp(θLow + θP Pt )
Equation 3, the value function involves an expectation of – ,
[1 + exp(θLow + θP Pt )]
the future value of program participation. When expected
future prices are high, customers expect to pay a significant and
premium to obtain the points needed for a reward. Expecta-
exp(θHigh + θLow + θP Pt )
tions of future prices have been considered by several mar- (7) Pr (Pt + 1 = PHigh ) = 1 − .
keting researchers (Kalwani et al. 1990; Narasimhan 1989) [1 + exp(θHigh + θLow + θP Pt )]
and have been estimated with dynamic programming mod-
els (Gönül 1999). The complete set of state and control variables is summa-
Expectations of future prices are modeled as a function rized in Table 2, which also includes the variables con-
of current prices with an ordered logit structure. Specifi- structed from the state variables and the levels the variables
cally, expectations of the future pricing environment are may take.
modeled by means of a discrete three-point distribution that
enables customers to anticipate an average, low (one stan- Utility Functions
dard deviation below the average), or high (one standard The single-period utilities associated with each of the J
deviation above the average) price in the next period. The options are given subsequently, where Rj,i(t) corresponds to
probabilities are estimated with the following equations, the utility to individual i of option j at time t:
where Pt is the current-period price, and θP, θLow, and θHigh
are parameters to be estimated: (8) R no,i ( t ) = β no + ε no,it ,

Table 1
SHIPPING AND HANDLING FEE STRUCTURES

Shipping and Small Basket: Medium Basket: Large Basket:

Handling Policy Description $0–$50 $50–$75 Greater Than $75
1 Mixed/decreasing $4.99 $6.97 $0
2 Increasing (high) 4.99 6.97 8.95
3 Increasing (low) 2.99 4.99 4.99
4 Free 0 0 0

Table 2
STATE AND CONTROL VARIABLES

State Variables and Related Covariates

Variable Description Levels
Cum Cumulative purchasing $33 dollar increments to $3,033 (91 levels)
FM(f) Indicates how to classify customers according to levels of Not a state variable; constructed from the cumulative spending variable
cumulative spending
L Indicates whether a loyalty reward is earned in period t Binary variable; not a state variable but a function of cumulative
expenditures and customers’ actions at time t
Dur Time (in weeks) since the previous purchase Ranges from 1 to 9 weeks (9 levels)
Rcn(k) Indicator for recency categories Not a state variable (4 levels)
P Price Continuous for the current period
C E-mail coupon Binary variable (2 levels)
SH(h) Shipping and handling schedule Four different shipping and handling fee schedules used during data
collection (4 levels)
T Individual-specific time (in weeks) relative to the program’s 5 < t ≤ 52 (Weeks 1–5 are initialization period)
finite (52-week) horizon (more formally, should be denoted as ti)
E(Pt + 1) Expectations of future prices Function of current-period prices

Control Variable
dj(t) Purchase quantity decision 4 levels (no buy, small, medium, and large)
286 JOURNAL OF MARKETING RESEARCH, AUGUST 2004

H In each period, it is assumed that customers select the

(9) Rsm , i ( t ) = βsm + β P , sm Pit + β C, sm Cit + ∑β h , sm SH (h) it alternative that maximizes utility over the remaining deci-
sion horizon. The probability that a customer selects an
h = 1
K 2 alternative at time t is defined in terms of the alternative
+ ∑ β Rcn(k) k it + ∑β f , sm FM(f ) it + β L L it specific value functions. Under the assumption that the ran-
dom terms are distributed extreme-value i.i.d., it is possible
k = 1 f = 1
to obtain the following closed-form expression for the
+ εsm , it , choice probabilities (Rust 1994):
(10) R med , i ( t ) = β med + β P , med Pit + β C, med Cit exp{v j [S( t )]}
(12) Pr[d j ( t ) = 1|S( t )] = .
H K J

+ ∑β h , med SH (h) it + ∑ β Rcn(k)

k it ∑ exp{v j′ [S( t )]}
h = 1 k = 1 j′ ∈ {no, sm, med, lrg}

2
In this expression, vj represents the deterministic portion of
+ ∑β f , med FM(f ) it + β L L it + ε med , it , the alternative specific value function. Therefore, model
f = 1 estimation requires the repeated solution of a dynamic pro-
and gramming problem to calculate the value functions. Under
the assumption of a homogeneous population, the log-
H

