12
Topics discussed
CHAPTER OBJECTIVES
After completing Chapter 12, you should be able to:
1. Describe the process followed in hypothesis testing.
2. Describe the concepts type I error, type n error, and statistical power.
Introduction
In Chapter 4 we discussed the sleps to be followed in hypothesis development and testing.
These steps are:
1. State the null and the alternate hypotheses.
2. Determine the level of significance desired (p = 0.05, or more, or less).
3. Choose the appropriate statistical test depending on the type of scales that have been used
(nominal, ordinal, interval, or ratio).
4. See if the output results from computer analysis indicate that the significance level is met.
When the resultant value is larger than the critical value, the null hypothesis is rejected, and
the alternate accepted. If the calculated value is less than the critical value, the null
hypothesis is accepted and the alternate hypothesis rejected.
In this chapter we will discuss hypothesis testmg. First, we will pay attention to type I errors,
type II errors, and statistical 1KJWer. Then, we w ill d iscuss various univariate and bivariate
statistical tests that can be used to test hypotheses. Finally, we w ill come back to the Excelsior
Enterprises case and test the hypotheses that were develo ped in the previou s chapter.
example introduced in Chapter 11 , a type I error would occur if we conclud ed, based on the
data, that burnout affects intention to leave when, in fact, it does not. The probability of type I
error, also known as the sigllificance level, is determined by the researcher. Typical significance
levels in business research are 5% 0.05) and 1 % 0.01).
A type U error, also referred to as beta ({3), is the probability of failing to reject the null
hypothesis given that the alternate hypothesis is actually true; e.g., concluding, based all the
data, that burnout does not affect intention to leave when, in fact, it does. The probability of
type. II error is inversely related to the probability of type I error: the smaller the risk of one of
these types of error, the higher the risk of the other type of error.
A third important concept in hypothesis testing is statistical power (1  /3). Statistical
power, or just power, is the probability of correctly rejecting the nuU hypothesis. In other
words, power is the probability that statistical Significance will be indicated if it is present.
Statistical power depends on:
1. Alpha (a): the statistical significance criterioJ1 used in the test. If alpha moves closer to zero
(for instance, if alpha moves from 5% to 1%), then the probability of finding an effect when
there is an effect decreases. This implies that the lower the a (that is, the closer a moves to
zero) the lower the power; the higher the alpha, the higher the power.
2. Effect size: the effect size is the size of a difference or the strength of a relationship ill tile
population: a large difference (or a strong relationship) in the population is more likely to be
found than a small difference (similarity, relationship).
3. The size of the sample: at a given lcvel of alpha, increased sample sizes produce more
power, because increased sample sizes lead to more accurate parameter estimates. Thus,
increased sample sizes lead to a higher probability of finding what we were looking for.
However, increasing the sample size can also lead to too much power, because even very
small effects will be found to be statistically significan t.
Along these lines, there are four interrelated components that affect the inferences you
might draw from a statistical test in a research project: the power of the test, the alpha, the
effect size, and the sample size. Given the values for any three of these components, it is thus
possible to calculate the value of the fourth. Generally, it is recommended to establish the
power, the alpha, and the required precision (effect size) of a test first, and then, based on the
values of these components, determine an appropriate sample size.
The focus of business research is usually on type I error. However, power (e.g., to
d etermine an appropriate sample size) and in some sihlations, type n error (e.g., if you
are testing the effect of a new drug) must also be given serious consideration.
Univariate techniques:
Testing a hypothesis on a single mean:
Mefr;c da la:
Nonmetric dala :
MannWhitney lJ.test
Rela ted samples
Metic data
Nonmetric data :
Multivariate techniques:
Dna metric dapenden! variable
Discriminant analysis
logistic regression
More than one metric dependent variable
Multivariate analysis of
Canonical correlation
variance
elaborately than others. A detailed discussion of all these techniques is beyond the scope of
this book.
average student stud ies 32 hours a week. From what you have observed so far, you think that
students from your university (the population from which your sample will be drawn) study
more. Therefore, you ask twenty class mates how long they study in an average week. The
average study time per week tums out to be 36.2 hours, 4 hours and 12 minutes more than the
stu dy time of students in general. The question is: is this a coincidence?
In the above example, the sample of students from your university differs from the typical
student. What you want to know, however, is whether your fellow students come from a
different population than the rest of the students. In other words, did you select a group of
motivated students by chance? Or is there a " true" difference between students from your
university and students in general?
In this example the null hypothesis is:
Ha: Tile. I1Ilmber of study hours of stude/1ts from ollr university is equal to the number of study
hours of students ill general.
The a lternate hy pothesis is:
H ,: The lIum!Jer of study II01Irs of students from Ollr university differs fro'" the number 0/ study
hours of students in general.
The way to decide whether there is a significant difference between students from your
university and students in general depends on three aspects: the value of the sample mean (36.2
hours); the value of the comparison standard (32 hours); and the degree of u ncertainty concerning
how weU the sample mean represents the population mean (the standard error of the sample mean).
Along these lines, the following formula is used to compute the tvalue:
XI'
t,, _1 =
s/.,fii.
Assume thai the observed standard deviation is 8. Hence, the tstatistic becomes:
t=
36.2  32
8/.;20
2.438
Having calculated the tstatistic, we can now compare the tvalue with a standard table of
Ivalues w ith 11  1 d egrees of freedom to determine whether the tstatistic reaches the
threshold of statistical significance. When the tstatistic is larger than the appropriate table
value, the null hypothesis (no significant difference) is rejected.
Our Istatistic (2.438) is larger than the appropriate table value 0.729). TIlis means that the
difference between 36.2 and 32 is statistically significant. The null hypothesis must thus be
rejected: there is a significant d ifference in study time between students from our university
and students in general.
