Sie sind auf Seite 1von 10

Non-Parametric Tests

PROS:
Distribution free methods
Appealing and intuitive
Can be used for non-quantitative data
CONS:
Less efficient and powerful than parametric
techniques (if both are applicable)

What are we to take up?

Sign Test
Signed Rank Test
Rank-Sum Test
Kruskal-Wallis Test
Runs Test

Sign Test
Used to test single population median or
difference of two population medians
~

H0: ~0 ~
~ ~
~ ~
H1: (a) 0 (b) 0 (c) 0
C.R.

(a) All x values such that P(X x when p=1/2) <


(b) All x values such that P(X x when p=1/2) <
(c) if x < n/2, all x values such that 2P(X x when p=1/2) < or
if x > n/2, all x values such that 2P(X x when p=1/2) <

Sign Test
Computation:
~
1. Replace each sample value > 0 with a + sign and each
~
sample value < 0 with a sign.
~
2. Exclude sample values equal to 0 and reduce sample
size.
3. Solve for corresponding P from binomial distribution or
if n>10 and np = nq >5 use normal approximation
(remember to adjust for continuity)
z

x np
npq

Signed Rank Test

Utilizes direction and magnitude of


difference
Applies only for symmetric continuous
distribution
To test Ho
0

H1

Compute
w+

w-

0
1 2

w
w+
w-

1 2

1 2

1 2

Signed Rank Test

Computation:
1. Subtract 0 from each sample value
2. Discard all differences equal to zero

3. Rank the differences regardless of sign. Asign rank of 1 to smallest


absolute difference. When the absolute value of 2 or more differences
is the same, assign to each the average of the ranks that would have
been assigned if the differences were distinguishable.
4. Sum up the ranks of all positive differences (w+) and all negative
differences (w-). Let w = min{w+, w-}.
5. For 5 n 30, C.R. is w values from Table A.16
Use also for paired observations . When n > 15 is, use normal
approximation (remember to adjust for continuity)
w

n(n 1)
4

2w

n(n 1)(2n 1)
24

w w
w

Rank Sum Test

For non-normal independent samples

To test Ho

1 2

H1

Compute

1 2

U1

1 2

U2

1 2

Computation:
1. Let n1 be the smaller sample size and n2 the larger
sample size
2. Arrange the n1 + n2 observations of the combined
samples in ascending order
3. Substitute a rank of 1,2, ... n1+n2 fpr each observation
4. Sum up the ranks corresponding to the n1 observations
(W1 ) and the ranks corresponding to the n2 observations
(W2)

Rank Sum Test


5. Compute

6. Let u = min {U1, U2}


7. For C.R. , u values from Table A.17
. When both n1 and n2 > 8, use normal approximation

Kruskal Wallis Test


Use to test equality of k > 2 population means
Generalization of the rank-sum test for k > 2 samples
HO: k independent samples are from identical populations
H1: The k means are not all equal
C.R. H > 2 with v = k -1
Computation:
1. Combine all k samples (with at least 5 observations each)
and arrange in ascending order
2. Substitute appropriate rank from 1, 2, ..., n for each
observation
3. Sum up ranks corresponding to the n, observations in the
ith samples (R, )
4. Compute
k r2
12
h
i 3( n 1)
n( n 1) i 1 ni

Runs Test
Based on the order in which the sample observations are obtained
Run - a subsequence of one or more identical symbols representing a common
property of the data
HO: observations have been drawn at random
C.R. All v such that P(Vv when Ho is true) < /2 or
All v such that P(Vv when Ho is true) < /2
Computation
1. Let n1 be the smaller sample size and n2 be the larger sample size with n=n1 +
n2
2. For n1 & n2 1 0, use Table A.18. Check if v< v or v> v to determine which
CR to use.
Can be used as an alternative to the Wilcoxon two-sample test
1. Combine observations from both samples and arrange them in ascending order
2. Assign letter A to each observation taken from one of the populations and letter
B to the other
. When n1 and n2 increase in size (n1 & n2 >10) normal approximation could be
used

2n1n2
1
n1 n2

v2

2n1n2 ( 2n1n2 n1 n2 )
( n1 n2 ) 2 ( n1 n2 1)

Das könnte Ihnen auch gefallen