∑β
likelihood function to be maximized consists of the sum of
(11) R lrg, i ( t ) = β lrg + β P , lrg Pit + β C, lrg Cit + h , lrg SH (h) it the logarithms of the choice probabilities defined in Equa-
h = 1 tion 12.
K 2 The estimation procedure involves nesting an algorithm
+ ∑ β k Rcn( k )it + ∑β f , lrg FM(f ) it + β L L it to solve the dynamic programming problem within a
maximum-likelihood routine. At each iteration, the dynamic
k = 1 f = 1
programming problem is solved by the use of the current
+ ε lrg, it .
estimate of the parameter vector associated with the reward
functions. The alternative specific value functions are used
Equation 8 represents the reward associated with the no- to compute the log-likelihood for the current iteration. The
purchase option, and Equations 9–11 represent the rewards maximum-likelihood algorithm then updates the parameter
associated with small, medium, and large orders. For exam- vector, and the process is repeated until convergence.
ple, in Equation 9, the utility of a small order is a function Because the loyalty program under study has a finite
of βsm, which is the constant associated with a small order, horizon, I solve the dynamic programming model used to
the price index, and whether an e-mail coupon was avail- generate the choice probabilities with backward recursion.
able at time t. The shipping charges enter through the sum- Given the annual time horizon of the program and the
mation of βh,sm and SH(h)it. For example, β1,sm is the coef- assumption of weekly decisions, the decision horizon
ficient for a small order (option j = 2) when the shipping involves 52 periods. The recursion procedure is executed as
schedule is of Type 1 (h = 1). The second summation term, follows: For a set of parameter values, I compute the alter-
which involves βk and the duration indicator variables native specific value functions for the final period (T = 52)
Rcn(k)it, captures the effects of time since the previous pur- for all combinations of the state space S:
chase. The third summation term involves the cumulative
purchasing variables, FM(f)it. In the term for the loyalty (13) Vj[S(52)] = E[Rj(52)|S(52)].
indicator, βL is the coefficient that describes the benefit of a
loyalty reward. The εjit terms correspond to the random I then use the values for Vj(52) to calculate the value func-
component of utility. In addition, for identification pur- tions for t = 51. Specifically, the alternate specific value
poses, the constant of the no-purchase option is normalized function for an arbitrary state at t = 51 is the expected
to zero (βno = 0). reward during that period plus the discounted expected
reward of an optimal action in the next period:
MODEL ESTIMATION Spot

The estimation of the parameters of the reward functions

is somewhat analogous to the approach used in standard
(14) Vj [S(51)] = E[R j (51)|S(51)] + α ∑ Pr[s|S(51), j]V[s(52)],
s
discrete-choice models. As in standard choice models, the
estimated probability of an observed choice is based on a where Spot is the set of potential states that can be reached in
comparison of the deterministic portion of the utilities of the next period, given the current state. The process continues
each feasible alternative. The major difference with the until the value functions are computed for all periods. I then
dynamic programming model is that the utility associated use the value functions to calculate the probability of each
with each alternative involves both an expression for observed choice in the data using Equation 12 (for more
current-period utility and an expectation of conditional details on the general estimation approach, see Rust 1994).
future benefits. Thus, for dynamic programming models, As stated, the model applies to a population with homo-
the estimation of choice probabilities is based on a compar- geneous preferences for the firm’s services and the offered
ison of the alternative specific value functions given in rewards. This is potentially problematic for a model of
Equation 3. response to a loyalty program. A more realistic assumption
Loyalty Programs and Customer Retention 287

