Chapter 9

Nonparametric Tests

9.1 Sign Test

The reader should recall that the procedures discussed in Chapter 5
for testing the null hypothesis that 𝜇 = 𝜇0 are valid only if the
population is approximately normal or if the sample is large. If n<30
and the population is decidedly nonnormal, we must resort to a
nonparametric test. The sign test is used to test hypotheses on a
population median. In the case of many of the nonparametric
procedures, the mean is replaced by the median as the pertinent
location parameter under test. The population counterpart, denoted
by𝜇̃, has an analogous definition. Given a random variable X,𝜇̃ is
defined such that P(X>𝜇̃) ≤ 0.5 andP (X<𝜇̃) ≤ 0.5. In the continuous
case, P(X>𝜇̃)=P(X<𝜇̃)=0 .5.

The appropriate test statistic for the sign test is the binomial random
variable X, representing the number of plus signs in our random
sample. If the null hypothesis that 𝜇̃ = 𝜇̃0 is true, the probability that
a sample value results in either a plus or a minus sign is equal to 1/2.
Therefore, to test the null hypothesis that𝜇̃ = 𝜇̃0 , we actually test the
null hypothesis that the number of plus signs is a value of a random
variable having the binomial distribution with the parameter p =1 /2.
P-values for both one-sided and two-sided alternatives can then be
calculated using this binomial distribution. For example, in testing
𝐻0 : 𝜇̃ = 𝜇̃0 , 𝐻1 : 𝜇̃ < 𝜇̃0 ,
we shall reject 𝐻0 in favor of 𝐻1 only if the proportion of plus signs is
suffi ciently less than 1/2, that is, when the value 𝑥 of our random
variable is small. Hence, if the computed P-value
 𝑃 = 𝑃(𝑋 ≤ 𝑥 𝑤ℎ𝑒𝑛 𝑝 = 1 /2)
is less than or equal to some preselected significance level α, we reject
𝐻0 in favor of 𝐻1 . For example, when n = 15 and𝑥 = 3, we findthat
3
1 1
𝑃 = 𝑃 (𝑋 ≤ 𝑥 𝑤ℎ𝑒𝑛 𝑝 = ) = ∑ 𝑏 (𝑥; 15, )
2 2
𝑥=0
3 𝑥 15−𝑥 3 15
15 1 1 15 1
= ∑( )( ) ( ) = ∑( )( )
𝑥 2 2 𝑥 2
𝑥=0 𝑥=0
15
1 15 15 15 15
= ( ) (( ) + ( ) + ( ) + ( ))
2 0 1 2 3
= 0.0000305(1 + 15 + 105 + 455) = 0.0176
so the null hypothesis 𝜇̃ = 𝜇̃0 can certainly be rejected at the
0.05level of significance but not at the 0.01 level. To test the
hypothesis
𝐻0 : 𝜇̃ = 𝜇̃0 , 𝐻1 : 𝜇̃ > 𝜇̃0 ,
we reject 𝐻0 in favor of 𝐻1 only if the proportion of plus signs is
suffi ciently greater than 1/2, that is, when the value 𝑥 of our random
variable is small. Hence, if the computed P-value
𝑃 = 𝑃(𝑋 ≥ 𝑥 𝑤ℎ𝑒𝑛 𝑝 = 1 /2)

is less than 𝛼, we reject 𝐻0 in favor of 𝐻1 .

Finally, to test the hypothesis
𝐻0 : 𝜇̃ = 𝜇̃0 , 𝐻1 : 𝜇̃ ≠ 𝜇̃0 ,
we reject 𝐻0 in favor of 𝐻1 when the proportion of plus signs is
significantly less than or greater than 1/2. This, of course, is
equivalent to x being sufficiently small or sufficiently large.
Therefore, if 𝑥 < 𝑛/2 and the computedP-value
𝑃 = 2𝑃(𝑋 ≤ 𝑥 𝑤ℎ𝑒𝑛 𝑝 = 1 /2)
is less than or equal to α, or if𝑥 > 𝑛/2 and the computed P-value
𝑃 = 2𝑃(𝑋 ≥ 𝑥 𝑤ℎ𝑒𝑛 𝑝 = 1 /2)
is less than or equal to α, we reject 𝐻0 in favor of 𝐻1 .

