Wilcoxon

Distribution-free testing
If the data are normally distributed, we may apply a z-

test or t-test when the parameter of interest is . But
what if this is the normal probability plot:
Or this:
1
The Box-plots also indicate non-normality: skewness
in the first case and more outliers than expected under
normality in the second case:
2
When the data are clearly non-normal, we need an
alternative: distribution-free testing. Also known
under the terms: nonparametric testing or rank tests.
Characteristic of such tests: they can be applied

irrespective of what the true probability distribution of
the data is (which we prove later).
Disadvantage: the power of these tests is (slightly)

smaller than those of normal-based tests if the data are
normally distributed.
Two of the most widely applied tests:
Wilcoxon rank sum test: test H0 : 1 = 2 for two

unpaired samples.
Wilcoxon signed-rank test: Either one sample and

test: H0 :  = 0 or two samples, paired, test H0 :
d = 0, with d = 1 -2.
3
Hypothesis testing, once again
Steps of hypothesis testing
1. Parameter of interest (, 2, p)? Assumptions?

Normal distribution, yes/no?
2. Hypotheses. One- or two-sided?
3. Testing with?
a) Computer: p-values
b) Table of critical values
c) Asymptotic distribution of test statistic
4. Reject null hypothesis if

a) p-value smaller or equal to 
b) Value of test statistic lies in critical region.
c) Asymptotic p-value is smaller or equal to 
4
Wilcoxon rank sum test
1. Parameter of interest:  = 1 - 2
2. H0 :  = 0 (or 1 = 2). Alternative is either one-

or two-sided
3. Computation of exact p- value with software

(StatXact, Mathematica), tables of critical values
are available in Statistisch Compendium (Dutch).
Asymptotic normal distribution of the test statistic

will be proven.
4. Reject null hypothesis if

a) p-value smaller or equal to 
b) Value of test statistic lies in critical region.
c) Asymptotic p-value is smaller or equal to 
5
Test statistic
1: Rank the observations from small to large. Equal

observations (ties) are assigned an average rank
number (so, if the 4th, 5th and 6th observations are tied
then they all correspond to rank (4+5+6) / 3 = 5.
Prothesis 1 Prothesis 2 Ranks 1 Ranks 2
3.2 2.3 8 7
4.3 4.3 9.5 9.5
5.6 1.3 11 5
1.2 2.2 4 5
0.6 0.7 2 3
0.4 1
Sum = 34.5
2: Test statistic W: sum of ranks of the smallest

sample, 34.5 in the example.
3: Find critical region for type I error .
Note: for one-sided testing, first multiply  by 2 to

obtain the critical value if the table is based on two-
sided testing.
6
Smallest sample, sample ‘1’. The other is sample ‘2’.
If
a) H1: 1 < 2 , use the left-critical value only
b) H1: 1 > 2 , use right-crititical value only
c) H1: 1  2 , use both.
In the example (n = 5 and m = 6) for  = 0.05:

a) look at  = 0.1: WL = 20, so reject H0 if W  20
b) WR = n(m + n + 1) – WL = 5*12 – 20 = 40. So, reject

H0 if W  40
c) look at  = 0.05. WL = 18, WR = 42. Reject H0 if W  18

or W  42.
7
Prothese 1 Prothese 2 Rangnrs 1 Rangnrs 2
3.2 2.3 8 7
4.3 4.3 9.5 9.5
5.6 1.3 11 5
1.2 2.2 4 5
0.6 0.7 2 3
0.4 1
Som = 34.5
4:Compare test statistic with critical region. One-sided

testing (first protesis is a new design) : H1: 1 > 2 ,
because you want to show that the new protesis is
better than the old one.
W = 34.5 < 40 so do not reject H0: we cannot say that

the new protesis is better than the old one.
m
Later we show that for n  , m  , N
 c,
where c > 0,
W  n(m  n  1) / 2
z0 
mn(m  n  1) / 12
is standard normal distributed, so use this for values

not in the Wilcoxon table.
8
Wilcoxon signed-rank test
Situation:
a) One sample and test: H0 :  = 0
b) Two samples, paired, test H0 : d = 0, with
d = 1 -2.
In both cases we create one sample of which the mean