is that the loyalty program alters the behavior of only some Table 3
fraction of the customer population. To extend the model to SUMMARY OF MODEL FITS
account for variability in customer preferences, I use a
latent class approach (Kamakura and Russell 1989). This Model Number of Parameters BIC
approach assumes that the population consists of M types,
Single Segment
where πm is the proportion of type m in the population. For Myopic 28 35,419.5
the finite-mixture approach, the sample likelihood is Dynamic 31 34,996.8
N M Two Segments
(15) ∏∑
i = 1m = 1
Ti |type = m )π m ,
Pr(d1mi , d 2mi , ..., d m Two myopic
Two dynamic
60
66
34,388.0
33,996.7
One dynamic and one myopic 63 34,230.5
where Three Segments
Three dynamic 95 34,003.9
T
Pr[d m
i (1), dm
i (2), ..., dm
n (T)] = ∏ Pr[d(t)|S(t), type = m].
t = 1 models. The addition of a third segment marginally
improves fit but yields several nonintuitive parameters.
The use of a finite-mixture model to account for customer Estimation results provided in Table 4 indicate two fairly
heterogeneity significantly increases the computational bur- different segments. In broad terms, Segment 1 purchases
den, because the dynamic optimization problem must be less frequently and favors larger order sizes, whereas Seg-
repeatedly solved for each type in the population. ment 2 purchases more frequently and favors smaller
Because the direct estimation of the discount factor α orders. To generate descriptions of the two segments, Table
tends to be difficult, this parameter is fixed before the esti- 5 reports results of simulation studies based on each seg-
mation of the other model parameters. The value of α used ment’s parameter values and a marketing policy similar to
in the estimation procedure plays a key role in the analysis. that employed by the firm. Segment 1 places an average of
By setting α equal to zero, the model is estimated under an 8.7 orders per year and spends more than $670 annually.
assumption that customers are myopic and do not consider Segment 2 spends an average of $586 dollars per year and
the value of future rewards in current decisions. If α is set to places approximately 14 orders. In terms of the loyalty
a high value (i.e., close to one), the model assumes that cus- rewards, approximately 15% of the Segment 1 population
tomers value potential future benefits. The empirical strat- earns at least one reward, whereas only approximately 1%
egy is to estimate both types of models. A comparison of of Segment 2 earns a loyalty reward. Table 5 also shows
model fits then provides evidence of whether customers how the two segments combine to an overall population and
have a dynamic response to the loyalty-based rewards. how this simulated population compares with the actual
sample.7
EMPIRICAL IMPLEMENTATION
An important difference in the segments is the levels of
The complete data set from the Internet retailer includes dispersion. Although the two segments have fairly similar
transaction histories for more than 30,000 customers and mean levels of annual spending, there is a significant differ-
spans a 13-month period from 1997 to 1998. For estimation ence in terms of variability. Segment 1 tends to be more
purposes, I selected a small subsample of the population. volatile, which is why it contains the majority of award
Because the primary focus is the effectiveness of the reward recipients. Figure 1 shows the distribution of annual pur-
program, the empirical analysis focuses on customers with chasing for both segments and illustrates the differences in
meaningful transaction histories. The selection criterion is the dispersion of the two groups. Although the BIC meas-
that customers make at least two purchases during their first ures help justify the incorporation of dynamic factors and
5 weeks in the database and that there are at least 30 weeks heterogeneity controls, measures that show how these
of observations. The final sample consists of 1058 cus- model enhancements lead to improved managerial forecasts
tomers who made an average of 10.45 purchases per year. are also important. Because I am interested in predicting
I estimated models for various single- and multiple- long-term behavior, I define measures that describe how
segment formulations and for different assumptions regard- well each specification can replicate the actual distribution
ing the discount rate. Fit statistics and assumptions regard- of cumulative purchasing. To construct the measures, I use
ing segment discount rates for single-, two-, and each model to simulate the distribution of annual spending
three-segment models are provided in Table 3. Table 4 pres- for 100,000 customers. I then compare the simulated distri-
ents the estimated coefficients for a two-segment mixture butions with the actual percentages of the population in
model in which both segments possess a dynamic orienta- terms of absolute deviations (Table 6). This provides an
tion.6 The two-segment dynamic model is the best fitting of indication of how well the specifications capture the vari-
all the two-segment models, according to the Bayesian ability in the population. To show how well the specifica-
information criterion (BIC) fit measure, and is the basis for tions capture the influence of the loyalty program, I also
subsequent analysis. In addition to providing a better fit include the percentage of customers who are predicted to
than the competing two-segment models, the two-segment earn a loyalty reward. In general, the dynamic specifications
dynamic model is superior to the homogeneous population help better account for the spikes caused by the reward pro-

6To improve readability, Table 4 does not report coefficients that are not 7I use simulation studies rather than posterior probabilities to describe
significant. However, when relevant, I mention such coefficients in the the segments in order to be consistent with the simulation studies described
text. Full results are available on request. in the “Policy Experiments” section.
288 JOURNAL OF MARKETING RESEARCH, AUGUST 2004

