Introduction

After completing the course, a student must demonstrate:

 Knowledge and understanding

- of how to formulate a relevant business problem based upon a customer data

- of how to construct a customer satisfaction survey and of how to plan and execute an experiment
(off-line or on-line)
- of statistical methods to answer questions related to acquisition, retention and development of
customers

 Skills

- to construct and carry out a customer satisfaction survey

- to plan and execute experiments either off-line or online
- to do customer segmentation based upon benefits or value
- to do perceptual mapping for product positioning
- to carry out a churn analysis
- to use different approaches to personalize offers for cross-selling / up-selling
- to analyse customer attitudinal and behavioral data of different types

 Competences

- To select, apply and reflect on a variety of statistical approaches to customer related business
problems
- To communicate the results of findings to management in an appropriate way.
 1. Course Organization

Aim of the course

Customer analytics – a use of a customer databases to enhance marketing productivity through more
effective acquisition, retention and development of relationships with customers.

Marketing productivity - Evaluate decisions made on economic terms (i.e. does it have an influence
on profits for a company?). does the plan you have for implementation of marketing activity, does it
have a better outcome in economic terms than an alternative.

Fundamental ways to improve profits when we talk about customers is (1) to be better at acquiring
customers, retaining them and/or developing relationship than the competition. (2) sell more to
existing customers, sell to new customers.

Reading material

https://blackboard.au.dk/bbcswebdav/pid-2955791-dt-content-rid-10792015_1/courses/BB-Cou-Hold-
41544/Customer_Analytics_460202U040_update%2003.03.pdf

most important ones:

Journal of Marketing; Journal of Marketing Research, Marketing Science, • Journal of Interactive

Marketing, Journal of Consumer Research.

Customer Data Sources

Sources that this course will focus on:

 Survey Data
 In-store and online sales data (transactions)
 Loyalty program data
 Web browsing data
 Experiments
 Social Networks data
 Text (Word of Mouth)

Survey data

Main advantage is that we get deep info about Disadvantages:

the customers:

 Customer values, attitudes and

 Limited customer group
behavioral intentions
 One-shot
 Customer lifestyle
 Have to be updated.
 Customer demographics
Questions that would be interesting to investigate using survey data:

 How big a fraction of total sales to a customer do we actuallu cover?

 Why does the customer prefer one product to another?
 Why did the customer churn?
Related to churn analysis is internal information investigation:

The information that we can collect from the customers (categories of databases in the company):

 Basic information: Name, address, ZIP code

 Demographic information (age, gender, number of people in a household)
 Psychographic information (values, interest, preferences, …)
 Derived information based upon transaction database (price sensitivity)

Addition info to include in the customer database:

Inactive customers -

 How were they initially acquitted?

 Why are they inactive?
The transaction database

Big Data = Transaction + Interactions + Observation

And you get a relatively small database (even though some companies think it is a very big database)

Transaction database and transaction history

What does it consists of?

 What type of product was purchased

 When did the customer make his last transaction? (R)
 How frequent is the customer purchasing? (F)
 How much was spent on the transaction? (M)
If we follow customers over time, we can get goal transaction history. The last three are the most
important and will be the focus of this course. R = recency F=frequency M= monetary value. RFM
are what is used over and over, that is the goal analysis.

An important distinction in transaction data (that will be used in future analysis) is the distinction
between contractual and non-contractual settings. Example:

These four different situations will be approached differently in terms of analysis.

Interesting question in a non-contractual and a contractual setting with respect to customer value

 Non contractual
- What is the probability that a customer is still alive?
- How many transactions can we expect in the future from a customer
 Contractual
Retention:
- What is the expected duration of a relationship?
Churn:
- How can we identify customers who are most likely to churn?
 Overall
- What is the value of the whole customer base?

How can we sell more to the existing customer and hence increase the value of our
customers?

 Transactions in baskets in binary format (basket analysis)

Market basket data can be represented in a binary format, where each row corresponds to a transaction
and each column corresponds to an item

This simplistic representation of real market basket data ignores certain important aspects of the data
such as the quantity of items sold or the price paid

 Transaction data used for a recommendation (based on user profile)

User based collaborative filtering. Bayesian networks might also be used for this.
 Web browsing data – clickstream data

 customer activity

 metadata about the product

 different states. All together

is a sequence. This is what we
analyse.

Sequential data:

In order to analyse it, we need to make the Markov assumption:

 The Markov property

xt is the state of the model at time t

Markov assumption: each state is dependent only on the previous one. Dependency given by a
conditional probability

Data from experiments – A/B testing

Concept is trivial

 Randomly split traffic between two (or more)

versions: A(control) and B(treatment)
 Collect metrics of interest (OEC)
 Analyze

Must run statistical tests to confirm differences are

not due to chance.Best scientific way to prove
causality, i.e. the changes in metrics are caused by changes introduced in the treatment(s).

User inference changes, enhancements to algorithms, changes to apps are tested.

Customer centricity

 To exploit customer heterogeneity (personalization)

 To develop better relationships with profitable customers
 To maximize the lifetime value for each individual customer to the firm, thereby increasing
firm profitability

From product-centric to customer-centric:

Customer Lifetime Value & Churn Management

2. Value-Based segmentation RFM and CLV

RFM analysis

This analysis is a way to obtain different segments. It is made of Recency, Frequency and Monetary
value, that comes from transactional data. It represents the classical way to segment customers. It is
the second most often way to analyse customer data by database marketers. It has a link to the value
of the customer.
How does it come about?

