Академический Документы
Профессиональный Документы
Культура Документы
610
As is the case with the t test for two dependent samples, the Wilcoxon matched-pairs
signed-ranks test can also be employed to evaluate a one-group pretest-posttest design. The
limitations of the one-group pretest-posttest design (which are discussed in Section VII of the
t test for two dependent samples and the Introduction) are also applicablewhen it is evaluated
with the Wilcoxon matched-pairs signed-ranks test.
It should be noted that all of the other tests in this text that rank data (with the exception
of the Wilcoxon signed-ranks test and the Mdes test for equal variability (Test 15)), rank
the original intervalhati0 scores of subjects. The Wilcoxon matched-pairs signed-rankstest,
however, does not rank the original intervaVratio scores, but instead ranks the intervalhati0
difference scores of subjects (or matched pairs of subjects). For this reason, some sources
categorize the Wilcoxon matched-pairs signed-rankstest as a test of intervalhatiodata. Most
sources, however (including this book), categorizethe Wilcoxon matched-pairs signed-ranks
test as a test of ordinal data, by virtue of the fact that a ranking procedure is part of the test
protocol.
11. Example
Example 18.1 is identical to Example 17.1 (which is evaluated with the t test for two
dependent samples). In evaluating Example 18.1 it will be assumed that the ratio data are
rank-ordered, since one or more of the assumptions of the t test for two dependent samples
have been saliently violated.
Example 18.1 A psychologist conducts a st+ to determine whether or not people exhibit
more emotionality when they are exposed to sexually explicit words than when they are exposed
to neutral words. Each of ten subjects is shown a list of 16 randomly arranged words, which
are projected onto a screen one at a timefor a period offwe seconds. Eight of the words on
the list are sexually explicit and eight of the words are neutral. As each word is projected on
the screen, a subject is instructed to say the w r d softly to himselfor herself: As a subject does
this, sensors attached to thepalms of the subject's hands recordgalvanic skin response (GSR),
which is used by thepsychologist as a measure of emotionality. Thepsychologist computes two
scores for each subject, one score for each of the experimental conditions: Condition 1:
GSWExplicit - The average GSR score for the eight sexually explicit words; Condition 2:
GSWNeutral- The average GSR scorefor the eight neutral words. The GSWfiplicit andthe
GSWNeutral scores of the ten subjectsfollow. (The higher the score, the higher the level of
emotionality.) Subject 1 (9,8); Subject 2 (2,2); Subject 3 (1,3); Subject 4 (4,2); Subject 5
(6,3); Subject 6 (4,O); Subject 7 (7,4); Subject 8 (8,5); Subject 9 (5,4); Subject 10 (1,O).
Do subjects exhibit dzFerences in emotionality with respect to the two categories of words?
H,: 8, = 0
(In the underlying populations represented by Condition 1 and Condition 2, the median of the
difference scores equals zero. With respect to the sample data, this translates into the sum of
the ranks of the positive difference scores being equal to the sum of the ranks of the negative
difference scores (i.e., CR+ = CR-).
Alternative hypothesis
H,: 8, + 0
(In the underlying populations represented by Condition 1 and Condition 2, the median of the
difference scores is some value other than zero. With respect to the sample data, this translates
Test 18
61 1
into the sum ofthe ranks ofthe positive differencescores not being equal to the sum ofthe ranks
of the negative difference scores (i.e., CR+ # CR-). This is a nondirectional alternative
hypothesis and it is evaluated with a two-tailed test.)
(In the underlying populations represented by Condition 1 and Condition 2, the median of the
difference scores is some value that is greater than zero. With respect to the sample data, this
translates into the sum of the ranks of the positive difference scores being greater than the sum
ofthe ranks of the negative difference scores (i.e., CR+ > CR-). The latter result indicates that
the scores in Condition 1 are higher than the scores in Condition 2. This is a directional alternative hypothesis and it is evaluated with a one-tailed test.)
H,: 8, < 0
(In the underlying populations represented by Condition 1 and Condition 2, the median of the
difference scores is some value that is less than zero (i.e., a negative number). With respect to
the sample data, this translates into the sum of the ranks ofthe positive difference scores being
less than the sum of the ranks of the negative difference scores (i.e., CR+ < CR-). The latter
result indicates that the scores in Condition 2 are higher than the scores in Condition 1. This
is a directional alternative hypothesis and it is evaluated with a one-tailed test.)