Table 4
TWO-SEGMENT MODEL RESULTS

Segment 1 Segment 2
Coefficient Standard Error Coefficient Standard Error
Small-Order Parameters
Intercept: βsm N.S. N.S.
Shipping policy 1: βh = 1,sm –1.02** .15 –.32** .07
Shipping policy 2: βh = 2,sm –.61* .26 –.28* .12
Shipping policy 4: βh = 4,sm N.S. N.S.
Price: βP,sm –1.90* .89 –1.38* .74
E-coupon: βC,sm N.S. .36** .11
FM1: βf = 1,sm N.S. .31* .13
FM2: βf = 2,sm N.S. .48** .08
Medium-Order Parameters
Intercept: βmed N.S. N.S.
Shipping policy 1: βh = 1,med –.95** .17 –.63** .24
Shipping policy 2: βh = 2,med –.62* .33 –.45* .24
Shipping policy 4: βh = 4,med N.S. N.S.
Price: βP,med –2.13* 1.14 –2.20* 1.18
E-coupon: βC,med N.S. .19* .11
FM1: βf = 1,med .94** .18 1.10** .34
FM2: βf = 2,med .91** .13 1.12** .18
Large-Order Parameters
Intercept: βlrg N.S. N.S.
Shipping policy 1: βh = 1,lrg .40** .05 1.02** .26
Shipping policy 2: βh = 2,lrg N.S. N.S.
Shipping policy 4: βh = 4,lrg .19* .09 N.S.
Price: βP,lrg –2.49** .76 –2.88** 1.15
E-coupon: βC,lrg .33** .12 .35* .17
FM1: βf = 1,lrg 1.59** .23 1.29* .55
FM2: βf = 2,lrg 1.90** .15 1.42** .29
Common Parameters
Reward: βL .75** .27 N.S.
Recency level 2: βk = 2 N.S. –.46** .07
Recency level 3: βk = 3 –.73** .08 –1.01** .11
Recency level 4: βk = 4 –1.73** .19 –1.84** .15
Price Expectation Model
Intercept 1: θLow N.S. N.S.
Intercept 2: θHigh N.S. N.S.
Price effect: θP –.65* .35 N.S.
Segment size: θseg1 .89** .11
*p < .10.
**p < .01.
Notes: I estimated the segment size parameter using a logit formulation such that π1 = exp(θseg1)/[1 + exp(θseg1)]. N.S. = not significant.

Table 5
SEGMENT DESCRIPTIONS

Segment 1 Segment 2 Simulated Population Actual Sample

Percentage of population 71% 29%
Order incidence rate 17% 27% 19% 20.1%
Annual orders 8.7 14 10.3 10.45
Annual spending $672 $586 $647 $637.28
Average order size $ 77 $ 42 $ 63 $ 60.98
Percentage earning reward 15% 1% 11% 19.7%

gram, and the heterogeneity controls help better reflect the and myopic homogeneous models are also notable.8 In the
variability in the population. dynamic model, the reward coefficient is positive and sig-
The segments are substantially different in terms of the nificant (.77), whereas the static model yields a smaller pos-
loyalty reward effects. Segment 1 yields a parameter of .75, itive estimate that is only marginally significant (.56). The
whereas the parameter for Segment 2 is just .22. In addition, static model yields a lower estimate for the reward parame-
the t-statistic for Segment 2’s response to the loyalty pro- ter because the specification does not consider the ability of
gram is just 1.3. To some extent, this is not surprising,
because the rewards are rarely relevant to members of Seg-
ment 2. The estimated reward parameters from the dynamic 8Full estimation results for the models are available on request.
Loyalty Programs and Customer Retention 289

Figure 1 category, –2.12 for the medium category, and –2.48 for the
COMPARISON OF SIMULATED POPULATIONS large category. In contrast, Segment 2’s coefficients are
–1.38, –2.20, and –2.88, respectively. These patterns are as I
expected, because higher prices have a larger absolute
25% impact on larger orders. In terms of segment differences,
the results suggest that higher prices tend to shift Segment 2
Percentage of Population

to smaller order sizes at a greater rate than Segment 1.

