Tải bản đầy đủ
7 Customer Analysis at Crédit Samouel (Case Study)
6 Hypothesis Testing & ANOVA
In an effort to control the campaign’s success and to align the marketing
actions, the management decided to conduct an analysis of newly acquired
customers. Specifically, the management is interested in evaluating the segment
customers aged 30 and below. To do so, the marketing department surveyed the
following characteristics of 251 randomly drawn new customers (variable names
– Gender (gender)
– Bank deposit in Euro (deposit)
– Does the customer currently attend school/university? (training)
– Customer’s age specified in three categories (age_cat)
Use the data provided in bank.sav (8 Web Appendix ! Chap. 6) to answer the
following research questions:
1. Which test do we have to apply to find out whether there is a significant
difference in bank deposits between male and female customers? Do we meet
the assumptions necessary to conduct this test? Also use an appropriate normality test and interpret the result. Does the result give rise to any cause for concern?
Carry out an appropriate test to answer the initial research question.
2. Is there a significant difference in bank deposits between customers who are still
studying and those that are not?
3. Which type of test or procedure would you use to evaluate whether bank deposits
differ significantly between the three age categories? Carry out this procedure
and interpret the results.
4. Reconsider the previous question and, using post hoc tests, evaluate whether
there are significant differences between the three age groups.
5. Is there a significant interaction effect between the variables training and
age_cat in terms of the customers’ deposit?
6. On the basis of your analysis results, please provide recommendations on how to
align future marketing actions for the management team.
1. Describe the steps involved in hypothesis testing in your own words.
2. Explain the concept of the p-value and explain how it relates to the significance
3. What level of α would you choose for the following types of market research
studies? Give reasons for your answers.
(a) An initial study on the preferences for mobile phone colors.
(b) The production quality of Rolex watches.
(c) A repeat study on differences in preference for either Coca Cola or Pepsi.
4. Write two hypotheses for each of the example studies in question 3, including
the null hypothesis and alternative hypothesis.
5. Describe the difference between independent and paired samples t-tests in your
own words and provide two examples of each type.
6. Use the data from the wishbird.net example to run a two-way ANOVA, including the factors (1) ease-of-use and (2) brand image, with sales as the dependent
variable. To do so, go to Analyze u General Linear Model u Univariate and
enter sales in the Dependent Variables box and image and ease in the Fixed
Factor(s) box. Interpret the results.
Field, A. (2013). Discovering statistics using SPSS (4th ed.). London: Sage.
An excellent reference for advanced types of ANOVA.
Hubbard, R., & Bayarri, M. J. (2003). Confusion over measure of evidence (p’s)
versus errors (α’s) in classical statistical testing. The American Statistician, 57(3),
The authors discuss the distinction between p-value and α and argue that there is
general confusion about these measures’ nature among researchers and
practitioners. A very interesting read!
Kanji, G. K. (2006). 100 statistical tests (3rd ed.). London: Sage.
If you are interested in learning more about different tests, we recommend
this best-selling book in which the author introduces various tests with information
on how to calculate and interpret their results using simple datasets.
Sawyer, A. G., & Peter, J. P. (1983). The significance of statistical significance tests
in marketing research. Journal of Marketing Research, 20(2), 122–133.
Interesting article in which the authors discuss the interpretation and value
of classical statistical significance tests and offer recommendations regarding
Boneau, C. A. (1960). The effects of violations of assumptions underlying the t test. Psychological
Bulletin, 57(1), 49–64.
Brown, M. B., & Forsythe, A. B. (1974). Robust tests for the equality of variances. Journal of the
American Statistical Association, 69(346), 364–367.
Cho, H. C., & Abe, S. (2012). Is two-tailed testing for directional research hypotheses tests
legitimate? Journal of Business Research, 66(9), 1261–1266.
Cohen, J. (1992). A power primer. Psychological Bulletin, 112(1), 155–159.
Field, A. (2013). Discovering statistics using SPSS (4th ed.). London: Sage.
Hubbard, R., & Bayarri, M. J. (2003). Confusion over measure of evidence (p’s) versus errors (α’s)
in classical statistical testing. The American Statistician, 57(3), 171–178.
6 Hypothesis Testing & ANOVA
Lilliefors, H. W. (1967). On the Kolmogorov–Smirnov test for normality with mean and variance
unknown. Journal of the American Statistical Association, 62(318), 399–402.
Welch, B. L. (1951). On the comparison of several mean values: An alternative approach.
Biometrika, 38(3/4), 330–336.
After reading this chapter, you should understand:
– What regression analysis is and what it can be used for.
– How to specify a regression analysis model.
– How to interpret basic regression analysis results.