 Division of customers into groups defined

by R, F and M
 Very often divided into 5 groups (5
quintiles) for RFM. (treat different groups
differently)
 Sorted and maybe the quintile of customer
with the most recent purchases (R=1)
corresponds to purchase within the last week
Arguments behind RFM
Assumptions:
Customers who have a) Purchased recently b) Made more purchases c) made larger purchases
Are more likely to respond than other customers. (recency is the most important one to calculate if the
customer will respond)

The basic steps in RFM used for direct marketing

1. Select a random sample of customers
2. Target a mail campaign to these customers
3. Analyse customer response and the correlation to RFM criteria
4. Calibrate a response and the correlation to RFM criteria
5. Target the mail campaign to customers with the highest likelihood to respond based upon
their values of R, F and M.

Analyzing a sample

Rank all customers according to recency and divide into 5 quintiles. For customers in recency quintile
1: rank by frequency and divide into 5 quintiles. Repeat step 2 for the other recency quintiles.
For each of the 25 recency-frequency groups rank by monetary value and divide into 5 quintiles.
e.g. we have 40 000 customers. Dividing them into groups multiple times, leaves us with 125 groups.
This is called the sequential approach (the traditional way) and is used most often.
The advantage is that you get the same amount
of customers in each group.
The interpretation of the frequencies will not be
the same across the different groups. That is the
disadvantage of this approach. Same with
monetary, the interpretation will be different for
different groups.

Evaluation of the RFM model Lift and Gains

Concept of LIFT
If we look at a random 10% of the potential customers, we expect to get 10% likely responders. Can
we select 10% of the potential customers and get more than 10% of likely responders? If so, we
realize “lift”. This is a key goal in data mining.
The cumulative lift chart
1. x – axis gives population percentile
2. y – axis gives lift
3. the top 10% of the scorers are 3 times
more likely to respond than a random
10% would be

the curve is RFM model, the line is random

model.
This curve describes how good the model is.

Over all conclusion – the cumulative gains chart

1. x-axis gives population percentile

2. y-axis gives % captured responses
3. customers with top 10% of the scores
account for 30% of responders

this curve is called the banana model, the fatter

the banana the better. This is used to describe
differences between models. We make different
charts and then compare them. Look for the fatter
banana

RFM characteristics
Simple and transparent, but

 Coding of RFM arbitrary

 Few transactions/customer may deliver unstable results
 Nothing about what’s customer to make purchase
 Not forward looking and hence cannot predict changes in customer behavior over time

Steps in a RFM analysis

1. Calculate response-rates 7. Calculate profit

2. Sort according to response-rates 8. Calculate cum_profit
3. Calculate average response rate 9. Make plots to evaluate the model
4. Calculate lift  Lift charts
5. Calculate cum_response and  Gains charts
cum_customers 10. Make profit
6. Calculate cum_lift 11. plot and calculate ROI
Critical points

 Profitability depends on response rates, costs estimates and revenue potential

 Each one impacts profitability
 The number we use are just estimates
 If we are off by a little here and there, our profit estimates could be off by a lot
 In addition, the same group of customers is chosen for multiple campaigns

CLV – customer lifetime value

- “The net present value of the profits linked to a specific customer once the customer has been
acquired after subtracting incremental costs associated with marketing, selling, production and
servicing over the customer’s lifetime” (HANS JØRN JUHL, lecture 2)

Individual level and aggregate level CLV

Individual level: Aggregate level

 Transaction data from each customer  Data from a segment (cohort) of

 Help firms to personalize strategies to customers or all customer
the profitability of the individual  Help firms in evaluating the overall
customer effectiveness of the marketing plan

Individual CLV
Contractual and non-contractual setting

The contractual setting

See excel for this

 Calculation of average CLV for a
customer in a contractual setting

Simplifying assumptions:

 Margins are constant over time and all cash flows take place at the beginning (or end) of the
period
 Estimated by (total gross margin in a period)/(no. of customers)
 Retention rate is constant over time and estimated based upon a cohorte
 Discount rate is constant
 Length of projection is infinite
 Only customers who are doing business with the firm
 Lost for good model

An important Distinction

 Expected lifetime value as-yet-to-be-acquired customer

 Expected lifetime value a just=acquired customer
 Expected residual lifetime value of an existing customer, E

Gupta & Lehmanns formula

Choose the type of case based on the

payment received from the customer.
Case 1 is usually used when we are
talking about acquiring customers

Application of the formula (Gupta & Lehmanns)

 Economics of improving retention rate (Maximum costs)

 Economics of acquiring customers (Maximum costs)
 Mergers and acquisitions (Value of the customer base)
  Difference in profitability among various market segments

Investment in increasing retention rates (Gupta et al. 2003)

Margin multiple:

If there is an increase in retention rate from 0.6 to

0.8, while the discount rate is 0.12, we can see the
change in margin:

2.5/1.15=1.17

If we want to increase retention rate, the discount

rate increases too. In this example we see that the customer is left with a value twise as big. It is a
good idea to check the increase in retention rates.

g = if we are able to sell more to customers

If margins increase by a constant rate g per period

then:

You can calculate how much the customer value increases based on these tables and these formulas.

3. CLV continuing

Customer-based valuation of a company

If we can:

 Estimate long-term value of a customer

 Forecast growth in number of customer
Then:
  We can estimate current and future value of a customer base as the sum of the value of
current and future customers. It can be seen as an estimate of the value of the firm

Model several cohorts:

calculating the n0 values. Who left and who came in.

In general the formula will look like this in discrete time:

c= cost (this can be derived from financial statements),

everything else will be given.

Modeling number of customers

It is assumed that number of customers are taking S curve from the graphs. The s curve can be
expressed:

 Cumulative number of customers:

 Number of customers acquired:

CLV in non-contractual settings

The challenge is that we cannot estimate any model of S(t) directly from the data. Migration models
are used to calculate non-contractual settings.