How does this work in SPSS?
Under the Analyze menu, choose Compa re Means, then OneSample T Test. Move
the dependent variable into the "Test Variable(s)" box. Type in the value you wish to
compare your sample to in the box called "Test Value."
..
..
......
".
. 
'
. ..'..
.... . ..
,'
......
..
......
.....
.. . ....
"
..,.'.
,...
,.
,.
,..
"
,.'.
....''...
'.....
.
. ..'. ,.
,.'.
.
..
'
,
.
'
.
,. . ... .." ,.,. ',. ... '.
,. ,. ,.
. .. ..
'
.
..',.. ..,.. ... ,..,.,... .''. .,. ..',.. ..,..,. .''..
.. ,. .. ..
..
.. ... ..'. '. ..
" .
..
..
.. ... ,.'. ,. ... ,.
. .. ... . .. ..'. .. .."
.... ..... .. .. .. ,.
..
.. .'.
"
"
samples ttest to test the null hypothesis that the average of the d ifferences between the before
and after measure is zero.
How does this work in SPSS?
Under the Ana lyze menu, choose Compare Means, then PairedSamples T Test.
Move each o f the two variables whose means you want to compare to the "Paired
Variables" list.
,.
,. . ,.'. ,.
,.,. ,.'.
.... '. .. ,.,.,.,.
".".".". ,.,.,.'. ,.,.,'... ,,,_.
,' '
.. ,.,. , _ .
.. ,. '.. '. ,.',..
""
 '. ,.''.. ..
>.
>.
,
,>
>
>.
'a
a
a
>.
>.
'
,'
,
>
'
>.
>.
>.
>.
>.
111111 la,
,>
>.
>.
,.
'. ''..
',.. '.
'.>.
>.
>.
>.
'''...
,.
 
'. '.
,.
'.
>.
,.>.
>.
>.
>.
>.
>.
>.
>.
>.
,.,.>.
,.'.
,.,.
,. ,.,. ,.,.,.
,.,. ,.
,.
''.. ,'.. .. '. ,.,.,.,. ,.,.'. ,'..
'. ..'. ,.
'0
>
>.
>.
>..
>
0
0
>
'
>
'
~_ J~
>.
>.
~'_'I
,,.
,
,.
,.>. ,.
~
~
,.
~
~
>.
>.
>.
>.
>.
'0
>.
>
,.
'''...
,.
,.
,.
''..
,.,.
,.
,,..
>.
>.
>
>.
>.
>.
>.
EXAMPLE
A university professor was interested in the effect of her teaching program on the
performance of her students. For this reason, ten students were given a math test in the
first week of the semester and their scores were recorded. Subsequently, the students
were given an equivalent test during the last week of the semester. The professor now
wants to know whether the students' math scores have increased.
Table 12.1 depiCts the scores of students on the math test in the first and in the last
week of the semester.
TABLE 12.1 Math scores of ten students in the first and last
week ofthe semester.
Math scores
Student
Difference
+ 20
75
55
......................................................................................................................
65
80
+ 15
2
...............................................
.................................................................................
75
+5
70
3
................................................................................................................................
+5
60
55
4
...............................................................................................................................
45
+5
40 ..........................................................................
5
.....................................................
 5
55
60
6
..........................................
...........................
...........................................................
75
 5
80
7
...............................................................................................................................
70
+ 35
35
8
.....................................
.. .........................
.. ..............................................................
1
75
55 .. ........................................................................
9
.......................................
.............
90
+ 30
60
10
..............................................................................................................................
+ 20
Average score
57.5
70
22.5
To find out if there is a significant difference in math scores we need a test statistic. The
test statistic is the average differencej sdifference j .,fij.
In th is example we get: 22.5j 13.79412j J1i) = 5.158.
H aving calculated the tstatistic, we can now compare the Ival ue with a standard
table of Ivalues with n  1 degrees of freedom to determine whether the tstatistic
reaches the threshold of statistical Significance. Again, when the Istatistic is larger than
the appropriate table va lue, the null hypothesis (no significant difference) is rejected.
Our tsta tistic is larger than the appropriate table value (1.83). This means that the
difference between 70 and 57.5 is statistically significant. The null hypothesis must thus
be rejected: there is a significant increase in ma th score.
The Wilcoxo n sig nedrank test is a non parametric test for examining significant differences between two related samples or repeated measurements on a single sample. It is used as
an alternative to a paired samples ttest when the population cannot be assumed to be
normally distributed.
.
 '. ''... ..'.. ''... ... . ...  ... ..'.  . ..'.
.... ..'"'. ,..
.
.
~ ,
'. '.
.
m

LJ0
,
.
..
..
,.
. .. '.

,
'
.
. .. ,,
,,
....... . ..,. ,. '.
, _.
.
,.
'" . . '.
,i
f
::
.
..,. ,.
'
.
'
.
,
.
.
.. ..,.,'.. ,.'.. '.. ,.,.,. ,.,.. ,. ,. .. ,'"... .. ....'" ..".
.
,. '. .
,.'. ,.
..
.
,.
,'.
.
'
,.
.. ... . . 
,.,
, ,.
.,
'"
_.
a
'
'.
'
0 __
,,
'
LJ!e . ....
'
'"
"
'"
'"
'
.s
'"
,'"
'"
'
'
G '
's
.s
'"
"'"
"
'
'
'
'
'
'
'"
McNemar's test is a non parametric method used on nominal data. It assesses the
significance of the difference between two dependent sa mples when the va riable of interest
is dichotomous. It is used primarily in beforeafter studies to test for an experimental effect.
