See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/312211457

Dynamically Managing a Profitable Email Marketing Program

Article in Journal of Marketing Research · December 2017

DOI: 10.1509/jmr.16.0210

CITATIONS READS
94 11,471

3 authors:

Xi Zhang V. Kumar
University of Toledo Georgia State University
3 PUBLICATIONS 291 CITATIONS 260 PUBLICATIONS 29,001 CITATIONS

SEE PROFILE SEE PROFILE

Koray Cosguner
Indiana University Bloomington
11 PUBLICATIONS 196 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Customer Engagement in Service, View project

Dynamic Pricing for New Products Using a Utility-Based Generalization of the Bass Diffusion Model View project

All content following this page was uploaded by Koray Cosguner on 16 February 2017.

The user has requested enhancement of the downloaded file.

 Dynamically Managing a Profitable Email Marketing Program

Xi (Alan) Zhang
V Kumar
Koray Cosguner

Xi (Alan) Zhang is an Assistant Professor in Marketing at College of Business and Innovation,

University of Toledo. Toledo, OH 43606; Tel: 419.530.5516; Email: [email protected].

V. Kumar is the Regents’ Professor, Richard and Susan Lenny Distinguished Chair & Professor
of Marketing, Executive Director of Center for Excellence in Brand & Customer Management,
and Director of the Ph.D. Program in Marketing at J. Mack Robinson College of Business,
Georgia State University. Atlanta, GA 30326; Chang Jiang Scholar, HUST, China, and Lee
Kong Chian Fellow, Singapore Management University, Singapore. Tel: 404.413.7590;
Email: [email protected].

Koray Cosguner is an Assistant Professor in Marketing at J. Mack Robinson College of

Business, Georgia State University. Atlanta, GA 30303; Tel: 404.413.7697; Email:
[email protected].

All authors contributed equally.

1
 Dynamically Managing a Profitable Email Marketing Program

ABSTRACT

Although email-marketing is highly profitable and widely used by marketers, it has received
limited attention in marketing literature. Extant research either focuses on customers’ email
responses or studies the “average” effect of emails on purchases. In this paper, we use data from a
U.S. home improvement retailer to study customers’ email open and purchase behaviors by using
a unified Hidden Markov and Copula framework. Contrary to conventional wisdom, we find that
email active customers are not necessarily active in purchases, and vice versa. Furthermore, we
find that the number of emails sent by the retailer has a non-linear effect on both the retailer’s
short- and long-term profitability. Through a counterfactual study, we provide a decision support
system to guide the retailer to make optimal email contact decisions. This study shows that sending
the right number of emails is vital for long-term profitability. For example, sending 4 (10) emails
instead of the optimal number of 7 emails can cause the retailer to lose 32% (16%) of its lifetime
profit per customer.

Keywords: Email marketing, Hidden Markov models, Dynamic programming, Copula models

2
 Introduction

Email marketing is a widely-used marketing tool by most B2B and B2C companies. A

survey study by Ascend2 (2015) shows that 82% of B2B and B2C companies use email technology

as part of their marketing strategies. In addition to its widespread usage, email marketing can be

quite profitable. According to a recent study published by the Direct Marketing Association (UK),

the average revenue-based return on investment (ROI) of email marketing increased to £38 for

each £1 spent compared to the average ROI of £24.93 reported in 2013 (DMA 2015). Based on

the same study, roughly 1 in every 5 companies report an ROI of more than £70 which is three

times higher than the figures from 2013. Without a doubt, firms aim to generate higher ROI levels

from email technology by launching effective email marketing programs. Although the firm’s goal

is evident, firms may find it difficult to manage effective email programs for multiple reasons.

First, customers who are active in opening a firm’s emails may not be active in purchasing

from the firm or vice versa. For example, some heavy buyers may not open the firm’s emails

frequently since they already know the firm and its offerings quite well. This implies that the firm

may not need to target such customers aggressively through email marketing since this type of

buyer may pay little to no attention to emails received. On the other hand, some light buyers might

be quite responsive in opening the firm’s emails since they seek to get information through emails.

Even though this type of customer opens emails more frequently, from the profit perspective,

targeting such customers might be suboptimal for the firm since the chances of converting such

customers to repeat-buyers are likely low.

Second, sending the right number of emails is critical for the firm’s profitability, especially

since most customers tend to complain about the large number of emails sent by firms. A survey

study by BlueHornet (2013), a marketing firm focusing on email solutions, finds that email

3
frequency might overtake content irrelevance as an important reason that drives customers away

from email marketing. Although sending the right number of emails is paramount, finding that

magic number is very challenging for a firm since not only do its customers have different intrinsic

preferences to emails (i.e., email open1 frequencies differ across customers), but customer

preferences might change dynamically over time (i.e., within same customer, email open

frequencies might differ over time). In other words, the right number of emails to send may differ

across customers and over time. Consistent with this argument, a study by Return Path (2015), an

industry expert company on email optimization, suggests that email frequency optimization should

depend on the engagement level of customers. However, the study does not provide an actionable

tool for firms to adopt to launch effective email marketing programs.

Third, managers tend to treat each email marketing campaign as an independent solicitation

process and fail to consider its long-term impact on the email open and purchase behaviors. In

addition to immediate impact, an email might also affect the future profitability of the firm by

shifting the customer-firm relationship level (Luo and Kumar 2013). For example, customers

opening emails might become more interested with the firm and its offerings, and ultimately, they

might start to purchase more frequently. In this case, ignoring the long-term effects of emails might

cause the firm to make suboptimal email marketing decisions.

Due to all these challenges, firms need guidance in the form of a decision support system

(DSS, hereafter) to decide the right number of emails to send to different customers over time. In

this study, our objective is to provide such a DSS to firms through 1) linking customers’ email

open behaviors with their purchase behaviors, 2) capturing heterogeneity in email open and

purchase behaviors across customers and over time, 3) incorporating the long-term effects of email

1
For this study, we define “email open” as the action of opening an email message.
4
marketing on firm profitability.

Methodologically, we propose a unified framework combining hidden Markov (HMM)

(Netzer, Lattin and Srinivasan 2008) and Copula models (Danaher and Smith 2011). In the hidden

Markov component of our model, we capture the dynamics in customers’ purchase and email

behaviors by allowing the latent relationship states to govern both behaviors. As later discussed,

modeling the evolution of purchase and email states jointly is important in the context of email

marketing since customers’ responsiveness in opening emails might also affect their relationship

with the firm. In this regard, our study differs from the other studies in the HMM domain which

only focus on the customer’s purchase behavior and ignores the fact that other non-purchase

activities may also affect the customer-firm relationship (except the study by Schweidel et al.

2014) level. In addition, customers’ email open and purchase behaviors might be correlated due to

time-invariant factors like customer unobserved heterogeneity and time-variant factors like the

unobserved customer-firm relationship states. In the Copula component of our model, we use a

bivariate Frank Copula to capture that correlation. We model the email open behavior using a

binomial distribution (BD, hereafter) and the purchase behavior using a zero-inflated negative

binomial distribution (ZINBD, hereafter). In this regard, the Copula component of our model also

differs from most Copula applications in marketing where Copulas are used to correlate variables

coming from continuous distribution families (except the study by Stephen and Galak (2012), in

which double Poisson variables are first converted into continuous ones and treated as continuous

in the estimation).

Substantively, we seek to answer the following questions:

(1) How the latent Markov states govern customers’ purchase and email open behaviors,

(2) Whether there is any correlation between customers’ email open and purchase behaviors,

5
(3) How email contacts (i.e., number of emails sent by a firm) affect customers’ purchase behavior

both in the short and long run, and

(4) What the optimal number of emails is to send to different customers over time to maximize the

long-term firm profit?

The empirical study identifies three customer-firm relationship states governing

customers’ and email open and purchase behaviors: Low Open – Medium Purchase, High Open –

Low Purchase and High Open – High Purchase states. In other words, email active customers may

not be active in purchases; and email inactive customers may still be relatively active in their

purchase behaviors. In addition, we find that the firm’s email contact policy has non-linear short-

and long-term effects on customers’ purchase and email open behaviors as well as on the retailer’s

profitability. Furthermore, from our Copula component, we identify a positive correlation between

the two customer behaviors. We next derive a dynamic email marketing resource allocation policy

using the estimates from our proposed hidden Markov and Copula model of customer purchase

and email open behaviors. This way, we provide a DSS to the firm (about the optimal number of

emails to send) to maximize its lifetime profit. Based on our DSS, we find that the optimal number

of emails to send not only differs across customers, but also differs within the same customer over

time significantly. In addition, we find that sending the right number of emails is very critical from

the firm’s profit perspective. For instance, sending 4 (10) rather than the optimal number of 7

emails per month causes the retailer to lose 32% (16%) of its lifetime profit.