Note: Only one of the above noted alternative hypotheses is employed. If the alternative
hypothesis the researcher selects is supported, the null hypothesis is rejected.
X,
x
2
2
3
4
8
2
1
4
5
6
4
7
3
0
4
D=X,-X2
RankofJDI
0
-2
2
3
4
3
10
1
1
SignedrankofIDI
4.5
4.5
7
9
7
-4.5
4.5
7
9
7
7
2
7
2
2
CR+
CR-
40.5
- 4.5
In Table 18.1, X, represents each subject's score in Condition 1 (sexually explicit words)
and X, represents each subject's score in Condition 2 (neutral words). In Column 4 of Table
Copyright 2004 by Chapman & Hal/CRC
612
18.1 a D score is computed for each subject by subtracting a subject's score in Condition 2
from the subject's score in Condition 1 (i.e., D = X, - X,). In Column 5 the D scores have
been ranked with respect to their absolute values. Since the ranking protocol employed for the
Wilcoxon matched-pairs signed-ranks test is identical to that employed for the Wilcoxon
signed-ranks test, the reader may find it usehl to review the ranking protocol described in
Section IV of the latter test. To reiterate, the following guidelines should be adhered to when
ranking the difference scores for the Wilcoxon matched-pairs signed-ranks test.
a) The absolute values of the difference scores (I Dl) are ranked (i.e., the sign of a difference score is not taken into account). Because absolute values are employed to represent the
difference scores, D = X2 - X, can also be employed to compute the value of D.
b) Any difference score that equals zero is not ranked. This translates into eliminating
from the analysis any subject who yields a difference score of zero.
c) When there are tied scores present in the data, the average of the ranks involved is
assigned to all scores tied for a given rank.
d) As is the case with the Wilcoxon signed-ranks test, when ranking difference scores for
the Wilcoxon matched-pairssigned-ranks test it is essential that a rank of 1 be assigned to the
difference score with the lowest absolute value, and that a rank of n be assigned to the difference
score with the highest absolute value (where n represents the number of signed ranks - i.e.,
difference scores that have been ranked).2
Upon ranking the absolute values of the differencescores, the sign of each differencescore
is placed in front of its rank. The signed ranks of the difference scores are listed in Column 6
of Table 18.1. Note that although 10 subjects participated in the experiment there are only n =
9 signed ranks, since Subject 2 had a differencescore of zero which was not ranked. Table 18.2
summarizes the rankings of the difference scores for Example 18.1.
Table 18.2 Ranking Procedure for Wilcoxon Matched-Pairs Signed-Ranks Test
Subject number
1
1
6
4
--
Rank of ID1
2 4 . 5 4 . 5 7
The sum of the ranks that have a positive sign (i.e., CR+ = 40.5) and the sum ofthe ranks
that have a negative sign (i.e., CR- = 4.5) are recorded at the bottom of Column 6 in Table 18.1.
Equation 18.1 (which is identical to Equation 6.1) allows one to check the accuracy of these
values. If the relationship indicated by Equation 18.1 is not obtained, it indicates an error has
been made in the calculations.
(Equation 18.1)
Employing the values CR+ = 40.5 and CRrelationship described by the equation is true.
= 40.5 in
Test 18
To5
TO*
Two-tailed values
Onotailed values
Since the null hypothesis can only be rejected if the computed value T= 4.5 is equal to or
less than the tabled critical value at the prespecified level of significance,we can conclude the
following.
In order for the nondiiectional alternative hypothesis HI: 0, z 0 to be supported, it
is irrelevant whether CR+ > CR- or CR- > CR+. In order for the result to be significant,
the computed value of T must be equal to or less than the tabled critical hivetailed value at
the prespecified level of significance. Since the computed value T = 4.5 is less than the tabled
critical twtailed .05 value To, = 5, the nondirectional alternative hypothesis HI: OD + 0 is
supported at the .05 level. It is not, however, supported at the .O1 level, since T= 4.5 is greater
than the tabled critical two-tailed .O1 value To, = 1.