– What the issues with, and assumptions of regression analysis are.
– How to validate regression analysis results.
– How to conduct regression analysis in SPSS.
– How to interpret regression analysis output produced by SPSS.
Adjusted R2 • Autocorrelation • Durbin-Watson test • Errors • F-test •
Heteroskedasticity • Linearity • Moderation • (Multi)collinearity • Ordinary
least squares • Outliers • Regression analysis • Residuals • R2 • Sample size •
Stepwise methods • Tolerance • Variance inflation factor • Weighted least
Agripro is a US-based firm in the business of selling seeds to farmers and
distributors. Regression analysis can help them understand what drives
customers to buy their products, helps explain their customer’s satisfaction,
and informs how Agripro measures up against their competitors. Regression
analysis provides precise quantitative information on which managers can
base their decisions.
M. Sarstedt and E. Mooi, A Concise Guide to Market Research,
Springer Texts in Business and Economics, DOI 10.1007/978-3-642-53965-7_7,
# Springer-Verlag Berlin Heidelberg 2014
Regression analysis is one of the most frequently used tools in market research. In its
simplest form, regression analysis allows market researchers to analyze relationships
between one independent and one dependent variable. In marketing applications, the
dependent variable is usually the outcome we care about (e.g., sales), while the
independent variables are the instruments we have to achieve those outcomes with
(e.g., pricing or advertising). Regression analysis can provide insights that few other
techniques can. The key benefits of using regression analysis are that it can:
1. Indicate if independent variables have a significant relationship with a dependent
2. Indicate the relative strength of different independent variables’ effects on a
3. Make predictions.
Knowing about the effects of independent variables on dependent variables can
help market researchers in many different ways. For example, it can help direct
spending if we know promotional activities significantly increases sales.
Knowing about the relative strength of effects is useful for marketers because it
may help answer questions such as whether sales depend more on price or on
promotions. Regression analysis also allows us to compare the effects of variables
measured on different scales such as the effect of price changes (e.g., measured
in $) and the number of promotional activities.
Regression analysis can also help to make predictions. For example, if we have
estimated a regression model using data on sales, prices, and promotional activities,
the results from this regression analysis could provide a precise answer to what
would happen to sales if prices were to increase by 5% and promotional activities
were to increase by 10%. Such precise answers can help (marketing) managers
make sound decisions. Furthermore, by providing various scenarios, such as
calculating the sales effects of price increases of 5%, 10%, and 15%, managers
can evaluate marketing plans and create marketing strategies.
Understanding Regression Analysis
In the previous paragraph, we briefly discussed what regression can do and why it is
a useful market research tool. But what is regression analysis all about? To answer
this question, consider Figure 7.1 which plots a dependent (y) variable (weekly
sales in $) against an independent (x) variable (an index of promotional activities).
Regression analysis is a way of fitting a “best” line through a series of observations.
With “best” line we mean that it is fitted in such a way that it minimizes the sum of
squared differences between the observations and the line itself. It is important to
know that the best line fitted with regression analysis is not necessarily the true line
(i.e., the line that holds in the population). Specifically, if we have data issues, or
fail to meet the regression assumptions (discussed later), the estimated line may be
Understanding Regression Analysis
Weekly sales in USD
Index of promotional activities
Fig. 7.1 A visual explanation of regression analysis
Before we introduce regression analysis further, we should discuss regression
notation. Regression models are generally noted as follows:
y ¼ α þ β 1 x1 þ e
What does this mean? The y represents the dependent variable, which is the
variable you are trying to explain. In Fig. 7.1, we plot the dependent variable on
the vertical axis. The α represents the constant (sometimes called intercept) of the
regression model, and indicates what your dependent variable would be if all of
the independent variables were zero. In Fig. 7.1, you can see the constant
indicated on the y-axis. If the index of promotional activities is zero, we expect
sales of around $2,500. It may of course not always be realistic to assume that
independent variables are zero (just think of prices, these are rarely zero) but the
constant should always be included to make sure that the regression model has the
best possible fit with the data.
The independent variable is indicated by x1. β1 (pronounced as beta) indicates
the (regression) coefficient of the independent variable x. This coefficient
represents the gradient of the line and is also referred to as the slope and is shown
in Fig. 7.1. A positive β1 coefficient indicates an upward sloping regression line
while a negative β1 indicates a downward sloping line. In our example, the gradient
slopes upward. This makes sense since sales tend to increase as promotional
activities increase. In our example, we estimate β1 as 55.968, meaning that if we
increase promotional activities by one unit, sales will go up by $55.968 on average.
In regression analysis, we can calculate whether this value (the β1 parameter)
differs significantly from zero by using a t-test.