To reiterate our contribution, to the best of our knowledge, this is the first empirical study

that jointly models customers’ email open and purchase behaviors over time. Methodologically,

our paper is the first study that combines the HMM and Copula approaches together in a unified

framework. Therefore, this research provides not only important implications for firms to

6
understand customers’ behavioral attitudes towards email marketing, but also an easy-to-

implement modeling framework to marketing researchers in the field. The proposed framework is

flexible and marketing researchers in the field can use it to study other non-purchase customer

behaviors (including but not limited to customers’ responses to other electronic and potentially

interactive communications) that are possibly correlated with purchase. The framework can

incorporate, across customers and over time, unobserved behavioral heterogeneity. Finally,

through our DSS, we provide substantive managerial guidelines to firms to implement an effective

email-marketing program. Similarly, our suggested DSS is also versatile and firms can use it to

manage marketing decisions such as direct mails, telephone and sales person contacts.

In the following sections, we first review the literature on email marketing and customer

relationship dynamics in email marketing. Second, we describe our data and present descriptive

statistics. Third, we discuss our modeling framework. Fourth, we present the empirical results.

Fifth, we derive the optimal email marketing contact policy for the firm. Sixth, we discuss the

robustness of our proposed model. Finally, we summarize our findings, and conclude with caveats

and directions for future research.

Literature Review

Email Marketing

There are several reasons for email marketing’s popularity. First, emails enable marketers

to send messages to their customers at very low cost. Chittenden and Rettie (2003) demonstrate

that the total cost for acquisition and retention campaigns of email to be $26,500 per 5,000

customers, as compared to that of direct mail at $69,600 per 5,000 customers. Second, email

messaging requires less preparation and execution time. Industry practice shows that an email

marketing campaign targeting 50,000 customers needs only 6 hours to prepare and run; while a

7
similar size direct mail campaign needs 17 days before it can reach a targeted customer’s mailbox.

Third, emails typically generate faster response and create opportunity for interactive

communication with customers. For example, a customer can respond to a email the moment

he/she receives it by clicking the hyperlinks that direct him/her to the sender firm’s website via

his/her computer or mobile device.

In this study, we focus on permission-based email marketing. Permission-based email

marketing requires marketers to seek the customers’ permission before sending them email

messages (Godin 1999; Kumar, Zhang and Luo 2014). This type of email marketing intends to

maintain a repeat purchase relationship with customers, rather than getting customers to buy one

time only. In line with this idea, previous research shows that email marketing has a positive effect

on customer loyalty. Tezinde et al. (2002) discover that email advertisements are useful to

customers by inducing them to visit the physical store. Merisavo and Raulas (2004) find that email

marketing can enhance customer attitudinal loyalty towards the brand. Their study shows that

customers tend to recommend the email messages to their friends if they find the messages

interesting and useful.

Although the overall influence of email marketing is positive, we argue that researchers

and practitioners should examine the customers’ response to email-based marketing messages

based on two perspectives. First, customers open and read an email simply to keep track of the

firm’s products and offerings. This behavior does not necessarily indicate that they are actively

looking for information to assist their purchase decisions. Bonfrer and Dreze (2009) study a series

of email marketing campaigns and propose a bivariate hazard model to predict when customers

open or click an email. Kumar, Zhang and Luo (2014) look at the total number of emails that are

opened and clicked and investigate their impact on the time the customers subscribe to the email

8
program. Second and more importantly, customers make purchases because of email marketing

messages. Sahni, Zou and Chintagunta (2016) analyze 70 randomized field-experiments and find

that email promotions not only increase customers’ average purchase spending during the

promotion window, but also carry over to a week later after the promotion expires. Kumar, Zhang

and Luo (2014) find that the average email open-rate has a positive effect on average purchase

spending while the effect of the average email click-rate is not significant.

Although there are studies that consider each of these two perspectives separately—email

open and purchase, it is surprising that there is no study investigating both behaviors together.

Bonfrer and Dreze (2009) do not consider the possible link between email open/click rate and

purchase behavior due to the limitation of their data. Kumar, Zhang and Luo (2014) capture the

“average” effect of the customer’s response to emails on purchase, but they do not consider the

dynamics and heterogeneity in both email open and purchase behaviors. Sahni, Zou and

Chintagunta (2016) conduct a post-hoc analysis of the experiments to show the aggregate-level

effects of emails on customer expenditure. They do not quantify how email open behavior affects

customer purchase nor do they consider the dynamic effects.

Customer Relationship in Email Marketing

Previous research shows that the customers’ relationship with the firm evolves over their

lifetime. Netzer et al. (2008) propose a hidden Markov model to study the transitions of customers

among their latent relationship states with the firm. Following Netzer et al. (2008), hidden Markov

models were widely used by marketing researchers across a wide range of marketing settings

(Montoya et al., 2010; Kumar et al., 2011; Li et al., 2011; Luo and Kumar, 2013). This stream of

research primarily examines the evolution of customer-firm relationships by only looking at the

purchase behavior, whereas we argue that other customer activities along with the purchase, such

9
as customers’ email open behavior, can also reveal useful information. Conceptually, we claim

that studying customers’ email open and purchase behaviors jointly is necessary for multiple

reasons. First, there is a possible correlation between the two behaviors. For example, customers

who are active in purchasing from a firm may also actively search and examine information about

the firm and its offerings by opening and reading emails. In contrast, customers who are inactive

in purchasing from the firm may not be interested in reading emails related to the firm either.

Second, the email open behavior might carry relevant information related to the level of customer-

firm relationship. In other words, for a customer, being active or passive in opening emails might

also affect her relationship level with the firm, and more importantly (through this effect on the

relationship level) affect her purchase behavior. For example, assume we have two customers at

the same relationship level with the firm and each make the same purchases in the current period.

If this was the case, the purchase only model would predict the same level for the customer-firm

relationship for the next period. However, if one of the customers is also active in opening emails,

due to her email activity, his/her customer-firm relationship level might be different compared to

the other customer. A purchase only model cannot capture this difference in the relationship level

of the two customers. Due to this reasoning, we characterize the customer-firm relationship by

jointly utilizing customer-level email open and purchase information.

To the best of our knowledge, Schweidel et al. (2014) is the only study that models the

customer-firm relationship by utilizing both customer-level purchase and non-purchase behaviors.

They study the dependence of two customer behaviors, i.e., digital purchase and digital posting,

and the evolution of the associated customer-firm relationship. One of the important findings is

there is a correlation between the latent attribution processes of the two activities, meaning a

customer’s involvement in one activity is informative of another. Methodologically, we differ from

10
Schweidel et al. (2014) since 1) in that we use a Hidden Markov model rather than a latent

changepoint framework; 2) we use a Copula model to capture the correlation between purchase

and email open behaviors rather than a multivariate choice framework. More importantly, our

research differs from Schweidel et al. (2014) in that we additionally provide a DSS that allows

firms to allocate resources by utilizing information from two or more customer-based activities.

Due to this, ours is the first empirical study modeling customer relationship over multiple customer

decisions and which provides a DSS to help the firm maximize its lifetime profit. We summarize

the characteristics of the aforementioned studies along with our study in Table 1.

Insert Table 1

To provide an effective DSS to guide a firm to make profitable email marketing decisions,

one needs a good predictive model (that will be an input to the DSS) of customer purchase and

email open behaviors. A good predictive model must accommodate the following three critical

factors. First, across customers, the intensity of purchase and email open behaviors might be

different. For example, some customers might regularly check emails (possibly to keep themselves

updated with new product offerings or recent promotional offers) but may not be active in

purchasing. On the other hand, some customers might be very active in purchases, but may not

open emails frequently (they possibly know the firm and its offerings quite well). Due to these

inherent differences in behavior across customers, capturing the heterogeneity in customers’ email

open and purchase behaviors is critical in predicting their behaviors.

Second, customers’ email responsiveness and purchase activeness may evolve over time

since customers have varying interests and needs at different times. For example, at the initial

period of joining the email program, customers could be extremely active in opening emails to

know more about the firm but their purchase activities may not catch up with their enthusiasm

11
with the emails. After spending some time and effort getting to know the firm, customers may

become familiar and favorable to the firm. Having that said, they might reduce the amount of

attention they pay to the emails but unsurprisingly increase their purchase activities. Therefore, it

is also critical for firms to understand such dynamics in both behaviors to predict the customer’s

email open and purchase levels. Finally, there is a possible correlation between customers’ email

open and purchase behaviors, and firms need to understand such dependence between the

behaviors to predict them accurately. Consequently, a good predictive model (of customers’

purchase and email behaviors) must incorporate 1) customer heterogeneity, 2) customer-

relationship dynamics; and 3) correlation among customers’ behaviors.

Our paper contributes to the customer relationship literature, by jointly studying customers’

email open and purchase behaviors in the same modeling framework. Methodologically, we

propose a unified Hidden Markov and Copula framework that can capture not only the correlation

between customer email open and purchase behaviors but also the evolution of the customer-firm

relationship states controlling these two behaviors. Our model is flexible that purchase activeness

does not necessarily align with email responsiveness. In other words, customers who are active in

purchases may not be active in opening emails, or vice versa. In addition, we account for

unobserved customer heterogeneity in both customer activities through our random coefficient

specification. After estimation of the proposed model, the estimated model becomes an input to

our dynamic programming model, in which we develop a DSS to guide the firm to make profitable

email marketing decisions. In the following sections, we first describe our dataset, and next, we

discuss our modeling framework.