should be approximately equal to ‘0’ if H0 holds.
a) Subtract specified 0 from each observation
Data: 3.9, 2.3, 4.0, 4.5, 1.5, 2.2, 1.7, 3.6, 6.1, 1.2, 5.3,
3.3, -0.6, 5.2, 0.2, 0.9, 2.6, 2.2, 3.4, 2.8
Hypothesis: H0:  = 2. Sample to which we will

apply the test:
Data:1.9, 0.3, 2.0, 2.5, -0.5, 0.2, -0.3, 1.6, 4.1, -0.8, 3.3,
1.3, -2.6, 3.2, -1.8, -1.1, 0.6, 0.2, 1.4, 0.8
b) Compute pairwise differences. These differences

are the sample to which we will apply the test .
9
Assumption: density f (x) of one-sample data is
symmetric.
Data: Before introducing a new beer on the market,

the brewery wants to know whether people appreciate
it more than an existing beer. Fifteen people give
marks to both beers (blind test).
The null hypothesis is H0 : d = 0, with d = nieuw –
bestaand and the alternative hypothesis is H1 : d > 0.
Person Mark, new Mark, old Difference

1 6 4 2
2 8 3 5
3 4 7 -3
4 8 6 2
5 9 5 4
6 6 8 -2
7 7 4 3
8 5 5 0
9 8 6 2
10 8 5 3
11 8 8 0
12 7 5 2
13 9 7 2
14 5 4 1
15 6 5 1
10
Data and ranks:
Person New Old Differ. Abs. Rank Abs Positive
1 6 4 2 2 7.5 yes
2 8 3 5 5 15 yes
3 4 7 -3 3 12 no
4 8 6 2 2 7.5 yes
5 9 5 4 4 14 yes
6 6 8 -2 2 7.5 no
7 7 4 3 3 12 yes
8 5 5 0 0 1.5 zero
9 8 6 2 2 7.5 yes
10 8 5 3 3 12 yes
11 8 8 0 0 1.5 zero
12 7 5 2 2 7.5 yes
13 9 7 2 2 7.5 yes
14 5 4 1 1 3.5 yes
15 6 5 1 1 3.5 yes
Test statistic W+: sum of ranks corresponding to

positive observations plus half of the sum of ranks
corresponding to ‘0’. Ties: average the ranks. In the
example: W+ = 99.
11
Example Alternative hypothesis was H1 : d > 0, so
one-sided test for  = 0.05. Right critical value WR =
n(n+1) / 2 – WL = 120 – 30 = 90.
W+ = 99 > WR = 90, so reject H0: d = 0: new beer is

significantly nicer than the old beer.
Normal approximation
 W   n(n  1) / 4 99  n( n  1) / 4
P (W  99)  P (  )
n(n  1)(2n  1) / 24 n(n  1)(2n  1) / 24
 P ( Z  2.15)  0.015.
12

Wilcoxon

Uploaded by

Copyright:

Available Formats

Wilcoxon

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wilcoxon

Uploaded by

Copyright:

Available Formats

Distribution-free testing

If the data are normally distributed, we may apply a z-

Characteristic of such tests: they can be applied

Disadvantage: the power of these tests is (slightly)

Two of the most widely applied tests:

Wilcoxon rank sum test: test H0 : 1 = 2 for two

Wilcoxon signed-rank test: Either one sample and

1. Parameter of interest (, 2, p)? Assumptions?

2. Hypotheses. One- or two-sided?

4. Reject null hypothesis if

2. H0 :  = 0 (or 1 = 2). Alternative is either one-

3. Computation of exact p- value with software

Asymptotic normal distribution of the test statistic

4. Reject null hypothesis if

1: Rank the observations from small to large. Equal

2: Test statistic W: sum of ranks of the smallest

3: Find critical region for type I error .

Note: for one-sided testing, first multiply  by 2 to

In the example (n = 5 and m = 6) for  = 0.05:

b) WR = n(m + n + 1) – WL = 5*12 – 20 = 40. So, reject

c) look at  = 0.05. WL = 18, WR = 42. Reject H0 if W  18

4:Compare test statistic with critical region. One-sided

W = 34.5 < 40 so do not reject H0: we cannot say that

is standard normal distributed, so use this for values

In both cases we create one sample of which the mean

Hypothesis: H0:  = 2. Sample to which we will

b) Compute pairwise differences. These differences

Data: Before introducing a new beer on the market,

Person Mark, new Mark, old Difference

Test statistic W+: sum of ranks corresponding to

W+ = 99 > WR = 90, so reject H0: d = 0: new beer is

You might also like