Statistics – 5th Week Michaelmas Term
Non-parametric tests of location
This week we are looking at non-parametric tests. You need to know how to carry out the following
tests:
• Rank sum test (sometimes called Mann-Whitney rank sum test or Wilcoxon rank sum test)
• Wilcoxon paired sample test
• Median test
• Sign test
You also need to know which tests are suitable for which data (paired vs unpaired), which are more
powerful and why (covered in the lecture).
Concepts
1. What assumptions must be met for a t-test to be used to compare two samples?
Which of these must be met to use the Rank-Sum test (also known as the Mann-Whitney U
test)?
2. One advantage of non-parametric tests is that they are robust to outliers. Why is this the
case?
3. Consider the following list of tests
Paired sample t-test
Independent samples t-test
Single sample t-test
Wilcoxon test for matched pairs
Rank sum test
Median test
Sign test
Which of the tests is suitable for which sort of data?
a) Two independent samples
b) Two dependent samples?
Which tests are
c) Parameteric
d) Non-parametric?
e) Imagine that our data really are two independent samples and the population data
distribution is Normal. List all the tests we could validly use and put them in order of
sensitivity (sensitivity means how likely they are to detect a difference between
samples, if there is one)
Give a brief explanation for your rankings.
f) Imagine our data are really two paired/dependent samples and the population data
distribution is Normal. Repeat the exercise in part e.
4. t-tests compare means between groups whilst non-parametric tests compare medians. Why
(hint – the median is the 50th centile, how does this relate to the idea of ranking data
common to many non-parametric tests?)
5. Two of the non-parametric tests discussed this week are closely related to the binomial
distribution. Which ones? Explain the relationship.
Calculations
(adapted from the text book by Howell)
6. Kapp, Frysinger, Gallagher and Hazelton (1979) have demonstrated that lesions in the
amygdala, a part of the brain’s limbic system, can reduce certain physiological responses
associated with fear. If fear is really reduced in animals with amygdala lesions, then it should
be more difficult to train these animals to avoid a frightening stimulus.
Consider two groups of rabbits: one group with amygdala lesions and another control group
without lesions. The following data represent the number of trials taken to learn an
avoidance response for each rabbit.
Rabbit number No. trials to learn fear response
Lesion group Control group
1 15 9
2 14 4
3 15 9
4 8 10
5 7 6
6 22 6
7 36 4
8 19 5
9 14 9
10 18
11 17
a) Plot the data and comment with regards to whether they Normally distributed
b) Compare them with two appropriate non-parametric tests
…In each case clearly state your hypotheses.
c) A researcher carries out an independent samples t-test on the same data and finds
t=3.69, dof=18, p=0.00084 (one tailed).
Which test gives the most significant result, out of the t-test and the two tests you
performed?
Which test is most appropriate?
7. It has been argued that first-born children tend to be more independent than later born
children. Suppose we score independence using on a 25 point scale where a higher score
means the person is more independent. We then sample 20 sibling pairs (in adulthood to
minimise effects of age at time of testing). Here are the data:
Sibling pair number Independence score
First born Second born
1 12 10
2 18 12
3 13 15
4 17 13
5 8 9
6 15 12
7 16 13
8 5 8
9 8 10
10 12 8
11 13 8
12 5 9
13 14 8
14 20 10
15 19 14
16 17 11
17 2 7
18 5 7
19 15 13
20 18 12
There are quite a few data points so this question will be easier if you enter the data into a
computer spread sheet (eg Excel, or use Google spread sheets online) and do the calculations
there. If you do so, though, you need to give your tutor some evidence of your working.
a) Plot the data in a way that shows the relationship between scores for sibling pairs.
b) Are the data Normally distributed?
c) Are the differences Normally distributed?
d) Analyse the data using two appropriate non-parametric tests. Clearly state your
hypotheses and conclusions in each case.
e) Which test is most sensitive?
f) Which test is best and why?
Tutor-marked questions
8. A new drug has been developed which the manufacturer claims reduced the frequency with
which patients with post traumatic stress disorder experience flashbacks. In a study, two
groups of patients were given either the drug, or a placebo. They were asked to log each
flashback over a one week period, using a mobile phone app. The total number of flashbacks
for each patient is shown in the table below
Patient ID 1 2 3 4 5 6 7 8 9 10
DRUG 2 3 0 1 0 9 11 12 1 10
Patient ID 11 12 13 14 15 16 17 18 19 20
PLACEBO 1 3 4 7 5 6 6 4 7 8
a) We would like to know if the drug reduces the number of flashbacks. Name three tests
that could be used to address this question
b) Plot the data for each group separately and comment, with respect to the
appropriateness of each test named in part a-
c) Carry out two appropriate tests to determine whether the drug reduces the number of
flashbacks. In each case state your hypotheses, test statistic, alpha level, critical value and
conclusions
d) Which of the two tests used in part c is more likely to detect a drug effect, if there is one?
Why?
Imagine that instead of two separate groups of patients, one group of patients experienced
both the drug and placebo conditions in different time periods (one period on the drug, the
other on the placebo). This is a within-subjects design.
e) Without calculation, state whether you would expect to be more likely to detect a small
reduction in flashbacks using a within-subjects design, as opposed two independent
groups. Give a reason.
9. The following table gives the scores on a happiness questionnaire for 8 individuals before and
after an intervention in which they spent 15min per day cuddling a kitten.
score
Patient ID before after
1 54 60
2 59 62
3 83 88
4 73 69
5 51 61
6 58 57
7 72 84
8 61 68
A kitten therapist hypothesizes that the happiness scores will be higher after the kitten
intervention.
a) Plot the data and comment
b) Carry out two non-parametric statistical tests to determine whether the hypothesis was
correct.
c) Comment on the results of the two tests