Data Description

Our database comprises information from a home improvement retailer in the United States.

12
The retailer sells products and services from multiple categories, such as Kitchen, Plumbing,

Electrical, Flooring, Paint, and Outdoors. The dataset consists of information on the purchase

transactions made by the customers2, the number of emails sent by the firm to these customers,

and the customer’s email open histories. The average inter-purchase time is approximately 6

months but typically ranges from a couple of weeks to a few months. Because of the variety and

complexity of the products, the large number of categories, and the lack of transactional data at

the category level, we study the overall purchase behavior instead of the purchase behavior at the

category level.

To form a sample of data for model estimation, we randomly select a cohort of 200

customers who opted-in to the retailer’s email program in February 2007. These customers have

remained in the email program throughout the observation window and have received emails from

the firm continuously3. Thus, we have data comprising customers’ email open and purchase

activities over a period of 39 months. Although we do not have information on the content of the

emails, according to the management team of the focal firm, the type of content contained in these

emails varies significantly. The emails could be informative of new products or sales events,

persuasive featuring key benefits of the products, or a combination of the two types. In any given

email, the firm generally features products from multiple categories such as Kitchen, Lumber and

Flooring.

In Table 2, we report the descriptive statistics and the correlation matrix of the customer

(purchase and email open) and firm (email contacts) level behaviors. On average, the retailer sent

6.90 emails to its customers per month. The customers open 1.64 emails and make 0.69 purchases

2
Although customers have the option to buy online and offline, 95% of the observed purchase transactions are
offline.
3
We observe that less than 1% of the customers in our database opted-out from the firm’s email marketing program.
We only include the customers who have stayed in the email program during the observation window.
13
on average per month. Note that we only count unique email opens because most the emails are

only opened once, if opened. The correlation matrix shows that the number of purchases has a

weak correlation with the number of emails sent and the number of emails opened.

Insert Table 2

To demonstrate the complexity of customers’ purchase and email open behavior, we

randomly select three customers from our database, and plot their purchase count and email open

frequency over the observation period of 39 months (see Figure 1). Customer 1 is not active in

either purchase or opening emails. Customer 1 only makes purchases in three months and opens

emails in two months. The time elapsed between the first and second purchase is 21 months.

Customer 1 ceases to purchase or open emails after month 26. In comparison, Customer 2 has a

more active purchase behavior over the course of 39 months. However, Customer 2 is not equally

active in opening emails as he/she only opens emails in two months—month 10 and 11. In addition,

we find that Customer 2 decreases his/her purchase behavior after month 19, demonstrating a time-

variant purchase behavior. Customer 3 is moderately active in both purchase and opening emails.

We observe that the average inter-purchase time of Customer 3 is approximately 4 months.

Insert Figure 1 here

Figure 1 shows that customers’ purchase and email open behaviors are both heterogeneous

and time-variant, and that purchase and email open behaviors are not perfectly aligned with each

other. The observed pattern indicates that the two behaviors may possibly correlate with each other

as well. In the subsequent section of our modeling framework, we discuss how we model the

heterogeneity, dynamics and the correlation in purchase and email open behaviors.

In addition, to understand the process of customer purchase and email open frequency, we

plot the distributions of both behaviors (see Figure 2). The distribution of purchase count shows

14
that a discrete distribution such as a Poisson process may be able to capture the data generating

process. However, the mean (0.68) and the variance (2.64) of the purchase count variable suggest

over-dispersion, which violates the underlying assumption of Poisson distribution. Furthermore,

we observe an excess of zero purchases (71%) that can affect the estimation of the Poisson model.

To account for both the over-dispersion and excess of zeros, we use the zero-inflated negative

binomial distribution (ZINBD) to model the purchase count variable.

The distribution of email open count also suggests a discrete distribution. The (unique)

email open count is the number of unique emails that are opened conditional on the total number

of emails received in each month. We observe that the maximum number of the emails sent (per

customer and per month) in our data set is 20 emails. Therefore, the number of unique emails

available to a customer to open cannot exceed 20. To capture this process, we use the typical

binomial distribution (BD) which captures the number of success (email open) in a sequence of

events (receiving emails).

Insert Figure 2 here

Modeling Framework

In line with previous research, we use a hidden Markov model to identify customer-firm

relationship states and the transitions among these states. An HMM describes a Markov process

using discrete latent states. HMM is a stochastic model used to capture the transition between these

latent states and translate these states to observed behaviors. HMMs are widely used in the

marketing literature to study customer-firm relationships (e.g., Netzer el al. 2008; Montoya et al.

2010; Kumar et al. 2011; Luo and Kumar 2013). In the context of email marketing, we use a HMM

15
to study two observed customer behaviors jointly: purchase and email open4. In our setting, the

latent states of the Markov process translate into different levels of customer-firm relationships

that yield different purchase and email open activeness for customers. In addition, since there is a

possible correlation between customers’ purchase and email open behaviors, we capture the

correlation between the two behaviors through a Frank Copula function (see Figure 3 for the

graphical illustration of our proposed hidden Markov and Copula model of customer purchase and

email open behaviors).

Insert Figure 3 here

In the following section, we first discuss the primitives of our HMM. Second, we specify

our conditional (on latent states of the Markov process) email open and purchase models. Third,

we illustrate how we capture the correlation between state-dependent email open and purchase

behaviors through our Frank Copula component. Fourth, we discuss the model estimation. Finally,

we discuss the model identification.

Overview of the Model

Let be the number of emails customer opens at time5 t. Let be the number of

purchases customer makes at time t. For customer i, we model the sequence of observations

, ,…, , using a HMM characterized by (1) the initial

state distribution , (2) a sequence of transition probabilities , and (3) a vector of

4
In addition to our earlier discussion (in the Customer Relationship in Email Marketing part of our literature review
section) related to the conceptual need of joint modeling, we provide two pieces of empirical evidence in our Web
Appendix showing that the joint model is indeed an empirical requirement. First, we conduct a simulation study in
which the joint model (bivariate email open and purchase) is compared with the purchase only model (univariate
purchase). We find that the estimated (from a simulated dataset generated from the bivariate model) bivariate model
recovers the assumed parameters more closely and efficiently compared to the estimated (from the same simulated
dataset) univariate model. Second, we show that based on the actual data fit, the estimated (from the observed
empirical purchase data) univariate model predicts the purchase behavior by 16% worse than the estimated (from the
observed empirical purchase and email open data) bivariate model. See our Web Appendix for further discussion.
5
The unit of time is a month in our study. Therefore, we use time t and month t interchangeably throughout the paper.
16
probabilities that relate the latent states to the observed purchase and email open behaviors .

The Initial State Distribution

At any given month t, let s denote the level of customer i’s relationship. Let be the

probability that customer is initially in relationship state s, where 0 and ∑ 1,

where NS is the number of the latent Markov states. In this study, we assume that all customers

′
start at the lowest relationship state in the first month. Therefore, , ,…, ,

1,0, … ,0 .

The Markov Chain Transition Matrix

In our proposed HMM framework, we allow customers to transit to any relationship

state 1,2, … , . Following Kumar et al. (2011), we use a multinomial logit specification to

formulate this transition process. We define the transition matrix as

State at t
State at t-1 1 2 … NS
1 , → , → … , →
(1)
, → 2 , → , → … , →

⋮ ⋮ ⋮ ⋱ ⋮
NS , → , → … , →

is the probability that customer moves from state at time 1 to state ′ at time ,
, →

where 0 1, ∀ , ′, and ∑ 1. We specify the indirect transition utility

, → ′ ′ , → ′

of customer i for transitioning from the relationship state at period 1 to state ′ at time

( , → ′) as follows:

(2)
, → ′ , → ′ , → ′

where , → ′ and , → ′ denote the stochastic and deterministic components of the above

17
indirect utility, respectively. We assume that , → ′ for i=1,…,N, t=1,…,T, s, s’=1,…,NS are

distributed i.i.d. Gumbel with location 0 and scale 1. We operationalize the deterministic

component, , → ′, as follows:

, → ′ → ′ → ′ , (3)

where → ′ is the intrinsic utility of transitioning from relationship state to s’ at time t. ,

contains the following variables: I[ , 0, , 0, , , , ; where I[A] is the

indicator function that takes the value of 1 when event A occurs and the value of 0 otherwise,

, is the lagged email open count, , is the lagged purchase count, , is the number

of emails sent by the firm at time t-1. → ′ , → ′, , → ′, , → ′, , → ′ is the vector

of corresponding response coefficients. We normalize the deterministic utility for customer at

time to transition to the lowest state ( , → ) to be zero for the identification purpose, i.e.,

, → , → , where , → for i=1,…,N, t=1,…,T, s =1,…,NS are distributed i.i.d. Gumbel

with location 0 and scale 1. Therefore, the transition probability for customer i, transitioning from

state to ′ at time t ( , → ′ ) becomes the well-known multinomial Logit share function as seen

below.

exp , → ′
, → ′ (4)
1 ∑ exp , →

Next, we discuss our conditional email count (CEOM, hereafter) and purchase count models

(CPM, hereafter).