20%
Response to the e-mail coupons is similar for both seg-
ments. The e-mail coupon coefficients are of the expected
15% sign and are significant for the large and small categories.10
Expectations about the future pricing environment are
10%
similar for both segments. The coefficients related to the
expectations suggest that an increase in current-period
prices creates an expectation of higher prices in future peri-
5% ods. This result suggests that high prices can deter current-
period ordering and negatively influence retention. This is
0% important for the operation of a loyalty program, because
$– $500 $1,000 $1,500 $2,000 $2,500 expectations of high prices can reduce the attractiveness of
program participation. However, this conclusion should be
Annual Cumulative Spending
viewed with some caution, because the coefficients for the
Segment 1 Segment 2 expectation terms do not yield significant t-statistics. The
terms are included in the model because they significantly
improve overall fit.11
The time duration parameters are similar for both seg-
Table 6 ments. These variables indicate that the probability of a cus-
ALTERNATIVE FIT MEASURES tomer placing an order decreases as the time since the pre-
vious order increases. A difference is that the effect of
Two-Segment Homogeneous Homogeneous recency accelerates more quickly for Segment 2. The coef-
(Dynamic) (Dynamic) (Myopic) ficients for the constructed cumulative purchasing variables,
Absolute FM1 and FM2, are also of the expected sign and structure.
deviations 65 84 90 For both segments, I observe the same pattern of positive
Percentage signs for all terms and increasing parameters for the larger
earning rewards 10.7% 9.8% .6%
sizes. The results suggest that customers with higher levels
of cumulative purchasing buy more frequently and in larger
customers to delay rewards. For example, a customer who quantities.
has sufficient cumulative purchases for a reward with an In addition, although time (t) and cumulative purchasing
appropriate purchase in the next period may decide to wait (Cum) are not elements of the reward functions, they are
or to make a smaller purchase. The reward may seem not to meaningful elements of the state space. The importance of
be valued, but the customer may actually be planning to these factors for Segment 1 is illustrated in Figure 2, which
claim the reward in the future. gives the probability of a purchase as a function of time and
In addition to the loyalty rewards, the results indicate the cumulative spending and illustrates how the pull of the loy-
importance of several other factors. Shipping fees have sig- alty program varies by proximity to a reward threshold and
nificant effects on both order incidence and order size.9 The time constraints for three levels of cumulative spending
policy (h = 1) that provides free shipping for large orders ($500, $700, and $900).12 For the customer with $900
has a significant, positive impact on the percentage of large worth of cumulative spending, the probabilities rise contin-
orders for both segments. Conversely, this policy results in uously as the expiration time approaches; at the $700 level,
significantly fewer small and medium-sized orders. The probabilities first rise and then decline. As time approaches
estimation results also suggest that Policy 2 (increasing/ approximately 35 weeks, the probabilities increase for the
high) results in significantly fewer small and medium-sized $700 level and exceed the probabilities for the $900 level.
orders than does Policy 3 (increasing/low). Notably, the free However, as the remaining time becomes short, the proba-
shipping promotion (Policy 4) has only a minor effect com- bilities for this group plunge and are close to the probabili-
pared with the base policy (Policy 3) for the population ties for the $500 level by the end of the horizon. This occurs
under study. In terms of segment comparisons, Segment 1
tends to favor large orders, and Segment 2 favors small
orders.
The price coefficients are negative and significant for
10Note that for several of the pricing parameters (e.g., shipping, e-mail
each order size and reveal important segment differences.