In order for the directional alternative hypothesis HI: 0, > 0 to be supported, CR+ must
be greater than CR-. Since CR+ > CR-, the data are consistent with the directional alternative
hypothesis HI: 0, > 0. In order for the result to be significant, the computed value of T must
be equal to or less than the tabled critical one-tailed value at the prespecified level of
significance. Since the computed value T = 4.5 is less than the tabled critical onstailed .05
value To, = 8, the diiectional alternative hypothesis HI: 0, > 0 is supported at the .05 level.
It is not, however, supported at the .Ol level, since T = 4.5 is greater than the tabled critical onetailed .Ol value To, = 3.
Copyright 2004 by Chapman & Hal/CRC
614
In order for the directional alternative hypothesis HI: 8, < 0 to be supported, the following two conditions must be met: a) CR- must be greater than CR+;and b) The computed value
of T must be equal to or less than the tabled critical one-tailed value at the prespecified level of
significance. Since the first ofthese conditions is not met, the directional alternative hypothesis
Hl : 8, < 0 is not supported.
A summary of the analysis of Example 18: 1 with the Wilcoxon matched-pairs signedranks test follows: It can be concluded that subjects exhibited higher GSR(emotiona1ity)scores
with respect to the sexually explicit words than the neutral words.
The results obtained with the Wilcoxon matched-pairs signed-ranks test are reasonably
consistent with those obtained when the t test for two dependent samples is employed to evaluate the same set of data. In the case of both tests, the analogous nondirectional alternative
hypotheses H,: 8, + 0 and HI: pl + p, are supported, but only at the .05 level. In the case
of the Wilcoxon matched-pairs signed-ranks test, the directional alternative hypothesis
HI: 8, > 0 is only supported at the .05 level, whereas the analogous directional alternative
hypothesis HI: p, > p, is supported at both the .05 and .Ol levels when the data are evaluated
with the t test for two dependent samples. The latter discrepancy between the two tests reflects
the fact that when a parametric and nonparametric test are applied to the same set of data, the
parametric test will generally provide a more powerfd test ofan alternativehypothesis. In most
instances, however, similar conclusions will be reached if the same data are evaluated with the
t test for two dependent samples and the Wilcoxon matched-pairs signed-ranks test.
(Equation 18.2)
Although Example 18.1 involves only nine signed ranks (a value most sources would view
as too small to use with the normal approximation), it will be employed to illustrate Equation
18.2. The reader will see that in spite ofemploying Equation 18.2 with a small sample size, it
will yield essentially the same result as that obtained when the exact table of the Wilcoxon
distribution is employed. When the values T= 4.5 and n = 9 are substituted in Equation 18.2,
the value z = -2.13 is computed.
Copyright 2004 by Chapman & Hal/CRC
Test 18
The obtained value z = -2.13 is evaluated with Table A1 (Table of the Normal Distribution) in the Appendix. In Table A1 the tabled critical two-tailed .05 and .0l values are
z,, = 1.96 and z,,, = 2.58, and the tabled critical one-tailed .05 and .O1 values are
z,,, = 1.65 and z,, = 2.33.
Since the smaller of the two values CR+ versus CR- is selected to represent T, the value
of z computed with Equation 18.2 will always be a negative number (unless CR+ = CR-, in
which case z will equal zero). This is the case, since by selecting the smaller value Twill always
be less than the expected value T,. As a result ofthis, the following guidelines are employed
in evaluating the null hypothesis.
a) If a nondirectional alternative hypothesis is employed, the null hypothesis can be rejected ifthe obtained absolute value ofz is equal to or greater than the tabled critical two-tailed
value at the prespecified level of significance.
b) When a directional alternative hypothesis is employed, one of the two possible
directional alternative hypotheses will be supported if the obtained absolute value ofz is equal
to or greater than the tabled critical one-tailed value at the prespecified level of significance.
Which alternative hypothesis is supported depends on the prediction regarding which ofthe two
values CR+ versus CR- is larger. The null hypothesis can only be rejected if the directional
alternative hypothesis that is consistent with the data is supported.
Employing the above guidelines, when the normal approximation is employed with
Example 18.1 the following conclusions can be reached.
