- ------------------------·

Chapter
Frequency Distribution, Cross-Tabulation,
and Hypothesis Testing

Overview Once the data have been prepared for analysis (Chapter 14), the researcher should con-
Frequency distribution and cross-tabulations are duct some basic analysis. This chapter describes basic data analysis, including frequency
distribution, cross-tabulation, and hypothesis testing. First, we describe the frequency
distribution and explain how it provides both an indication of the number of out a/-
basic techniques that provide rich range, missing, or extreme values as well as insights into the central tendency, variability,
and shape of the underlying distribution. Next, we introduce hypothesis testing by
describing the general procedure. Hypothesis-testing procedures are classified as tests
__ insights into the data and lay the of associations or tests of differences. We consider the use of cross-tabulation for under
standing the associations between variables taken two or three at ;:i time. Although
the nature of the association can be observed from tables, statistics are available for
_foundation for more advanced analysis. examining the significance and strength of the association. Then, we present tests
for examining hypotheses related to differences based on one or two samples. Finally,'"~
discuss the use of software in frequencies, cross-tabulations, and hypotheses testing.
Laurie Hurrington, Marketing Manager--Custorner Retention, Help for running the SPSS and SAS Learning Edition programs used in this chapter is
First Choice Power provided in four ways: (1) detailed step-by-step instructions are given later in the chapte,,
(2) you can download (from the Web site for this book) computerized demonstration
movies illustrating these step-by-step instructions, (3) you can download screen captures
with notes illustrating these step- by-step instructions, and (4) yuu can refer to the Study
Guide and Technology Manual, a supplement that accompanies this book.
Objectives [ At the end of this chapter, the student should be able to:]
Many commercial marketing resec1rch projects do not go beyond basic data analysis.
1, Describe the significance of preliminary data analysis and the insights that These findings are often displayed using tables and graphs, as discussed further in Chapter 23.
can be obtained from such an analysis. Although the findings of basic analysis are valuable in their own right, they also provide
guidance for conducting multivariate analysis. The insights gaineU frorn the basic analysis are
2, Discuss data analysis associated with frequencies, including measures of also invaluable in interpreting the results obtained from more sophisticated statistical
location, measures of variability, and measures of shape. techniques. To provide the reader with a flavor of these techniques, W(~ illustrate lhe use of
3. Explain data analysis associat'::d with cross-tabulations and the associated cross-tabulation, chi-square analysis, and hypothesis testing.
statistics: chi-square, phi co(:;1cient, contingency coefficient, Cramer's V, and
lambda coefficient.
4. Describe data analysis associated with parametric hypothesis testing for one Real Research Commercial Battle of the Sexes
sample, two independent samples, and paired samples.
5. Understand data analysis associated with nonparametric hypothesis testing A comparison of television advertising in Australia, Mcxi..:o, and the United States focused on the analysis
for one sample, two independent samples, and paired samples. of sex roles in advertising. Results showed differences 111 the portrayal of the sexes in diffcn.:nt countries.
Australian advertisements revealed somewhat fewer, and Mcxic~rn adve11isernenb slightly more, scx-rnl\'
differences than U.S. advertisements. Cruss-tahulation and chi-square analy:-..is p1ovided the following
infonnat1011 for Mexico.
These results indicate that in Mexican commercials, women appeared in commercial.~ for products used
by women or hy either sex but rarely in commercials for men's products. Men appeared in commercials tor

Persons Appearing in the AD(%)

Product Advertised
Used b\f Women Men
Females 25.0 4.0
Males 6.8 11.8
Either 68.2 84.2
 CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 483
482 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

As can be seen, a ,'iignificantly (p < 0.01) higher overall amount of perceived risk was altached tu prod-
Cross-tabulations have been ucts purchased by catalog as compared to those purchased from a retail store. Although this stmly reveals
used to analyze sex roles in risk associated with catalog purchasing, terrorist threats, time shortage, anJ increased convenience have
advertising in Austru!ia,
Mexico, and the United
increased the amount of products that are purchased from catalogs as well as online_ 2 •
States.
These two examples show how basic data analysis can be useful in its own right The cross-
tabulation and chi-square analysis in the international television advertising example, and the
paired t tests in the catalog shopping example, enabled us to draw specific conclusions from the

•§.sas
data, These and other concepts discussed in this chapter are illustrated in the context of explaining
Internet usage for personal (nonprofessional) reasons, Table 15.1 contains data for 30 respondents
giving the sex (l = male, 2 = female), familiarity with the Internet (I = very unfamiliar, 7 = very
familiar), fnternet usage in hours per week, attitude toward the Internet and toward technology,
SPSS Data File both measured on a 7-point scale (I = very unfavorable, 7 = very favorable), and whether the
respondents have done shopping or banking on the Internet ( I = yes, 2 = no), !-'or illustrative
purposes, we consider only a small number of observations. In actual practice, frequencies, cross-
tabulations and hypotheses tests are performed on much larger samples, such as that in the Dell

SAS Data File

:l.:itWm
Internet Usage Data

Respondent Internet
Attitude
Toward
Attitude
Toward
Usage of
Internet:
Usage of
Internet:
l
,

products used hy either sex. These differences were abo found in the U.S. ads, although to a lesser extent,
Number Sex Familiarity Usage Internet Technology Shopping Banking iI
but were not found in Australian ads. Therefore, U.S. consumer prodrn..:ts companies should not be I LOO 7,00 14.UO 7.00 6.00 LOO LOO
advertising in Mexico in the same ways in which they advertise to the U.S. market. In the United States, the 2 2,00 2.00 2,l)I) 3,00 3.00 2,00 2.00
increasing population of Hispanic Americans has turned many advertisers' attention to Spanish-language 2,00 3 01) 3.00 4,00 3.00 LOO 2.00
television advertising. Sex roles in the Hispa111c culture show women as trad1t1onal homemakers, conserva-
4 2,00 3.00 3,00 7,00 5,00 LOO 2,0()
tive, and dependent upon men for support, but many Hispanic families in the UnileJ States do not fit
l,00 7,0() 11.00 7.00 7,00 LOil LOO
this traditionally held view. In 2009, more than half of Hispanic women worked outside the home, which
almost matched the proportion of women in the Anglo population that worked outside the home in the 6 2,00 4.00 6,00 5,00 4,01) LOO 2.00
United States. 1 • 7 2,(J() 2.00 2.()() 4,()() 5,00 2.00 2.00
2,00 3.0() 6.00 5,00 4.00 2.rm 2,00
9 2,00 3,()() 6.00 6.00 4,00 LOO 2.00
10 LOO 9.00 15,00 7,00 6,00 LOO 2.00
Real Research Catalogs Are Risky Busin•:, s II 2.00 4.00 3,00 4.00 3.00 2,00 2.00
12 2,00 5,0() 4,00 6.00 4.00 2,0() 2.00
Twclvt:: product categories were examined to compare catalog to store shopping. The null hyrothesis that
there is no significant difference in the overall amount of risk perceived when buying products by catalog 13 I.OIi 6.00 9.00 6.00 5.00 2 00 l,()t)
compared to buying the same products in a retail store was rejected. The hypothesi,<, was tested by comput- 14 LOO 6.00 ~,00 3,00 2.00 2.00 2,00
ing 12 (one for each product) paired-observations t tests. Mean scores for overall pen.::eiw.d risk for some of 15 I.OIi 6,00 5,00 5.00 4,00 LIJO 2,00
the products in both buying situations are presented in the following tahle, with higher sl:ore:-. indic;.iting 2,00 4.00 3.00 4.00 3.00 2,llO 2.00
16
greater ri:.k. 17 UlO 6.00 9.00 5,00 :l.OO LOO l.00
18 ],l)(J 4,0() 4,00 5,()0 4.00 LOO 2.00
Mean Scores of Overall Perceived Risk 19 LOO 7,IX) 14.00 6,00 6.00 1.00 LOO
for Products by Purchase Mode
20 2,00 6,00 6.00 6.00 4.00 2.00 2.00
Overall Perceived Risk 6,00 9,IXl 4,00 2,00 2.00 2,00
21 LOO
Product Catalog Retail Store 22 1.00 5,00 5,00 5,00 4,00 2,00 l,00

4L98' 23 2.00 3.00 2,00 4,00 2.00 2,00 2.00

Stereo hi-fi 48.89
32,65 28,74' 24 LOO 7,00 15,00 6,00 6,00 LUO I ,00
Music album
58,60 50,80' 25 2.00 6.00 6.00 5.00 3,()1) 1.01) 2.00
Dress shoes
TV 48.53 40,91' 26 LOO 6,00 13.00 6,00 6.00 LOO I.Oil
Athletic socks 35.22 30,22' 27 2.00 5,00 4,00 5,00 5,00 1.00 1,00
Pocket PC/PDA 49,62 42,00' 28 2.00 4,0() 2,00 :1.00 2,00 2.00 2,00
Digital camera 48,13 39,52' 4.00 4,00 5,00 J,00 LOO 2.00
29 1.00
Perfume 34,85 29,79'
JO LOO 3,01) 3,00 7,00 5,00 I 00 2,00
*Significant at 0.01 level
484 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 15 • FREQUENCY DISTRIBUl'ION, CROSS-TABULATION, AND HYPOTHESIS TESTING 485
running case and other cases with real data that arc presented in this book. As a first step in the FIGURE 15,1
analysis, it is often useful to examine the frequency distributions of the relevant variables.
Frequency
Histogram
Frequency Distribution
Marketing researchers often need to answer questions about a single variable. For example:

• How many users of the brand may be characterized as brand loyal? c

C

•§.sas
" What percentage of the market consists of heavy users, medium users, light users, and 1l
nonusers? l 4
" How many customers are very familiar with a new product offering? How many are famil-
iar, somewhat familiar, and unfamiliar with the brand'/ What is the mean familiarity rating?
ls lhere much variance in the extent to which customers are familiar with the new product? SPSS Output File
• What is the income distribution of brand users? Is this distribution skewed toward low-
income brackets?

The answers to these kinds of questions can be determined by examining frequency distrib- ()
fn:;quency distrihution utions. In a frequency distribution, one variable is considered at a time. The objective is to 4
A mathematical d1stnbut1on obtain a count of the number of responses associated with different values of the variable. The SAS Output File flamiliarity with the lnlernd
whose obJect1ve 1s to obtain relative occurrence, or frequency, of different values of the variable is then expressed in percent-
a count of the nurnber of ages. A frequency distribution for a variable produces a table of frequency f..'.Ounts, percentages, or relative frequencies of the values ai·e placed along the Y-axis. Figure 15. I is a histogram uf the
1esponse;. associated with and cumulative percentages for all the values associated with that variable. frequency data in Table 15.1. From the histogram, one coukl examine whether the ohserveJ
ci1ffere11t values of one
Table 15.2 gives the frequency distribution of familiarity with the Internet. In the table, the first distribution is consistent with an expected or assumed distrihution, such as the normal distribution.
variable dlld to express
column contains the labels assigned to the different categoric-; of the variable~ and the second column
these count,;; in perce11tage
indicates the codes assigned to each value. Note that a code of 9 has been assigned to missing values.
terms Real Research Basic Analysis Yields Olympic Results
The third column gives the number of respondents checking each value. For example, three respon-
dents checked value 5, indicating that they were somewhat familiar with the Internet. The fourth
For the 1996 Olympic games in Atlanta, lll()re than 2 million unique visitors came to the games and more
column displays the percentage of respondents checking each value. The next column shows
than 11 million tickets were soJJ. Jn Sydney, at the 2000 Olympic game-s, as well as in Athens .:it the 2004
percentages calcu1ated by excluding the cases with missing values. [f there were no missing values,
Olympic games, 5 million tickets were sold. ll is ohvious that thi"' is a potential target market that cannot be
columns 4 and 5 would be identical. The last column represents cumulative percentages after ignored. Rcse.archers at the University of Colorado at Boulder decided to find whal motivated the interna-
a~iusting for missing cases. As can be seen, of the 30 respondents who participated in the survey, tional and domestic lraw.lcrs to come to the Olympic games. A survey wa~ developed and administered to
l 0.0 percent checked value 5. If the one respondent with a missing value is excluded, this percentage visitors to Olympic games via personal interviews. Three hundred twenty surveys were completed con·ectly
t:hanges to l 0.3. The cumulative {ccentagc corresponding to the value of 5 is 58.6. In other words, and were use<l m the data analysis.
58.6 percent of the respondents with valict responses indicated a familiarity value of 5 or less. The r~sults (see the following table) showed that the top three factors that motivated people to uttend
A frequency distrihution helps determine the extent of item nonresponse ( I respondent out the games were a once-in-a-lifetime opportunity, availability of housing, and availuhility of tickr1s. The
of 30 in Table 15.1 ). It also indicates the extent of illegitimate responses. Values of O and 8 would results of this study helped planners for the 2008 Olympic games in Beijing find what spec1fo.: charactcris-
he illegitimate responses, or errors. The cases with these values could be i<lenti1ied and corrective tks the city rweded to improve. For instance, from this research, BL:ijing put funds into pmjed~ that ..iddcd
hotel rooms to the city. ThL:y also constructed state-of-the-art transportation and unique venu~s (Olympic
action taken. The presence of outliers or cases with extreme value~ can also be dct~cte<l. In the case
p<:1rks, .stadiums, tourist i>itcs) sn that visi1ors truly felt that they were gelling ,1 once-in-a-lite.time experi-
of a frequency distribution of household size, a few isolated families with household sizes of 9 or
ence. As this survey continues to evolve over the years, the Jata received will become very valuable to the
more might be consiLicrcd outliers. A ftl~4uency distribution also indicates the. shape of the empiri- next host city. The planning for the 2012 Olympic games in London was also guided hy these: f1ndings. 3
t.'.al distribution of the variahle. The fre4uenL:y data may be used to construct a histogram, or a
vertical bar chait, in which the values of the variable are portrayed along the X-axis and the absolute