The migration model:

 Always a share model
 Customers don’t necessarily purchase in each period
 Customers are said to be in a given state dependent upon their purchase history
Preliminary work before calculation

Defining the discrete states and the number of states

 Recency is normally used as criterion for the

definition of a state, assuming that the
probability of another purchase(p) in a given
period depends upon recency.
Let us assume 4 states, where the last state is a so-called absorption state

 It defines the possible transitions

 From the absorption state you can not return to one of the other states.
Markov property:
Purchasing will not depend upon the path the customer took to reach the current recency state but only
on his present state.
Assumptions
Purchases are made at most once a period
Periods have equal length. The company uses a per-period discount rate
For each period a customer is considered to be active the company will spend an amount M in order to
remarket to the customer
At the end of each period a customer may end up in a given number of states.

A Matrix formulation (migration model)

Input to the model:

Critical parameters:

 NC: net contribution

 M: marketing costs per period per active customer
 Pj: probability the customer purchases in this period given that he is in state j (assumed to be
constant over time)

A formal description of the migration model

Vector G – contributions from different customers belonging in different states.
Profit contribution from a given state.

The expected profit after 1 period. Multiple with probabilities.

Discounted expected profit after T periods

Discounted expected profit after infinite periods. Easier formula, and

gives similar enough results.

Transition probability matrix (4 recency states):

Calculation of P:

The example of these formulas:

Marketing cost is 0 (M=0)

231.08 – the value of a customer starting out in a
recency 1. Someone who just bought smth.
If a customer starts in recency 2 then its value will be
57,29 and so on

4. Customer base analysis in a contractual setting

Contractual setting

Continuous time

Discrete time
Survival rate and retention rate

Main points to take away (Fader & Hardie, 2010):

Remember to take heterogeneity (not the same retention rates) into account when you calculate the
value of the customers and the value of the whole customer base
Although a given customer has an unchanging unobserving retention rate from period to period, the
overall retention rate in the customer base may increase over time.

Why do retention rates in a customer database increase over time?

Individual-time dynamics (e.g. increasing loyalty as the customer gains more experience with the
firm).
A sorting effect in a heterogeneous population. (the customers with low retention rates will leave
earlier than the customers with high and then over time, the average retention rate will increase)

How to project customer retention

“Future retention rates for a specific cohort of customers in a discrete time contractual setting should
take customer heterogeneity into account”

 The S-BG model (shifted beta geometric model)

Assumptions:
An individual remains a customer for the firm with constant unobserved retention probability (1-0).
ϴ - the churn rate. (1- ϴ mean that the customer will survive)
 probability that the customer will leave in period t
 survival function
Heterogeneity in ϴ follows a Beta distribution with pdf

B is the beta function

Beta distributions

Very flexible distribution, because can reflect

many different compositions of customers
(therefore, heterogeneity is good to assume)

Fundamental property of the S-BG model

Retention rates will be increasing over time due to the fact that high-churners drop out early in the
period. Here is the model:

There are no variables that varies through time. any analyst wishing to do so would need to forecast.
We assume that this basic model refers also to future periods (therefore this model can be used for
forecasting). Also, it is assumed that the consumer will follow geometric distribution and
heterogeneity will follow the basic distribution.
(for visualization check R code)
You can evaluate through graphs (r code) or check results:

Distribution of clv across regular customers

We assume customers buy on average for 100

If we look at the regular customer, we will expect that 37,4% of the customers will only be there for 1
year. 37.4% has a CLV equal to 100

Based upon the estimated

parameters we can calculate:
 With this type of model, you can get a
good description of our customer base. All
you need is alpha and beta.

(The last bar is a mistake)

Calculation of CLV for a customer base – residual lifetime value

Including not only the period had, but also

the future values too.

Proactive churn Management

When we have a contractual setting, it is good idea to check churn.

Churn rate is an important KPI for companies operating in a contractual setting. A churn model that
predicts a customer’s likelihood of churning (voluntary).
The model could be a

 Logistic regression model

 Decision trees
 Neural networks
 Support vector machines
 Ensemble methods (random forests, bagging, boosting etc)
How to select a model?
Typically:

 Evaluated on a validation sample on its top decile lift

 Proportion of actual churners in the 10% of customers that the model predicts to (have the
highest risk of churning
(Hansen, 2015)
• “ The business objective is never to predict the churners it is to reduce the value of and rate of churn.
For that predicting the churner is simply a first step to taking action to dissuade the potential churner.
That may sound like two distinct steps but do not waste time improving identification of churners,
focus on identifying those that can be dissuaded.”
(Ascarza, 2016)
The firm should target those customers with the highest sensitivity to the intervention
Economics of a retention campaign

 things to think about when thinking

about churn analysis.

5. Customer Base Analysis In A Non-Contractual Setting

Probabilistic models for the analysis of transaction data

Building a probabilistic model (CTD)

1) Determine the marketing decision problem/information needed
2) Identify the observable individual-level behavior of interest. (We denote this by x. It is
usually a timing problem[when is a customer leaving] or a counting that we want to consider).
3) Select a probability distribution that characterizes this individual-level behavior
This is denoted by f(x|ϴ)
We view the parameters of this distribution as individual-level latent traits.
4) Specify a distribution to characterize the distribution of the latent trait variable(s) across the
population
We denote this by g(ϴ)
This is often called the mixing distribution
5) Derive the corresponding aggregate or observed distribution for the behavior of interest

6) Estimate the parameters (of mixing distribution) by fitting the aggregate distribution to the
observed data
7) Use the model to solve the marketing decision problem/provide the required information
Individual level – Timing
A Bernoulli trial is a probabilistic experiment in which there are two possible outcomes, “success” (or
1) and “failure” (or 0), where ϴ is the probability of success.
Repeated Bernoulli trials lead to the geometric and binomial distributions.