Conditional Email Open Count Model (CEOM)

We assume that the number of emails customer (in state s) opens at time follows a BD

with parameters and | , given by

18
 | | 1 | (5)

where | is the conditional probability that customer (in state s) opens an email in month . We

model | as a function of customers’ past email open behavior as follows.

exp | ln
| (6)
1 exp | ln

where | and capture the intrinsic utility of opening an email and the effect of duration

dependence, given state , respectively. We capture the duration dependence through time from

the last email open variable, i.e., . We use the natural logarithm of to capture the

diminishing effect. For identification purpose, we impose the following restrictions: 1) |

| exp ∆ | , where ∆ | is a parameter to estimate from the data, 2) is state-

invariant. These two restrictions guarantee that customers in a higher relationship state, all else

being equal, have a higher probability to open emails than those in a lower state. In addition, we

allow both | and to be customer-specific to control for the unobserved customer

heterogeneity. We assume that | and are distributed with the following normal

distributions.6

| |
(7)

where ~N 0, and ~N 0, .

Conditional Purchase Count Model (CPM)

6
We acknowledge that the random coefficient specification we implemented here might not fully capture all the cross-
sectional heterogeneity that might exist. Using customer-level fixed effects might better tease out the unobserved
customer-level heterogeneity. However, we believe that the use of random coefficient specification is a reasonable
compromise due its high parsimony over the customer fixed effect specification.
19
 Conditional on being in the relationship state at time , we assume that the number of

purchases that customer makes follows a ZINBD with parameters | , | and . For each

observation , ZINBD assumes that there are two data generating processes (based on whether

the outcome is equal to zero or is greater than zero) which are defined as:

|
| 1 | 1 0 (8)

|
1 | 1 1 0
! |

where is a dispersion parameter that is assumed not to depend on covariates. The conditional

mean and variance of the ZINBD are given as | 1 | | and |

1 | | 1 | | | ⁄ , respectively. | and | capture the conditional

zero-inflated probability and conditional expected purchase count for customer (in state s) at time

, respectively.

To account for the excess of no purchases, we model the conditional zero-inflation

component of the ZINBD | as

1
| (9)
1 exp | | ln

where | and | capture the intrinsic utility of making a purchase and the effect of duration

dependence, given state , respectively. We capture the duration dependence through time from

the last purchase variable, i.e., . We use the natural logarithm of to capture the diminishing

effect. To account for the unobserved heterogeneity, we allow | (for 0,1) to be customer-

specific. We assume that | (for 0,1) are normally distributed across customers as follows.

20
 | | , 0,1 (10)

where ~N 0, for 0,1.

We model the second component of the conditional purchase model, i.e., conditional

expected purchase count ( | ), as a function of the number of emails sent by the firm ( ), that

is given by

| exp | , | , | (11)

Conditional on state , | is the intrinsic propensity to make purchases, , | and , |

are the corresponding response parameters. Note that, unlike the CEOM, we do not impose any

identification restrictions on the parameters of the CPM. In other words, we do not make any

restrictions such as | | ⋯ | and/or | | ⋯ | . This

restriction free specification (in the CPM) allows our model to be flexible such that, for example,

customers in a more active email open state may be less likely to be active in purchases, or vice

versa. In other words, instead of pre-imposing restrictions such that email active customers must

be active on purchases, we let the data to speak about whether this is really the case.

To account for the unobserved heterogeneity, we allow | , , | and , | to be

customer-specific. We assume that | and , | (for 1,2) are normally distributed across

customers as follows:

| | ∆
(12)
, | , | ∆ , , 1,2

where ∆ ~N 0, and ∆ , ~N 0, ,
for k=1,2.

The Correlation between Purchase and Email Open Behavior

In each month, both the number of emails opened and purchases made by a customer might

21
possibly indicate the customer’s interest to and level of interactions with the firm. Thus, we argue

that there is a possible correlation between purchase and email open count distributions. Note that

both the purchase count and the email open count follow a discrete distribution. It is not

easy to find a bivariate distribution that can capture the correlation between the ZINBD and the

BD.

Danaher and Smith (2011) pioneer the use of Copula approach in marketing to link two

marginal distributions that are not from the same family. Copula models are used in various studies

to study multi-dimensional marketing problems (e.g., Park and Gupta 2012; Stephen and Galak

2012; Schweidel and Knox 2013; Kumar, Zhang and Luo 2014; Glady, Lemmens, and Croux

2015). However, most studies focus on using Copula models to correlate variables coming from

continuous distribution families. Stephen and Galak (2012) model multivariate count variables

using a double Poisson model, but they convert these discrete variables into continuous ones first

and use the Gaussian Copula to correlate these converted variables. In the context of this study,

we are dealing with two discrete variables that are characterized by two distinct distributions—

ZINBD and BD. It is uncertain whether Stephen and Galak (2012)’s approach will work with such

distinct discrete distributions. In addition, for efficiency, modeling the correlation between the two

discrete variables directly rather than converting them to continuous ones first and treating them

as continuous is more desirable.

To tackle the challenge of correlating discrete distributions, we turn to the literature of

mathematics and statistics. Based on Sklar’s theorem (1959), in a bivariate case as an example, the

cumulative distribution functions (cdfs) of any two variables can be connected using a Copula

function, and such Copula function is unique if the two variables are continuous. With two

continuous variables, the bivariate density can be derived from the partial derivatives of the chosen

22
Copula function. With two discrete variables, such as the bivariate count data in this study,

although we cannot rely on partial derivatives, we can still obtain the bivariate probability mass

function (pmf) using finite differences of the chosen Copula function. There are several statistical

applications of the idea to model bivariate count data. Lee (1999) develop a bivariate negative

binomial distribution to model rugby league scores using Frank Copula. Song (2000) develops a

multivariate dispersion model generated from Gaussian Copula. Nikoloulopoulos and Karlis

(2010) capture the association between the purchase counts of certain product categories.

Following Nikoloulopoulos and Karlis (2010) and others, we use a Copula to construct the

bivariate pmf of and as follows:

, , 1 ,
(13)
, 1 1 , 1

where ∙ is the Copula function, and are the distribution function of and ,

respectively. We use Frank Copula (e.g., Frank 1979; Genest 1987) in this context because of its

flexibility to capture the full range of correlation. The Frank Copula function is given by

1 1 1
, ; log 1 (14)
1

where and are the distribution functions (ZINBD and BD in our case) and is the Frank

Copula correlation parameter.

Endogeneity Correction

There is a possibility that the firm is targeting customers by sending different number of

emails based on their past purchase activities and inherent characteristics (that are unobserved to

the researcher). In other words, the email contact variable (i.e., EM) might be endogenous. To

solve this potential endogeneity problem in EM, we use a control function approach (Petrin and

23
Train 2010, the idea of which is like the approach in Villas-Boas and Winer (1999). This involves

running a first-stage linear regression of EM on instruments7, which include customer-level

purchase recency, purchase frequency (to capture customers’ past purchase activeness) and

customer level fixed effects (to capture every other customer-level characteristics that are possibly

observed by the firm (to do targeted emailing), but not the researcher). From the first-step

regression, we obtain an R2 of 48.5%. In addition, we find the coefficients of purchase recency and

frequency to be positive and statistically significant. In the second step, we incorporate the

residuals from the first-stage regression as additional control variables into the HMM transition

(Equation 3) and CPM (Equation 11) components of our proposed customer email open and

purchase model.

Model Estimation

There are four sets of parameters to be estimated from our model: 1) parameters of the the

transition matrix (Equation 3); 2) parameters of the CEOM (Equations (6-7); 3) parameters of the

CPM (Equations 9-12); 4) the correlation parameter of the Frank Copula function (Equation 14).

Following Netzer et al. (2008), we write the vector of the bivariate probability mass function as a

diagonal matrix . Given the proposed HMM structure, the likelihood function for a sequence

of observations , ,…, , can be expressed as

, ,…, ,

(15)
′

7
Since the variation in EM variable is large, we treat it as a continuous variable rather than discrete in our
application. In addition, we run a Poisson regression and find that residuals from that count regression and the linear
regression are highly correlated (r=0.97).
24
where is an 1 vector of ones.

Model Identification

Since all model components (transition matrix, CEOM and CPM) always enter the

likelihood calculation (see Equation 15) as a bundle, it is not possible to separately identify the

coefficients of a specific variable if that variable is used repeatedly across different equations.

Consequently, we choose and use a distinct set of variables in different components of our model.

The chosen variables represent customers’ past activities and the firm’s past and current marketing

activities; and their choice is in line with the CRM literature (e.g., Netzer et al. 2008; Kumar et al.

2011; Li et al. 2011). Specifically, we model customer i’s transition probability (from period t-1

to t) among the states of the Markov process as a function of interactions between customer i and

the firm at period t-1. We use , , , to capture the recent interaction of the customer i with

the firm, and , to capture the firm’s recent interaction with customer i. We model the

CEOM component as a function of the customer i’s time from the last email open variable, i.e.,

. We model the CPM component as a function of customer i’s time from the last purchase

variable ( ) and the firm’s email marketing contact policy ( ). Also, note that we use lagged

and contemporaneous levels of customer and firm level behavioral variables in the transition and

conditional behavior (CEOM and CPM) components of our model. Since we have sufficient within

variable variation (for these behavioral variables) over time, we would be able to separately

identify all three components of our model, i.e., the transition, CEOM and CPM.