In the following example, a researcher wants to determine whether the use of a new training
method (called CA RE) has an effect on the performance of athletes. Counts of individual athletes
are given in Table 12.2. The performance (average/good) before the treatment (the new training
method) is given in the columns (244 athletes delivered an average perfonnance before they
trained with the CARE method, whereas 134 athletes delivered a good perfom,ance before they
adopted th is method). You can find the performance after the treatment (average/good) in the
rows (190 athletes delivered an average performance after using the new training method; the
number of athletes that delivered. a good performance increased to 188).
The cells of Table 12.2 can be represented by the letters n, b, c, and d. The totals across rows
and columns are marginal totals (a + b, c + d~a + c, and b + d). The grand total is represented by
II, as shown in Table 12.3.
McNemar's test isa rather straightforward technique to test marginal homogeneity. Mnrgillni
homogeneity refers to equality (or the lack of a significant difference) between one or more of the
average
After
good
totals
average
good
totals
112
132
244
78
56
134
190
188
378
average
After
"
good
totals
average
good
totals
b
d
b+ d
a+b
c+ d
c
a +c
n./
marginal row totals and the correspond ing marginal column totals. [n this example, marginal
homogeneity implies that the row totals are equal to the corresponding column totals, or
c + d = b+d.
Ma rginal homogeneity would mean there was no effect of the treatment. In this case it
wou ld mean that the new training method would not affect the performance o f athletes.
The McNemar test uses the x2 distribution, based on the formula: (I b  c[ _ 1)2) I (b + c) .
l is a s tatistic with 1 degree of freedom (# rows  1 x # columns  1). The marginal frequencies
are I/ot homogeneous iJ the X2 result is significant at p < 0.05.
The 'x? value in this example is: (178  1321 1)2)/(78 + 132) = 532 /210 = 13.376.
The table of the dis tribution of ch isquare, with 1 degree of freedom, reveals that the
difference between samples is significant: at the 0.05 level: the critical value of chisquare is
3.841. Since 13.376 com puted for the example above exceeds this valu e, the difference between
samples is significant. Hence, we can conclude that the new training method has a positive
effect on the performance of athletes.
Note that if band / or e are s mall (b + e < 20) then X2 is not approximated by theChisquare
d istribution. lns tead a sign test should be used.
.~
..
''...
..',.,,...
..'"'..
',.'..
,.
".
,.
,.
,.
,'
,.
.
o.
0 __
e

"
...
..
..
. .. .,..
'... ..'.. ''. ..,.
'.... . '. ..'..
.. . ..,..
. ,
',... ,..'.. ,.''.. ,.,..
,. ,. , ..
. ,.,.. '. ,....
,.,..'.
,'. ,. ,'.. ,.,.'. ,.,.
,.
,.'. '. ,.'. ,.
,. ...'.
,. ,. ,.
.....,. ............
,. ,. ,.
,.
,." ,.,., ,.
,.
,.
,.
,.
 ,, ,.,. ,.
~
'
[;r!  !  I ~
,,.
,.
,.
r1
,.'
,.
,
,
,. ,. ,
.,.
',.
,.

..
.
,.
..
.
,.
,. ,. ..
,.
,. .. '.
n
ow
.....I....
(IJI _ M ...
"
"
'
'
.... Ic::!
..,.'. ,..
...,.,. ''..
,.
'
,.'
,.
,.
,
'
'
'
"
"
'
,.
,.
'
accounting department; younger and older employees) is tested to see if there is a significant
mean difference between the two split groups on a dependent variable, w hich is measured on an
interval or ratio scale (for instance, extent of wellbeing; pay; or comprehension level).
.'. '.
.. '.
.. ....'.
.
..
........ ..,. ....,.
....... ,..... ....,.
..... ..
.... ....,. ..''..
,.
,.
'
,,
'
""
'
'
'"
'"
,.
'
,.
,.'. ,.
,.,. ,.
,
,
,.
,,.,_.
.'1' ~ ~:
'
G
O
=
'...
Br
,.
,....,.
,.
,.
)~
,.
I 
,.
,.
,... ,.. ,. .
,. ,.,,.
,
'
'
'
'
,.
'
'
'
'
'
'
'
'
or ratioscaled dependent variable. For example, is there a significant difference in the amount
of 5<1Ies by the following four groups of salespersons: those who are sent to training schools;
those who are given onthejob training during field trips; those who have been tutored by the
sales manager; and those who have had none of the above? Or is the rate of promotion
significantly d ifferent for those who have assigned mentors, choose their OWll mentors, and
have no mentors in the organizational system?
The results of ANOVA show whether or not the means of the various groups a re
significantly different from one another, as indicated by the F sta tistic. The F s tatistic s hows
whether two sample variances differ from each other or are from the same population. The F
d is tribution is a probability distribution of sample variances and the family of distributions
changes w ith changes in the sample size. Details of the F s tatistic may be seen in Table IVat the
end of the book.
,.
.
".".
".
".
".,.
..
,,.
,.
~
o. ,.
..
,. ,.
,.o..
,. ,.
..
..
,.,. '.o.. ..,.'". '"..o. o.. .... ... o.. ..,". ,.,. . ..
.. ..'. ,... ...." ... '..... ..'.. ..,.''.. ....'.
..
,..
',.,.. . '.
,.
''.. '.....
~
,.
,.
,.
~
'
'
'
o
o
'
'
..'. ~ Ik
om
'
,.
~
o
"
o.
'. ..
'. ,.'.
..,.
..
..
~
'
'.
1110

'
'w
,
.
~.
,.
'
'.
/II
..',... ..,.
w
~G[ I ~~~~~~~~,~
E ~ ~
..
. .
'
Elfi:.
1
'P
".
L~:](E:L!:iJ(=:1LJ ae___
~ _
'.