Individual level – Building Blocks

Let the random variable x be the number of independent and identically distributed Bernoulli trials
required until the first success: X is a (shifted) geometric random variable, where

The (shifted) geometric distribution can be used to model discrete-time timing data.

Capturing Heterogeneity in Latent Trails

The beta distribution:

B(α,β) is the beta function, which can be expressed in terms of gamma functions:

The beta distribution is a flexible distribution … and is mathematically convenient

Illustrative Beta Density Functions

Steps to the S-BG model

This is the basic level of

probabilistic model.

Non-Contractual Settings – continuous

Continuous – many transactions during a period. The most typical situation in practice.
Stylized pattern of purchasing over time by a cohort of customers in a non-contractual setting and
latent attrition

Customers disappear over time. it happens at a different point in time. they might leave forever. We
assume that customers buy at a steady rate (same expected number of transactions per period) and
then they will leave under unobserved time. these are the two assumptions that we make.

Questions from a manager

 How many of the customers can be viewed as being active?

 Which customers are most likely to be inactive?
 How many repeat transactions can be expected in the next periods by those customers listed
in the database, both individually and collectively?

The link to the value of a specific customer

x- nr of repeat transactions, tx-
recency, T – the period that we have
observed the customer.
DERT: Discounted Expected Residual Transactions
“Classes” of models

 Timing ---- when?

 Counting ----- How many?
The toolbox contains “simple” models for each behavioral process.
More complex behavioral phenomena can be captured by combining models from each of these
processes and it leads to what is called “integrated models”.

Individual level - Counting (number of transactions per period x)

 Individual-level building blocks

Count data arise from asking the question, “How many?”. As such, they are non-negative integers
with no upper limit.
Let the random variable X be a count variable:
X is distributed Poisson with mean λ if
E(X)= λ

The biggest problem is that the Poisson distribution is memoryless. In both models for the continuous
non-contractual settings. Regardless of this, the model has been proven to work.

 Capturing Heterogeneity in Latent Traits

The gamma distribution

- Γ(·) is the gamma function – because of mathematical convenience we use this distribution.
(flexible)
- r is the “shape” parameter and α is the “scale” parameter
- The gamma distribution is a flexible (unimodal) distribution … and is mathematically
convenient.

 Illustrative Gamma Density Functions

The Aggregate Distribution – The Observed Distribution

If we combine the two mixing distribution and the assumption of the individual level then we get the
aggregate distribution.
Counting negative binomial distribution

 Negative Binomial Distribution (NBD)

The individual-level behavior of interest can be characterized by the Poisson distribution when the
mean λ is known.
We do not observe an individual’s λ but assume it is distributed across the population according to a
gamma distribution
X=x : number of customers with x transactions
per period. That will be the only thing we will
need to know to calculate this.

[The first milestone paper in the study of buyer behavior was “The Pattern of Consumer Purchases”
(Ehrenberg 1959) which showed that the Negative Binomial Distribution (NBD) provided a very good
description of the numbers of purchases of individual brands of consumer goods in a single time
period. From ”Ehrenberg’s contribution to social and marketing research” ]

the average transaction rate.

Calculating the probability and the expected number of customers having zero purchases
Summary of probability models and integrated models

The difference in assumptions between Pareto/NBD and BG/NBD

  this is better because it is
more flexible

BG/NBD

Formulas important for lifetime-value at the individual level

Data material for estimation of BG/NBD or PARETO/NBD

Necessary data:
We need to know
a) Recency (time of last transaction (t(x)) )
b) Frequency ( (x) number od repeat transactions in a time interval (0,T) )
c) Total time between first purchase and the end of the observation window (T)
We don’t need to know
a) Information on when each of the x transactions occurred

Basic input and output from the integrated probabilistic models for a continuous non-contractual
setting:

PARETO/NBD

Application area

 Non-contractual setting where the opportunities for transactions are continuous as customers
can make as many transactions as they want and any time in each period
 Attrition is unobserved. The company does not know if a silent period means that the
customer has left the company
Main questions to answer with the model

 The probability that a customer is alive at a given moment

 The expected number of repeat transactions for a customer in a future period
 CLV for a given customer
Assumptions:

 As long as a user is alive, the number of transactions follows a Poisson process with
transaction rate λ. Thus, if we left X(t) be a random variable which denotes the number the
probability of observing x repeat transactions in time interval (0,t) is:
The assumption also implies that the time between repeat transactions follows an exponential
distribution with transaction rate λ:

Where tj is the time of the jth repeat transaction

 Although the number of repeat transaction for each user follow a Poisson process, the
heterogeneity in the users is captured in the heterogeneity in the transaction rate λ across
users, which follows a gamma distribution with the probability density function

Where r is the shape parameter and α is the scale parameter. From the gamma distribution we can find
the mean transaction rate as E(λ)= r / α and the variance of the transaction rate as V(λ)= r / α2

 About timing. For modeling lifetime, the model assumes that each user has an unobserved
“lifetime” of length τ, after which the user is inactive. This unobserved lifetime has an
exponential distribution with a dropout rate μ:

From the exponential distribution, the mean of the lifetime is thus

The heterogeneity is the users regarding lifetime is captured in the heterogeneity in the dropout rate μ
across users, which follows a gamma distribution with the probability density function:

Where s is the shape parameter and β is the scale parameter. From the gamma distribution we can find
the mean dropout rate as E(μ)= s / β and the variance of V(μ)= s / β2

 Finally, we assume that the transaction rate λ and the dropout rate μ vary independently
across users. This assumption is debatable. We would e.g. expect that a user with a high mean
number of repeat transaction is a week, a large transaction rate λ, also has a longer expected
lifetime, a lower dropout rate μ, as the user is most likely an active and engaged customer.

6. BTYD (buy til you die)

Questions from a manager

 How many of the customers can be viewed as being active?