,.. ....
lli7.'f, Motivational Factors That Influenced the Decision to Attend the Olympic Games
Motivational Factor ,l:ro..-.,
..cyu,;:;"'-Y
,onr\l
Frequency Distribution of Familiarity with the Internet Percentage
Once-in-a-lifetime opportunity 95 29 7
Valid Cumulative
Value Label Value Frequency (n} Percentage Percentage Percentage Availability of housing 36 11.2
Availability of tickets 27 8.4
Very unfamiliar I 0 0.0 0.0 ()0
Distance away from home 24 7.5
6.7 6.9 69 Bt1siness/employrne11t 17 SJ
20.0 20.7 27.6 Availability of money----{)verall expenses 17 5.3
20.0 20.7 48.3 Availahility of time 12 3.8
SPSS Output File 5 3 [(JO 10.3 58.6 Personal relationship with participant or official 8 2.5
6 8 26.7 27.6 86.2 Other motivational factor 2.5

§.sas.
SAS Output J<'ile
Very familiar
Missing
TOTAL
7
9
4
I
30
13.3
3.3
100.0
13.8

100.0
100.0 Vi::.;it Olympic city
Security
Did not respond
Total
69
320
4 t.3
0.9
21.6
100.0 •
486 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING
487
Note that the numbers and percentages in the preceding example indicate the extent to which the
various motivational factors attract individuals to the Olympic games. Because numbers are of central tendency for interval or ratio data. The mean makes u.se of all the information available
involved, a frequency distribution can be used to calculate descriptive or summary statistics. because all of the values are used in computing it. However, the mean is sensitive to extremely
small or extremely large values (outliers). When there are outliers in the data, the mean is not a
good measure of central tendency and it is useful to consider both the mean and the median.
Statistics Associated with Frequency Distribution In Table 15.2, since there are no extreme values and the data are treated a.s interval, the mean
As illustrated in the previous section, a frequency distribution is a convenient way of looking at value of 4.724 is a good measure of location or central tendency. Although this value is greater
different values of a variable. A frequency table is easy to read and provides basic information, than 4, it is still not high (i.e., it is less than 5). If this were a large and representative sample, the
but sometimes this inl'ormation may be too detailed and the researcher must summarize it by the interpretation would be that people, on the average, are only moderately familiar with the
use of descriptive statistics. The most commonly used statistics associated with frequencies are Internet. This would call for both managerial action on the part of Internet service providers and
measures of location (mean, mode, and meuian), measures of variability (range, interquartile public policy initiatives on the part of governmental bodies to make people more familiar with
range, standard deviation, and coetficient of variation), and measures of shape (skewness and the Internet and increase Internet usage.
kurtosis). 4
Measures of Variability
Measures of location measures of
variability The measures of variability, which are calculated on interval or ratio data, include the range,
rnP«l'.>Hre'J of location The measures of location that we discuss are measures of central tendency because they tend to
A statistic that indicates the inte111uartile range, variance or ,i;;tandard deviation, and coefficient of variation.
A statistlC that describes a describe the center of 1he distribution. If the entire sample is changed by adding a fixed constant d1stnbut1on's dispersion
locat1011 within a data set to each observation, then the mean, mode, and median change hy the same fixed amount. RANGE The range measures the spread of the data. It is simply the difference hctwecn the
Mea-,ures of central range largest and smalkst values in the ,ample. As such, the range is directly atfocted by outliers.
tendency describe the MEAN The mean, or average value, is the most commonly used measure of central tendency. It is The difference between the
center of the, d1stribut1un used to estimate the mean when the data have heen collected using an interval or ratlo scale. The largest and smallest values Range = X1argesl - Xsmalle~t
data should display some central tendency, with most of the respon.i;es Jistrihuted around the mean. of a d1stnbut1on.
met1n [fall the values in the data are multiplied by a conslant, the range is multiplied by the same
The averdge, that value
The mean, X, is given by constant. The range for the data in Table 15.2 is 7 - 2 = 5.000.
obtdtnL'd by summing r1ll n
interquartile range
INTERQUARTILE RANGE The interquartile range is the difference between the 75th and 25th
elPmenb m c1 set and X= LX,ln The range of a distribution
d1wi1ng by the number i-c.-l encompassing the middle percentile. For a set of data points ,manged in order of magnitude, the rth percentile is the value
where 50 percent of the that hasp percent or the data points below it and (100 - p) percent ahove it. If all the data points
ot eiements.
ob,;rrvat1ons
X, ~ observed values of the variable X are multiplied by a constant, the interquartile range is multiplied by the same constant. The
interquartile range for the data in Table I 5.2 is 6 - 3 = 3.000.
n = number of observations (sample size)

lf there arc nu outliers, the mean is a robust measure and does not change markedly as data variance VARIANCE AND STANDARD DEVIATION The difference between the mean and an observed
values arc added or deleted. For the frequencies g{,:;n in Tahle 15.2, the mean value is calculated The mean squared value is called the duviationfmm the mean. The variance i., the mean squared deviation from the
as follows: dev1at1on of all the values mean. The variance can never be negative. When the data points arc clustered around the mean,
from the mean. the variance is small. When the data points are scattered, the vanance is large. rr al\ the data
X=OX2+6X3+6X4+3X5+8X6~4X7W~ standard deviation values arc multiplied by a constant, the variance is multiplied by the square of the constant. The
= (4 + 18 + 24 + 15 + 48 + 28)/29 The square root of the standard deviation is the square root of the variance. Thus, the standard deviation is expressed
variance in the same units as the data, rather than in squared units. The standard deviation ofa sample, s,
= 137/29 i::. calculated as:
·= 4.724

mode MODE The mode is the value that occurs most frequently. ll represents the highest ptoak of the (f~.--~7.i
A measure of central
tendency given rl'> the value
distribution. The tnode is a good measure of location when the variable i8 inherently categorical
or has otherwise been grouped into categories. The mode in Tahle 15.2 is 6.000.
s = Vi-2_1-;;-=1·---
that occurs the most in a
sarnp\e d1stribut1on MEDIAN The median of a sample is the middle value when the data are arranged in ascending we divide by n - 1 instead of n because the sample is drawn from a population and we arc
or descending order. lf the number of data pointi;; is even, the median is usually estimated as the lrying to determine how much the responses vary from the mean of the entire population.
median However, the population mean is unknown; therefore the sample mean is used instead. The use
midpoint between the lWl) middle values-by adding the two middle values and dividing their
A measure of centr a\
sum by 2. The mcJ.ian is the 50th percentile. The median is an appropriale measure of central of the sample mean makes the sample seem less variable than it really is. By dividing by n. - I,
tendency given ac; the value
tendency ror ordinal data. In Table 15.2, the median is 5.000. instead of n, we compensate for the smaller variability observed in the sample. for the data given
above which half of the in Table 15. 2, the vari<1ncc is calculated as follows:
value':> f.1\I and below which For the data in Table l 5.1, the three measures of central tendency for tbis distribution are
hall o! the values ta\\ different (mean= 4.724, mode= 6.000, median = 5.000). This is not smp1ising, because each 2
measure defines central tendency in a different way. So which measure should be used? If the
/2 X (2 - 4.724) + 6 X (3 - 4.724)2 + 6 X (4 - 4.724)2 + 3 X (5 4.724)2
variable is measured on a nominal scale, the mode should be used. [f the variable is measured on 2 + 8 X (6 - 4.724) 2 + 4 X (7 -- 4.724)2/
,Y = -~- --~---~---~--·--·-- -
an ordinal scale, the median i.s appropriate. If the variable is measured on an interval or ratio 28
scale, the mode is a poor measure of central tendency. This can be seen from Table 15.2. 14.840 + 17.833 + 3.145 + 0.229 I 13.025 + 20.721 J
- - - - - - - - - - ~ ----- ---·--
Although the modal value of 6.000 has the highest frequency, it represents only 27 .6 percent of 28
the sample. In general, for interval or ratio data, the median is a better measure of central 69.793
tendency, although it too ignores available information about the variable. The actual values of 28
lhe variable above and below the median are ignored. The mean is the most appropriate measure
= 2.493
 CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 489
488 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

positive, then the distribution is more peaked than a normal <lislribution. A negative
The standard deviation, therefore, is calculated as:
value means that the distribution is flatter than a normal distribution. The value of this statistic
s = \!2.493 for Table 15.2 is -1.261, indicating that the distribution is flatter than a normal distribution.
1.579 Measures of shape are important, because if a distrihution is highly skewed or markedly peaked
COEFFICIENT OF VARIATION The coefficient of variation is the ratio of the standard deviation or flat, then statistical procedures that assume normality should be used with caution.
coefficient of
variation to the mean, expressed as a percentage, and it is a unitless measure of relative variability. The
A 1J<;eful expression 1n coefficient of variation, CV, is expressed as:
ACTIVE RESEARCH
sampling theory for the s ---~---- --~--------·---~~-- -------------------~---
standard deviation as a CV==
X Wendy's Customers: Who Are the Heavyweights?
percentdge of the mean
The coefficient of variation is meaningful only if the variable is measured on a ratio scale. It Visit www.wendys.com and conduct an Internet search using a search engine and your library's online
remains unchanged if all the data values are multiplied by a constant. Because familiarity with dalabase to obtain information on the heavy users of fast-food restaurants.
the Internet is not measured on a ratio scale, it is not meaningful to calculate the coefficient of As the marketing director for Wendy's. how would you target the heavy users of fast-food restaurants?
variation for the data in Table 15.2. From a managerial viewpoint, measures of variability are In a survey for Wendy's, information was obtained on the number of visits to Wendy's per month.
How would you identify the heavy users of Wendy's and what statistics would you compute to summarize
important because if a characteristic shows good variability, then perhaps the market could be
the number of visits to Wendy's per month?
segmented based on that characteristic.

Measures of Shape Introduction to Hypothesis Testing

[n addition to measures of variability, measures of shape are also useful in understanding the nature Basic analysis invariably involves some hypothesis testing. Examples of hypotheses generated in
of the distribution. The shape of a distribution is assessed hy examining skewness and kurtosis.
marketing research abound:
SKEWNESS Distributions can be either symmetric or skewed. In a symmetric distribution, the values
• The department store is being patronized by more than l O percent of the households.
skewness on either side of the center of the distribution are the same, and the mean, mode, and median are
• The heavy and light users of a brand differ in terms of psychographic characteristics.
A charactenst1c of a equal. The positive and corresponding negative deviations from the mean are also equal. In a skewed
d1stnbut1on that assesses its • One hotel has a more upscale image than its close competitor.
distribution, the positive and negative devialions from the mean are unequal. Skewness is the • Familiarity with a restaurant results in greater preference for that restaurant.
symmetry about the mean.
tendency of the deviations from the mean to be larger in one direction than in the other. It can be
thought of as the tendency for one tail of the distribution to be heavier than the other (see Figure 15.2). Chapter 12 covered the concepts of the sampling distribution, standard errm of the mean or
kurtosis
A rneasure of the relative The skewness value for the dala ofTahle 15.2 is -0Jl94, indicating a slight negative skew. the proportion, and the confidence interval. 5 All these concepts arc relevant to hypothesis tc:-;ting
and shoul<l be reviewed. Now we describe a general procedure for hypPthesis testing that can he
peab~dness or flrttness of
the curve defined by the
KURTOSIS of
Kurtosis is a measure tne relative peakedness or flatness of the curve defined by applit:.·d to test hypotheses about a wide range of p~rameters.
fr{~quency distribution
the frequenl·y distrihution, The kurtosis of a normal distribution is zero. If the kurtosis is