1/11
""
,.
''''
'/11
>111
...
' '''
./11
!J:II
1/11
" "
IJ')
'.
!J:II
'<11
""
. ...
<
...
.""
t'"
~_
0o_
__
'115
IJ>
..,.
. ........
'..
'.
,
...
~:
UII "0"_"'____' ,.
1/11 0 _ ...
'
::Ql.__ ..
II.
1!1f
'111
'
,/I>
'l1li
!J:II
' /If
1m
' /11
' /11
UII 
'I"
f: ,
~
'/II ~ ~
....'..,..
When significant mean differences among the groups are in d ica ted by the F statistic, there
is no way of knowing from the ANOVA results alone where they lie; that is, whether the
significant difference is between Groups A and B, or between Band C. or A and C, and so on. It
is therefore unwise to use multiple itests, taking two groups at a time, because the greater the
number of itests done, the lower the confidence we can place o n results. For example, three ttests d one simultaneously decrease the confidence level from 95% to 86% (0.95)3. However,
several tests, such as Schefte's test, Duncan Multiple Range lest, Tukey's test, and StudentNewmanKeul's test are available and can be used. as appropriate, to detect w here exactly the
mean differences lie.
Regression analysis
Simple regression an alysis is used in a situation where one independent variable is hypothesized to affect one dependent variable. For instance, assume that we propose that the
propensity to buy a product depends only on the perceived quality of thal product. I In
this case we would have to gather information on perceived quality and propensity to buy a
product. We could then plot the data to obtain some first ideas on the relationship between
these variables.
From Figure 12.2 we can see that there is a linear relationship between perceived qua lity and
propensity to buy the product. We can model this linear relationship by a I~st squares fllnctio1l.
A simp le line"r regression equation represents a straight line. Indeed, to summarize the
relationship between perceived quality and propensity to buy, we ca n drtlw a straight line
through the data points, as in Figure 12.3
We can also express this relationship in an equation:
The parameters f30 and fh are called regression coefficients. They are the iJltercept (f30) and
the slope (.8t) of the straight line relating propensity to buy (Y) to perceived quality (XI)' The
slope call be interpreted as the number of units by which propensity to buy would increase if
perceived quality increased by a unit. The error term denotes the error in prediction or the
difference between the estimated propensity to buy and the actual propensity to buy.
In this example the intercept (Po) was not Significant whereas the slope (/31) was. The
unstanda rdized regression coefficient f31 was 0.832. This means that if perceived quality is
rated 2 (on a fivepoin t sca le), the estimated propensity to buy is 1.664. On the other hand, if
perceived qua li ty is rated 4 (on a fivepoint scale), the estimated propensity to buy is 3.328.
I In reality, any effort to model the effect of perceived quality 011 propensity to buy a product without
careful attention to other factors that affect propensity to buy would cause a serious statistical problem
5,00
,>
4,00
>
3,00
<
!l
2,00
1,00
2,00
3,00
Percei ved Quality
4,00
5,00
5,00
2,00
1,00
A Sq liner = 0,519
o
~~~~~~
5,00
1,00
2,00
3,00
4,00
Perceived Qu ality
be ex plained by the regression model In this case, the regression model fits the data pooriy. In
the aforementioned example, R2 fo r the model is 0.519. This means that almost 52% of the
"
"
.
".".
".
..
..".
...,.
....
..
.
. ... ..
'. ,.'. '. .. ..'. ,. '. ..,.,. .. ''.. ,.,'". ,.'. ,.,.'.
.
..
.
..
'
,.
,.
.
'. .
.'. ''.. ..'.. ''.... '.. '... .. ....' ''... ,'... '..'... ,.''... ,.,'..
'. .
~
,
'
~
~
,.
,.
"
,.
'
,.
"
II
,.
,.
r(EJ
,.
'
J
~~
'
'
'
'
'
'
,.
'
,,
'
'
'
,.
,.
,.
,.
'
'
..::=
,'
'
'
,.
,'
'
'
,.
'm
'
"
,.
,.
the prediction of the dependent variable. For example, suppose that a resea rcher believes that
the variance in performance can be explained by four independent variables, A, B, C, and D (say.
pay, task difficulty, supervisory support, and organizational culture). When these variables are
joinUy regressed against the dependent variable in an effort to explain the variance in it, the sizes
of the individual regression coefficients indicate how much an increase of one unit in the
independent variable would affect the dependent variable, assuming that all the other independent variables remain unchanged. What's more, the individual correlations between the
independent variables and the d ependent variable collapse into what is called a multiple r or
multiple correlation coefficient. The square of multiple r, Rsquare, or R2 as it is commonly
known, is the amount of variance explained in the dependent variable by the predictors.
/ TABLE 12.4
Work shift
Original code
DummyD!
DummyD2
First shift
Second shift
Next, the dummy variables D\ and D2 have to be induded in a regression model. This
would look like this:
In this example workers from the third shift ha ve been selected as the reference ca tegory.
For this reason, this category has not been included. in the regression equation. For workers in
the third shift, DJ and D2 assume a value of 0, and the regression equation thus becomes:
Yi = /JO+ ei
For workers in the first shift the equation becomes:
The coefficient PI is the difference in predicted job satisfaction for workers in the first shift,
as compared to workers in the third shift. The coefficient fJz has the same interpretation. Note
that any of the three shifts could have been used as a reference category.
Now do Exercises 12.1 and 12.2
EXERCISE
Provide the equation for workers in the second shift.
EXERCISE
Use the data of the Excelsior Enterprises case to estimate the effect of work shift on job
satisfaction.