The nondirectional alternative hypothesis H,: 0, + 0 is supported at the .05 level. This
is the case, since the computed absolute value z = 2.13 is greater than the tabled critical twotailed .05 value z,, = 1.96. The nondirectional alternative hypothesis Hi:
8, + 0 is not
supported at the .O1 level, since the absolute value z = 2.13 is less than the tabled critical twotailed .Ol value z,,, = 2.58. This decision is consistent with the decision that is reached when
the exact table of the Wilcoxon distribution is employed to evaluate the nondirectional
alternative hypothesis H I : 8, z 0 .
The directional alternative hypothesis H I : 8, > 0 is supported at the .05 level. This is
the case, since the data are consistent with the latter alternative hypothesis (i.e., CR+ > CR-),
and the computed absolute value z = 2.13 is greater than the tabled critical one-tailed .05 value
z,, = 1.65. The directional alternative hypothesis HI:
8, > 0 is not supported at the .O1
level, since the obtained absolute value z = 2.13 is less than the tabled critical one-tailed .O1
value z,,, = 2.33. This decision is consistent with the decision that is reached when the exact
table ofthe Wilcoxon distribution is employed to evaluate the directional alternative hypothesis
H,: 0, > 0 .
The directional alternative hypothesis H,:8, < 0 is not supported, since the data are
not consistent with the latter alternative hypothesis (which requires that CR- > CR+).
It should be noted that, in actuality, either CR+ or CR- can be employed to represent the
value of T in Equation 18.2. Either value will yield the same absolute value for z. The smaller
of the two values will always yield a negative z value, and the larger of the two values will
always yield a positive z value (which in this instance will be z = 2.13 if CR+ = 40.5 is
employed to represent T). In evaluating a nondirectional alternative hypothesis the sign of z
is irrelevant. In the case ofa directional alternative hypothesis, one must determine whether the
data are consistent with the alternative hypothesis that is stipulated. If the data are consistent,
Copyright 2004 by Chapman & Hal/CRC
616
one then determines whether the absolute value ofz is equal to or greater than the tabled critical
one-tailed value at the prespecified level of significance.
2. The correction for continuity for the normal approximation of the Wilcoxon matchedpairs signed-ranks test As noted in the discussion of the Wilcoxon signed-ranks test, a
correction for continuity can be employed for the normal approximation of the Wilcoxon test
statistic. The same conection for continuity can be applied to the Wilcoxon matched-pairs
signed-ranks test. The correction for continuity (which results in a slight reduction in the
absolute value computed for z) requires that .5 be subtracted fiom the absolute value of the
numerator of Equation 18.2. Thus, Equation 18.3 (which is identical to Equation 6.3)
represents the continuity-corrected normal approximation of the Wilcoxon test statistic.
(Equation 18.3)
Employing Equation 18.3, the continuity-corrected value z = 2.07 is computed. Note that
as a result of the absolute value conversion, the numerator of Equation 18.3 will always be a
positive number, thus yielding a positive z value.
The result of the analysis with Equation 18.3 leads to the same conclusions that are
reached with Equation 18.2 (i.e., when the correction for continuity is not employed).
Specifically, since the absolute value z = 2.07 is greater than the tabled critical two-tailed .05
value z,,, = 1.96, the nondirectional alternative hypothesis H,: 8
, + 0 is supported at the .05
level (but not at the .O1 level). Since the absolute value z = 2.07 is greater than the tabled
critical one-tailed .05 value z,, = 1.65, the directional alternative hypothesis H,: 8
, > 0 is
supported at the .05 level (but not at the .O1 level).
3. Tie correction for the normal approximation of the Wilcoxon test statistic Equation
18.4 (which is identical to Equation 6.4) is an adjusted version of Equation 18.2 that is
recommended in some sources (e.g., Daniel (1990) and Marascuilo and McSweeney (1977))
when tied difference scores are present in the data. The tie correction (which is identical to the
one described for the Wilcoxon signed-ranks test) results in a slight increase in the absolute
value of z. Unless there are a substantial number of ties, the difference between the values of
z computed with Equations 18.2 and 18.4 will be minimal.