coupon), the coefficients for the different order sizes often are not dramat-
Segment 1’s price coefficients are –1.89 for the small ically different.
11Restricting these terms to P
low = 1/3, Pave = 1/3, and Phigh = 1/3 yields
9Note that order-size intercepts are defined in terms of the third shipping a model with a log-likelihood of –16,628.31. A model comparison test
fee schedule. A chi-square test of restricting the shipping coefficients to results in a chi-square statistic of 37, which is significant at p < .001.
zero results in a t-statistic of 465.9 (18 degrees of freedom), which is sig- 12Probabilities are calculated with a fixed state for all variables with the
nificant at p < .001. exception of time and cumulative buying.
290 JOURNAL OF MARKETING RESEARCH, AUGUST 2004

Figure 2 Figure 3 illustrates the effects of the loyalty program in

PURCHASE PROBABILITIES AS A FUNCTION OF TIME AND more detail. The series coded as “Loyalty Program” shows
CUMULATIVE BUYING (SEGMENT 1) the predicted distribution of annual ordering for the popula-
tion when loyalty awards are available, and the series coded
as “No Loyalty Program” shows the prediction without loy-
alty rewards. With loyalty awards, a spike occurs at an
.45 annual level of purchasing of slightly more than $1,000.
Probability of Purchase

.4 This corresponds to customers who make sufficient pur-

chases to earn an award. In contrast, there is no incentive to
.35
reach purchasing targets without the loyalty program, so the
.3 model predicts a single modal distribution.
.25
In addition to the policy experiment of eliminating the
reward program, many other analyses are possible. Policy
.2 experiments provide a means for evaluating the long-term
.15 impact of short-term promotions such as e-mail coupons.
5 10 15 20 25 30 35 40 45 50 For a point of comparison, Table 8 reports the results of an
Week experiment that evaluates the effect of an incremental
e-mail coupon. This experiment suggests that an additional
$500 Cumulative spending
e-mail coupon has a slightly smaller overall impact than the
$700 Cumulative spending loyalty program. A notable contrast between the experi-
$900 Cumulative spending ments is that whereas the loyalty program substantially
affects only one segment, the e-mail coupon leads to more
buying by both segments.
The e-mail coupon experiment also highlights the
because the likelihood of earning a reward is similar for the model’s decision-support capabilities in that it captures
two levels ($500 and $700) when only 1 or 2 weeks remain. interactions between short-term and dynamic promotions.
POLICY EXPERIMENTS An additional coupon stimulates demand in two ways. First,
it effectively lowers prices and thus directly increases
An issue associated with dynamic programming models demand in the current period, but a more subtle benefit
of behavior is that their complexity makes it difficult to
interpret model coefficients. To understand the model’s
implications more fully, the results may be evaluated with Table 7
simulation or policy experiments. This section examines the LOYALTY PROGRAM IMPACT
effects of changing or eliminating the loyalty program. I
conducted the policy experiments using the estimated Scenario 1: Scenario 2:
parameters to simulate customer behavior over an extended Loyalty No Loyalty
Program Rewards
time period. I evaluated the consequences of eliminating the
program by comparing a simulation that uses the full Purchase incidence rate 19.7% 19.2%
dynamic finite-mixture model structure with a simulation Mean annual orders per customer 10.25 9.98
Average customer revenue $647 $634
that removes the incentives associated with cumulative
buying.13
I performed the simulations using the estimated coeffi- Figure 3
cients to compute the probabilities of each alternative (no POLICY EXPERIMENT: COMPARISON OF ANNUAL
purchase, small, medium, or large orders) for each possible
PURCHASING WITH AND WITHOUT A LOYALTY PROGRAM
customer state. I then simulated consumer purchases using
a random-number generator. For the purpose of the simula-
tions, a small order is valued at $33, a medium order at $67,
Percentage of Population