Symmetric Di!')tribution
A General Procedure for Hypothesis Testing
F'IGUltE 15.2
The following steps are involved in hypothesis testing (Figure 15.J).
Skewness of a
Distribution 1. formulate the null hypothesis H0 and the alternative hypothesis H,.
2. Select an appropriate statis!ical technique and the corresponding test statistic.
3. Choose the ll!vel of significance, a.
4. Determine the sample size and coJlect the data. Calculate the value of the test statistic.
5. Determine the probability associated with the test statistic under the null hypothesis, using
the sampling distribution of the test statistic Alternatively, dctennine the critical value~
associated with the test statistic tllat divide the rejection and nonri.::jcction regions.
6. Compare the probability associated with the test statistic with the level of significance
Mean
Median specified. Alternativdy, determine whether the test statistic has fallen into the rejection or
Mode the nonrejcction region.
(a) 7. Make the statistical decision to reject or not reject the null hypothesis.
null hypothesis
8. Express the statistical <leci~ion in terms of the marketing research prohlcm.
Skewed Distrihution A staternent in which no
difference or effect 1s
expected. If the null
Step 1: Formulate the Hypotheses
hypothesis l'.i not rejected, The lirst step is t" formulate the null and alternative hypotheses. A null hypothesis is a state-
110 changes will be made ment of the status <.JlH\ one of no difference or no effect. If the null hypothesis is not rejected, no
changes will he made. An alternative hypothesis is one in which some difference or effect is
alternative hypotf1esis
expected. Ac.:ceptjng the alternative hypothesis will lead to changes in opinion1- or actions. Thus.
A statement thdt some
difference or eftect 1s
the alternative hypothesis is the opposite of the null hypothesis.
expected. Accepting the The null hypothesis is always the hypllthesis that is tested. The null hypothesis refers lo a
alternative hypothesis will specified value of the population parameter (e.g.,µ, u, 1r), not a sample statistic (e.g., X, s, p).
Mean Median Mode le.Jd to changes in opinions A null hypothesis may be rejected, but it can never be accepted based on a single test. A statistical
(b) or actions. test can have one of two outcomes. One is that the null hypothesis is rejected and the alternative
 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 491
490
Step 2: Select an Appropriate Test
FIGURE 15.3 Step l Fo;mulate Ho and Ht·
To test the null hypothesis, it is necessary to select an appropriate statistical technique. The
A General i researcher should take into consideration how the test statistic is computed and the sampling distrib-
Procedure for Step 2 Select an appropriate test. test statistic ution that the sample statistic (e.g., lhe mean) follows. The test statistic measures how close the
Hypothesis Testing
i A measure of how dose the
sample has come to the null
sample has come to the null hypothesis. The test statistic often follows a well-known distribution,
Choose the level of significance', a. such as the normal, t, or chi-square distribution. Guidelines for selecting an appropriate test or statis-
Step 3
i ... hypothesis. It often follows tical technique are discussed later in this chapter. In our example, the 2 statistic, which follows the
a well-known distribution. standard normal distribution, would be appropriate. This statistic would be computed as follows:
Collect data and calculate tile such as the normal, t, or
Step 4 p -
test statistic, TScAL·;, chi-square distribution. 1f
z=---
<rp

I where

De~r:hine. th~·probability
l
. Det~r~i~e°th~ ~ijtic~l
er =
fJ
f;{i_ =_,,;.)
v~--n
Step 5
associated with the :value ofthe,iest, Step 3: Choose Level of Significance, a

,· r·
test statistic; TScAL- statisiic;TS<;:R.
i Type I error
Whenever we draw inferences ahout a population, there is a risk that an incorred cone lusion wiJI
be reached. Two types of errors can occur.

Step 6 D~t;;:niirie ifTScAL TYPE I ERROR Type I error occurs when the sample results lead to the rejection of the null
Com~are pro~a~ilit; · Also known as alpha error,
falls intr rejection ·. hypothesis when it is in fact true. In our example, a Type I error would occur if we concluded, based
· with level · .w nonrejection. it occurs when the sample
, ,,,of Signifi~anc~, ·a. results lead to the reJectton on the sample data, that the proportion of customers preferring the new service plan was greater
<rel!ion;·
. f of a null hypothesis that 1s
in fact true
than 0.40, when in fact it was less than or equal to 0.40. The probability of Type I error (a) is also
called the level of significance. The Type l error is controlled hy establishing the tolerable level of
~r dp~ot reje~t Ho, risk of rejecting a true null hypothesis. The selection of a particular risk level should depend ou the
Step 7
. i . . level of significance
The probability of making d
cost of making a Type I error.
Type I error TYPE II ERROR Type II error occurs when, based on the sample results, the null hypothesis is
Step 8 .••. 'bra\'/ a :n·arketing research 'tonclusiort.
not rejected when it is in fact false. In our example, the Type JI error would occur if we
Type II error
condudcd, based on sample data, that the proportion uf customers preferring the new service
Also known as beta error,
it occurs when the sample
plan was less than or equal to 0.40 when, in fact, it was greater than 0.40. The probability of
hypothesis is accepted. The other outcome is that the null hypothesis is not rejected based on the
evidence. However, it would he incorrect to conclude that because the nuJI hypothesis is not results lead to the Type ll error is denoted by {3. Unlike a, which is specified by the rese.ircher, the magnitude
rejected, it can be accepted as valid. Jn classical hypothesis testing, there is no way to determine nonreject1on of a null of f3 depends on the actual value of the population parameter (proportion). The probability of
hypothesis that 1s in fact Type I error (c,) and the probability of Type II error (/3) are shown in Figure I 5.4. The
whether the null hypothesis is true.
In marketing research, the null hypothesis is formulated in such a way that its rejection leads false. complement (I - f3) of the probability of a Type 11 error is called the power ofa statistical te.rt.
to the acceptance nf thl'. desired conclusion. The alternative hypothesis represents the conclusion
for which evidence i~ sought. For example, a major department store is considering tbe introduc-
tion of an Internet shopping service. The new service will be introduced if more than 40 percent FIGURE 15.4
of the Internet users shop via the Internet. The appropriate way to formulate the hypotheses is:
Type I Error (a) and
Ho: 1r s 0.40 Type II Error (/3)
95%of
Hr: 1r > 0.40 Total Area

If the null hypothesis H is rejected, then the alternative hypothesis H 1 will be accepted and the c,= 0.05
0
new Internet shopping service will he introduced. On the other hand, if H 0 is not rejected, then
the new service should not be introduced unless ad<litional evjdence is obtained.
This test of the null hypothesis is a one-tailed test, because the alternative hypothesis is =0.40
one-tailed test
A te'..t of the null hypothesis expressed directionally: The proportion of Internet users who use the Internet for shopping is ~ Zn= l.645
where the alternative greater than 0.40. On the other hand, suppose the researcher wanted to determine whether the pro-
hypothe'.>IS 1<:. expressed portion of Internet users who shop via the Internet is different from 40 percent. Then a two-tailed Critical Value
d1rertionc1lly test would be required, and the hypotheses would be expressed as: of z

two-tailed test Ho: 1r = 0.40

A test of the null hypothesis Ht: 1r -fc 0.40 fi= 0.01
whNe thP dlternat1ve
hypothesis 1s not expressed In commercial marketing research, the one-tailed test is used more often than a two-tailed test.
directionally Typically, there is some preferred direction for the conclusion for which evidence is sought. = 0.45
For example, the higher the profits, sales, and product quality, the better. The one-tailed test is more
powerful than the two-tailed test. The power of a statistical test is discussed fu1ther in step 3.
'fl= -2.330 4
 CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 493
492 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

significance of 0.05. Hence, the null hypothesis is rejected. Alternatively, the calculated value of
power of ·1e,,t POWER OF A TEST The power of a test is the probability (l ~ {3) of rejecting the null
the test statistic z = 1.88 lies in the rejection region, beyond the value of 1.645. Again, the same
The probability of reJecting hypotl1esis when it is false and should be rejected. Although {3 is unknown, it is related to a. conclusion to reject the null hypothesis is reached. Note that the two ways of testing the null
the null hypothesis when it
An extremely low value of a (e.g., = 0.00 I) will result in intolerably high {3 errors. So it is hypothesis are equivalent but mathematically opposite in the direction of comparison. If the
is in fact false and ':.hould
be reiected necessary to balance the two types of errors. As a compromise, a is often set at 0.05; sometimes probability associated with the calculated or observed value of the test statistic (TSCAt) is less
it is 0.0 I; other values of a are rare. The level of a along with the sample size will determine the than the level of significance (a), the null hypothesis is rejected. However, if the absolute value
level of {3 for a particular research design. The risk of both a and {3 can be controlled by of the calculated value of the test statistic is greater 1han the absolute value of the critical value
increasing the sample size. For a given level of a, increasing the sample size will decrease {3, of the test statistic (TScR), the null hypothesis is rejected. The reason for this sign shift is that
thereby increasing the power ot' the test. the larger the absolute value of TSCAL' the smaller the probability of obtaining a more extreme
value of the test statistic under the null hypothesis. This sign shift can be easily seen:
Step 4: Collect Data and Calculate Test Statistic if probability ofTSCAL < significance level (a), then reject l/ 0 ,
Sample size is determined after taking into account the desired a and /3 errors and other qualita-
but
tive considerati,lns, such as budget constraints. Then the required data are collected and the value
of the test statistic computed. In our example, 30 users were surveyed and 17 indicated that they if ITSCALI > ITSCRI, thenrejectH0 .
used the Internet for shopping. Thus the value of the sample proportion is p = l 7 /30 = 0.567.
The value of up can be determined as follows:
Step 8: Marketing Research Conclusion
"P = JJr_(i/~}i The conclusion reached by hypothesis testing must be expressed in terms of the 1narkcting
research problem. In our example, we conclude that there is evidence that the proportion uf
Internet users who shop via the Internet is significantly greater than 0.40. Hence, the recommen-
= no~~))(0-:-60)
dation to the department store would be to introduce the new Internet shopping service.
'y 30
As can be seen from Figure 15.6, hypotheses testing ca11 be related to either an ~xamination
= 0.089 of associations or an examination of differences. ln tests of associations, the null hypothesis is
that there is no association he~ween the variables (l/0 : .. is NOT related to ... ). ln tests of dif-
The test statistic z can he calculated as follows: ferences, the null hypothesis is that there is no difference (H0 : ... is NOT different from ... ).
p - 7T: Tests of differences could relate to distrihutions, means, proportions, medians, or rankings. First,
2 ~
we discuss hypotheses related to associations in the context of cross·tahulations.
{TJJ

0.567 - 0.40
-~--~-
0.08') Cross-Tabulations
Although answers to questions related to a single variable .u·e interesting, they often raise additional
1.88 questions about how to link that variable to other variables. To introduce the h·equen<.:y distribution,
we posed several representative marketing research questions. For each of these, a researcher might
Step 5: Determine the Probability (or Critical Value) pose additional questions to relate these variables to other variables. For example:
Using standard normal tables (Table 2 of the Statistical Appendix), the prohabilily of ohtaining a cross-tabulation
z value of 1.88 can be calculated (see Figure 15.5). The shaded area between - <XJ and 1.88 is A stat1st1cal technique that • How many bran<l~loyal users are n1ales?
0.9699. Thcref'ore, the area to the right of z = 1.8H is 1.0000 - 0.9699 = 0.0301. This is also describes two or more • ls product use (mca:mre<l in terms of heavy users, medium usen,, light users, and nonusers)
variable~ simultaneously and related to interest in outdoor activities (high, medium, and low)?
p value called the p value and is the probability of observing a value of the test statisti(.'. as extreme a.s, or
This 1s the rrobc1bil1ty uf
results in tables that reflect • [s familiarity with a new product re.lated to age and education levels?
more extreme than, the value actually observed, assurning that the null hypothesis is true.
the Joint d1st1ibution of two • Is product ownership related to income (high, medium, and low)?
observing c1 vc1lue of the Alternatively, the critical value of z, which will give an area to the right si<le of the critical
test stat1st1c a5 extreme as, or more variables that have
value of (l.05, is between 1.64 and 1.65 and equals 1.645. Note that in determining the critical a limited number of The answers to such questions can he determined by examining cross-tabulations. Whereas
or more extreme them, the value of the test statistic, the area in the tail beyond the critic:al value i.s either (X or o:/2. lt is£~ for
v,,lue actually observed, categories or distinct values. a frequency distribution describes one variable at a time, a cross-tabulation dcscri bes two or
a one-tailed test and a/2 for a two-tailed test.
a',sum1nq that the null
hyµothes1s 1s true
Steps 6 and 7: Compare the Probability (or Critical Value) and Make the Decision FIGURE 15.6

~ --
The probability associated with the calculated or observed value of the lest statistic is 0.030 I. A Broad
This is the prohahility of gettmg a p value of 0.567 when 7T = 0.40. This is Jess than the level of Classification of
Hypothesis Tests r------ i
1-IGWH: '15.5 Tests of Tests of
Associati-0n Differences
Probability of z
with a One·Tailed Shaded Area Unshaded Area
Test = 0.9699 = 0.030 I

L l l i
Medians/
Distributions Means Proportions
Rankings
0 z = 1.88
494 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 495

SPSS Output File

Sex and Internet Usage

Internet Usage
Light (l)
Heavy (2)
Male
5
10
Sex
---
Female
10
5
"'

Row Total
15
15
..-. .