Example 9.1
The following data represent the number of hours that a rechargeable
hedge trimmer operates before a recharge is required:
1.5,2.2,0.9,1.3,2.0,1.6,1.8,1.5,2.0,1.2,1.7.
Use the sign test to test the hypothesis, at the 0.05 level of
significance, that this particular trimmer operates a median of
1.8hours before requiring a recharge.
Solution
1. 𝐻0 : 𝜇̃ = 1.8.
2. 𝐻1 : 𝜇̃ ≠ 1.8.
3. 𝛼 = 0 .05.
4. Test statistic: Binomial variable X with 𝑝 = 1 /2.
5. Computations: Replacing each value by the symbol “+” if it
exceeds 1.8 and by the symbol “−” if it is less than 1.8 and discarding
the one measurement that equals 1.8, we obtain the sequence
1.5 2.2 0.9 1.3 2.0 1.6 1.8 1.5 2.0 1.2 1.7
, , , , , , , , , ,
− + − − + − − + − −
for which n = 10, x = 3, and n/2 = 5. Therefore, the computed P-value
is
 3
1
𝑃 = 2𝑃(𝑋 ≤ 3 𝑤ℎ𝑒𝑛 𝑝 = 1 /2) = 2 ∑ 𝑏 (𝑥; 10, )
2
𝑥=0
3 𝑥 3
10 1 1 10−𝑥 10 1
10
= 2∑( )( ) ( ) = 2∑( )( )
𝑥 2 2 𝑥 2
𝑥=0 𝑥=0
10
1 10 10 10 10
= ( ) (( ) + ( ) + ( ) + ( ))
2 0 1 2 3
= 2 × 0.0000977(1 + 10 + 45 + 120) = 0.3438 > 0.05
6. Decision: Do not reject the null hypothesis and conclude that the
median operating time is not significantly different from 1.8 hours.

We can also use the sign test to test the null hypothesis 𝜇̃1 − 𝜇̃2 =
𝑑0 for paired observations. Here we replace each diff erence, 𝑑𝑖 , with a
plus or minus sign depending on whether the adjusted difference,
𝑑𝑖 − 𝑑0 , is positive or negative. Throughout this section, we have
assumed that the populations are symmetric. However, even if
populations are skewed, we can carry out the same test procedure, but
the hypotheses refer to the population medians rather than the means.

9.2 Signed-Rank Test

The reader should note that the sign test utilizes only the plus and
minus signs of the differences between the observations and 𝜇̃0 in the
one-sample case, or the plus and minus signs of the differences
between the pairs of observations in the paired-sample case; it does
not take into consideration the magnitudes of these differences. A test
utilizing both direction and magnitude, proposed in 1945 by Frank
Wilcoxon, is now commonly referred to as the Wilcoxon signed-
rank test. The analyst can extract more information from the data in a
nonparametric fashion if it is reasonable to invoke an additional
restriction on the distribution from which the data were taken. The
Wilcoxon signed-rank test applies in the case of a symmetric
continuous distribution. Under this condition, we can test the null
hypothesis 𝜇̃ = 𝜇̃0 . We first subtract 𝜇̃0 from each sample value,
discarding all differences equal to zero. The remaining differences are
then ranked without regard to sign. A rank of 1 is assigned to the
smallest absolute difference (i.e., without sign), a rank of 2 to the next
smallest, and so on. When the absolute value of two or more
differences is the same, assign to each the average of the ranks that
would have been assigned if the differences were distinguishable. For
example, if the fifth and sixth smallest differences are equal in
absolute value, each is assigned a rank of 5.5. If the hypothesis 𝜇̃ =
𝜇̃0 is true, the total of the ranks corresponding to the positive
differences should nearly equal the total of the ranks corresponding to
the negative differences. Let us represent these totals by 𝑤+ and𝑤− ,
respectively. We designate the smaller of 𝑤+ and𝑤−by w. In selecting
repeated samples, we would expect 𝑤+ and𝑤− , and therefore w, to
vary. Thus, we may think of 𝑤+,𝑤− and 𝑤 as values of the
corresponding random variables 𝑊+,𝑊− andW. The null hypothesis
𝜇̃ = 𝜇̃0 can be rejected in favor of the alternative 𝜇̃ < 𝜇̃0 only if 𝑤+ is
small and 𝑤− is large. Likewise, the alternative 𝜇̃ > 𝜇̃0 can be
accepted only if 𝑤+ is large and 𝑤− is small. For a two-sided
alternative, we may reject 𝐻0 in favor of 𝐻1 if either 𝑤+ or𝑤−, and
hence w, is sufficiently small. Therefore, no matter what the
alternative hypothesismay be, we reject the null hypothesis when the
value of the appropriate statistic 𝑊+,𝑊−, or W is sufficiently small.

Two Samples with Paired Observations

To test the null hypothesis that we are sampling two continuous
symmetric populations with 𝜇̃1 − 𝜇̃2 for the paired-sample case, we
rank the differences of the paired observations without regard to sign
and proceed as in the single-sample case. The various test procedures
for both the single- and paired-sample cases are summarized in Table
9.2.