We identify the parameters of each of the abovementioned component as follows. For the

transition component itself, we observe a significant within customer behavior (email open and

purchase) variation over time. This implies that customers’ relationship levels with the firm vary

significantly over time. This is the key for us to empirically identify parameters of the transition

25
component (i.e., the parameters in Equation 3). Within relationship levels, we also observe

customers behaving differently, and this helps us identify the unobserved customer heterogeneity

parameters (i.e., the parameters of the heterogeneity distributions in Equations 7, 10 and 12). For

the CEOM and CPM components, we observe a significant within time variation across email open

and purchase behaviors of customers. For example, at a given time, some customers are active in

one, some active in both behaviors, and some passive in both. In addition, within customers, the

intensity of the two behaviors significantly vary over time. These variations help us identify the

parameters of CEOM and CPM components (i.e., the parameters in Equations 6, 9 and 11). In the

next section, we discuss our empirical results.

Empirical Results

We estimate our HMM model using the maximum likelihood estimation (MLE) method.

We maximize the simulated sample log-likelihood to estimate our model parameters. We decide

on the number of HMM states based on the Bayesian information criterion (BIC). In our

application, we compare the performance of HMMs of up to 4 states (see Table 3). We find that

the HMM with 3 states provides the best fit to the data as it gives the lowest BIC value. Thus, we

choose the 3-state HMM for further analysis.

Insert Table 3 here

In the following section, we first discuss the parameter estimates from our proposed hidden

Markov and Copula Model of customer purchase and email open behaviors. Second, we illustrate

how the firm’s email contact decision would affect the customer-firm relationship states.

Specifically, we discuss how the firm’s email contact decision affects customers’ transitions

among different relationship states. Third, we discuss the distribution of customers across the three

customer-firm relationship states. Specifically, we illustrate the proportion of customers that

26
belong to different relationship states over time.

Parameter Estimates

Table 4 reports the parameter estimates for the 3-state HMM. In order to label the states in

terms of their purchase and email open behaviors, we calculate the predicted purchase and email

open counts (at the average observed levels of customer and firm level behavioral variables such

as LY, LO, EM etc.). We find those customers in the relationship states 1, 2 and 3 open 0.215,

2.745 and 2.764 emails (on average) per month, respectively. In terms of the purchase behavior,

we find those customers in the relationship states 1, 2 and 3 make 0.338, 0.148 and 1.219 purchases

(on average) per month, respectively. Based on that calculation, we label the three latent

relationship states, which govern the frequency of customers’ email open and purchase in each

month as “Low Open – Medium Purchase”, “High Open – Low Purchase”, and “High Open –

High Purchase”8. The results suggest that in State 3, the conventional wisdom, i.e., the more

customers make purchases, the more they open emails, holds. However, such conventional wisdom

does not hold in states 1 and 2. Specifically, email active customers might be inactive in purchases

(State 2), and email inactive customers may still be reasonably active in purchases (State 1). This

result underlines the importance of our assumption related to not restricting the parameters of the

purchase model along with the email open model (i.e., imposing the following restriction: higher

the state, higher the email open and higher the purchase levels).

Insert Table 4 here

Table 4 shows that customers in all relationship states show a negative purchase duration

dependence. The longer they have not made a purchase, the less frequently they make purchases.

8
For ease of presentation, we use purchase only labels, i.e., Low, Medium and High Purchase to refer states 2, 1 and
3, when we have a discussion related to purchases only. We use the full labels Low Open – Medium Purchase, High
Open – Low Purchase and High Open – High Purchase to refer states 1, 2 and 3, when we have a discussion related
to both purchase and email open behaviors.
27
The effect is strongest for the Low Purchase customers (-2.235), and it is the weakest for the High

Purchase customers (-1.294). Similarly, customers also show a negative duration dependence for

the email open behavior (-0.637), i.e., the longer they have not opened an email, the less frequently

they open emails. The results suggest that customers making purchases (opening emails) in the

previous month increases (decreases) the transition probability to the High Open – High Purchase

state. The results also suggest that, across all relationship states, the number of emails sent by the

firm initially increases, but then decreases (inverse U-shape) the short-term purchase count. In

other words, the email contact has a nonlinear effect on the purchase count. In addition, the peak-

point of the inverse U-shaped curve varies across different relationship states, i.e., customers

respond in diverse ways to email contacts depending on their relationship state. This implies that

there are different optimal numbers of emails to send to maximize the short-term purchase count

across the three relationship states.

Results suggest that two out of the five endogeneity correction parameters are significant.

In other words, it is important to control for the endogeneity of EM to be able to recover the short-

and long-term effects correctly of EM in customers’ purchase behaviors and transitions among the

relationship states. We also find significant heterogeneity in both the estimates of intercepts and

the response coefficients of the CEOM and CPM components of our proposed model.

Transition Probability Matrix

The transition matrix from the HMM model shows how customers evolve across different

relationship states. We calculate the transition probabilities of a “typical” customer using

Equations (1-4). We vary the number of emails received in the previous period and check the effect

of email contacts on the state transitions (see Table 5).

Insert Table 5 here

28
 When there is no email contact, the customers from Medium Purchase and Low Purchase

states do not tend to move, while those from High Purchase state tend to move down to a Medium

Purchase state or stay within the High Purchase state. Table 5 shows that email contact has diverse

effects on the transition probabilities of the customers who are from different purchase states. For

example, ten email contacts per month increases the likelihood that customers from the Low

Purchase State move up to the Medium Purchase state (from 4.53% to 21.42%). However, one

email contact per month only marginally increases such likelihood to 5.67%. For the customers

who are from the Medium Purchase state, email contact increases their probabilities of staying in

the same state. In summary, email contact has a positive long-run effect on either making

customers be more engaged with the firm if they were not so or making them stay engaged if they

have already been so. However, email contact has a negative long-run effect on the customers who

have already been in a good relationship with the firm. For the customers who are from the High

Purchase state, excessive email contacts move them down to lower purchase states. While the

likelihood of them staying in the High Purchase state is 46.62% in the no-email scenario, such

likelihood decreases significantly to 25.97% in the 10-emails scenario. This non-linear dependence

between the email contacts and the transition probabilities among the purchase states motivate us

for our policy simulation (in the next section) in which we determine the optimal number of emails

to send to maximize the firm’ lifetime profit per customer.

In Figure 4, we plot the average probabilities of customers residing in the three purchase

states over time. We calculate the state membership distribution of each customer using the

filtering approach (Montgomery et al. 2004; Netzer et al. 2008). Since we assume that each

customer starts from the lowest state, we drop the first five periods as initialization periods and

plot the states evolution over the rest of the 33 periods. We find that, on average, 74%, 12% and

29
14% of customers started in the Medium-, Low- and High-Purchase states, respectively. Over the

course of 33 months, the majority of the customers remained in the medium purchase state while

the rest moved up or down to different states. In addition to understanding the aggregate

distribution of customers across the different relationship states (Figure 4), the firm might use the

filtering approach to understand the state membership distribution at the customer-level. Then, the

firm might rely on this information to make targeted email contact decisions. We illustrate this

idea in our policy simulation section next.

Insert Figure 4 here

Optimal Email Marketing

The optimal marketing contact strategy is the focus of several earlier studies (e.g., Khan,

Lewis, and Singh 2009; Kumar et al. 2011; Li, Sun and Montgomery 2013). In this study, since

we find that sending more emails is not necessarily good for the firm to encourage its customers

to make more purchases, determining the optimal number of emails might provide substantive

profit gains for the firm. Therefore, our objective in this section is to provide a DSS that managers

can rely on to maximize the long-run firm profit by sending the right number of emails to their

customers over time.

In this setting, at any given time t, the firm has to make the decision of how many emails

to send. Given our estimated customer response function, the email contact decision has both short-

and long-term implications on the customer behavior. Short-term effect derives from the direct

effect of emails on customer purchase and email open behaviors at time t (see Equations 5 and 11).

Long-term effect comes from two sources: (1) customers’ purchase and email open behaviors at

time t affect the evolution of customer relationship state from time t to t+1 (see Equation 3); (2)

the email contacts at time t influence the relationship state transition from time t to t+1 (see

30
Equation 3). Due to these long-term effects, determining the optimal number of emails to send by

the firm requires us to solve a dynamic programming problem.