II
SPSS Output File
Internet

Sex
Male
Female
by Sex

Light
33.3%
66.7%
Internet Usage
Heavy
66.7%
33.Yi<J
Total
100.0%
100.0%

§.sas
15

§.sas
Column totals 15
------------ Which of these tables is more useful? The answer depends on which variable will be considered
as the independent variable and which as the dependent variable. The general rule is to compute
more variables simultaneously. A cross-tabulation is the merging of the frequency distribution of the percentages in the direction of the independent variable, across the dependent variable. In our
SAS Output File two or more variables in a single table. It helps us to understand how one variable sueh as brand SAS Output File
analysis, sex may be considered as the independent variable and Internet usage as the dependent
loyalty relates to another variable such as sex. Cross-tabulation results in tables that reflect the variable, and the correct way of calculating percentages is as shown in Table 15.4. Note that
joint Jistribution of two or more variables with a limited number of categories or distinct values. whereas 66.7 percent of the males arc heavy users, only 33.3 percent of females fall into this
The categories of one variable are cross-classified with the categories of one or more other category. This seems to indicate that males are more likely to be heavy users of the Internet as
variables. Thus, the fre4uenc.:y distribution of one variable is subdivided acconling to the values compared to females.
or categories of the other variables. Nute that computing percentages in the direction of the dependent variable across the indepen-
Suppose we are interested in determining whether lntemet usage is related to sex. For the dent variable, as shown in Table 15.5, is not meaningful in this case. Table 15.5 implies that heavy
purpose of cross-tahulation, respondents are classified as light or heavy users. Those reporting Internet usage causes people to he males. This latter finding is implausible. It is possihle, however,
S hourn or les,-. usage are classified as light users, and the remaining are heavy users. The cross- that the association between Internet usage and sex is medialed by a third variahle, such as age or
tabulation is shown in Table 15.3. A cross-tabulation inc1udes a cell for every combination of the income. This kind of possibility points to the need to examine the effect of a third vmiahle.
categories of the two variahles. The number in each cell shows how many respondents gave that
combination of responses. In Table 15.3, LO respondents were females who reported light Three Variables
lnternet usage. The marginal totals in this table indicate that of the 30 respondents with valid Often the introduction of a third variable clarifies the initial association (or lack of it) obscrvcJ
responses on both the variables, 15 reported light usage and 15 were heavy users. In terms of sex, between two variables. As shown in Figure 15.7, the introduction of a third variable can result -in
15 respondents were females and l 5 were males. Note that this information could have been four possibilities.
ohtained from a separate frequency <listrihution for each variable. In general, the margins of a
cross-tabulation show the same information as the frequency tables for each of the variables. 1. It can refine the association ohserved hetwecn the two original variables.
contingency t~ble Cross-tabulation tables are also called contingency tables. The data arc considered to he qualita- 2. It can indicate no association between the two variables, although an association was
A cro':i':i~tabulation table It tive or categorical data, because each variable is assumed to have only a nominal scalc. 6 initia11y observed. In other words, the third variahle indicates that the initial association
contain) a cell tor every Cross-tabulation is widely used in co111mercial marketing research, because ( 1) cro!:i,'--tabu \ation between the two variables was spurious.
c.ornbinat1or1 of categorie':i analysis and results can be easily interpreted and understood by managers who arc not statistically 3. It can reveal some association hetween the two variables, although no association was
uf thP two vanables oriented; (2) the clarity of interpretation provides a stronger link between research results and initially ob,r.;erved. In this case, the third variable reveals a ,-.upprcs.scd associatinn between
managerial action; (3) a series of cross-tabulations may provide greater insights into a complex the first two variables: a suppressor effect.
phenomenon than a single multivariate analysis; (4) cross-tabulation may alleviate the problem of 4. lt can indicate no change in the initial association.8
sparse cells, which could he serious in discrete multivariate analysis; and (5) cross-tabulation These cases arc exphiine.J with examples based on a sample of l,000 resrondcnts. Although
analysis is simple to conduct and appealing to less sophisticated researchers. 7 these examples arc contrived to illustrate spec.;ific <.:ases, such cases arc not uncona11on in
C<)mn1ercial marketing tl:search.
Two Variables
Cross-tahulation with two variahles is also known as bivariate cross-tuhulation. Consider again the FIGURE 15.1 rlgi1111l Two armbles
cross-classilication of Internet usage with sex given in Table 15.3. Is usage related to sex? lt appears
The Introduction of
to be from lable 15.3. We sec that disproportionately more of the respondents who me males are
a Third Variable in 1
heavy Internet users as compared to females. Computation of percentages can provide more insights.
Cross-Tabulation J --i

II
Because two variables have been cross-classified, percentages could be computed either
colurnnwise, based on column totals (Table 15.4), or rowwise, based on row totals (Table 15.5). Some Association No Association
Between the Two Between the 1\vo
Variables Variables
Ffi,:J:f~
Sex by Internet Usage
J J
Introduce a Introduce a
Sex Third Variable Third Variable
SPSS Output File
Internet Usage Male Female
J l r-1---i
sas
SAS Output File
Light
Heavy
Column totals
33.3%
66.7%
100.0%
66.7%
33.3%
1000%
Refined Association
Between the Tua
Original Variables
No Association
Between the 1\vo
Original Variables
No Change in
the Initial
Pattern
Some Association
Between the Two
Original Variables
496 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 497

II rn:l{,..,J'l,
Purchase of Fashion Clothing by Marital Status
Current Marital Status
Ownership of Expensive Automobiles by Education and Income Levels
.,.

Income
_______ ____ ,_

Purchase of Fashion Clothing Married Unmarried Low Income Education High Income Education
High 31% 52% Own Expensive College No College College No College
Low 69% 48% Automobile Degree Degree Degree Degree
Column totals 100% 100% Yes 20% 20cM) 40% 40%
Number of respondents 700 300 No 80% 80% 60% 60%
Column totals 100% 100% 100% 100%
Nurnher of respondents 100 700 150 50
REFINE AN INITIAL RELATIONSHIP An examination of the relalionship between the purchase of
fashion clothing and marital status resulted in lhe data reported in Table 15.6. The respondents
were classified into either high or low categories based on their purchase of fashion clothing.
with college degrees own an expensive automobile, as compared to 21 percent of those without
Marital status was also measured in terms of two categories: currently married or unmarried. As
college degrees. The researcher was tempted to conclude that education intlucnced ownership
can be seen from Table 15.6, 52 percent of unmarried respondents fell in the high-purchase
of expensive automobiles. Realizing that income may also be a factor, the researcher decided to
category, as opposed to 31 percent of the married respondents. Before concluding that unmarried
reexamine Lhe relationship between education and ownership of expensive automobiles in light
respondents purchase more fashion clothing than those who are married, a third variable, the
of income level. This resulted in Table 15.9. Note that the percentages of those with and without
buyer's sex, wa.,;; introduced into the analysi8.
coHcgc degrees who own expensive automobiles are the same for each of the income groups.
The buyer's sex was selected as the third variable based on past research. The relationship
When the data for the high-income and low-income groups are examined separately, the
between purchase of fashion clothing and marital status was reexamined in light of the third vari-
association between education and ownership of expensive autornnhiles disappears, indicating
able, as shown in Table 15.7. In the case of females, 60 percent of the unmaJTied fall in the high-
that the initial relationship observed between these two variables was spurious.
purchase category, as compared to 25 percent of those who are married. On the other hand, the
percentages are n1uch closer for males, with 40 percent of the unmarried and 35 percent of the REVEAL SUPPRESSED ASSOCIATION A researcher suspected desire to travel nbroad may be
married falling in lhe high-purchase category. Hence, the introduction of sex (third variable) has influenced hy age. However, a cross-tabulation of the two variables produced lhe results in
refined the relationship between marital status and purchase of fashion clothing (original Table 15.10, indicating no association. When sex was introduced as the third variable,
variables). Unmarried respondents are more likely to fall in the high-purchase category than Table 15.11 was obtained. Among men, 60 percent of those under 45 indicated a desire to
married ones, and this effect is much more pronounced for females than for males. travel ahroaLl, as compared to 40 perccut of those 45 or older. The pattern was reverse<l for
women, where 35 percent of those under 45 indicated a desire to travel abroad, as opposed to
INITIAL RELATIONSHIP WAS SPURIOUS A researcher working for an advertising agency
65 percent of those 45 or ol<lcr. Because the association between desire to travel abroad and
promoting a line of autrnnobifos co.-.ting more than $30,000 was attempting to explain the
age runs in the· opposite direction for males and females, the relationship hetwcen these two
ownership or expensive automobiles (see Table 15.8). The table shows that 32 percent of those
variables is masked when the data are aggregated across ~ex, as in Table 15.10. But when the

$
Purchase of Fashion Clothing by Marital Status and Sex ··--------
Desire to Travel Abroad by Age
Sex Age
Male Marital Status Female Marital Status Desire to Travel Abroad Less Than 45 45 or More
Purchase of Fashion Clothing "'''-'"'
Marrio:,.rl llnm::irrit:1,'"1
. . . ,,,,, .... ,, .... ,.... 1\/1::i.rrit:1,rl
.................... Unmarried Yes 50% .10'111
High 35<x) 40% 25% 60% No 50% )0%
Low 65% 60% 75% 40% Column totals IOO'fh 1oon4J
Column totals l ()()'Yo 100% 100% \()()<l(J Number of respondent~ 500 soo
Numher of cases 400 ]20 3110 180 ------------------·-------

m:Jl~: Desire to Travel Abroad by Age and Sex

Ownership of Expensive Automobiles by Education Level Sex
Education Male Age Female Age
Own Expensive Automobile College Degree No College Degree Desire to Travel Abroad <45 ,,;;45 <45 ;e:45
Yes 32% 21°.kJ
Yes 60'-l<; 40<:0 35% 65%
No 68% 79%
No 40% 6fVYv 6:i% 35%
Column totals 100% \ll0%
Column totals 100% [()()0/(J 100% I00 1fti
Number of cases 250 7SO
Number of cases 300 300 201) 20(1

--------------
498 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOlHESIS TESTING 499

degree of association, is important from a practical or substantive perspective. Generally, the

strength of association is of interest only if the association is statistically significant. The
r;;,:!:Frequen.tly in "._::t-foo~ Restaurants by Family~ strength of the association can be measured by the phi correlation coefficient, the contingency
Family Size coefficient, Cramer's V, and the lambda coefficient.
Eat Frequently in Fast-Food Restaurants Small Large
Chi-Square
Yes 65% 65% 2
chi-square statistic The chi-square statistic (X ) is used to test the statistical significance of the observed association
No 35% 35% The statistic used to test the in a cross-tabulation, It assists us in determining whether a systematic association exists between
Column total.s 100% 100% statistical significance ot the the two variables. The null hypothesis, H 0 , is that there is no association between the variables.
Number of cases 500 500 observed association in a The test is conducted by computing the cell frequencies that would be expected if no association
cross-tabulation. It assists us were present between the variables, given the existing row and column totals. These expected cell
in determining whether a
frequencies, denoted./~. are then compared to the actual observed frequencies,.(,, found in
systematic cissonation exists
the cross-tabulation to calculate the chi-square statistic. The greater the discrepancies between

r
between the two variables
$ the expected and actual frequencies, the larger the value of the statistic. Assume that a cross-
Eating Frequently in Fast-Food Restaurants by Family Size and Income tabulation has r rows an<l c columns and a random sample of n observations. Then the expected
frequency for each cell can be calculated by using a simple formula:
Income
Low Income High Income n,- nc
1; = --;,--
Family Size Fa~ily Size
Eat Frequently in where
Fast-Food Restaurants Small Large Small Large
nr = total number in the row
65% 65% 65% 65%
nc = total number in the cDlurnn
35% 35% 35% 35%
' n = total sample size
100% JOO% 100% 100%
250 25(1 250 250 For the Internet usage data in Table 15.3, the expected frequencies for the cells, going from
Jeft to right and from top to bottom, are:

15 X -~5 = 7.50 15 X IS = 7.50

30 30
effect of sex is controlled, as in Table 15. 11, the suppressed associatinn between desire to
travL:I abroad and age is revealed ftlr the separate categories of males anJ females.
15 X 15 = 7.50 15 X .!2 = 7.'itl
30 ~
NO CHANGE IN INITIAL RELATIONSHIP In some cases, the introduction of the third variable
Then the value of X2 is calculated as follows:
does not change the iuitial relatiqnship oh.served, regardless uf whether the original variables
were associated. This .suggi:sts that the third variable does not influence the relationship
hetween the first two. Consider the cross-tabulation of family size and the tendency to eat out
X2 = 22 <ct.,_=-lJ__2
frequently in fast-food restaurants, as shown in T.-tble 15.12. The respondents were classified c~\Js le
into small and large family size categories based on a mct.lian split of the distrihution, with
500 respondents in each category. No association is observed. The respondents were further For the data in Table 15 .3, the value of X2 is calculated as:
classified into high- or low-income groups baset.l on a median split. When income was
introduced as a third variable in the analysis, Tahlc 15.13 was nhtained. Again, no association (5 -- 7.5)' (10 - 7.5)2 (10 - 5 7.5)2
X2 =---------+-----------+ ----- +
was observnl. 7.5 7.5 7.5 7.5
= 0.833 + U.833 + 0.833 + 0.8:l3
General Comments on Cross-Tabulation
= 3.333
More than three variables can be cross-tahulated, hut the interpretation is quite complex. Also,
because the number of cells increases multiplicatively, maintaining an adequate number of To determine whether a systematic a.ssociation exists, the probability of obtaining a value of
respondents or cases in each cell can be problematic. As a general rule, there should be at least chi-square as large as or larger than the one calculated. from the crm,s-tabulation is csti1nale<l.
five expected observations in each cell for the statistics computed to be reliable. Thus, cross- An important characteristic of the chi-square statistic is the number (Jf degrees of freedom (df) asso-
tabulation is an inefficient way of examining relationships when there are several variables. ciated with it. ln general, the number of degrees of freedom is cqlrnl to the number of observntions
chi-square distribution less the number of constraints needed to cakulatc a l'Jtatistical term. In the case of a chi-square statis-
Note that cross-tabulation examines association between varlables, not causation. To examine A skewed d1stnbut1on
causation, the causal research design framework should he adopted (see Chapter 7). tic associated with a cross-tabulation, the number of degrees of freedom i~ equal to the product of
whose '>hape depends solely
number of row~ (r) less one and the nmnber of columns (c) less one. That is, df = (r - 1) X
on the number of degrees
of freedom. As the number
(c - I ). 9 The null hypothesis (H0) of no association between the two variables will be rejected only
Statistics Associated with Cross-Tabulation of degrees of freedom when the calculated value of the test statistic is greater lhan tlie ciitical value of the chi-square distri-
We will discuss the statistics commonly used for assessing the statistical significance and increases, the chi-square bution with the appn)priate degrees of freedc>m, as shown in Figure I S.S.
strength of association of cross-tabulated variables. The statistica1 significance of the observed distribution becomes more The chi~square distribution is a skewed distrihution whose shape depends solely on the
association is commonly measured by the chi-square statistic. The strength of association, or symmetrical. number of degrees of freedom. 10 As the number of degrees of freedom incr~ases, the chi-square.
 500 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 501

HGmu: ·1s.n phi value. However, for the purpose of illustration, we show how the values of phi an<l other
Chi-Square Test measures of the strength of association would be computed. The value of phi is:
of Association
Do Not cp = l\1?
= 0.333
Reject H0
Thus, the association is not very strong. [n the more general case involvrng a table Qf any size,
- the strength of association can be assessed by using the contingency coefficient.
Critical x2
Value Contingency Coefficient
conting~ncy Whereas the phi coefficient is specific to a 2 X 2 table, the contingency coefficient (C) can he
coefficient (C) used to assess the strength of association in a table of any size. This index is also related to chi-
A mea~ure of the strength square, as follows:
distribution becumes more symmetrical. Table 3 in the Statistical Appendix contains upper-tail
areas of the chi-square distnbution for different degrees of freedom. In this table, tbe value at the
top of each column indicates the area in the upper pmtion (tbe right si<le, as shown in Figure 15.8)
of association 1n a table of
any size
c ' J·---~-2--
= ~--·-
ot the chi-square distribution. To illustrate, for I degree of freedom, the value for an upper-tail
x2 + II

area of 0.05 is 3.841. This indicates that for I degree of freedom, the probability of exceeding a The contingency coefficient varies between O and I. The O value occurs in the case of no associ-
chi-square value of 3.841 is 0.05. In other words, at the 0.05 level of significance with J degree of ation (i.e., the variables are statistically independent), Out the maximum value of I is never
freedom, the critical value of the chi-square statistic is 3.841. achieved. Rather, the maximum value of the contingency coefficient depends on the size of the
For the cross-tabulation given in Table 15 .3, there are (2 - I) X (2 - I) = I degree of table (number of rows and number of columns). For this reason, it should be used only tu
free,Jom. The calculated chi-square statistic had a value of :1.333. Because this is Jess than the compare tables of the same size. The vnluc of the contingency coefficient for Table 15.3 is:
critical value of 3.841, the null hypothesis of no association cannot he rejected, indicating that
the association is not statistically significant at the (l.05 level. Note that this lack of /3.333
significance is mainly due to the small sample size (30). If, instead, the sample size were
C = 'y D33+-30
300 and each entry ofTahle 15.3 were multiplied by 10, it can be seen that the value of the chi- = 0.316
square statistic would he multiplied by 10 and would he 33.33, which is significant at the 0.05
This value of C indicates that the association is not very strong. Another statistic that can he
level.
calculated for any table is Cramer's V.
The chi-square statistic can also be used in goodness-of-fit tests to determine whether
certain models fit the oh~erved data. These tests are conducted by calculating the significance
Cramer's V
of sample deviations from assumed theoretical (expected) distributions, and can be
Cramer's V Cramer's Vis a modified version of the phi correlation coefficient,(/), and is used in tables !urger
performed on cross-tabulations as well as on frequencies (one-way tabulations). The calcula-
A measure of the strength than 2 X 2. When phi is calculated for a tahle larger than 2 X 2, it has nu upper limit. Cramer's Vis
tion of the chi-square statistic and the determination of its significance is the same as of assoc1at1or1 u~ed in tables
ii Just rated here. obtained by a<ljw;ting phi for either the number of rows or the number nf column.sin the tahle,
larger than 2 x 2.
The chi-square statistic should be estimated only on counts of data. When the data are in based on which nfthe two is smaller. The adjustment is such that Vwill range frnm II to I. A large
percentage form, they should first be converted to absolute counts or numbers. ln addition, an value of V merely indicates a high degree of association. 1t dues not indicatl.': how the variables are
assuciateU. F(lf a tabit! with r rows and c columns, tht; relationship bctwt.:en Crnmcr's Vand the phi
underlying assumption of the chi-square test is that the observations arc drawn independently. As
a general rule, chi-square analysis should not be conducted when the expected or theoretical crnTelatic)n cc1cfficicnt is expressed as:
----~-----
,,,2
frequencies in any of the cells is less than five. If the number of observations in any cell is Jess
than IO, or if the table has two rows and two columns (a 2 X 2 tahle), a correction factor should
be applied. 11 With the correction factor, the value is 2.133, which is not significant at the
or
V=
& (r - 1),(c - I)

0.05 level. The calculation of the correction factor is complex but can be conveniently done using
appropriate software. In the case of a 2 X 2 table, the chi-sL1uare is related to the phi coefficient.
V- r-_-X2i11 __ ·_
Phi Coefficient \J;nin(r - I), (c -- I)

phi coe!iicient The phi coefficient (c/>) is used as a measure of the strength of association in the special case of The value of Cramer's V for Tahle l 5.3 is:
A measure of the strength a table with two rows and two columns (a 2 X 2 tahle). The phi coenicicnt is proportional to the
of assoc1at1on in the speC1al
case of a table with two
square root of the chi-square statistic. For a sample of size n. this statistic is calculated as: V = J3 if3~1<2
rowc; and two column'.>
(ci 2 x 2 table)
,1,=fJ asymmetric lanibda
A measure of the
percentage 1mprovemer1t 1n
= 0.333

Thus, the association is not very strong. As can be seen, in this case V =</>.This is always the
predicting the value of the
It takes the value of O when there is no association, which would be indicated by a chi-square case for a 2 X 2 table. Another statistic commonly estimateJ is the lamh<la coefficient.
depenc:ler1t variable, given
value of Oas well. When the vaiiables are perfectly associated, phi assumes the value of I and all the value of the independent
the observations fall just on the main or minor diagonal. (In some computer programs, phi variable 1n contingency tab!e
Lambda Coefficient
assmnes a value of ·~ I rather than 1 when there is perfect negative association.) In our case, analysis Larnhda also vanes Lambda assumes that the variables are measured on a nominal scale. Asytnmetric lambda mea~ures
because the association was not significant at the 0.05 level, we would not normally compute the between O and 1 the percentage improvement in predicting the value of the Jependent variable, given the value of the
 502 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 503

independent variable. Lambda also varies between O and I. A value of O means no improvement in
prediction. A value of I indicates that the prediction can be made without error. This happens when Hypothesis Testing Related to Differences
each independent variable category is associated with a single categ01y of the dependent variable. The previous section considered hypothesis testing related to associations. We now focus on
Asymmetric lambda is computed for each of the variables (treating it as the dependent hypothesis testing related to differences. A classification of hypothesis-testing procedures for
variable). In general, the two asymmetric lambdas are likely to be different because the marginal examining differences is presented in Figure 15.9. Note that Figure 15.9 is consistent with the
sy1n1t1etric lambda distributions are not usually the same. A symmetric lambda is also computed, which is a kind of classification of univariate techniques presented in Figure 14.6. The major difference is that
The symmetnc lambda does average of the two asymmetric values. The symmetric lambda does not make an assumption Figure 14.6 also accommodates more than two samples and thus deals with techniques such
not make an assumption
about whkh variable is dependent. It measures the overall improvement when prediction is done as one-way ANOVA and K-W ANOVA (Chapter 14), whereas Figure 15.9 is limited to no
about which variable is
in both directions. 12 The value of asymmetric lambda in Table 15.3, with usage as the dependent more than two samples. Also, one-sample techniques such as frequencies, which do not
dependent It measures the
overall 1rnprover11ent when
variah!e, is 0.333. This indicates that knowledge of sex increases our predictive ability by the involve statistical Lcsting, are not covered in Figure 15.9. Hypothesis-testing procedures can
prediction 1s done in both proportiot1 of 0.333, that is, a 33.3 percent improvement. The symmetric lambda is also 0.333. be broadly classified as parametric or nonparametric, based on the measurement scale of the
d1rect1ons. parametric tests variables involved. Parametric tests assume that the variables of interest are measured on at
Other Statistics Hypothes1s-test1ng least an interval scale. Nonparametric tests assume that the variahlcs are measured on a
Note that in the calculation of the chi-square statistic, the variables are treated as being measured on procedures that assume nominal or ordinal scale. These tests can be further classified based on whether one, two, or
only a nominaJ scale. Other statistics such as tau b, tau c, and gamma are available to measure asso- thdt the var1ablPs of interest more samples are involved. As explained in Chapter 14, the number of samples is determined
ciation between two ordinal-level variables. All these statistics use information about the ordering are measured on at least an based on how the data are treated for the purpose of analysis, not based on how the data were
interval scale
of categories of variable:::. hy considering every possible pair of cases in the table. Each pair is exam- collected. The samples are independent if they are drawn randomly from different popula-
ined tu determine if its relative ordering on the first variable is the same as its relative ordering on nonparametric tests tions. For the purpose of analysis, data pertaining to different groups of respondents, for
the second variable (concordant), if the OrLlering is reversed (disconJant), or if the pair is tied. The Hypothesl'>- testing example, males and females, are generally treated as independent samples. On the other
manner in which the ties are treated is the basic difference between these statistics. Both tau band procedures that assume hand, the samples are paired when the data for the two samples relate lo the same group of
tau b tau c adjust for ties. Taub is the most appropriate with square tables, in which the number of rows that the variables are respondents.
Test statistic that measures
and the number or columns me c4ual. Its value varies between + 1 and ~ 1. Thus the direction (pos- measured on a norrnnal The most porular parametric test is the t test, conducted for examining hypotheses ahout
thP association between or ordinal scale
itive or negative) as well as the strength (how close the value is to 1) of the relationship can be means. The t test could be conducted on the mean of one sample or two smnples of observations.
two ord1n..il-l0vel variables
determined. For a rectangular tahle in which the numher of rows is different from the number of In the case of two samples, the sam,.ples could he independent or paired. Nonparametric tests based
It make'.:> an ddJU:>tment for
t1t1:,, and 15 most appropriate
columns, tau c shoulU be used. Gamma does not make an adjustment for either ties or table size. on observatjons drawn from one sample include the Kolmogorov-Smirnov test, the chi-s4uarc test,
wlwn tlw ldble of variables Gamma also varies between + I and - I anJ generally has a higher numerical value than tau b or the runs test, and the binomial test. In case of two independent samples, the Mann-Whitney U test,
I':, SQUJre tau c. For the data in Tahle l:5.J, as sex. is a nominal variable, it is not appropriate to calculate orc.Ji- the median test, anJ the Kolrnogorov-Smirnov two-sample test are used for examining hypotheses
nal statistics. All these statistics can be estimated by using the appropriate computer programs for about location. These te~ts are nonparametric counterparts of the two-group t test. The. chi-s4uare
tau L cross-tabulation. Other 5;tatistics for measuring the strength of as:mciation, namely product moment test can also he conducted. For paired samples, nonparametric te,-;ts inc]ude the Wilcoxon matched-
Te'.>t '.>tat1st1c that measures correlation an<l nonrnetric correlation, are Jiscussed in Chapter 17. pairs signed-ranks test and the sign test. These tests are the counterparts of the paired t test. In
the dS'.,OC 1r1tior1 between
addition, the McNcmar and chi-square tests can also he used. Parametric: as we.II as nonpant1T1etric
two ordinal-level variables
tests arc also available for evaluating hypotheses relating to more than two samples. These tests arc
It make5 dr1 cldJustment for Cross-Tabulation in Practice con~iclcrcd in later chapters.
tie'.> and i':> rnost appropriate
Vv'hcn conducting cross-tabulation analysis jn practice, it is useful to proceed along the following
when the table of variable~
1s not squrJre but a
steps.
rectdngle
1. Test the null hypothesis that there is no association between the variables using the chi-