Multicollinearity
Multicollinearity is an often encoun tered statistical phenomenon in which two or marc
independent variables in a multiple regression model are highly correlated. in its most severe
case (if the correlation between two independent variables is equaJ to 1 or 1) multicollinearity makes the estimation of the regression coefficients impossible. In all other cases it
makes the estimates of the regression coefficients unreliable.
The simplest and most obvious way to detect multicollineari ty is to check the correlation
matrix for the independent variables. The presence of high correlations (most people consider
correlations of 0.70 and above high) is a ftrst sign of sizeable multicollinearity. However, when
mu lticollinearity is the result of complex relationships among several independent variables,
it may not be revealed by this approach. More common measures for identifying
multicollinearity are therefore the tolerance value and the variance infIatioll faclor (VIF  the
inverse of the tolerance valu e). These measures indicate the degree to which one independent
variable is expla ined by the other independent variables. A common cutoff value is a tolerance
value of O.ID, which corresponds to a VIF of 10.
How does this work in SPSS?
Under the Analyze menu, choose Regression, then Linear. Move the dependent
variable into the "Depend ent" box. Move the independent variables into the "Independent(s)" list. Select "Statistics" by clicking the button on the righthand side. Select
"CoIHnearity diagnostics" and dick continue. Then dick OK.
I'.
"
""
'...
,'.
.,.,.'...
...
,,'.
.
,,.
.
..
,.
,.
,.
,'
,
,
,.,.
,.
..''... ',,,
. EI
'"
I.
I.
'.' =
.. ,.,... ""''
,.
G..
"."" . , E1r  =
..". .... ,",,, E1rGr... ..''.. ',.,.'.. ,. ,.,... ,......_..._ I~
.... . ..,. .... ..',.. ,.... ,.... ,... ... _.,
. ' .. . .
'
,.
i_I
,.
"w
'
"
"
I.
"
6
I~
~_
0> _ _
0
~~
0._
0._ _
'
J ~ D
..
,OJ
"
'
'
'
'
'
.~
..
'
F _ .........
i
( c
0 __ _
__ 0 0_
i ......
'
....
..,.
'
,.
'
Note that multicollinearity is /lot a serious problem if the purpose of the study is to predict
or forecast future values of the dependent variable, because even though the estimations of the
regression coefficients may be unstable, multicollinearity does not affect the reliability of the
foreca st. However, if the objective of the study is to reliably estimate the individual regression
coefficients, muJticollinearity is a problem . In this case, we may use one or more of the
fo llowing methods to reduce it:
Reduce the set of independent variables to a set that are not collinear (note that this may lead
to omitted variable bias, which is also a serious problem).
Use more sophisticated ways to analyze the data, such as ridge regression.
Crea te a new variable that is a composite of the highly correlated variables.
;11
tile libmry.
Now suppose that we also believe that, even though this relationship will hold for all
students, it will be nonetheless contingent on computer ownership. That is, we believe that the
relationship between the judgment of computers in the library and the judgment of the libra ry
is affected by computer ownership (indeed com puter ownership is a dummy variuble).
Therefore, we hypothesize that:
JI}: Tile relatiol1slrip betwcclI tile judgment of tire libmry and judgmcllt of complIfers i" tire library
is moderated by computer ownership.
The reJationship between the judgment of the library and judgment of computers in the
library can be modelled as follows:
(1)
We have also hypothesized that the effect of XI on Y depends on X2 This can be modelled
as follows:
(2)
Adding the second equation into the fITS ! one leads to U,e following model:
(3)
Model (3) sta tes that the slope o f model (1) is a function of variable X2 Although this model
allows us to test moderation, the following model is better :
(4)
You may have noticed that model (4) includes a direct effect of Xl on Y. This .. !lows u s to
dLfferenti .. te between pure moderatioll and qllasi modem /iol' , (compare Sharma, Durand and
GurArie, 1981) explained next.
If YI = 0, And Y2 i 0, X2 is not a moderator but simply an independent predictor variable. If
YI 1: 0, Xl is a moderator. Model (4) allows us to differentiate between pure moderators and
quasi moderators as fo llows: if Yl t 0, and Y2 = O,X2 is a pure moderator; that is, X2 moderates
the relationship between Xl and Y, but it has no direct effect on Y.Jf Yr # 0, and Y2 1: 0, X2 is a
Xl
x Xu)
(5)
o.
EXERCISE
Is computer possession a pure moderator or a quasi moderator? Explain.
EXERCISE
Provide a logical explanation for the negative moderating effect of computer possession.
The previous example shows that dummy variables can be used to allow the effect of one
independent variable on the dependent variable to change depending on the valueof the dummy
variable. It is, of course, also possible to include metric variables as moderators in a modeL In
such cases, the procedure to test moderation is exactly the same as in the previous example.
In this section, we have explained how moderation can be tested with regression analysis.
Note that it is also possible to test mediation with regression analysis. We will explain this later
on ill this chapter using the Excelsior Enterprises data .
...
...
...
....
..'.
'
 ....
..
..'...
....
...
....
...
'
..
...
..
..
..
.
..
.. .... .. .. ..... ........ '..... '... .......
.... .... ..... ..... ...... .. ..... ...
.. . ... ...... ....... ,..... ........ ....
..... .. .. ... ..
'"
"
"
>.
>
>.
'
>.
>
>.
'
>
...
...
. ...
..... ...
. ...
. ,.
.. .....
..
'
"
>
'
.
.
.. ..
....... ...
.. ....
.. ....
. .'.
. .
..
.. ..