z =
n(n
Copyright 2004 by Chapman & Hal/CRC
1)(2n
1)
(Equation 18.4)
Ct3
Ct
Test 18
617
Table 18.4 illustrates the application of the tie correction with Example 18.1. In the data
for Example 18.1 there are three sets of tied ranks: Set 1 involves three subjects (Subjects 1,9,
and 10); Set 2 involves two subjects (Subjects 3 and 4); Set 3 involves three subjects (Subjects
5,7, and 8). The number of subjects involved in each set oftied ranks represents the values of
t in the third column of Table 18.4. The three t values are cubed in the last column ofthe table,
after which the values C t and C t are computed. The appropriate values are now substituted
in Equation 18.4.4
Table 18.4 Correction for Ties with Normal Approximation
Subject
Rank
t3
The absolute value z = 2.15 is slightly larger than the absolute value z = 2.13 obtained
without the tie correction. The difference between the two methods is trivial, and in this
instance, regardless of which alternative hypothesis is employed, the decision the researcher
makes with respect to the null hypothesis is not affected.s
Conover (1980, 1999) and Daniel (1990) discuss andlor cite sources on the subject of
alternative ways of handling tied difference scores. Conover (1980, 1999) also notes that in
some instances retaining and ranking zero difference scores may actually provide a more
powerfid test of an alternative hypothesis than the more conventional method employed in this
book (which eliminates zero difference scores fiom the data).
4. Sources for computing a confidence interval for the Wilcoxon matched-pairs signed
ranks test Conover (1980,1999), Daniel (1990), and Marascuilo and McSweeney (1977) describe procedures for computing a confidenceinterval for the Wilcoxon matched-pairssignedranks test - i-e., computing a range of values within which a researcher can be confident to
a specified degree (or that the probability is) that a difference between two population medians
falls.
618
References
Bell, C. B. and Doksum, K. A. (1965). Some new distribution-fiee statistics. Annals of Mathematical Statistics, 36,203-214.
Conover, W. (1980). Practical nonparametric statistics (2nd ed.). New York: John Wiley &
Sons.
Copyright 2004 by Chapman & Hal/CRC
Test 18
619
Conover, W. (1999). Practical nonparametric statistics (3rd ed.). New York: John Wiley &
Sons.
Daniel, W. (1990). Applied nonparametric statistics (2nd ed.). Boston: PWS-Kent
Publishing Company.
Fisher, R. A. (1935). The design of experiments (7th ed.). Edinburgh-London: Oliver &
Boyd.
Hollander, M. and Wolfe, D. A. (1999). Nonparametric statistical methods. New York: John
Wiley & Sons.
Marascuilo, L. and McSweeney, M. (1977). Nonparametric and distribution-free methods
for the social sciences. Monterey, CA: BrookdCole Publishing Company.
Sheskin, D. J. (1984). Statistical tests and experimental design: A guidebook. New York:
Gardner Press.
Siegel, S. and Castellan, N., Jr. (1988). Nonparametricstatistics for the behavioral sciences
(2nd 4.). New York: McGraw-Hill Book Company.
Van der Waerden, B. L. (195211953). Order tests for the twesample problem and their power.
Proceedings Koninklijke Nederlandse Akademie van Wetenshappen(A), 55 (Indagationes Mathematicae 14),453-458, and 56 (Indagationes Mathematicae, 15), 303-3 16
(corrections appear in Vol. 56, p. 80).
Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics, 1, 80-83.
Wilcoxon, F. (1949). Some rapid approximate statistical procedures. Stamford, CT: Starnford Research Laboratories, American Cyanamid Corporation.
Endnotes
1. Some sources note that one assumption ofthe Wilcoxon matched-pairs signed-ranks test
is that the variable being measured is based on a continuous distribution. In practice, however, this assumption is often not adhered to.
2. When there are tied scores for either the lowest or highest difference scores, as a result of
averaging the ordinal positions of the tied scores, the rank assigned to the lowest difference
score will be some value greater than 1, and the rank assigned to the highest difference score
will be some value less than n.
3. A more thorough discussion of Table A5 can be found in Section V ofthe Wilcoxon signedranks test.
x=l
The computed value of 54 is the same as the corresponding value (Ct - Ct) = 62 - 8
in Equation 18.4 through use of Table 18.4.
= 54 computed
620
5. A correction for continuity can be used in conjunction with the tie correction by subtracting
.5 fiom the absolute value computed for the numerator of Equation 18.4. Use of the correction for continuity will reduce the tie-corrected absolute value of z.