From the retailer’s perspective, the variable of interest is the number of times a customer

purchases per month from the store. Under the assumptions of constant purchase amount per each

purchase transaction and fixed gross margin for the retailer9, the purchase count is directly

translated into the firm’s profit (per customer and per transaction). For the firm’s dynamic

optimization problem, the payoff relevant state variables are: (1) the probabilities that the customer

exists in each of the customer-firm relationship states ( , , 1 ), (2) the time

since last purchase ( ), and (3) the time since last email open ( ). Therefore, the state vector at

time t becomes , , , . Following Kumar et al. (2011), we assume the timing of

the email contact decisions as follows. At the beginning of each month t, the firm predicts the

probability that the customer exists in each of the three relationship states, , and 1

. Next, based on the predicted , and , as well as and , the firm decides

how many emails to send to the customer. We use a multinomial logit share function to capture

the state membership probabilities , and and relate them to two parameters and

as follows:

exp
1 exp exp
exp (16)
1 exp exp
1

9
For confidentiality reasons, customer-level purchase amount is not observed in our dataset. However, based on our
communication with the managers of the firm, we are advised that the average purchase amount for each customer
transaction is around $80-$110. We are also advised that the average profit margin is around 18-22% across different
product categories. Thus, in our profit and lifetime value calculations, we assume an average purchase amount of $100
per customer, per transaction; and 20% profit margin for each customer purchase occasion. This assumption yields
$20 net profit per customer at each purchase occasion.
31
 Let , , , denote the state vector at time t. Time from last purchase and

open states, and , evolve based on whether the customer makes purchases and opens emails

at time t. If the customer makes purchases at time t, the corresponding state becomes 1, and

if he or she doesn’t make any purchases at time t, becomes 1. Similarly, if the

customer opens emails at time t, the corresponding state becomes 1, and if he or she doesn’t

open any emails, becomes 1. Since the purchase and email open processes of

customers are modeled with ZINBD and BD, the time from last purchase and open states, and

, evolves in a stochastic manner as follows

1, 0
1, 0
(17)
1, 0
1, 0

The evolution of the first two state variables , and , conditional on the firm’s

email contact decision are given as

, | 0, 0 , → 1,1,

0, 0 , → 0,1,

(18)

0, 0 , → 1,0,

0, 0 , → 0,0,

where , → 0, 0, is the transition function from Equation 4 which is

used to calculate the probability of transitioning customers from state at time t to s at time t+1

32
conditional on email contacts, purchase and email open indicators at time t.

At each time t conditional on the state vector , the objective of the firm is to determine

the optimal number of email contacts to maximize the discounted sum of expected future profits.

Under some regularity conditions this objective can be written in the following Bellman equation.

, | , (19)

where is the discount factor, . is the per period profit and the expectation is over all the future

states and actions of the firm.

In order to solve this dynamic optimization problem, we discretize the state space with 10

levels for each state dimension , , and yielding 10,000 state combinations. We use

the value iteration algorithm (Rust 1987) to find the optimal mappings of the firm’s email contacts

to our chosen state combinations. Due to the discretization of the first two state dimensions, the

value functions for the other points in the state space are computed via interpolation (Keane and

Wolpin 1994). After we calculate the vector of the optimal mapping of email contacts (through the

value iteration algorithm), we fit a multinomial logit model (MNL)10 to predict optimal email

contacts as a flexible function of the chosen state space. Next, we use this fitted MNL policy

function to predict the optimal number of emails to send for any state combinations out of the

chosen state space.

Based on the optimal mapping of email contacts to states from the value iteration algorithm,

we find that the optimal email contact number ranges from 5 to 14. We see a lot of heterogeneity

in the ranges of optimal number of emails sent based on customers’ relationship states. For

instance, if the firm has a belief that with more than 30% probability that the customer is in the

10
The choice of the MNL functional form for the policy function is because the optimal mapping of the email contacts
takes discrete and finite number of values (ranging between 5 and 14 emails to send as discussed in the next paragraph).
33
High Purchase state, the optimal number of emails to send ranges between 5 and 7 (based on the

value of the remaining state combinations). If the firm’s belief becomes such that the customer is

in the Medium Purchase state with more than 30% probability, then the optimal number of emails

to send ranges between 6 and 10. If the firm has a strong belief (more than 90% probability) that

the customer is in Low Purchase state, the optimal number of emails to send ranges between 12 to

14 emails. Whereas, if the opposite is the case, i.e., the firm has a strong belief that the customer

is not in the Low Purchase state, the optimal number of emails to send ranges between 5 and 7.

This heterogeneity in the optimal number of emails to send suggests that the firm should

target different customers (by sending different number of emails) based on their observed

behaviors (LY and LO) and its belief about their relationship states. To illustrate how the firm

might use our DSS to target its customers through email marketing, we randomly pick two

customers from our database. We then calculate their HMM state membership probabilities over

time through the filtering approach, and use our optimal policy function to set the number of email

contacts (based on their observed LY and LO levels in the data). We plot the optimal number of

email contacts over time for these two customers in Figure 5. As can be seen in Figure 5, the

optimal number of emails to send varies significantly not only across these two customers, but also

within the same customer over time. This exercise suggests that our proposed DSS can easily be

used as an actionable management tool by the firm to dynamically manage its email contact

decisions in a profitable manner.

Insert Figure 5 here

Next, we use our optimal policy function to simulate the firm’s email contact decisions

along with the customer purchase and email open responses over a long-time horizon for three

representative customers in Low, Medium and High Purchase states, respectively. This forward-

34
simulation exercise helps us understand the CLV (customer lifetime value) of customers differ

across three purchase states. Based on our assumptions (see Footnote 9), we find that CLVs of

customers in Low, Medium and High Purchase states as $1,333, $1,411, and $1,465, respectively.

This exercise also reveals that, in the steady-state, the optimal number of emails to send to each

customer becomes 7 emails per month. In addition, the discounted sum of lifetime purchase and

email open counts become 73 and 118 per customer, respectively.

Lastly, we conduct a what-if simulation study to measure how much profit the firm would

leave on the table if it deviates from the recovered optimal email policy function (see Figure 6).

We use the steady-state distribution as the starting state combinations and test the alternative

scenarios in which the firm deviates from sending the optimal number of 7 emails. We test the

scenarios where the firm sends 4, 5, 6, 8, 9, and 10 emails instead of 7. Figure 6 shows that sending

sub-optimal number of emails might cause the firm to lose significant amount of profit. For

instance, sending 4 (10) emails instead of the optimal level of 7 emails causes the firm to lose 32%

(16%) of its lifetime profit per customer. This result suggests that sending the right number of

emails is highly critical for the profitability of the firm’s email marketing program.

Insert Figure 6 here

Robustness Checks

To check whether our model and findings are robust to alternative model specifications,

we conduct four sets of robustness checks:

1) Using alternative distributions to model the conditional purchase counts,

2) Using different Copula functions to model the correlation between customers’ purchase and

email open behaviors,

3) Imposing HMM identification restriction on the CPM, and

35
4) Including holiday dummies as additional control variables in the CPM model.

As our first robustness check, we estimate three alternative model specifications where the

conditional purchase count is assumed to be distributed as Poisson (PD), Zero-Inflated Poisson

(ZINPD), and Negative Binomial (NBD), respectively. We find that the estimates from these three

alternative specifications look qualitatively similar to our CPM with ZINBD specification.

However, based on the BIC criteria, we determine that our proposed ZINBD specification

outperforms the alternative models (BIC 30,658.11 , BIC 33,440.23 , BIC

32,201.11 , BIC 30,994.36 ). Therefore, we assume the conditional purchase count is

distributed by ZINBD.

As our second robustness check, we estimate two alternative model specifications.

Specifically, we estimate the alternative models with Clayton (1978) and Gumbel (1960) Copula

functions, which have been widely used in the statistics literature. Although, all three

specifications yield qualitatively very similar estimates, based on the BIC criteria, we find that the

Frank Copula (BIC = 30,658.11) slightly outperforms Clayton (BIC = 30,676.25) and Gumbel

(BIC = 30,682.59) Copula specifications. Therefore, we use Frank Copula as our proposed Copula

specification, and use the Frank Copula specification in the rest of our analysis.

For identification purpose, we could have imposed HMM restrictions on the CPM instead

of the CEOM. However, the former restriction requires us to impose 6*(NS-1) restrictions,

including NS-1 intercepts and NS-1 response coefficients in the component, and NS-1

intercepts and 3*(NS-1) response coefficients in the component, while the latter one only

requires 2*(NS-1) restrictions (NS-1 intercepts and NS-1 response coefficients in the

component). Due to this, we prefer to impose restrictions on the CEOM rather than the CPM. As

our third robustness check, we estimate the alternative specification in which the identification

36
restriction is imposed on the conditional purchase model. Based on the BIC criteria, our proposed

model (BIC = 30658.11) outperforms that alternative specification (BIC = 31308.9). Therefore,

we use our proposed model with HMM restrictions on CEOM as our main model in the rest of the

analysis.

Finally, as our last robustness check, we incorporate the holiday dummies into our CPM

as additional control variables. Essentially, customers might be more active in their purchase

activities during the holidays. Since our data is at the monthly level, we define the months

corresponding to Memorial Day, Fourth of July, Labor Day, Thanksgiving and Christmas as

holiday months, and dummy coded this holiday variable. From the comparison of our proposed

model and the model with holiday dummies, we find that 1) all parameters look qualitatively very

similar under both specifications; 2) coefficients for the holiday dummies are statistically

insignificant; and 3) based on the BIC criteria our proposed model slightly outperforms the

alternative model (30658.11 versus 30661.85). As a result, we use our proposed model without

holiday dummies as our main model in the rest of the analysis. We discuss our empirical findings

next.