The last element of the notation, the e denotes the error (or residual) of the
equation. The term error is commonly used in research, while SPSS uses the term
residuals. If we use the word error, we discuss errors in a general sense. If we use
residuals, we refer to specific output created by SPSS. The error is the distance
between each observation and the best fitting line. To clarify what a regression
error is, consider Fig. 7.1 again. The error is the difference between the regression
line (which represents our regression prediction) and the actual observation.
The predictions made by the “best” regression line are indicated by y^ (pronounced
y-hat). Thus, the error for the first observation is:1
e1 ¼ y1 À y^1
In the example above, we have only one independent variable. We call this
bivariate regression. If we include multiple independent variables, we call this
multiple regression. The notation for multiple regression is similar to that of
bivariate regression. If we were to have three independent variables, say index of
promotional activities (x1), price of competitor 1 (x2), and the price of competitor
2 (x3), our notation would be:
y ¼ α þ β 1 x 1 þ β 2 x 2 þ β3 x 3 þ e
We need one regression coefficient for each independent variable (i.e., β1, β2, and
β3). Technically the βs indicate how a change in an independent variable influences
the dependent variable if all other independent variables are held constant.2
Now that we have introduced some basics of regression analysis, it is time to
discuss how to execute a regression analysis. We outline the key steps in Fig. 7.2.
We first introduce the data requirements for regression analysis that determine if
regression analysis can be used. After this first step, we specify and estimate the
regression model. Next, we discuss the basics, such as which independent variables
to select. Thereafter, we discuss the assumptions of regression analysis, followed by
how to interpret and validate the regression results. The last step is to use the
regression model, for example to make predictions.
Conducting a Regression Analysis
Consider Data Requirements for Regression Analysis
Several data requirements have to be considered before we undertake a regression
analysis. These include the following:
– Sample size,
– Variables need to vary,
– Scale type of the dependent variable, and
Strictly speaking, the difference between predicted and observed y-values is e^.
This only applies to the standardized βs.
Conducting a Regression Analysis
Consider data requirements for regression analysis
Specify and estimate the regression model
Test the assumptions of regression analysis
Interpret the regression results
Validate the regression results
Use the regression model
Fig. 7.2 Steps to conduct a regression analysis
126.96.36.199 Sample Size
The first data requirement is that we need a sufficiently large sample size. Acceptable
sample sizes relate to a minimum sample size where you have a good chance of finding
significant results if they are actually present, and not finding significant results if these
are not present. There are two ways to calculate “acceptable” sample sizes.
– The first, formal, approach is a power analysis. As mentioned in Chap. 6 (Box
6.2), these calculations are difficult and require you to specify several parameters,
such as the expected effect size or the maximum type I error you want to allow for
to calculate the resulting level of power. By convention, 0.80 is an acceptable
level of power. Kelley and Maxwell (2003) discuss sample size requirements.
– The second approach is through rules of thumb. These rules are not specific to a
situation but are easy to apply. Green (1991) proposes a rule of thumb for sample
sizes in regression analysis. Specifically, he proposes that if you want to test for
individual parameters’ effect (i.e., if one coefficient is significant or not), you
need a sample size of 104 þ k. Thus, if you have ten independent variables, you
need 104 þ 10 ¼ 114 observations.3
Rules of thumb are almost never without issues. For Green’s formula, these are that you need a
larger sample size than he proposes if you expect small effects (an expected R2 of 0.10 or smaller).
In addition, if the variables are poorly measured, or if you want to use a stepwise method, you need
a larger sample size. With “larger” we mean around three times the required sample size if the
expected R2 is low, and about twice the required sample size in case of measurement errors or if
stepwise methods are used.
188.8.131.52 Variables Need to Vary
A regression model cannot be estimated if the variables have no variation. Specifically, if there is no variation in the dependent variable (i.e., it is constant), we also
do not need regression, as we already know what the dependent variable’s value is.
Likewise, if an independent variable has no variation, it cannot explain any variation in the dependent variable.
No variation can lead to epic fails! Consider the admission tests set by the
University of Liberia. Not a single student passed the entry exams. Clearly in
such situations, a regression analysis will make no difference!
184.108.40.206 Scale Type of the Dependent Variable
The third data requirement is that the dependent variable needs to be interval or
ratio scaled (scaling is discussed in Chap. 2). If the data are not interval or ratio
scaled, alternative types of regression need to be used. You should use binary
logistic regression if the dependent variable is binary and only takes on two values
(e.g., zero and one). If the dependent variable consists of a nominal variable with
more than two levels, you should use multinomial logistic regression. This should,
for example, be used if you want to explain why people prefer product A over B or
C. We do not discuss these different methods in this chapter, but they are intuitively
similar to regression. For an introductory discussion of regression methods with
dependent variables measured on a nominal scale, see Field (2013).