 Which customers are most likely to be inactive?
 How many transactions can be expected next period by those customers listed in the database,
both individually and collectively?
 How big is the CLV for a given customer?
Answers based on PARETO/NBD

 Number of active customers ( Sum of P(alive| Information) )

  Which individuals on this list most likely represent active customers? ( Ranking of P(alive|
Information) )
 What level of transactions should be expected next year by those on the list, both individually
and collectively?
 Can also be predicted by the model CLV=DERT*margin
See R_code for further elaboration and practical exercises.

Summary

RFM CLV

 Based upon transactional history

 Based upon transactional history
 Level of recency and frequency reflect
 Forward looking
tendency to buy again and hence value
in an implicit way

 Not forward looking and hence cannot

predict changes in customer behavior
over time

CLV contractual

 Naïve approach (average). The expression depends upon when the company receives their
payments
 Heterogeneity. The retention rate and hence the residual lifetime value depends upon the
number of renewals a customer has made. S-BG model
CLV non-contractual
Discrete Continuous
Migration models: Pareto/NBD or BG/NBD

 The tendency buy depends upon the recency

 Up to one transaction per period
 Customer satisfaction and loyalty
7. Concepts and models of CS, NPS

Why customer satisfaction and customer loyalty?

Customer Satisfaction have a leading indicator to revenue and profits.

Contagion perspectives

 Satisfied customers buying adjacent offerings and further motivating others to purchase
through word of-mouth
Affective-state perspectives

 Satisfied customers developing positive affinities and enhanced product and brand loyalties
leading to future purchases
Market-free perspectives

 Earned reputation for satisfying customers creating barriers to entry for non-incumbents
leading to market share gains and additional scale effects

Customer loyalty (retention) as a leading indicator for profits

m – profit margin, i – discount rate, r – proportion of retained
customers (retention rate)

we can see that the affect is exponential, if you can increase the loyalty expressed by retention rate,
then value of the customer will increase in an exponential way.
The link between satisfaction and loyalty

It is not a linear
relationship.
More realistic is
the s shaped
curved.

The link from satisfaction to the economic effects 0 different performance measure

 Accounting measures or performance (profit, net operating cash flows)

 Market measures of financial performance (stock price)
 Marketing measures of performance (market share and revenue growth)

Definition

 Satisfaction –
Is the customer’s fulfillment response. It is a judgement that a product or service feature, or the
product or service itself, provided (or is providing) a pleasurable level of consumption related to
fulfillment, including levels of under- or over fulfillment (Oliver, 1996)
Satisfaction as the post consumption consumer judgement of whether the good or service provided a
pleasurable level of overall usage-related fulfillment (Oliver, 2020)
Satisfaction is the post-consumption evaluation of a product and service (Mittal)
Micro or Macro:
The expectancy-disconfirmation model of satisfaction

There are two ways that we can measure expectations and performance and compare them two each
other. The two ways:
Direct and inferred approach

Criticism of the inferred approach

1) It supposes that everyone has precise expectations before the service experience
2) If the customer has little information about the service, it may be problematic
3) Works better with repurchases because the expectations are more precise
4) Expectations may mean different things to different people (different standards)
5) Calculation of P-E may be misleading. Zone of indifference may explain why customers with
unmet expectations may report satisfaction.
6) Expectations are typically rated higher than performance

The expectation standard

Levels of desire:

 Can be
 Will be
 Must be
 Should or ought to be
It is very important that when you ask your questions you make a
standard expectation definition otherwise the answers will be a mess
Servqual
- Model that is developed to compare expectations and perceptions. The idea is to subtract
expectation from perception across all levels (gaps).

Dimensions of service quality

 Reliability  Responsiveness
 Perform promised service dependably and  Willingness to help customers promptly
accurately  Example: avoid keeping customers
 Example: receive mail at the same time waiting for no apparent reason
each day  Quick recovery, if service failure occurs

 Assurance  Empathy
 Ability to convey trust and confidence  Ability to be approachable, caring,
 Give a feeling that customers’ best interest understanding and relating with customer
is in your heart needs
 Example: being polite and showing respect  Example: being a good listener
for customer

 Tangibles
 Physical facilities and facilitating goods
 Example: cleanliness

Purpose
To develop a scale for measuring perceived service quality based on a disconfirmation approach
P-E (P – perception (what a specific service provider offers), E – expectation (what the customer
would expect from ‘excellent’ service) )
Example – context effect

A telephone company

In general P-E lead to negative scores. And that’s why inferred approach is not preferable.

Overall perceived service quality

A typical satisfaction study importance – performance analysis

A number of question where features are evaluated one by one on a scale of importance and on
performance scale. (better than expectations. In practice its opt to ask for importance)
ALTERNATIVE approach often used in practice to obtain important scores
1) Have all features evaluated on a performance scale
2) An addition overall satisfaction measure for the whole product/service
3) Regression analysis
National index satisfaction models

There are a lot of companies that occupy with this. The main purpose of these national index
satisfaction models is to measure customer satisfaction at (1)company level (2) section level (3)
national level
Demands of these models:

 Should be standardized, sensitive and flexible

 Measure a report satisfaction as an index of multiple measures (several questions)
 Link satisfaction to firm’s controllables or inputs
 Estimate the consequences of an increase or decrease in satisfaction
How customer satisfaction (CS) and customer loyalty (CL) are defines by national index satisfaction
models
Three indicators for satisfaction:
1) Overall satisfaction
- Considering all your experience with company X, how satisfied are you in general
- Completely dissatisfied to complete satisfied
2) Live up to expectations
- To what degree did company X fulfill your expectations?
- Much less than expected, much more than expected?
3) The ideal company
- Imagine a company that is perfect in all aspects. How close to this ideal do you
consider company X to be?
- Very far away to very close
Two indicators for loyalty:
1) Will chose the company next time
2) Will recommend the
company

Swedish CS model
This model is problematic because it assumes some causal relationships. (e.g. that CS leads to CL.
These should be measured at different points in time).
The next problem is with customer complains is that probably not a lot of customers complained and
therefore there are too few obs. For complains.
A third problem is with customer expectations. All other things are determined by the model (there
are no arrows to it) meaning if you want to improve your model you need to improve customer
expectations outside the model.
This model is called structural equation model.