~
'
"
"
'
'
'
"
'
"
>
Discriminant analysis
Discriminant analysis helps to identify the independen t variables that d iscriminate a
nominally scaled dependent variable of interest  say those who are high on a variable
from those who arc Iowan it. The linear combination of independ ent variables indicates the
discriminating fun ction showing the large difference that exists in the two group means. In
other words, the independent variables measured on an interva l or ratio scale discriminate the
grou ps of interest to the study.
logistic regression
Logistic regression isa lso used when the dependent variable is nonmetric. However, when the
dependent variable has only two groups, logistic regression is often preferred, because it does
not fa ce the strict assumptions that discriminant analysis faces and because it is very similar to
regression analysis. Although regression analysis and logistic regression analysis are very
different from a statistical point of view, they are very much alike from a practical viewpoint
Both methods proouce prediction equations and in both cases the regression coefficients
measure the predictive capability of the independent variables. Thus, logistic regression allows
the researcher to predict a discrete outcome, such as "will purchase the product/ will not
purchase the product" , from a set of variables that may be continuous, discrete, or dichotomous.
How does this work in SPSS?
Under the Analyze menu, choose Regression, then Binary Logistic. Move the
dependent varia ble into the "Dependent" box. Move the independent variables into
the "Covariate(s)" list and click OK.
..
.
.
.
.
.
..
.
.
..
.. . .. . ....
.. .. .... .. .
..
.... ... ....... .... .. .. .. .....
.. ... ,.... .. '.... .. .... ......
.... ..,. ..,. .... .... '.... ... .".
.... . '. ..... ... .....
..
..
 ....'.
"
.. ..
...
...
....
"
...
"
....
"
'
'
'
'.
,
'
'
"
'
'
'
"
Conjoint analysis
Conjoint analys is is a statistical tech nique tllat is used in many fields induding marketing,
product management, and operations research. Conjoint analysis requ ires participan ts to
make a series of tradeoffs. In marketing, conjoint analysis is used to understand how
consumers develop preferences for products or services. Conjoint analysis is built on the
idea that consumers evaluate the value o f a product or service by combining the value that is
provided by each attribute. An attribu te is a general feature of a product or service, such as
price, product quality, or deJiveryspced. Each attribute has specific levels. For instance, for the
attribute " price", levels might be 249, 279, and 319. Along these lines, we might d escribe a
mobile telephone using the attributes " memory", "battery life", "camera" and "price". A
specific mobile phone would be d escribed as follows: memory, 12 Mbytes; battery life, 24
hours; camera 5 mega pixels; and price 249.
Conjoint analysis takes these attribute and level descriptions of products and services and
uses them by asking participants to make a series of choices between different products. For
instance:
Memory
Battery liCe
Camera
Price
Phone X
12 Mbytes
24 hou rs
5 megapixels
249
Phon e Y
16 Mbytes
12 hours
8 megapixcls
319
By asking for enough choices, it is possible to establish how important each of the levels is
relative to the others; this is known as the utility of the level. Conjoint ana lysis is traditionally
carried out with some form of multiple regression analysis. More recently, the use of
hiera rchical Bayesian analysis has become widespread to develop models of individual
cons umer decisionmaking behavior.
Twoway ANOVA
Twoway ANOV A can be used to examine the effect of two nonmetric independent variables
on a single metric dependent variable. Note that, in this context, an independen t variable is
often referred to as a factor and this is why a d esign that aims to examine the effect of two
nonmetric independent variables on a single metric d ependent va riable is often caUed a
factorial design. The factorial design is very popular in the social sciences. Twoway ANOVA
enables us to examine main effects (the effects of the independent variables o n the d ependen t
variable) but also interaction effects that exist between the ind ependent va riables (or factors) .
An interactio n effect exists when the effect of one independent variable (or o ne factor) on the
d ependent variable depends on the level of the other independent variable (factor).
MANOVA
MANOVA is sim ilar to ANOVA, with the difference that ANOVA tests the mean differences
of more than two groups on one d ependent variable, whereas MANOVA tests mean differen ces among groups across severnl dependent variables simultaneously, by using s ums of
squares and crossproduct matrices. Just as multiple Itests would bias the results (as
explained earLier), multiple ANOVA tests, using one dependent variable at a time, would also
bias the results, since the dependent variables a.re likely to be interrelated. MAI"OVA
circumvents this bias by simultaneously testing all the dependent variables, canceJJjng out
the effects of any intercorrelations among them.
In MANOVA tests, the independent variable is measured on a nominal scale and the
dependent va riables on an interval or ratio sca le.
The null hypothesis tested by MANOVA is :
Ho: J.d = 1'2 = j.l.3 ... 1'11.
Canonical correlation
Ca non ica l correla tion examines the relationship between two or more dependent variables
and several independent variables; for example, the correlation between a set of job beha viors
(such as engrossment in work, timely completion of work, and number of absences) and their
influence on a set of performance factors (such as quality of work, the output, and rate of
rejects). The focus here is on delinea ting the job behavior profiles associated w ith performance
that result in highquality production.
In sum, several univariate, bivariate, and multivariate techniques are available to analyze
sample da ta. Using these techniques allows us to generalize the resul ts obtained from the
sample to the population at large. It is, of course, very important to use the correct statistical
techniqu e to test the hypotheses of your study. We have explained earlier in this chapter that
the choice of the appropriate statistical technique depends on the number of variables you are
exa mining. on the scale of measurement of your variable(s), on whether the assumptions of
parametric tests are met, and on the size of your sample.
 >
>.
>
. 
>
>.
o.
".
"'
'M
o.
"
.
>
>.
>.
>.
..,.
>
,.
'.
>
,.
,.
>.
'
>.
>
'
>.
>.
,.
OR
>
>.
>.
>
,
,
,.
,.
>
>
>.
>.
>.
>
>
.'. ..
.. ....
. .... ..,. . .. '.
>.
,.
>.
>
>.
>.
>.
,.
>.
>.
,.
">
>.