and a large order at $100. I then determined customers’ 18%

future states on the basis of the variable definitions and the 16%
14%
laws of motions described in the “Model Estimation” sec-
12%
tion. The results reported in this section are based on simu- 10%
lations of 100,000 customers who make 51 sequential deci- 8%
sions using a marketing policy (e.g., prices, e-mail coupons) 6%
that is similar to that employed by the firm. 4%
Table 7 presents summarized results from a simulation 2%
experiment that eliminates the loyalty program. The simula- 0%
tion suggests that removal of the loyalty program will $– $500 $1,000 $1,500 $2,000 $2,500
decrease the purchase incidence rate from 19.7% to 19.2% Annual Purchasing
per week. This change translates to a predicted drop in rev-
enue of approximately $13 per customer. No loyalty program
Loyalty program
13Specifically,
I removed the loyalty program by setting the coefficients
(βL) associated with the rewards equal to zero.
Loyalty Programs and Customer Retention 291

The main complication of modeling a program with a

Table 8 rolling time horizon is that it is necessary to retain the cus-
POLICY EXPERIMENT 2: ADDITIONAL E-MAIL COUPON
tomer’s purchase history for the length of the specified time
horizon. With a rolling structure, the customer’s relevant
cumulative purchases may change from period to period for
Scenario 2:
Additional two reasons. First, as mentioned previously, the customer’s
Scenario 1: E-mail investment in the program may increase in response to
Base Case Coupon current-period buying. Second, with a rolling time horizon,
Purchase incidence rate 19.7% 19.9% it is possible that investment in the program decreases
Mean annual orders per customer 10.25 10.37 through the expiration of previously earned points. From
Average customer revenue $647 $657 the perspective of the firm, customer transitions between
levels of cumulative purchasing could include both a proba-
bilistic element related to current-period buying decisions
occurs in an interaction between the short-term coupon pro- and a deterministic element related to the number of points
motion and the long-term loyalty promotion. Second, by that are about to expire.
inducing customers to purchase, the e-mail coupon has the A potential difficulty with this type of program is the
effect of increasing a customer’s cumulative buying or maintaining of computational tractability. The issue is the
investment in the reward program. As the customer moves degree to which previous purchase history must be main-
closer to attaining a reward, the pull of the loyalty program tained. Theoretically, it is necessary to maintain the cus-
is increased, which stimulates additional purchases in future tomer’s complete transaction history as elements of the cus-
periods.14 tomer’s state space. However, in practical terms, it is likely
LIMITATIONS AND FURTHER RESEARCH that most of these data could be approximated or summa-
rized, particularly because it is doubtful that customers
Although loyalty programs are a common marketing themselves have perfect memory of their transaction
instrument, empirical examinations of their effectiveness histories.
are limited (Drèze and Hoch 1998; Sharp and Sharp 1997). Related to the issue of computational tractability are
This is largely because the programs are difficult to evaluate questions about potential extensions and refinements of the
with standard techniques, because the rewards for cumula- model. For example, it may be useful to consider the inter-
tive purchasing can influence behavior over an extended action between a customer’s most recent purchase and
period. Although the dynamic programming method over- recency. This would potentially help disentangle inventory-
comes this difficulty, there are opportunities for further based inactivity from attrition-based inactivity. The diffi-
research. culty with this extension is that it requires the augmentation
Methodologically, there are opportunities to develop of the state space to include the customer’s previous order
techniques for programs with structures different from the size. Although this is a minor extension to the state space
one evaluated in this article. A salient characteristic of the (one variable with three levels), the end result is a tripling
program under study is that it has a strictly defined finite of the total state space. Another potential avenue of study is
horizon. The selection of a program’s time horizon is an interactions of the marketing variables with the customer’s
important element of loyalty program design. In contrast to level of cumulative purchasing. For example, although the
the program under study, many programs have rolling or model accounts for the effects of time and cumulative pur-
even indefinite time horizons.15 Moreover, programs may chasing, it is conceivable that these factors interact with
be designed as hybrids so that some elements have strict marketing-mix elements, such as price.
finite time horizons, and other elements do not include The modeling of shipping fees as categorical rather than
expirations. For example, airline programs often determine dollar values also merits discussion. There are potentially
a traveler’s classification (e.g., gold, silver) on the basis of significant benefits to using the dollar values of shipping
miles flown in a calendar year, but the miles earned from fees as covariates in the model. With dollar values, analyses
travel may expire after a few years or never. Reward pro- that quantify the effects of changes to such fees could be
grams with these types of time horizons can be evaluated performed. An understanding of the elasticities associated
with dynamic programming models, but they require adjust- with shipping fees is important, but it is left to further
ments to the solution approach and additional data. research for two reasons. First, the development of an
Programs with rolling or indefinite time horizons may be appropriate specification may not be straightforward,
evaluated by modeling the customer’s objective function as because the choice of a given order size may be a function
an infinite-horizon dynamic programming problem. of multiple shipping fees. As a case in point, consider the
Infinite-horizon problems can be solved with techniques customer response to the free shipping policy and the policy
such as value or policy iteration (Bertsekas 1996). Although offering free shipping for large orders. Although both
such techniques can be more computationally intensive than charge the same price for a large order, they result in differ-
the backward recursion used in the finite case, they are still ent average order sizes.16 Modeling of the impact of ship-
feasible for fairly large problems. ping fees requires a specification that includes direct and
relative shipping cost terms. Second, the use of categorical
14The results of an experiment that simultaneously removes the loyalty variables helps maintain computational feasibility. In the
program and adds an incremental e-mail promotion suggest that the aver-
age annual purchasing increases from $634 to $642.
15By “rolling,” I mean programs that involve points that expire accord- 16The average order size in the free shipping policy is $44; in Policy 1,
ing to the date that each point is acquired. the average size is $65.
292 JOURNAL OF MARKETING RESEARCH, AUGUST 2004