ACSI model

The only new thing is the perceived

value.
The way perceived value is structured
here is problematic, since it is hard to
define perceived value in one side of
the model…

EPSI model

This is better than the

others. Cause it helps
explain what happens when
you change the input, how
the satisfaction/loyalty
changes with input

Epsi is European. And is based upon 250 interviews per company

 Used first of all for benchmarking

 Used to describe variation in CS over time (Leads to conservatism, because you can compare
between the years)
Definition

 Customer loyalty –
A deeply held commitment to re-bui or re-patronize a preferred product or service consistently in the
future.

 four types of loyalty

High/Low in terms of happiness with the good or
service

Net Promoter Score (NPS)

In 2003, Reichheld published an article in HBR, in which he claims his Net Promoter Score, is the
only number you need to grow, and the only number you need to manage customer loyalty.
A lot of companies use this score (Microsoft and others)
Reichheld’s claim: By substituting a single question for the complex black box of typical customer
satisfaction survey, companies can actually put consumer survey results to use and focus employees
on the task of stimulating growth (2003)
The single question answered by all respondents is “How likely is that you would recommend
Company X to a colleague or a friend? 0 – 10”
The calculation of NPS
Calculate the percentage of customers who respond with nine or ten (promoters) and the percentage
who respond with zero through six (detractors).
Subtract the percentage of detractors from the percentage of promoters to arrive at your net-promoter
score.
(check NPS in R)
Evaluation criteria

 Is the measure valid?

−Do we actually measure what we claim to measure?
 Is the measure reliable?
−Do we measure with the best possible precision?
 Is the measure robust?
−Is the measure invariant to changing conditions?
 Is the measure part of a structural system?
−Does the measure give guidance concerning where to improve?
−Is the measure part of a causal system?
 Is the measure relevant?
−Does the measure have a strong relation to the business result?

An attempt to evaluate validity

We compare the outcomes of the NPS question to a question of whether the customer is still a
customer a year from now. This is measured on a 10-point scale, where 1=”no, absolutely not” and
10=”yes, absolutely”. This question of stated loyalty has proved to be very good indicator of actual
loyalty in the financial sector.
All values for both scales have been transformed into 0 or 100 scale.
Conclusions

 The main argument for measuring satisfaction and loyalty is that they are leading indicators
for business results
 It is important to distinguish between transaction specific and cumulative satisfaction
 It is important to distinguish between attitudinal loyalty and behavioral loyalty (retention)
 The national index models are often used models as a starting point for a customer
satisfaction survey
 The main advantage of national index models is that they provide possibilities for
benchmarking and measuring of developments over time
 The NPS is a very popular indicator in business because it is simple and easy to communicate
 The academic literature does not include many studies supporting NPS as an indicator for
revenue growth.
 8. Survey

- A tool for understanding the workings of the human mind and the dynamics of social
interaction. Survey is one of the sources or primary data.
Type of data we can analyse
1. Customer identification data
2. Demographic data
3. Psychographic or lifestyle data (survey data)
4. Transaction data
5. Marketing action data
6. Other type of data (e.g. financial, competitive data)

Sources of customer information

1. Internal secondary data
2. External secondary data
3. Primary data – survey, focus groups, in depth-interviews, observational techniques (the most
costly, time consuming)

Innovation in survey designs

Main issues:
Q1: to what extent do you think it is possible to ask consumers what they want and need in the future
and to get meaningful responses? (A: sometimes people do not know what they want and need)
Q2: Even if consumers do spend time to think the issues through and give a response, how accurate
will it be? (A: what a person wants now can be different from a week from now. There’s an issue of
reliability. Did the respondent was truthful?)
One possible way to improve reliability is through incorporation of experiments.
Surveys – attitudinal data.
Customer satisfaction survey – CSS

Sampling and response rate

 Representative sampling is essential to permit generalization from a sample to a population

 To maximize representativeness, representative sampling need to be complemented by:
a) High response rate (a real problem for surveys)
b) Statistical weighting procedures (Krosnick, p539)
 Evidence:
a) Young and old adults, males and people with the highest income levels are underrepresented
b) People with the lowest education levels are overrepresented
c) People who do not participate in surveys are likely to live in big cities and work long hours
Sampling techniques
 we don’t want non-
probability sampling.
Therefore we choose
probability sampling
techniques. (random)

Probability sampling methods:

Representativeness
A sample will be nearly perfectly representative of a population if a probability sampling methods is
used and if the response rate is 100%
Representatives increases monotonically with increasing response rate. NOT TRUE. Surveys with
low response rates can be more accurate than surveys with much higher response rates (Visser et al.,
1996)
e.g. telephone surveys would forecast election outcomes more accurately by accepting lower response
rates, rather than aggressively pursuing high response rates (Greenwald et al.)
Who are the respondents?
People who agree to be interviewed are likely to believe that:

 It is their social responsibility to participate in surveys

 They could influence government and the world around them
 Are likely to be happy with their lives
(Cialdini et al.)
Conclusion 1

 Aggressively pursuing high response rates may lead to inaccurate results

 Inherent limitations of nonprobability sampling methods
 When probability sampling methods are used, it is no longer sensible to argue that lower
response rates necessarily signal lower representativeness