The last data requirement is that no or little collinearity is present. Collinearity is a
data issue that arises if two independent variables are highly correlated. Multicollinearity occurs if more than two independent variables are highly correlated.
Perfect (multi)collinearity occurs if we enter two (or more) independent variables
with exactly the same information in them (i.e., they are perfectly correlated).
Perfect collinearity may happen because you entered the same independent
variable twice, or because one variable is a linear combination of another
(e.g., one variable is a multiple of another variable such as sales in units and
sales in thousands of units). If this occurs, regression analysis cannot
estimate one of the two coefficients and SPSS will automatically drop one
of the independent variables.
Conducting a Regression Analysis
In practice, however, weaker forms of collinearity are common. For example, if
we study how much customers are wiling to pay in a restaurant, satisfaction with the
waiter/waitress and satisfaction with the speed of service may be highly related. If
this is so, there is little uniqueness in each variable, since both provide much the
same information. The problem with having substantial collinearity is that it tends
to disguise significant parameters as insignificant.
Fortunately, collinearity is relatively easy to detect by calculating the tolerance
or VIF (Variance Inflation Factor). A tolerance of below 0.10 indicates that (multi)
collinearity is a problem.4 The VIF is just the reciprocal value of the tolerance.
Thus, VIF values above ten indicate collinearity issues. We can produce these
statistics in SPSS by clicking on Collinearity diagnostics under the Options button
found in the main regression dialog box of SPSS.
You can remedy collinearity in several ways. If perfect collinearity occurs, SPSS
will automatically delete one of the perfectly overlapping variables. SPSS indicates
this through an additional table in the output with the title “Excluded Variables”. If
weaker forms of collinearity occur, it is up to you to decide what to do.
– The first option is to use factor analysis (see Chap. 8). Using factor analysis, you
create a small number of factors that have most of the original variables’
information in them but which are mutually uncorrelated. For example, through
factor analysis you may find that satisfaction with the waiter/waitress and satisfaction with the speed of service fall under a factor called service satisfaction. If
you use factors, collinearity between the original variables is no longer an issue.
– The second option is to re-specify the regression model by removing highly
correlated variables. Which variables should you remove? If you create a
correlation matrix (see Chap. 5) of all the independent variables entered in the
regression model, you should focus first on the variables that are most strongly
correlated. Initially, try removing one of the two most strongly correlated
variables. Which one you should remove is a matter of taste and depends on
your analysis set-up.
Specify and Estimate the Regression Model
To conduct a regression analysis, we need to select the variables we want to include
and decide on how the model is estimated. In the following, we will discuss each
step in detail.
220.127.116.11 Model Specification
Let’s first show the main regression dialog box in SPSS to provide some idea of
what we need to specify for a basic regression analysis. First open the dataset called
The tolerance is calculated using a completely separate regression analysis. In this regression
analysis, the variable for which the tolerance is calculated is taken as a dependent variable and all
other independent variables are entered as independents. The R2 that results from this model is
deducted from 1, thus indicating how much is not explained by the regression model. If very little
is not explained by the other variables, (multi) collinearity is a problem.
Fig. 7.3 The main regression dialog box in SPSS
Sales data.sav (8 Web Appendix ! Chap. 7). These data contain information on
supermarket sales per week in $ (sales), the (average) price level (price), and an
index of promotional activities (promotion), amongst other variables. After opening
the dataset, click on u Analyze u Regression u Linear. This opens a box similar to
For a basic regression model, we need to specify the Dependent variable and
choose the Independent(s). As discussed before, the dependent variable is the
variable we care about as the outcome.
How do we select independent variables? Market researchers usually select
independent variables on the basis of what the client wants to know and on prior
research findings. For example, typical independent variables explaining the supermarket sales of a particular product include the price, promotional activities, level
of in-store advertising, the availability of special price promotions, packaging type,
and variables indicating the store and week. Market researchers may, of course,
select different independent variables for other applications. A few practical
suggestions to help you select variables:
– Never enter all the available variables at the same time. Carefully consider which
independent variables may be relevant. Irrelevant independent variables may be
significant due to chance (remember the discussion on hypothesis testing in Chap.
6) or can reduce the likelihood of determining relevant variables’ significance.
– If you have a large number of variables that overlap in terms of how they are defined,
such as satisfaction with the waiter/waitress and satisfaction with the speed of
service, try to pick the variable that is most distinct or relevant to the client.
Alternatively, you could conduct a factor analysis first and use the factor scores as
input for the regression analysis (factor analysis is discussed in Chap. 8).