specification herein, the shipping fee schedules are repre- Drèze, Xavier and Stephen J. Hoch (1998), “Exploiting the
sented as four distinct states, and the use of dollar values Installed Base Using Cross-Merchandising and Category Desti-
would require three distinct variables with three levels. nation Programs,” International Journal of Research in Market-
Another area for further research pertains to the effective- ing, 15 (5), 459–71.
Eckstein, Zvi and Kenneth Wolpin (1989), “The Specification and
ness of alternative reward types. A limitation of the current
Estimation of Dynamic Stochastic Discrete Choice Models,”
data set is that rewards are limited to a single type. With Journal of Human Resources, 24 (4), 562–98.
appropriate data, it may be possible to measure the relative Erdem, Tülin, Susumu Imai, and Michael Keane (2003), “Brand
effectiveness of different rewards or the benefits of offering and Quantity Choice Dynamics Under Price Uncertainty,”
reward choices. Quantitative Marketing and Economics, 1 (1), 5–64.
——— and Michael Keane (1996), “Decision-Making Under
CONCLUSION Uncertainty: Capturing Dynamic Brand Choice Processes in
This article presents a dynamic programming model of Turbulent Consumer Markets,” Marketing Science, 15 (1),
1–20.
customer response to a loyalty program and other marketing Gönül, Füsun (1999), “Estimating Price Expectations in the OTC
tactics. The model measures the influence of rewards by Medicine Market: An Application of Dynamic Stochastic Dis-
considering customers’ sequences of purchases as a solution crete Choice Models to Scanner Panel Data,” Journal of Econo-
to a dynamic optimization problem. A primary strength of metrics, 89 (1–2), 41–56.
the approach is that the dynamic programming framework ——— and Mengze Shi (1998), “Optimal Mailing of Catalogs: A
provides a rich structure for modeling customer behavior. New Methodology Using Estimable Structural Dynamic Pro-
The model includes both the influence of previous behavior gramming,” Management Science, 44 (9), 1249–62.
in terms of cumulative purchases and forward-looking fac- ——— and Kannan Srinivasan (1996), “Estimating the Impact of
tors, such as expectations of future prices and loyalty Consumer Expectations of Coupons on Purchase Behavior: A
rewards. The inclusion of forward-looking behaviors is Dynamic Structural Model,” Marketing Science, 15 (3),
262–79.
computationally expensive but intuitively appealing for Kalwani, Manohar U., Chi Kin Yim, Heikki J. Rinne, and Yoshi
modeling customer response to a loyalty program. Given Sugita (1990), “A Price Expectations Model of Consumer
the prevalence of loyalty programs, this is a salient topic for Brand Choice,” Journal of Marketing Research, 27 (August),
researchers and practitioners. 251–62.
The model estimates the effects of e-mail coupons, pric- Kamakura, Wagner and Gary J. Russell (1989), “A Probalistic
ing changes, shipping fees, and the loyalty program. The Choice Model for Market Segmentation and Elasticity Struc-
sign and statistical significance of the parameter associated ture,” Journal of Marketing Research, 26 (November),
with the reward in the dynamic model is evidence that the 379–90.
loyalty program effectively increases repeat-purchase rates. Kim, Byung-Do, Mengze Shi, and Kannan Srinivasan (2001),
Further evidence of the program’s effectiveness is the rela- “Reward Programs and Tacit Collusion,” Marketing Science, 20
(2), 99–120.
tive fit of the dynamic model compared with a static formu- Kivetz, Ran and Itamar Simonson (2002), “Earning the Right to
lation in which customers do not consider future benefits. Indulge: Effort as a Determinant of Customer Preferences
The results should provide solace to advocates of loyalty Toward Frequency Program Rewards,” Journal of Marketing
programs. Research, 39 (May), 155–70.
The model also provides a platform for conducting what- Kopalle, Praveen and Scott A. Neslin (2003), “The Economic Via-
if analyses of customer retention. Simulation studies can be bility of Frequency Reward Programs in a Strategic Competitive
used to study the relative power of each marketing instru- Environment,” Review of Marketing Science, (accessed April
ment to increase long- and short-term purchasing. Such 16, 2004), [available at http://www.bepress.com/romsjournal/
studies are useful because the estimation results indicate vol1/iss1/art1].
that multiple instruments can stimulate repeat buying. Fur- Lemon, Katherine, Tiffany Barnett White, and Russell Winer
(2002), “Dynamic Customer Relationship Management,” Jour-
thermore, given appropriate cost information, it is a simple nal of Marketing, 66 (January), 1–14.
extension to use these experiments to estimate the prof- Narasimhan, C. (1989), “Incorporating Consumer Price Expecta-
itability of alternative policies for customer relationship tions into Diffusion Models,” Marketing Science, 8 (4), 343–57.
management. O’Brien, Louise and Charles Jones (1995), “Do Rewards Really
Create Loyalty?” Harvard Business Review, 73 (May–June),
REFERENCES 75–82.
Bellman, Richard (1957), Dynamic Programming. Princeton, NJ: Rust, John (1994), “Structural Estimation of Markov Decision
Princeton University Press. Processes,” in Handbook of Econometrics, Vol. 4, D. McFadden
Bertsekas, Dimitri (1996), Dynamic Programming and Optimal and R. Engle, eds. Amsterdam: North-Holland, 3081–3143.
Control. Belmont, MA: Athena Scientific. Sharp, Byron and Anne Sharp (1997), “Loyalty Programs and
Boulding, William, Ajay Kalra, Richard Staelin, and Valarie Zei- Their Impact on Repeat-Purchase Loyalty Patterns,” Interna-
thaml (1993), “A Dynamic Process Model of Service Quality: tional Journal of Research in Marketing, 14 (5), 473–86.
From Expectations to Behavioral Intentions,” Journal of Mar- Soman, Dilip (1998), “The Illusion of Delayed Incentives: Evalu-
keting Research, 30 (February), 7–27. ating Future Effort-Money Transactions,” Journal of Marketing
Bult, Jan Roelf and Tom Wansbeek (1995), “Optimal Selection for Research, 35 (November), 427–37.
Direct Mail,” Marketing Science, 14 (4), 378–95. Sun, Baohong, Scott A. Neslin, and Kannan Srinivasan (2004),
Deighton, John (2000), “Frequency Programs in Service Industries,” “Measuring the Impact of Promotions on Brand Switching
in Handbook of Services Marketing and Management, T. Swartz When Consumers Are Forward Looking,” Journal of Marketing
and D. Iacobucci, eds. London: Sage Publications, 401–408. Research, 40 (November), 389–405.
Dowling, Grahame R. and Mark Uncles (1997), “Do Customer Zhang, Z. John, Aradhna Krishna, and Sanjay K. Dhar (2000),
Loyalty Programs Really Work?” Sloan Management Review, “The Optimal Choice of Promotional Vehicles,” Management
38 (Summer), 71–82. Science, 46 (3), 348–62.