Pretesting

Interviewers conduct with a small number of interviews

Objective: to identify questions requiring further explanation or wording that was confusing
(1) Conventional pretesting
Limitations:
- Debriefings are relatively unstructured
- Interviewer’s opinions may be subjective

(2) Modern pretesting

Behavior coding – an observer monitors pretest interviews. He or she notes events that occur during
interactions between interviewers and respondents that constitute deviations from the script
Cognitive pretesting – which involves asking respondents to think aloud while answering questions,
verbalizing whatever comes to mind as they formulate responses

Conclusion 2
The three pretesting methods differ in terms of the kinds of problems they detect and in the reliability
with which they detect these problems:

 Conventional testing and cognitive interviews tend to exhibit low reliability

 Behavior coding is the most consistent in detecting apparent respondent difficulties and
interviewer problems
Rigid vs Conversational interview
Principle: the same questionnaire should be administered identically to all respondents
Q: if respondents’ express uncertainty and ask for help, interviewers avoid interference by saying
something like “it means whatever it means to you” is this correct? In terms of data quality?
A: No, in fact, when interviewers were free to clarify the meanings of the questions and response
choices, the validity of reports increased substantially (Schober and Conrad, 1997). The shift to
computer assisted and self-administered interview proved to be problematic in light of Schober and
Conrad (1997) evidence that conversational interviewing can significantly improve data quality.
Conclusion 3
Evidence shows that conversational interviewing can significantly improve data quality

Questionnaire design

Open vs closed questions?

 The quality of data collected by each method seemed equivalent in some studies
 Recent studies found that the reliability and validity of open-ended questions exceeded that of
closed ended questions
Closed questions (disadv.)

 Respondents tend to confine their answers to the choices offered

 The close questions need to be comprehensive (include all possible answers) – difficult to
assure
Open questions (disadv.)

 Some responds feel difficulty when expressing their thoughts and feelings
 Respondent’s laziness and lack of time
Conclusion 4
Open-ended questions seem to be more viable research tools in some studies while closed-ended
questions are more viable in other studies. No clear black and white answer.

Category labels and intervals between categories

 When there are only numeric labels for the categories between the verbally labeled endpoints,
respondents probably assume the categories are equidistant
 Providing verbal labels for all the categories, both endpoints and intermediate ones, produce
more reliable measurement (not al all, slightly, somewhat, pretty, very, extremely)
“very” which commonly appears as an anchor is not intense enough fo most applications
Conclusion 5
Reliability and validity can be significantly improved if all points on the scale are labeled with words,
because they clarify the meanings of the scale points.
Numbers should reinforce the meanings of the words rather than communicate conflicting meanings

Social desirability bias, overreporting and underreporting

When individuals answer, they do so just because they think externally they should.
Belli et al.(unpublished manuscript) significantly reduce overreporting by explicitly alerting
respondents to potential memory confusion and encouraging them to think carefully to avoid such
confusion.

Models of the process of answering

Stages (not completely sequential and independent):

In words:
1. Respondents must interpret the question
1.1. bring the sounds of the words into their working memories
1.2. break the words down into groups, each one representing a single concept, settling on the
2. meaning of each
2.1. build the meaning of the entire question by establishing the relations among the concepts
2.2. discard the original words of the question and retain their interpretations as they begin to
3. formulate an answer.
4. They must search their memories for relevant information
5. Then integrate that information into a single judgment (if more than one consideration is
recalled).
6. Finally, they must translate the judgment into a response by selecting one of the alternative
offered.
Types of respondents:
Optimizing – great deal of cognitive work required to generate an optimal answer. All four steps are
executed DILLIGENTLY
Weak Satisficing – Instead of generating the most accurate answers, respondents settle for merely
satisfactory ones. All four steps executed, but each one less diligently
Strong Satisficing – respondents may interpret each question superficially and select what they believe
will be a reasonable answer to the interview and researcher.
3 Conditions that foster satisficing
a) Task difficulty
b) Respondent’s inability
c) Respondent’s demotivation to optimize
Response order effects

 Primacy effects (response choices presented early were most likely to be selected) (Satisfiers
are heavily affected)
 Recency effects (response choices presented last were more likely to be selected)
OR

 No order effects
Facts

 Weak satisficing seems likely to produce primacy effects under conditions of visual
presentation
 When response alternatives are presented orally, memory factors promote recency effects
more than primacy effects
 Items presented at the beginning and end of a list are more likely to be recalled after the
question is read, particularly if the list is long
 A primacy effect was stronger for people with more limited cognitive skills
 More response order effects in questions involving more words and words that were difficult
to comprehend
Conclusion 6

 Respondents can be classified as maximizers, weak satisfiers and strong satisfiers

 The answers of satisfiers are less diligent and biased by response order effects
 Survey is dominated by “techniques for satisficing” – for conserving time and energy
Explaining acquiescence
ACQUIESCENCE = the tendency to endorse any assertion made in a question, regardless of its
content
Agree/disagree true/false, and yes/no questions are very popular, appearing in numerous batteries
developed for attitude and personality measurement
Fact
When people are asked to agree or disagree with pairs of statemains stating mutually exclusive views,
e.g. “I enjoy socializing” and “ I don’t enjoy socializing” answers should be strongly negatively
correlated.
But across more than 40 studies, the average correlation was only -.22
Explanation
Agreeableness Courtesy and respect Satisficing People may simply choose to agree because it seems
like the commanded and polite action to take.
Conclusion 7
  Yes/No questions may manifest this tendency less than Agree/Disagree or True/False items
 Acquiescence is more common among people with more limited cognitive skills
Ratings – rankings
Rankings: Ratings:
Explicitly ask respondents to make choices by Ask people to rate each object individually,
rank ordering a set of alternative allowing the researcher to derive the rank
order implied by the ratings