~
FIGURE 15.9
9arnrna square stt1tistic. ff you faj l to reject the null hypothesis, then there is no relationship.
Test stafr,t1c th<1t mensures 2. If H 0 is rejected, then determine the strength of the association using an appropriate statistic Hypothesis Tests
the association ht?tween (rhi coefficient, contingency coefficient, Cramer's V, lambda coellicicnt, or other statistics). Related to
. ~---~ -- I
t11vu ordinal-level variables 3. If H 0 is rejected, interpret the pattern of the relationship by computing the percenlages in Differences
It does not make an the direction of the independent variable, anoss the dependent variable. Parametric Tests Nonparametric Tests
adjustment tor ties (Metric Data) (Nonmetric Data)
4. If the variahles are treated as ordinal rather than nominal, use tau b, tau c, or gamma as

l
the test statistic. If f/0 is rejected, then determine the strength of the association using l - -·1
,~~ l
the magnitude, and the direction of the relationship using the sign of the test statistic.
---i
One Two One Two
5. Translate the results or hypothe~is testing, strength of association, and pattern of associa- Samples Samples
Sample
tion into managerial implications and recommendations where meaningful. Sample
• t test • Chi-square
• z test , K-S
ACTIVE RESEARCH i Runs
lndependenl Paired Binomial
Analyzing Usage of Cosmetics Is No Cosmetic Effort Samples Samples
Visit www.loreal.com and conduct an Internet search using a search engine and your library's online database Independent Paired
• Two-group • Paired Samples Samples
to obtain information on the heavy users, light users, and nonusers of cosmetics.
t lest t test
How would you analyze the data to determine whether the heavy, light, and nonusers differ in terms • z test • Chi-square • Sign
of demographic characteristics? • Mann-Whitney • Wilcoxon
As the marketing director for L'OreaJ, what marketing strategie~ would you adopt to reach the heavy • Median • McNemar
users, light users, and nonusers of cosmetics'? • K-S • Chi-square

I
504 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 505

Parametric Tests The degrees of freedom for the t statistic to test the hypothesis about one mean arc n - I.
Parametric tests provide inferences for making statements about the means of parent populations. In this case, n - I '" 29 - 1 m 28. From Table 4 in the Statistical Appendix, the prnbabi lity ol
tP.-',t A t test is commonly used for this purpose. This test is based on the Student's I statistic. The getting a more extreme value than 2.471 is less than 0.0."i. (Alternatively, the critical I value for
A univariate hypothesis test t statistic assumes that the variable is normally distributed and the mean is known (or assumed to 28 degrees of freedom and a significance level of 0.05 is 1.7011, which is less than the calculated
using the t d1'.>tnbution, be known), and the population variance is estimated from the sample. Assume that the random vari- value.) Hence, the null hypothesis is rejected. The familiarity level docs exceed 4.0.
which 1s used when the able X is normally distributed, with mean µ. and unknown population variance <r 2, which is Note that if the population ,...;tandard deviation ww1 assumed to he known as 1.5, rather th.an
standard deviation 1s e.~timated by the .iample variance ,1 2, Recall that the standard deviation of the sample mean, X, is
2 estimated from the samrle, a z test would be appropriate. In this case, the value of the z. statistic
unk11own dnd the sample A univariJte l1ypollit•11s 1e,i Wllllid h~:
estimated ass x ~ s/Vn. Then t = (X - µ.)/.,'xis t distributed with n - l degrees of freedom.
siLe 1s small u'>ing the stanrJard normdl
The t distribution is similar to the normal distribution in appearance. Both distributions are d1stribut1on z = (x - µ.)l"x
t ~tat!st!( bell shaped and symmetric. However, as compared to the normal distribution, the t distribution
where
A statistic that assume'> th;:it has more area in the tails and less in the center. This is hecause population variance <r 2 is
the variable has a unknown and is estimated by the sample variance s 2. Given the uncertainty in the value of s2, the (TX= u1 \/29 = 1515.385 = o.n<J
syrnmetric bell-'.-.haped observed values oft are more. variahle than those of z. Thus, we must go a larger numher of stan~ and
d1strihut1on, the mean 1s dard deviations from Oto encompass a certain percentage of values from the t distribution than is
known (or a':.sumed to be
z. = (4.724 - 4.0)/0.279 ~ 0.724/0.279 cc 2.595
the case with the normal distribution. Yet, as the number of degrees of freedom increases, the
known), and the µopulat1on From Table 2 in the Statistical Appendix, the probability of getting a more extreme value of
t dL"itribution approaches the normal distribution. In fact, for large samples of 120 or more, the
varii:ince i'> est1111ated from z than 2.595 is less than 0.05. (/\lternativcly, the critical z value for a one-tailed test and a signif-
t distribution and the nmmal distribution arc virtually indistinguishable. Table 4 in the Statistical
thP sample.
Appendi, shows selected percentiles of the I distribution. Although normality is assumed, the icance level ortl.05 is I.MS, which is less than the calculated value.) Therefore, the null hypoth-
t rfr,tributim1 t test is quite rohust tu departures from normality. esi.<.; is rejected, reaching the same conclusion arrived at earlier by the t test.
A symmetric bell--;hapcd The proce<lun: for hypothesis testing, for the special case when the t statistic is useJ, is as follows. The procedure for testing a null hypothe1')is with resp~ct LD a proportion for a .single sample
d1i;t11but1011 that 1s U'.:ieful was illustrated earlier in this chapter when we introduced hypothesis testing.
for 'illldli '>amplP (n < 30) 1.Formulate the null (/i0) and the alternative (H 1) hypotheses.
test11HJ, when the mean is 2.Sekel th~ appropriate formula for the l statistic. Two Independent Samples
know11 and the population 3.Select a significance level, a, for testing H 0 . Typically, the 0.05 level is selected. J3
Several hypothese,<.; in marketing· relate to parameters from two different populations: for exam-·
variance 1s est1nv1ted trom 4.Take one nr two samples and compute the mean and standard deviation for each sample.
pie, the users und nonusers of a brand differ in terms of their perceptions of the brand, the
Ille sarnµle 5.Calculate the t statistic assuming H 0 is true.
high-income consumers spend more on entertainment than low-income consumers, or the pro-
6.Calculate the degrees of freedom and estimate the probability of getting a more extreme
portion ol' brand-Joyal users in segment l is more than the proportion in segment II. Samples
value of the statistic from Table 4. (Alternatively, calculate the critical value of the I statistic.)
drawn randomly from di ffcrent populations arc termed independent smnples. As in the case for
7. If the prohahihty cornputed in step 6 is smaller than the significance level selected in step 3, Two Sdmples that are not one sample, the hypotheses could relate to means or proportio11s.
reject H 0 . If the probability is larger, do not reject H0 . (Alternatively, if the absolute value of exrPmnentally relc1ted. The
the: cak.:ulate<l t .statistic in step 5 is larger than the absolute critical value dcten11ined in step 6, measun,•ment of one ME'.AN.5 In Lill.' case of 111eans for two independent satnples, the hyp{)thcscs take thL~ /'oll<iwi11g form.
reject H 0 . ff the absolute calculated value is smaller than the absolute critical value, do not sc.1rnplr} ha~, no etfect on th\'
VdltH:', o1 the set ond llo: /LJ -- 1-'2
reject H0 .) Failure to reject H 0 docs not necessarily imply !hat H 0 is true. It only means that
the true state is not significantly different fmm that assumed hy H 0 . 14 sample1 H1: µ1 * ,<2
8. Express the conclusion reached hy the t Lest in terms of the marketing research problem. TI1e two popul;1tions arc sampled J.nd tile means and variances computed based on samples
of sizes n 1 and n 1 . If both populations are found to have the ,<.;<une variance, a pookd varianl:c
One Sample estimate is compt~tcd from the two sample variance~ as follows:
In marketing research, the researcher is often interested in making statements about a single vari- 111 fl2

ahle against a known nr given standard. Examples of such statements include: The market share for I(X,1 X1) 2 + I<X,2 ·- ,\;d
a new product wi 11 exceed 15 percent; at least 65 percent of customers will like a new package 2 1' I i I
,\' -----.-~ -----
design; 80 percent of dealers will prefer the new pricing policy. These statements can be translated ni + n2 2
to null hypotheses that can he tested using a one-sample lest, such as the l test or the z test. In the or
case of at test for a single mean, the researcher is interested in testing whether the population mean
conforms to a given hypothc.sis (f/0 ). For the data in Table 15. I, suppose we wanted to tesl the (n, -- I
s-, = ----··--
)s; -1--- (n2____
-- I ).1·~
,_ --
hypothesis that the mean familiarity rating ex..:eeds 4.0, the neutral value on a 7-prnnt scale. fl1 + n2 ~ 2
A significance level of a = 0.05 is selected. The hypotheses may he fonnulated as: Tlw slan<lard deviation of the test statistic can he estinialed as:
Ho:µ
H1 .µ. > 4.0
~ 4.0

(x - µ.l
'x, - x,
The appropriate value oft can be calculated as:
~ +~E~ t)
= -~,;-
(x,
t= - ~ - ~ - -
--x,) -(µ, -µz)
- - ·--
.S\ s ..
Vri Xi - X2

1.579/'\/29 = 1.579/5.385 = 0.293 The degrees of freedom in this case are (n 1 \ n -· 2).
S:( ..::: 2
If the two populations have unequal variances, an exact t cannot he computed for the differ-
t = (4.724 - 4.0) I 0.293 = 0.724 I 0.293 = 2.471 ence in sample mean~. Jn,<,,lead, an approximation to I is compuk'.d. The numher of degrees of

I
506 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 507
fre~<lum in this case is usually not an integer, hut a reasonably accurate probability can be
ohtained by rounding to the nearest intcger. 15 conduct the t test both ways. Instead of the small sample of 30, if this were a large and representative
An F test or sample variance may be performed if ,t is not known whether the two popula- sample, there are profound implications for Internet SC[\/ice providers such as AOL, .EarthLink, and
A stdt1stical test of the tions have equal variance. 111 this case the hypotheses are: the various telephone (e.g., Verizon) and cahle (e.g_, Comcast) companies. ln order lo target the
equality of the variances of heavy Internet users, these companies should fo<.:us on males. Thu.<.;, more advcrlislng dollars should
two populations. H0: r.r? = (Ti he spent on magazines that cater to male audiences than those that target female<:,.
H1:(r12*-cri
f ~tatistit The Fstatistic is computed from the sample variances as follows:
The F stat1st1c is computed Real Research
as the rJt10 of two sample
Ferri \),(11 1 - I)
~L Stores Seek to Suit Elderly to a "t"
variances .)°~