>.
,.
>.
..'... .. ..
'.. ,.'. ,.
''.. ,. .. ..'. . '. '... ...
.. .. .. .. . .. . ..... ...>.. ....
,.
,..
,....
.
,.
,.
.
..
.. .. .... .. . .... ...
>.
~
o.
""
'..
'
,
'"
~
,.
'">
,.
'"
>
'"
,.
,.
>
>
>.
>.
>.
>.
>.
' M
>.
">.
,.
>.
>.
w
>.
>.
>.
,.
>.
"
,.
>.
' ,.
,.
'"
'>.
,. >.>M
,.
, ,.
,. >>. ,.
,.'
'
'"
..
"
,
,.
"
'"
"
"
>.
,.
>.
>.
>.
>.
>
,.
>
'M
,.
,.
,
,.
'
>
'
'
These hypotheses call for the use of mediated regression analysis (a ll the variables are
measu red at an interval level). The results of these tests and their interpretation are d iscussed
below.
To test the hypothesis that job satisfaction mediates the effect of perceived justice,
burnout, and job enrichment o n employees' intentions to leave, three regression models
were estimated, following Baron and Kenny (1986): model 1, regressi ng job satisfaction on
perceived justice, burnout, and job enrichment; model 2, regressing intention to leave on
perceived justice, burnout, and job enrichment; and model 3, regressing employees' intentions to leave on perceived justice, burnout, job enrichment, and job satisfaction. Separate
coefficients for each equation were estimated and tested . To establish mediation the
following cond itions must hold: perceived justice, burnout, and job enrichment must affect
job sa tisfaction in model 1; perceived justice, burnout, and job enrichment must be shown to
impact employees' intention to leave in model 2; and perceived justice, burnout, and job
enrichment must affect employees' intention to leave in model 3 (while controlli ng for job
sa tisfa ctio n). If these conditions all hold in the predicted direction, then the effect of
perceived justice, burnout, and job enrich men t must be less in model 3 than in model 2.
Perfect mediation holds if perceived justice, burnout, and job enrichment have no effect
when the efiect of job 5<'ltisfaction is controlled for (model 3).
The R square of the first regression model (model I ) was 0.165 and the model was statistically
significant. In this model, perceived justice and burnout were sign ificant predictors of job
satisfaction. whereas }Db enrichment was not. The R square of the second regression model
(model 2) was 0.393 and this model was also statistically significant. Model 2. as depicted in
Table 12.5, indic;'lted that perceived justice, bumout, and job enrichment affected employees'
intention to leave. The R square of the last model (model 3) was 0.487 and again the model was
statistically significant. Perceived justice and burnout were significant predictors of intention to
leave when job satisfaction was controlled for. The effect of perceived justice and burnout on
intention to leave was less in the third model than in the second model. Thus, all conditions for
perfect med iation were mel for perceived justice and burnout. lob enrichment was related to
neither job satisfa ction nor intention to leave (when job satisfaction was controlJed for).
We performed followup analyses to test for the indirect effect of perceived justice and
burnout on intention to leave via job satisfaction. Baron and Kenny (1986) provide an
approximate significance test for the indirect effect of perceived justices and burnout on
employees' intentions. The path from , respectively, perceived justice and burnout to job
satisfaction is denoted n and its standard error S~; the path from job sa tisfaction to intention to
leave is denoted b and its standard error Sq. The product ab is the estimate of the indirect effect
of perceived justice and burnout on employees' intentions to leave. TIle sta ndard error of nb is:
The ratio ab/SE.." can be interpreted as a z statistic. Indirect effects of perceived justice
< 0.05) and burnout (2.985, p < 0.01) were both significant.
(2.175, p
pvalue
0.000
O.D1S
0.000
0.332
..... .... .. .............. .......... ...................... . .... .. ...... ..... .... .. ..... . .......................... .................. ... .
Model fit = 0.165
.. ............ .... ... .. .. ........ .. ............... ...... .. ..... .. ....... ...... .. ................. ......................... .... ... ... . ...
Coefficient
1.840
 0.307
0.643
 0.165
pvalue
0.000
0.000
0.000
0.039
Note. Parameters are unstandardized regression weights, with significa nce levels of tva.lues. Twosided tests. N = 174.
remedying burnout may require both individual and organizational change. To solve the
problem of burnollt, the president may need to change the work environment a nd educate
workers on how to adapt and cope better to the stresses of the workplace.
The fact that only 50% of the variance in "intention to leave" was explained by the four
independent variables considered in this study still leaves 50% unexplained . In other words,
there are other add itional variables that are important in explaining ITL that have not been
cons idered in this study. Further research might be necessary to explain morc of the variance
in ITL, if the president wishes to pursue the matter further.
EXERCISE
Discuss: what do the unstandardized coefficients and their pvalues in the first model
imply? In other words, what happens to job sa tisfa ction if perceived justice, burnout,
and job enrichment change by one unit?
EXERCISE
Provide the tolerance values and the variance inflation factors for all the independent
variables in model t. Discuss: do we have a multicollinearity problem?
EXERCISE
Does work shift moderate the relationship between job satisfa ction and intention to
leave for Excelsior Enterprises employees?
We have now seen how different hypotheses ca n be tested by applying the appropriate
statistical tests in data analysis. Based on the interpretation of the results, the research report is
then written, making necessary recommendations and discussing the pros and cons of each,
together with costbenefi t analysis.
by identifying hidden relations and patterns in the data stored in it For instance, data mining
makes it possible to trace retail sales patterns by ZIP code and the time of day of the purchases,
so that optimal stocking of items becomes possible. Su ch " mined" data pertaining to the vital
areas of the organization call be easily accessed and used for different purposes. For example,
staffing for different times of the day can be planned, as can the number of checkout counters
that need to be kept open in retail stores, to ensure efficiency as well as effectiveness. We can
see that data mining helps to clarify the underlying patterns in different business activities,
which in tum faci litates decision making.
