Non-parametric methods

Biostat 2 _module 2
 Non-parametric methods
• Although confidence intervals and hypothesis testing are rather
different approaches to statistical analysis, they have a close
mathematical link for the majority of the statistical methods.

• This is true because they are both based on the same statistical
model and the same assumptions about sampling distributions.

• Theoretical distributions are described by quantities called

parameters, notably the mean and standard deviation.

• These methods are called parametric because we estimate the

parameters of the underlying Normal distribution.

• Methods which do not assume a particular family of distributions for

the data are said to be non-parametric or distribution-free methods.
 Non-parametric methods
• As they do not usually involve any distributional
assumptions, non-parametric methods are most
often used to analyze data which do not meet
the distributional requirements of parametric
methods.

• Skewed data are commonly analyzed by non-

parametric methods, and methods using ranks
are especially suitable for data which are scores
rather than measurements.
The Wilcoxon signed rank sum test
• As with the paired t-test the Wilcoxon matched-pairs
test is used when the groups are not independent (in
situations when the differences do not follow a normal
distribution).

The method has 3 steps

1. Calculate the difference between each observation and
the value of interest
2. Ignoring the signs of the differences, rank them in
order o magnitude
3. Calculate the sum of the ranks of all the negative (or
positive) ranks, corresponding to the observations
below (or above) the chosen hypothetical value
The Wilcoxon signed rank sum test
• For small samples (up to 25) P-values can be obtained
from the Table of the Wilcoxon one sample (or matched
pairs) test.

• For larger samples the test statistics has an

approximately Normal distribution, with mean n(n+1)/4
and variance n(n+1)(2n+1)/24

• Zero differences are omitted from the calculations, so in

this formula n is the number of non-zero differences, and
so may be less than the sample size
The Wilcoxon signed rank sum test
The following Table shows the dietary intakes of 11 women together with the
differences from the recommended intake

Daily energy intake of 11 healthy women with rank order of differences

(ignoring their signs) from the recommended intake of 7725 kJ
subject Energy intake Difference from Ranks of
(kJ) 7725 kJ differences
1 5260 2465 11
2 5470 2255 10
3 5640 2085 9
4 6180 1545 8
5 6390 1335 7
6 6515 1210 6
7 6805 920 4
8 7515 210 1.5
9 7515 210 1.5
10 8230 -505 3
11 8770 -1045 5
Altman D., Practical statistics for medical research, 1999
The Wilcoxon signed rank sum test
• The sum of the ranks of the two observed intakes above
the recommended 7725 kJ is 3+5 = 8
• From the Table of the Wilcoxon one sample (or matched
pairs) test, we get P<0.05.
• We could equally have used the sum of the ranks of the
intakes below the recommended intake, which is
1.5+1.5+4+6+7+8+9+10+11=58
• From the same Table, we get P <0.05.
• It is always worth checking that the ranks have been
calculated correctly, which is easy because the sum of
all the ranks is n(n+1)/2.
• Here we have 11 X12/2 = 66 and also 8+58 = 66.
 parametric or non-parametric
• In practice we usually perform only one
analysis of a set of data, choosing
between parametric or non-parametric
alternatives.

• We usually use a parametric method

unless there is some clear indication that it
is not valid, that is if the underlying
assumptions are not met.
Non-parametric comparison of two
groups- the Mann-whitney test
• There is a non-parametric alternative to
the t-test for comparing data from two
independent groups.
• There are two derivatives of the test, one
due to Wilcoxon and the other to Mann
and Whitney.
• In order to avoid the confusion (with the
paired test also due to Wilcoxon) we call
this method the Mann-Whitney test.
 The Mann-whitney test
• The Mann-whitney test requires all the
observations to be ranked as if they were
from a single sample.

• Then the sum of the ranks in one group is

calculated and a P-value is obtained from
the Table of the Mann-whitney test
(Wilcoxon two sample test).
 The Mann-whitney test
Twenty four hour total energy expenditure (MJ/day) in groups of lean and
obese women (Prentice et al., 1986)

Lean (n=13) Obese (n=9)

6.13 8.79
7.05 9.19
7.48 9.21
7.48 9.68
7.53 9.69
7.58 9.97
7.90 11.51
8.08 11.85
8.09 12.79
8.11
8.40
10.15
10.88
Calculations for the Mann-Whitney test on
energy expenditure (EE) data (MJ/day)
Lean (n = 13 ) Obese (n = 9)
Rank EE Rank EE
1 6.13
2 7.05
3.5 7.48
3.5 7.48
5 7.53
6 7.58
7 7.90
8 8.08
9 8.09
10 8.11
11 8.40
12 8.79
13 9.19
14 9.21
15 9.68
16 9.69
17 9.97
18 10.15
19 10.88
20 11.51
21 11.85
22 12.79

Sum = 103 Sum = 150

 The Mann-whitney test
• The sums of the ranks in the two groups
are 150.

• We can check our calculations by noting

that the sum of all ranks of N observations
must be N(N+1)/2, which here is 253.

• We can use two alternative statistics, T

and U
 The Mann-whitney test
• The statistic T (due to Wilcoxon) is simply the sum of the
ranks in the smaller group (i.e.,150 in this example).
• Either group can be taken if they are of the same size.
• The statistic U (due to Mann and Whitney) is relatively
more complicated.
• The Table of the Mann-whitney test (Wilcoxon two
sample test) gives us the critical values of the statistic T,
showing that with sample sizes of 9 and 13 the rank sum
of 150 is outside the P=0.01 range of expected rank
sums under the null hypothesis but not outside the P =
0.001 range.
• Therefore, we write P < 0.01