Fact

 Researchers have typically preferred to use rating questions rather than ranking questions.
 But, a number of studies indicate that rankings yield higher-quality data than ratings.
Rankings are more reliable and manifest greater validity.
 But what are the disadvantages of raking data? how close is 1st from 2nd in the ranking?
Example
- For example, when asked to rate the importance of a series of values (e.g. equality, freedom,
and happiness) on a scale from extremely important to not at all important, a satisficing
respondent can easily say they are all very important
- = Non-differentiation
- Nondifferentiation is more common among less educated respondents and is more prevalent
toward the end of a questionnaire
- Nondifferentiation is particularly pronounced among respondents low in verbal ability, for
whom fatigue is presumably most taxing

Decision for rating: bipolar versus unipolar

“Overall, are you satisfied or dissatisfied?”
OR
“Are you only a little satisfied, slightly satisfied, somewhat satisfied, very satisfied, or extremely
satisfied?” ( unipolar)
Usually researchers choose bipolar

“Do not know”

 DO NOT KNOW option should increase the quality of data obtained by a questionnaire.
Respondents would be discouraged from offering meaningless opinions.
 People sometimes offer such responses because
- they feel ambivalent about the issue
- do not understand the meaning of a question or the answer choices;
- think that they must know a lot about a topic to legitimately express an opinion
- reduced motivation at the end of survey

 Krosnick(1999) found no difference in data quality between filtered and standard versions in
any of their experiments.

Number of categories

 The standard advice has been to use 5 to 9 categories (Miller 1956, Cox 1980),
 Magnitude scaling and feeling thermometers, offer a very large number of numerical options,
but respondents usually choose answers that are multiples of 5, 10, or 25(at the low end of the
continuum) and 50, 100, or 1000 (at the higher end), so that the number of categories used is
less than one might expect
Conclusion 8
Ratings are mostly used in practice because they are easier to analyze. Rankings have some
advantages, by forcing the respondents to think deeper and decide.
Regarding using “Do not know” option, some studies found no difference in data quality between
filtered and standard versions. But no black and white answer until more studies available.
Question type
1. Questions an attitudes and values
2. Questions on events and behaviors
Questions on events and behaviors
Do members of the target population have encoded the information?
e.g. Researchers may want to ask parents about their children’s immunization history, but some
parents will be unaware of what, if any, immunizations their children have received (Lee et al. 1999).
Major errors: omissions and intrusions when memory is altered

Defining the event:

“During the past 12 months since July 1st 1987, how many times have you seen or talked with a
doctor or a medical assistant about your health? Do not count any times you might have seen a doctor
while you were a patient in a hospital, but count all the other times you actually saw or talked to a
medical doctor of any kind about your health.“
Defining the reference period
Researchers must decide how (and how often during a series of questions) to specify the reference
period they have selected.
The reference period should usually be given at the beginning of a question and that it should be fully
specified at the beginning of a line of questioning and then given in abbreviated form, and in a parallel
location, in subsequent questions
Example:
“ The next questions are about the amount of child support you actually received between January 1
and December 31, 2001. In 2001, did you receive any payments for child support?”

Overreporting errors
They can be reduced by using a panel survey with bounded interviews
In an initial bounding interview, respondents are asked to report events; the list of events reported at
time 1 is then consulted during the time 2 interview. If an item is reported at time 2, the interviewer
verifies that it is a new item and not a duplicate of the item reported at time 1

Response dimensions

 Occurrence (simple yes-no items, closed frequency items, and open frequency items)
 Absolute frequency (recommended open questions to avoid the answer is affected by the
response categories)
 Relative frequency (“very often”, “seldom” …vague quantifiers)
 Regularity or periodicity
 Date of occurrence
Issues of relevance
Respondents may be annoyed at being asked a question that does not apply to them
Use filter questions to avoid this
Choose between:

 “What is the number of servings of eggs you eat on a typical day?”

 “On days when you eat eggs, how many eggs do you usually have?“
The second version recognizes the difficulty of averaging across days when the respondent did and
did not eat eggs and is thus easier to answer

Threatening behaviors
Strategies to reduce voting claims:
1. Ask the respondent to remember details about the vote
2. Presented multiple response categories for “didn’t vote” instead of the usual single one, and
phrased the “did vote” category in definite terms
e.g. “I thought about voting this time but didn’t”
3. Self-administered modes of administration produce both higher reports of socially
undesirable behaviors and lower reports of socially desirable ones
Questions about subjective phenomena
Focus groups are often used during questionnaire development to identify appropriate names and
relevant dimensions.

V. Determine sample size statistically

Precision level (D)

The maximum permissible difference between the statistic and the parameter

Confidence interval and level

The range into which the true population parameter will fall

The probability that a confidence interval will include the population parameter

Evaluate a population mean based on a survey

Required n?

If the sample represents 10% of the population or more, apply the finite population correction to n
Example:
Conduct a survey to estimate the monthly amount invested in savings schemes so that the estimate
will be within ± 5.00€. What should be the sample size (n)?

The std. dev. Of the population (sigma) may be known from secondary studies, if not, it can be
estimated by conducting a pilot study; or researcher’s judgement.
If the population standard deviation is unknown, then, after the sample is drawn, we can estimate the
confidence interval for mean by employing sample standard deviation.

Multiple parameters
In most business research projects, however, several characteristics are of interest. If all variables are
equally important, the most conservative approach is either to take a sufficiently big sample (> 1000-
2000) or to evaluate the required sample statistically for each parameter and to select the largest value
for the required sample size.
Fact
n practice, in order to achieve the calculated sample size, a much greater number of potential
respondents have to be contacted. Why?
- Incidence rate of the phenomenon investigated
- Response rate

Conclusion 9