A study bascd on a national sample of 789 respondents who were age 65 or older attefllpted to determine
where the effect that !.:tck of mobility has on patronage behavior. A major re.search question related lu the difkr
n 1 = size uf sample I c1wes in the physical requirements ot dependent and self~relmnt elderly persons. That is, did the two group~
require different things to get to the store nr after they aJTived at the store'! A more <lctailcd analy:--.i:,., of the
n 2 = size of sample 2 physical requirements conducted by two-independent-sample t te.sh (.<;hown in the .tccornpanying l<1blc)
n 1 -, J = degrees of freedom for sample 1 indicated that dependent elderly pcn,ons are more likely to look for stort's that offer home delivery mid
phone orders, and stores to which they have accessible transportation. They are also more likely to look fo1
n2 ~ 1 = degrees of freedom for sample 2
a variety of stores located cluse together. Retailers, now more than ever, are realizing the .\ales potential in
sf = ,.;ample variance for sample I the elderly market. With the baby-boomer generation nearing retirement in 2010, stores such as W,i!-Mart,
s? = sample variance for sample 2 Coldwater Creek, and Williams-Sonoma .\ee ·'the icing on the cake." The elderly shoppers are more likely
to spend more money and be<..:ome patrons of a store. However, to attract them, stures should offer home
f di:.lnbut1on As can he seen, the critical value or the F distribution depends upon two sets of degrees or delivery and phone orders, and arrange accessible tran~portation. 1(1
A f1.--,quer,cy cli:,t!ibution freedom--those i11 the numerator nnd those in the denominator. The critical values of F for vari-
thdt dPpend'> upon two sets ous degrees of freedom for the numcrawr and Jenominator are given in Table 5 of the Statistical
uf degrees of freedorn---the Appcmlix. If the probability of Fis greater than the significance levd <¥, J-/ 0 is not rejected, and Differences in Physical Requirements Between Dependent and Self-Reliant Elderly
deqrE'PS of freedom m thi~
t based on the pooled variance estimate Gan be used. On the other hand, if the probability of Fis Meana
11um(-'r<1tor dlHl the d1.=grePs
less than or ~qual to (Y, H 0 i,.., re.1ectcd and t hascd on a separate variance estimate is used. _f'_hysical Requirement Items Self-Reliant
of freedom 1r1 thE-' Depend_ent t Test Probability
Using the data of Table 15. J, suppose we wanted tu determine whether Internet usage was De livery to hollle
1Je1HJT(ll!lci!Of 1.787
different for males as compare<l to females. A two-independent-samples t test was conducted. The 2.000 IJ.023
Phone-in order 2 030
results are presented in 1llble I 5. 14. Note that the F test of sample vaiiances has a probability that is 2335 IJ.003
Tran8portat1on to store 2.188
less than 0.()5. Accordingly, H 0 is rejected, and the t test bas~d on "'equal variances not assumed" 1.098 0.1100
Convenient parking 4.001
should be used. The t value is -4.492 and, with 18.014 degrees of rrcccl<Jm, this gives a probability 4.m, 0.305
Location close to /ll)mc 3.177
or ll.ll(X), which is less than the significance level of ll.05. Therefore, 01e null hypothesis of equal 3.:125 II. I:l7
Variety o! .~tores clo.<,,e together 3.4So
means is rejected. Because the mean w-.agc for males (sex = l) is lJ.333 and lhat for fr.mules (sex = 2) 3.1,81 0.1121
is 3.867, we conclude that males use the Internet to a .-.ignlticantly greater extent than females. We "Mh1sured on a 5-pomt scale from not important (1) to very 1mrurtarit (5). •
also show the t test assuming equal variances because most computer progr.:ims can automatically

In this example, we tested the difforencc hetwccn means. A similar test is available f'lll" testing the
l,: diffcrenc~ hctween proportions for two independent samples.
Two-Independent-Samples t Test
PROPORTIONS The case involving proportions for two independent :-.amplcs is alsu illustrnted
Statistics
using the data of Tah\c 15.1, which gives the number of males and females who use the lntcrnet
Number Standard for shopping. ls the prnport ion of respondents using the Internet for .<.;hopping the same for males
of Cases Mean Error Mean and females? The null and altern.:itive hypotheses are:
SPSS Output File Male 15 9.333 1.137
Hu:,.,-,= Tr2
0.435
Female 15 .l.867
*
sas.
H,: ,.,-, 1T2
F Test for Equality of Variances
A z test is used as in testing the prorortion for one sample. However, in this case the test :-.ta-
F Value 2-Tail Probability tistic is given by:

SAS Output File 15.507 0.000

1'1 -1'2
tTest z = s:::----;,2
Equal Variances Assumed Equal Variances Not Assumed ln the test stati.stic, the rnnncrator is the difference between the proportions in the two samples,
Degrees of 2-Tail Degrees of 2-Tail p 1 and p 2 . The denominator is the standard error of the difference 111 the two proportions and is
given hy:
t Value Freedom Probability t Value Freedom Probability
-A.492 28 (l.000
--· ·---.. -----------M------·--------·-----··---·.......J
--4.492 18.014 0.000
,\i;I fl2
Y·
/p(I - p)[_I_
flt
-;_1-J
n2
 CHAPTER 1S • FREQUENCY DISTRIBUTION, CROSS-TABULATION, AND HYPOTHESIS TESTING 509
508 PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING

II
where f!l:1118
Paired Samples t Test
"11'1 + n2P2
p= Number of Standard Standard
n1 + n2 SPSS Output File
Variable Cases Mean Deviation Error
A significance level of" = 0.05 is selected. Given the data of Table 15. l, the test statistic can
he calculated as: §.sas I lnternel Attitude
Technology Attitude
30
30
5.167
4.100
·-------
1.234
l.398
0 225
()255
SAS Output File
p
1
- p2 = (11/15) - (6/15) Difference = Internet - Technology
= 0.733 - 0.400 = 0.333 Difference Standard 2-Tail Degrees of 2-Tail
Standard
p = (15 X 0.733 + 15 X 0.4)/(15 + 15) = 0.567 Mean Deviation Error Correlation Probability t Value Freedom Probability

s,, I .. P =
2
r;;;;,~--:
\/U.::>O
0.433 .15 (l. + '-) ]5
= 0.181
1067 0.828 0.1511 0 809 IJ.000 7.0.19 29 0.000

z= 0.333/0.181 =l.84 themselves to be technologically savvy. Another application is provided in the context or
Given a two-tail lest, the area to the right of the critical value is a/2 or 0.025. Hence, the critical determining the relative effectiveness of 15-second versus 30-'iecond tclevi,-;ion commercials.
value of the test statistic is 1.96. Because the calculated value is less than the critical value, the null
hypothesis cannot he rejected. Thus, the proportion of users (0.733) for males and (0.400) for
females is not significantly different for the two samples. Note that although the difference is Real Research Seconds Count
substantial, it is not statistically significant due to the small sample sizes ( 15 in each group).
A survey of 81 ml'dia directors of tile largest Canadian aJvcrtising agencies was conducted to determine the
Paired Samples relative effectiveness of 15-second versus 30-sccond ..:ommerciaJ advertisement.:.. Usrng a 5-point rating
In rnany marketing research applications, the ohservations for the two groups are not selected scale ( I hcing ex.cc.I lent and 5 heing poor), 15- and JO-second commercials were rateU by each respondent
for hrand awareness, main idea recull, persuasion, and ability to tell .in emotional story. Tbc w..:cornpanyini!
frum independent samples. Rather, the observations relalc to paired samples in that the two sets
paired samples table indicates that 30-sccond com'mercials were rated more favorably on all the dimensirnu;. Pairet.l t test:)
In hypot~wc,1s testing, the of observations relate to the same respondents. A sample of respondents may rate two competing
indicated that these differences were significant, and the 15-sccond commercials wen: evaluated as less
observations are paired c;o brands, indicate the relative importance of two attributes of a product, or evaluate a brand at two cfft'ctivc Thus, 15-second commercials may not be the answer marketers arc looking for. Actually, today,
th,it the two set'> ot different times. The difference in these case~ iH examined hy a paired samples t test. T~) com- the problem may not be how effective television commL-rcials are, but whethl'r the consumers actually will
obsPrvatHJf1'..> 1elate to the pute t for paired samples, the paired difference variable, denoted hy V, is formed an<l its mean be watching the commercial:'.. Une 111 Jive users never wnk'hed a commcrc1:il m 2008, and there is a threat
,ame respo11denb and variance calculated. Then the t statistic is computed. The degrees of free<l~im arc n - J, that this number will increase in the future. Heavy aJvertisers such as Cfcneral Motors will have to comt' up
where n is the number of pairs. The relevant formulas are: with more cffc'-'.live and cn~ativc wayii to :,,how their commercials. 17
paired samples t test
A te'it tor differences 1n the Ho:/LD = 0 Mean Rating of 15- and 30-Second Commercials on the Four Communication Variables
r11edr1s of p,iired Sdmple'>
H1:µ, 0 *0
l5 - /.Ln Brand Main Idea Ability to Tell
ln - I = Awareness Recall Persuasion Emotional 51
SJ"!_

\In 15 30 15 30 15 30 15 JO
2.5 1.9 2.7 20 3.7 2.l 4.3 1.9 •
where
"
~D; The difference in proportions for paired samples can he tested by using the TvkNemar test or the

·~,,--~
1--1
[j ~
chi-square test, as explained in tl1e following section on nonparametriL' tests.

SI) -
_
J ~(D,_::_ 15)'
fl ·- l ACTIVE RESEARCH

Users and Nonusers: Are Their Evaluations of Reebok Different?

s[J c;. rn
SIJ
Visit www.reebok.com and conduct an Internet search using a search engine and your lihrary1s onlme datahase
to obtain infonnation on the factors consumers use to evaluate competing brands of athletic shoes.
In the Internet usage example (Table 15.1), a paired I test could be used to determine if As the marketing director for Reebok, how would you improve the image and competitive position-
the respondents differed in their attitude toward the Internet and attitude toward technology. ing of your brand?
The resulting output is shown in Table 15.15. The mean attitude toward the Internet is 5.167 The users and nonusers of Reebok evaluated the brand on five factors using Like1t-type scales. How
and that toward technology is 4.10. The mean difference between the variables is 1.067, with. would you analyze these data using two independent samples and paired samples t tests?
a slandard deviation of 0.828 and a standard error of 0.151 I. This results in a I value of
(1.067/0.1511) 7.06, with 30 - I = 29 degrees of freedom and a probability of less than
Nonparametric Tests
0.00 I. Therefore, the respondents have a more favorable attitude toward the Internet as com-
pared to technology in general. An implication, if this were a large and representative sample, Nonparametric tests are used \.vhcn the independent variables are nonmctric. Like parametric
would be that Internet service providers should not hesitate to market their services to tests, nonp..irametric tests are available for testing variahle1-. from one. sample, two independent
consumers who do not have a very positive attitude toward technology and do not consider samples, or two related samples
 PART 111 • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS·TABULATION, AND HYPOTHESIS TESTING 511
510

One Sample ~,:

Sometimes the researcher wants to test whether the observations for a particular variable could LMann-Whitney U Test
reasonably have come from a particular distribution, such as the normal, uniform, or Poisson
distribution. Knowledge of the dislribution is necessary for finding probabilities corresponding
i Mann·Whitney U-Wilcoxon Rank Sum W Test
Internet Usage by Sex
to known values of the variable or variable values corresponding to known probabilities (see
Kolmogorov-Srnir nov
(l<·S) une·S:llnple test
Appendix 12A). The Kolmogorov-Smirnov (K-S) one-sample test is one such goodness-of.fit
test. The K·S compares the cumulalive distribution function for a variable with a specified distri- SPSS Output File
I
I
Sex

Male
Mean Rank
20.93
Cases
15
A one-sample bution. A, denotes the cumulative relative frequency for each category of the theoretical ! Female 1().(17 IS
nonparametric goodness-
uf-f1t test that compares the
cumulative distribution
function for a variable with
(assumed) distribution, and 0, the comparable value of the sample frequency. The K-S test is
or
based on the maximum value the absolute difference between A, and O,. The test statistic is:

K = Max I A, - 0, I
.sas.
SAS Output File
!
i
Total

u w z
30
Corrected for
Ties 2·tailed
a specified d1stribut1on
The decision to reject the null hypothesis is based on the value of K. The larger the K is, the I 31.1100 151.()1)1) -3.406 0001
more confidence we have that HQ is false. For" = 0.05 , the critical value of K for large samples Ij Note
(over 35) is given by i .36/Vn. 1 Alternatively, K can be transformed into a normally distributed . U = Manr1·Whitney tr.s1 stc-1t1:,t1<
z statistic and its associated probahility detcnninc<l. W - Wilcoxon Wstdt1':it1c
In the rnntext of the lnternct usage example, suppose we wanted to test whether the distrib· Lz___ ""'· .U.t1·anstor1Md 1r1to
1
de normally d1str1butPd l ',tdtist1c
ution of Internet usage was normal. A K-S one-sample test is conducted, yielding the data shown
in Table I 5. 16. The largest absolute di fferencc between the observed and normal distribution was
K = 0.222. Although our sample size is only 30 (less than 35), we can use the approximate
formula and the critical value for K is 1.36/v':lo = 0.248. Because the calculated value of K is In the Mann-Whitney lJ test, the twn samples are comhincd and the cases are ranked in
smaller than the critical value, the null hypothesis cannot be rejected. Alternatively, Table 15.16 order of increasing size. The tc.~t statistic, U, is computed as the nurnhcr of times a score from
indicates that the probability of observing a K value of 0.222, as determined by Lhe normalized sample l or group I precedes a score rrnm group 2. If the samples arL: from the same popula-
nms te5t
A test of randornness tor a
z statistic, is 0.103. Because this is more than the significance level of 0.05, the null hypothesis tion, the distribution of scoreS from the two groups in the rank list should be nmdom. An
rlichotornous var1dble cannot be rejected, leading to the same conclusion. Hence, the distribution of Internet usage does extreme value of U would indicate a nonrandom pattern, pointing to the inequality of the twn
not deviate significantly from the normal distribution. The implication is that we are safe in using groups. For samples of less than 30, the exact significance level for[) is computed. For larger
binomial te~t stJtistical tests (e.g., the z tcsl) an<l procedures that assume the normality of this variahle. samples, U is transformed into a normally distrihutcd z statistic. This z. can be corrcctcJ for
A 0oodru:c;<,-of-t1t :.l<1.t1st1cal As mentioned earJicr, the chi-square test can also be performed on a single variable from one tics within ranks.
test for d1chotornous sample. In this context, the chi-square serves as a gooclness·of.fit test. It tests whether a signifi· tvvo- :>dmplt: We examine again the difference in the Internet w.:,age of males and females. Thi.-. time, though,
variables !t tii:''>l'> thP
cant difference exists between the ohserved number of cases in each category and the expected
te::.t the Mann-Whitney U test is used. The results arc given in Table 15.17. Again, a significant differ-
goodne:is ot fit ot th0 Nor1pdr amPtric tPst :, tatist1c
number. Other one-sample nonparametric tests include the runs test am! the binomial test. The ence is found hctween the two groups, corrohorating the results of the two-indepcndent-smnples
ob:ierved rn1mber of thc1t rlctermines w~wther
runs test i.s a test of randomness for the dichotomous variables. This test is conducted by t lest reported earlier. Because the ranks are <.1ssigneJ from the smalle~t observation to the largest,
ob'>ervJt1ons 1n each two qroupc,. are Jrawr1 tmin
cate(_Jory to the number
determining whether the order or sequence in which observations are obtained is random. The populritron:, w11h the same the higher mean rank (20.93) of males indicates thal they use the Internet to a greater extent than
exrected under n c,peufied binomial test is .-1lso a goodness-of-fit test for dichotomous variahles. It tests the goodness of fit rned1<1n Thi:, tesi is not as !emalcs (mean rank = I0.07).
ll1riom1al d1stnbL1t1rn1 of the ohserved number of observations in each category to the number expected under a powerful a', the Mann- Researchers often wish to test for a significant difference i11 proportiuns ohtainc.d from two
specified binomial distribution. For rnorc information on these tests, refer to standard statistical Wh1tr1ey U independent samples. /\s an alternative to the parametric z. test L'onsidcrcd earlier, one con kl alsu
Mann- Whitney U test literature. 19 use the cro~s-tahulation procedure to CL1t1duct a chi-square test. 21 ln this case, we will have a 2 X 2
A stdti<,tical tl::'St for a l(olmofJOt'ov-~mi: 11<1v table. One variahle will he used to denote the sample and will assume thc va!UL'. I for s.a111ple 1 and
variable measured on a11 two·s.urnµh:~ te~t
Two Independent Samples the value of 2 for sample 2. The other variable will he the binary variahle of interest.
ord111al c;cale, rornpdr1ng Nonpardrnt1lric te:,t statistic
When the difference in the location of two populations is to be compared ba....,cd on observations from Two other indcpendc11t-:•"1mplcs nonparametric tests arc the median test and Kolrnogoruv-
the difference 111 the that dt:te1 mines whPther
two indcpcmlcnt sampli:s, and the variable is measured on an ordinal scale, the Mann-Whitney U two distr1but1or)s ar-e the
Smimov test. The two-sample median test determines whether the two groups arl'. drawn from
loc at1on of two populations
bo,ed on observc1t1ons from test can be uscd. 20 This test corresponds to the two-independent-sample t test for interval scale :.ame It takes into ,-Kcount
populations with the same medi;rn. lt is not as powerful as the Man11-Whitncy lJ test hecause it
two 1ndepende11t samples v<u-iahlcs, when the variances of the two populations arc assumed equal. any differences 1n the two merely uses the localion of each observation relative to the median, and not the rank, uf each
di:,\r1but1ons, including ohscrvation. The Kolmogorov-Smirnov two-sample test examines whether the 1wo distribu-
mediari, d1:,per<;.ion, dnd tions arc the same. It takes into account any differences hetween the two distributions, including
•f!1:t•a1o"J111• skewness the median, dispersion, and ,.-..kewness, as illustrated by tile following example.
K-S One-Sample Test for Normality for Internet Usage
Test Distribution-Normal
Mean: 6.600 Real Research Directors Change Direction
Standard Deviation· 4.296
llow do marketing research directors .1ml users in F·orrutu.' 500 1rnrnufac:turing firms perceive the role of
SPSS Output File Cases: 30 marketing n:scarch in initiating change'> in mm keling s!ralegy formulation'! It was found that the ma1 kcting

sas
researt:h director<; were more strongly in favor uf init1a!mg changes in strategy and less in favor of holding
Most Extreme Differences
back than were users of marketing research. The percentage respom,cs to 1mc of the items, "Initiate change
Absolute Positive Negative K·S z Hailed p in the marketing strategy of !lie firm whenever pos!-iible," are given in the following table. Using tht:
0.222 0.222 -0.142 1.217 0.103 Kolmogorov-Smirnov (K-S) test, lhese differences o! role ddinttion were statistically significant al the 0.05
level, as shown in the table.
SAS Output File !__·--·------------------
 CHAPTER 15 • FREQUENCY DISTRIBUTION, CROSS-T.ABULATION, AND HYPOTHESIS TESTING 513
PART Ill • DATA COLLECTION, PREPARATION, ANALYSIS, AND REPORTING
512
The user~ uf marketing research hecame even more reluctant to initiate rnarkcling strategy changes dur-
Decision Research General Mills' Curves Cereal: Helping Women Achieve Their Curves
ing the uncertain economy of 2009. In today's business climate, however, the reluctance of these marketing
research users must be overcome to help gain a better understanding of the huyer's power. Thus, marketing
research firms should Jcvote considerable effort in convincing the users (generally, the marketing man-
The Situation
22 Stephen W. Sanger, CEO of General Mills (www.generalmills.com), is constantly faced with the challenge
agers) of the va!uc of marketing rescarcb.
of how to keep up with consumers' changing tastes and preferences. General Mills recently conducted thor-
The Role of Marketing Research in Strategy Formulation
.~~~~~~~~~~~~~~~~~~~
ough focus-group rcsearrh on the most important consumers in grocery stores today: women. It is a known
fact that three nut of every four grocery shoppers in the Llni1ed States an; women, and many of these women
Responses(%)
are tixiusing ll10!'('. on theil' health anJ the nutritiuus valui: of foods. Allh1Htgh there are many cereals on the
Preferably Mayor Preferably Absolutely market with the same amount of valuable vitamins and rniuerals, such as Total or Kellogg's Smart Start,
Absolutely
Should May Not Should Not Must Not General Mills decided to design a product specifically for women.
Sample N Must
43 19 l)iel1cian Roberta Duyff claims thal women do not get enough nutrients such a~ calcium or folic acid
1) 77 7 26
16 from Jay to day. According to Duy ff, "It is great that a woman can now increase her intake of these impor-
2 15 32 35
68 tant nutrients with a simple bowl of cereal for breakfast, and if you add milk, the vitamin iJ in milk makes
--- the calcium in both the fortified cereal and milk itself more absorhahlc." This is oni..: way in which Cicneral
K-S ':>ign1ficance = 0 05 Mills saw an advantagc----convenicnce for the woman. She can grab a howl in the mmning and start off he1
• D = directors, U =- users •
day with the nutrients she needs. Not only is the convenience of the product an incentive to market it, but
focus group findings also indicated that women like to have a prn<luct of their own. In fact, according to
In this example, the marketing research directors an<l users comprised two independent samples. Megan Nightingale, assistant nwrketmg manager at General Mills, "Our research has shown that wrnm~n
However, the samples are not always independent. In the case of paired samples, a different set are looking for something tlrnt's nutritious, fast, convenient, an<l has a good taste."
of tests should be used. ln 2007, <ieneral Mills, in partnership with Curves International, launched the Curves Cere.il. This new
cereal of lightly sweetened to;istcd flJkes of whole grain rice and wheat is available m two dclic1ou.s flavors,
Whole (Jrain Crunch and Honey Crunch. Both have fewer than 200 calorics per serving am! cnntain at least
Paired Samples 31 percent or the recommended amounts of whole grains and 2 grams of fiher. They abo are an exrcllent
An important nonparametric test for examining differences in the location of two populations
source of several impo,t;mt vita.mins and minenils.
based on paired observations is the Wilcoxon matched-pairs signed-ranks test. This test ana- A telephone smvcy was COlH.luCtcd to determine the preference for and crrnsumption of ( 'urvcs and the
Wil(OXOrl rnatthed-
pdlr'> ')igneU~ranks te~t
lyzes the differences between the paired observations, taking into account the magnitutle of the relative importance that women attached to a c.:erca\ being nutritious, fast, convenicnt, and good tasting.
A nonparametric test that differences. So it requires that the data are measured at an interval level of measurement.
dndlyze;, tl1P differences However it docs not require assumptions ahout the form of the distribution of the measurements.
The Marketing Research Decision
hetwl'en the pdired It should therefore he used whenever the distrihutional assumptions that underlie the t test cannot
1. What is the relative importance uf the four variables (nutrititllts, fast, convenient, ,md good taste) in
observation'.>, tdk1ng into be satisfied. This test computes the differences between lhc pairs of variables and ranks the
influencing women to buy Curws Cereal? What type of aualysi~ should he conducted?
d\lUUnt the mc:1<:JriitudP of absolute differences. The nex.t step is to sum the positive and negative ranks. The test statistic, z, is
a. Frequl'ncy distribution of the importance atlached lo the fuur factors
the d1ffen:,n(e) computed from the positive and negative rank sums. Under the null hypothesis of no difference,
b. Mean levels 11f importance of the f,n1r vanable.s
z is a standard normal variate with mean O and variance I for large samples. This test cnrrcsponds c. Cross-tabulation of rurves Cereal pun:hascs with the importance or the four variables:
to the raired t test considered earlier?' chi-square an:..dys1s
The example considered for the paired t test, whether the respondents differed in terms of d. Crnss-tahulation uf" Curve~ ('crcal 1rnrchases witl1 th"e importance of the f<nir variables·
attitude toward the Internet and attitude toward technology, i.-; considered again. Suppose we Cramer's V
assume that hoth these variables are measured on ordinal rather than interval scales. e. All of the ubovc
Accordingly, we use the Wilcoxon test. The results arc shown in Tahle 15.18. Again. a signifi· 2. Discuss the role of the type of d..tla analysis you recommend 1n enahling Stephen W. Sanger to
cant difference is found in the variables, and the results are in accordance with the concJusion undcrstanll women's preference fn1 and consumption of Cl1rves Cereal.
reached by the paired t test. There arc 23 negative differences (attitude toward technology is less
favorable than attitude toward Internet). The mean rank of these negative differences is 12.72. On The Marketing Management Decision
the other hand, there is only one positive difference (attitude toward technology is more favor- 1. Advertising for Curves Cereal should stress which of th~ four factors'!
able than attitude toward Internet). The mean rank of this difference is 7.50. There are si.'\. ties, or a. Nutrition
observations with the same value for both variahles. These numbers indicate that the attitude b. Quick consumption
toward the Internet is more favorable than toward technology. furthermore, the probability c. Convenience
associated with the z statistic is le.ss than 0.05, indicating that the difference is indeed significant. d. Good taste
e. Al! of the uhovc
2. Discuss how thl' management decision action that you recommend to Stephen W. Sanger i'> influ-
enced hy the type of data analy~i:-. you suggested earlier and by the tinding.s of that a11alysii... 24 •
Wilcoxon Matched-Pairs Signed-Rank Test sign test
Internet with Technologz A nonparametric test for
Another paired sample nonparametric test is the sign tcst. 25 This test is not as powerful as the
exam1rnng differences in the
(Technology-Internet) Cases Mean Rank Wilcoxon matched-pairs signed-ranks test, as it compares only the signs of the differences between
location of two populations,
12.72 pairs of variables without taking into account the ranks. In the special case of a binary variable where
SPSS Output File Ranks 23 bdsed on paired
lhe researcher wishes to test differences in proportions, tbe McNcmar test can be used. Alternatively,
l 7.50 observations, that compares

·sas
t Ranks the chi-square test can also be used for hi nary variables. The various parametric and nonparametric:
6 only Hw '.>1gns ot the
Ties test;; for differences are summarized in Table 15.19. The tests in Table 15.1 Y can he easily related t{)
differences between pairs of
Total 30
varidbles without taking those in Figure l 5. 9. Tahle 15. l 9 classifies the tc:sts in inure dctai I because parametric tests (based <)11
Z c· -4.207 Hailed fl - 0.0000 into a(rnunt the r;inks metric data) are classified separately for means and proportion.). Likewise, nonrarnmetric tests
SAS Output File