Conclusions, Limitations and Future Research

The email marketing programs are used extensively in various industries to engage with

customers. The general industry practice in measuring the effectiveness of an email-marketing

program is to examine customer responsiveness to emails such as email open rate. However, we

show that considering only the email open rate could be misleading. Our empirical study shows

that some of the very high email openers are the least active ones in their purchase behaviors. In

addition, our findings show that some email inactive customers are relatively active in their

purchase behaviors. Thus, if firms solely focus on the email open rate to allocate its resources, they

37
might potentially overlook a pool of customers who are inactive in responding to emails but are

relatively active in purchases.

To the best of our knowledge, substantively, this is the first empirical study that models the

customer’s email open and purchase behaviors jointly. In addition, methodologically, this is the

first study that combines the hidden Markov and Copula models in a unified framework. In our

HMM specification, we model the latent customer-firm relationship states that govern customers’

both purchase activeness and email responsiveness. In our Copula specification, we use a Frank

Copula to correlate the customer’s email open and purchase behaviors.

Note that, the purpose of this study is not diverting firms’ attention from email open rate.

This study shows that, on average, there is a positive correlation between email open and purchase

behaviors. Instead, we recommend firms to look at customers’ purchase behaviors in addition to

customers’ response rates to emails. If the goal is to maximize the long-term profitability, firms

should be informed about the optimal level of email contacts they should make to their customers.

Along these lines, we calculate the optimal email marketing contact policy by solving the firm’s

dynamic optimization problem. Eventually, we propose an implementable framework to study an

important substantive problem that can save firms millions of dollars.

In our specific application, although we use email open as the non-purchase customer

behavior, our framework is quite flexible that it is applicable to other customer-level non-purchase

behaviors. In addition, our DSS is designed to guide the firm about how many emails to send,

however, other firm-level marketing decisions might also be guided with the proposed structure

without significant issues. Furthermore, as more information that might affect the purchase and

non-purchase behaviors becomes available, one might incorporate that information through

constructing additional control variables entering purchase and non-purchase models. Among such

38
variables, the ones, which are relevant for the firm’s decision in the supply side, might be

incorporated into the operationalization of the proposed DSS. This will allow the firm to target

different customers at a given time, and the same customers over time differently based on

additionally observed customer level characteristics.

One of the limitations of this study is that we do not observe the content of the emails.

Therefore, we do not focus on the emailing strategy that intends to match with customers’

contemporary needs, such as cross- and up selling. Previous research has investigated this

important issue (e.g., Kumar, George and Pancras 2008; Li, Sun and Wilcox 2005). If the email

content data is available, future research might not only consider the effect of emailing content on

customers’ response to emails and purchase behavior, but also provide guidance to the firm related

to targeted email marketing based on personalized customization of the email content.

Another limitation of this study is that we do not observe the information from the

competitors. Customers’ lack of response to emails could simply be due to subscribing to many

email programs from different firms. Each email delivered to the customer’s inbox is a load of

information. Customers who are not capable of processing the information will be overwhelmed

and stop responding. If the firm is aware of its customers’ inbox activity, incorporating this

information into the study is imperative. However, due to the sensitivity of such information the

likelihood that the retailer can obtain this information is unlikely. Future studies could consider

conducting field experiments to understand how competing emails affect customers’ reactions to

the firm’s emails.

39
 REFERENCES

Ascend2 (2015), “Marketing Technology Strategy Survey Summary Report,” [available at

http://ascend2.com/home/wp-content/uploads/Marketing-Technology-Strategy-Survey-Summary-
Report-150813.pdf].

BlueHornet (2013), “One More Time: Email Frequency Chief Culprit in Unsubscribes,” [available
at http://www.marketingcharts.com/online/one-more-time-email-frequency-chief-culprit-in-
unsubscribes-27944/].

Bonfrer, André and Xavier Drèze (2009), “Real-Time Evaluation of E-mail Campaign
Performance,” Marketing Science, 28 (2), 251-263.

Chittenden, Lisa and Ruth Rettie (2003), “An Evaluation of E-mail Marketing and Factors
Affecting Response,” Journal of Targeting, Measurement and Analysis for Marketing, 11 (3),
203-217.

Clayton, D. G. (1978), “A Model for Association in Bivariate Life Tables and Its Application in
Epidemiological Studies of Familial Tendency in Chronic Disease Incidence,” Biometrika,
65 (1), 141-151.

Danaher, Peter J. and Michael S. Smith (2011), “Modeling Multivariate Distributions Using
Copulas: Applications in Marketing,” Marketing Science, 30 (1), 4-21.

Direct Marketing Association (UK) Ltd (2015), “National Client Email 2015,” [available at
http://dma.org.uk/uploads/ckeditor/National-client-email-2015.pdf].

Frank, M. J. (1979), “On the Simultaneous Associativity of F x, y and x y

F x, y ,”Aequationes Mathematicae, 19, 194-226.

Genest, Christian (1987), “Frank’s Family of Bivariate Distributions,” Biometrika, 74 (3), 549-
555.

Glady, Nicolas, Aurélie Lemmens, and Christophe Croux (2015), “Unveiling the Relationship
between the Transaction Timing, Spending and Dropout Behavior of
Customers,” International Journal of Research in Marketing, 32 (1), 78-93.

Godin, Seth (1999), Permission Marketing: Turning Strangers into Friends and Friends into
Consumers. New York: Simon & Schuster.

Gumbel, E. J. (1960), “Distributions des Valeurs Extremes en Plusieurs Dimensions,”

Publications de l’Institute de Statist´ıque de l’Universit´e de Paris, 9, 171–173.

Keane, Michael P. and Kenneth I. Wolpin (1994), “The Solution and Estimation of Discrete
Choice Dynamic Programming Models by Simulation and Interpolation: Monte Carlo
Evidence,” The Review of Economics and Statistics, 648-672.

40
Khan, Romana, Michael Lewis, and Vishal Singh (2009), “Dynamic Consumer Management and
the Value of One-to-One Marketing,” Marketing Science, 28 (6), 1063-1079.

Kumar, V., Morris George, and Joseph Pancras (2008), “Cross-buying in Retailing: Drivers and
Consequences,” Journal of Retailing, 84 (1), 15-27.

Kumar, V., Sriram S., Anita Luo, and Pradeep K. Chintaguta (2011), “Assessing the Effect of
Marketing Investments in a Business Marketing Context,” Marketing Science, 30 (5), 924-
940.

Kumar, V., Xi (Alan) Zhang, and Anita Luo (2014), “Modeling Consumer Opt-In and Opt-Out in
a Permission-Based Marketing Context,” Journal of Marketing Research, 51 (4), 403-419.

Lee, Alan (1999), “Applications: Modelling Rugby League Data via Bivariate Negative Binomial
Regression,” Australian & New Zealand Journal of Statistics, 41 (2), 141-152.

Li, Shibo, Baohong Sun, and Alan L. Montgomery (2011), “Cross-Selling the Right Product to the
Right Consumer at the Right Time,” Journal of Marketing Research, 48 (4), 683-770.

Li, Shibo, Baohong Sun, and Ronald T. Wilcox (2005), “Cross-selling Sequentially Ordered
Products: An Application to Consumer Banking Services,” Journal of Marketing Research,
42 (2), 233-239.

Luo, Anita and V. Kumar (2013), “Recovering Hidden Buyer-Seller Relationship States to
Measure the Return on Marketing Investment in Business-to-Business Markets,” Journal of
Marketing Research, 50 (1), 143-160.

Merisavo, Marko and Mika Raulas (2004), “The Impact of E-Mail Marketing on Brand Loyalty,”
Journal of Product & Brand Management, 13 (7), 498-505.

Montgomery, Alan, Shibo Li, Kannan Srinivasan, and John C. Liechty (2004), “Modeling Online
Browsing and Path Analysis using Clickstream Data,” Marketing Science, 23 (4), 579-595.

Montoya, Ricardo, Oded Netzer, and Kamel Jedidi (2010), “Dynamic Allocation of
Pharmaceutical Detailing and Sampling for Long-Term Profitability,” Marketing Science, 29
(5), 909-924.

Netzer, Oded, James M. Lattin, and V. Srinivasan (2008), “A Hidden Markov Model of Consumer
Relationship Dynamics,” Marketing Science, 27 (2), 185-204.

Nikoloulopoulos, Aristidis K. and Dimitris Karlis (2010), “Regression in a Copula Model for
Bivariate Count Data,” Journal of Applied Statistics, 37 (9), 1555-1568.

Park, Sungho and Sachin Gupta (2012), “Handling Endogenous Regressors by Joint Estimation
Using Copulas,” Marketing Science, 31 (4), 567-586.

41
Petrin, A., & Train, K. (2010). A control function approach to endogeneity in Consumer choice
models. Journal of marketing research, 47(1), 3-13.

Return Path (2015), “Frequency Matters: The Keys to Optimizing Email Send Frequency,”
[available at http://returnpath.com/wp-content/uploads/2015/06/RP-Frequency-Report-
FINAL.pdf].

Rust, John (1987), “Optimal Replacement of GMC Bus Engines: An Empirical Model of Harold
Zurcher,” Econometrica: Journal of the Econometric Society, 999-1033.