Table 9.2: Signed-Rank Test

𝑯𝟎 𝑯𝟏 Compute
𝜇̃ = 𝜇̃0 𝜇̃ < 𝜇̃0 𝑤+
{𝜇̃ > 𝜇̃0 𝑤−
𝜇̃ ≠ 𝜇̃0 𝑤
𝜇̃1 = 𝜇̃2 𝜇̃1 < 𝜇̃2 𝑤+
{𝜇̃1 > 𝜇̃2 𝑤−
𝜇̃1 ≠ 𝜇̃2 𝑤

It is not difficult to show that whenever n<5 and the level of

significance does not exceed 0.05 for a one-tailed test or 0.10 for a
two-tailed test, all possible values of 𝑤+ ,𝑤− , orw will lead to the
acceptance of the null hypothesis. However, when 5 ≤ n ≤ 30, Table
Signed-Rank Test shows approximate critical values of 𝑊+ and 𝑊−
for levels of significance equal to 0.01, 0.025, and 0.05 for a one-
tailed test and critical values of W for levels of significance equal to
0.02, 0.05, and 0.10 for a two-tailed test. The null hypothesis is
rejected if the computed value 𝑤+ ,𝑤− , or w is less than or equal to the
appropriate tabled value. For example, when n = 12, Table Signed-
Rank Test shows that a value of 𝑤+ ≤ 17 is required for the one-
sided alternative 𝜇̃ < 𝜇̃0 to be significant at the 0.05 level.

Example 9.2:Rework Example 9.1 by using the signed-rank test.

1. 𝐻0 : 𝜇̃ = 1.8.
2. 𝐻1 : 𝜇̃ ≠ 1.8.
3. 𝛼 = 0 .05.
4. Critical region: Since n = 10 after discarding the one measurement
that equals 1.8, Table Signed-Rank Test shows the critical region to
be w ≤ 8.
5. Computations: Subtracting 1.8 from each measurement and then
ranking the differences without regard to sign, we have
𝑑𝑖 -0.3 0.4 -0.9 -0.5 0.2 -0.2 -0.3 0.2 -0.6 -0.1
Ranks 5.5 7 10 8 3 3 5.5 3 9 1
Now 𝑤+ = 13 and 𝑤− = 42, so w = 13, the smaller of 𝑤+ and 𝑤− .
6. Decision: As before, do not reject H0 and conclude that the median
operating time is not significantly different from 1.8 hours.

The signed-rank test can also be used to test the null hypothesis that
𝜇̃1 − 𝜇̃2 = 𝑑0 . In this case, the populations need not be symmetric. As
with the sign test, we subtract 𝑑0 from each difference, rank the
adjusted differences without regard to sign, and apply the same
procedure as above.

Example 9.3:
It is claimed that a college senior can increase his or her score in the
major field area of the graduate record examination by at least 50
points if he or she is provided with sample problems in advance. To
test this claim, 20 college seniors are divided into 10 pairs such that
the students in each matched pair have almost the same overall grade-
point averages for their first 3 years in college. Sample problems and
answers are provided at random to one member of each pair 1 week
prior to the examination. The examination scores are given in Table
9.3.
 Table 9.3: Data for Example 9.3
Pair
1 2 3 4 5 6 7 8 9 10
With Sample Problems 531 621 663 579 451 660 591 719 543 575
Without Sample Problems 509 540 688 502 424 683 568 748 530 524

Test the null hypothesis, at the 0.05 level of significance, that sample
problems increase scores by 50 points against the alternative
hypothesis that the increase is less than 50 points.
Solution:
Let 𝜇̃1 and 𝜇̃2 represent the median scores of all students taking the
test in question with and without sample problems, respectively.
1. 𝐻0 : 𝜇̃1 − 𝜇̃2 = 50.
2. 𝐻1 : 𝜇̃1 − 𝜇̃2 < 50.
3. 𝛼 = 0 .05.
4. Critical region: Since n = 10, Table Signed-Rank Test shows the
critical region to be 𝑤+ ≤ 11.
5. Computations:
Pair
1 2 3 4 5 6 7 8 9 10
𝑑𝑖 22 81 -25 77 27 -23 23 -29 13 51
𝑑𝑖 − 𝑑0 -28 31 -75 27 -23 -73 -27 -79 -37 1
Ranks 5 6 9 3.5 2 8 3.5 10 7 1

Now we find that 𝑤+ = 6 + 3.5 + 1 = 10.5.

6. Decision: Reject 𝐻0 and conclude that sample problems do not, on
average, increase one’s graduate record score by as much as 50
points.
Exercises: 16.1, 16.5, 16.8, 16.10, 16.12