Operations research (OR) or management science (MS) is another sophisticated tool used to
simplify and thus clarify certain types of complex problem that lend themselves to quantification. OR uses higher mathematics and statistics to identify, analyze, and ultimately solve
intricate problems of great complexity fa ced by the manager. It provides an additional tool to the
manager by using quantification to supplement personal judgment. Areas of problem solving
that easily lend themselves to OR include those relating to inventory, queuing, sequencing,
routing, and search and replacement. OR helps to minimize costs and increase efficiency by
resorting to decision trees, lillear programming, network analysis, and mathematical models.
Other information systems such as management information systems (M IS), the decision
stlpport system, Ihe executive informatiolT system, and the expert system are good decisionmaking
aids, but not necessarily involved with data coUection and ana lyses in the strict sense.
In su m, a good information system collects, mines, and provides a wide range of pertinent
information relating to aspects of both the external and internal environments of the
organiza tion. By using the wide variety of tools and tedmiques available for solvin g problems
of differing magnitude, executives, managers, and others entrusted with responsibility for
results at various levels of the organization can find solutions to various concerns merely by
securing access to these data available in the system and analyzing them.
It should be ensured that the data in the information system are errorfree and are frequenUy
updated. After all, decision making can only be as good as lhedata made ava ilable to managers.
variables and manifest (observed) variables. You ca n also use U SREL to carry out exploratory
factor analysis and confi rmatory factor analysis.
MA TLAB is a computer program that was originally designed to simplify the implementation of numerical linea r algebra routines. It is used to implement numerical algorithms for a
wide range of applications.
SAS is an integrated system of software products. capable of performing a bro.:,d range of
statistical analyses such as descriptive statistics, multivariate techniques, and time series
analyses. Because of its capabilities, it is used in many disciplines, including medical sciences,
biological sciences, social sciences, and education.
SPSS (Statistica l Package for the Social Sciences) is a d"t<l management and analysis
program designed to do statistical data analysis, including descriptive sta tistics such as plots,
frequencies, charts, and lists. as well as sophisticated inferential and multivariate statistical
procedures like analysis of variance (ANOVA), factor analysis, cluster analysis, and categoric<l l data analysis.
Stata is a general purpose statistical software package that supports various statistical and
econometric methods, graphics, and enhanced features for data manipulation, programming,
and matrix manipulation.
In this chapter we covered the procedwe for hypolhesis testing. We have discussed type I
errors, type U errors, and statistical power. We have observed various statistical analyses and
tests used to examine different hypotheses to answer research questions. We discussed the
use of dummy variables, multicollinearity, and moderated regression analysis. Through the
example of the research on Excelsior Enterprises, we observed hypothesis testing using
mediated regression analysis and learned how the computer results are interpreted.
1. What kinds of biases do you think could be minimized or avoided during the data
analysis stage of research?
2. When we collect data on the effects of treatment in experimental designs, which
statistical test is most appropriate to test the treatment effects?
3. A tax col'\SUltant wonders whether he should bemoreselectiveabout the class of clients he
serves so as to maximize his income. He usually deals with four categories of clients: the
very rich. rich, upper middle class, and middle class. He has records of each and every
client served, the taxes paid by them, and how much he has charged them. Since many
particulars in respect of lhe clients vary (number of dependants, business deductibles,
etc.), irrespective of theca tegol)' they belong to, he would like a n appropriate ana lysis to be
done to see which among the four categories of clientele he should choose to continue to
serve in the future. Wlwt killd a/al1alysis should be dOl1e irl tltis case and wily?
EXERCISE
Open the file "resmethassignmentl" (you created this file doing the exercise from the
previous chapter). Answer the following questions.
Is the exam grade significantly larger than 75?
Are there significant differences in the exam grade for men and women?
Is there a significant difference between the exam grad e and the paper grade?
Are there significant differences in the paper grade for the fOUT year groups?
Is the sample representative for the IQ level, for which it is known that 50% of the
popu lation has an IQ below 100, and 50% has an lQ of 100 or higher?
f. Obtain a correlation matrix for all relevant variables and discuss the results.
g. Do a multiple regression analysis to explain the variance in paper grades using the
independent variables of age, sex (dummy coded), and IQ, and interpret the
results.
a.
b.
c.
d.
e.
EXERCISE
Below are Tables 12A to 12D, summarizing the results of data analyses of research conducted
in a sales organization that operates in.50 differentdties of the country, and employs a total
sales force of about 500. The number of salespersons sampled for the study was 150.
a. Interpret the information contained in each of the tables in as much detail as possible.
b. Summarize the results for the CEO of the company.
c. Make recommendations based on your interpretation of the resu lts.
Mean
Std. d eviation
Minimum
Maximum
75.1
8.6
45.2
97.3
10.3
5.2
6.1
15.7
Salespersons
Population
Income
Advertisement
1.0
, Ad . expenditure
0.56
0.21
0.11
1.0
0.68
0.16
0.36
0.23
/ TABLE 12C
1.0
p = 0.05.
pS
0.001.
Source of va riation
Sums of squares
df
Mean squares
Significance of F
552.5
150
0.411 73
df
(5.144)
5.278
TABLE 12 0 (Continued)
...~~~!~p.~~. ~.........................................?:~.~~~......................................................... .
...~ .~~?~~............................................?:~~?~.......................................................... .
0.41173
df
(5. 144)
5.278
...~.i~....................................................
?:0?!? ........................................................... .
0.47
4.54
0.00001