Sahni, Navdeep, Dan Zou, and Pradeep K. Chintagunta (2016), “Do Targeted Discount Offers
Serve as Advertising? Evidence from 70 Field Experiments,” Management Science,
forthcoming.

Schweidel, David A., Young-Hoon Park, and Zainab Jamal (2014), “A Multiactivity Latent
Attrition Model for Consumer Base Analysis,” Marketing Science, 33 (2), 273-286.

Schweidel, David A., and George Knox (2013), “Incorporating Direct Marketing Activity into
Latent Attrition Models,” Marketing Science, 32 (3), 471-487.

Sklar, A. (1959), “Fonctions de répartition à n dimensions et leurs marges,” Publications de

l’Institut de Statistique de L’Université de Paris, 8, 229–231.

Stephen, Andrew T., and Jeff Galak (2012), “The Effects of Traditional and Social Earned Media
on Sales: A Study of a Microlending Marketplace,” Journal of Marketing Research, 49 (5),
624-639.

Tezinde, Tito, Brett Smith, and Jamie Murphy (2002), “Getting Permission: Exploring Factors
Affecting Permission Marketing,” Journal of Interactive Marketing, 16 (4), 28-39.

Trivedi, Pravin K. and David M. Zimmer (2005), “Copula Modeling: An Introduction for
Practitioners,” Foundations Trends Econometrics, 1 (1), 1-111.

Villas-Boas, J. M., & Winer, R. S. (1999). Endogeneity in brand choice models. Management
Science, 45(10), 1324-1338.

Xue‐Kun Song, Peter (2000), “Multivariate Dispersion Models Generated from Gaussian
Copula,” Scandinavian Journal of Statistics, 27 (2): 305-320.

42
 Figure 1: Purchase and Email Open Count of Three Selected Customers
10
8
Montly Count

6
4 purchase count
2 email‐open count

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Months

Customer 1
10
8
Monthly Count

6
4 purchase count

2 email‐open count

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Months

Customer 2
10
8
Monthly Count

6
4 purchase count
2 email‐open count

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Months

Customer 3

43
 Figure 2: Distributions of Purchase and Email Open Count
Distribution of Purchase Count

Percentage

Purchase Count

Distribution of Email Open Count

Percentage

Email Open Count

44
Figure 3: Hidden Markov and Copula Model of Customer Purchase and Email Open
Behaviors

Figure 4: Distribution of Customers’ State Membership over Time

Note: On average, Low, Medium and High Purchase customers make 0.148,
0.338 and 1.219 purchases per month, respectively.

45
Figure 5: Within Customer Optimal Email Contacts Over Time
Customer 1
Customer 2

12
Optimal Number of Emails to Send

10
8
6

5 10 15 20 25 30 35

Months

Figure 6: Steady-State CLV versus Number of Email Contacts

1400
Steady-State Consumer Lifetime Value (in Dollars)

1300
1200
1100
1000

4 5 6 7 8 9 10

Number of Email Contacts

46
 Table 1: Comparison of Existing Studies
Type of Models Modeling Customer Modeling Customer
Type of Providing
Studies Dynamics Marginal Distribution Relationship over Relationship over
Data a DSS
Component Component Purchase Multiple Decisions
Netzer et al. (2008) Donation HMM Logit Yes No No
Montoya et al. (2010) Prescription HMM Binomial Yes No Yes
Kumar et al. (2011) B2B HMM Multivariate Tobit Yes No Yes
Li et al. (2011) Banking HMM Multivariate Probit Yes No Yes
Luo and Kumar (2013) B2B HMM Multivariate Tobit Yes No No
Schweidel et al. (2014) E-commerce Latent Changepoint Bivariate Choice Yes Yes No
Copula of Zero-Inflated
This Study Retailing HMM Negative Binomial and Yes Yes Yes
Binomial Distributions

Table 2: Descriptive Statistics & Correlation Matrix11

Descriptive Statistics Correlation Matrix
Mean Std. Dev. Lower 5% Upper 95% Number of Number of Number of
Purchases Emails Sent Emails Opened
Number of Purchases (per month, per customer) 0.69 1.63 0 4 1
Number of Emails Sent (per month, per customer) 6.90 4.91 0 15 0.013 1
Number of Emails Opened (per month, per customer) 1.64 3.16 0 10 0.050 0.425 1
Time from Last Purchase (in months) 6.33 7.16 1 22
Time from Last Email Open (in months) 5.45 7.23 1 22.05
Indicator of Open (per month, per customer) 0.40 0.49 0 1
Indicator of Purchase (per month, per customer) 0.29 0.45 0 1

Table 3: Selecting the Number of States

HMM States Log-Likelihood BIC
1 -17956.41 36064.73
2 -17152.83 34582.67
3 -15119.06 30658.11
4 -15098.16 30777.15

11
The monthly descriptive statistics are calculated from our data consisting N=200 randomly chosen customers (from the retailer’s customer database) who are
observed for a period of T=39 months.
47
 Table 4: Estimation Results for the 3-State Hidden Markov Model
Estimates Standard Error
Transition Matrix
Intercept for transition (State 1 to 2) -2.647*** 0.315
Intercept for transition (State 1 to 3) -1.432*** 0.248
***
Intercept for transition (State 2 to 2) 2.829 0.352
Intercept for transition (State 2 to 3) -3.815 4.554
*
Intercept for transition (State 3 to 2) -2.398 0.953
Intercept for transition (State 3 to 3) 0.061 0.274
I[Lagged Purchase > 0] on transition to State 2 0.173 0.218
*
I[Lagged Purchase > 0] on transition to State 3 0.363 0.150
I[Lagged Open > 0] on transition to State 2 0.422 0.376
I[Lagged Open > 0] on transition to State 3 -0.486* 0.216
Lag Email Sent on transition to State 2 -0.245** 0.088
Lag Email Sent on transition to State 3 -0.110 0.064
Lag Email Sent Square on transition to State 2 0.007 0.005
Lag Email Sent Square on transition to State 3 0.001 0.004
***
Lag Email Sent Residual on transition to State 2 0.144 0.038
Lag Email Sent Residual on transition to State 3 0.003 0.033
Email Open Frequency (Binomial)
Intercept State 1 -2.367*** 0.065
***
Intercept (additional State 2, exp) 1.105 0.016
Intercept (additional State 3, exp) -4.450*** 0.969
Time since last open (log) -0.637*** 0.067
***
Variance for the intercept 1.817 0.064
***
Variance for Time since last open 0.388 0.062
Conditional Purchase Frequency (ZINBD)
Purchase Count Equation
Intercept (State 1) -0.716*** 0.105
***
Intercept (State 2) -2.951 0.312
Intercept (State 3) -0.413 0.317
***
Email Sent (State 1) 0.196 0.031
Email Sent (State 2) 0.470*** 0.011
***
Email Sent (State 3) 0.321 0.091
***
Email Sent Square (State 1) -0.013 0.002
***
Email Sent Square (State 2) -0.020 0.001
Email Sent Square (State 3) -0.027*** 0.007
*
Email Sent Residual (State 1) -0.039 0.016
Email Sent Residual (State 2) 0.049 0.037
Email Sent Residual (State 3) -0.002 0.036
Variance for the intercept 1.021*** 0.049
Variance for Email Sent 0.020* 0.009

48
 Variance for Email Sent Square 0.000 0.001
Dispersion, exp 0.655*** 0.071
Excess of Zeros Equation
Intercept (State 1) 2.135*** 0.066
Intercept (State 2) 3.203*** 0.710
Intercept (State 3) 3.447*** 0.761
Time since last purchase (State 1) (log) -1.528*** 0.073
Time since last purchase (State 2) (log) -2.235*** 0.391
Time since last purchase (State 3) (log) -1.294*** 0.338
Variance for the intercept 2.008*** 0.208
Variance for Time since last purchase 0.168 0.091
Correlation (Email Open and Purchase)
Frank Copula correlation coefficient 0.324*** 0.067
Likelihood -15119.06
BIC 30658.11

Legends: * p-value<0.05, ** p-value<0.01, *** p-value<0.001

Labels of the Markov states – State 1: Low Open - Medium Purchase, State 2: High Open - Low
Purchase, State 3: High Open - High Purchase

Table 5 Transition Probability Matrix of the HMM

To Low Purchase To Medium Purchase To High Purchase
From Low Purchase
Without Email Contact 95.38% 4.53% 0.09%
With One Email Contact 94.22% 5.67% 0.10%
With Five Email Contacts 87.93% 11.92% 0.14%
With Ten Email Contacts 78.42% 21.42% 0.16%
From Medium Purchase
Without Email Contact 6.75% 76.53% 16.72%
With One Email Contact 5.49% 79.02% 15.48%
With Five Email Contacts 2.66% 86.20% 11.14%
With Ten Email Contacts 1.40% 91.27% 7.33%
From High Purchase
Without Email Contact 5.43% 47.95% 46.62%
With One Email Contact 4.55% 50.99% 44.46%
With Five Email Contacts 2.45% 61.93% 35.62%
With Ten Email Contacts 1.43% 72.60% 25.97%

49

View publication stats