BES - LAB 6

Non-parametric Tests 1
1. Objectives
- Explain the Mann-Whitney-Wilcoxon test and Kruskal Wallis test procedure
- Provide instructions on how to use R to run the tests.
- Understand and interpret R outputs and make decisions based on p-value.
2. Procedure

We will use the function [Link] to perform the Mann-Whitney-Wilcoxon test to compare 2
independent samples. The format of the function is:

Ø [Link](x,y, data = name of data frame, alternative =

"[Link]", paired = FALSE, exact = NULL)

where x and y are numeric vectors of data values.

Here are meanings of other arguments:

data name of your data frame that contains the data for the test.
alternative a character string specifying the alternative hypothesis
paired a logical indicating whether you want a paired-samples test. We are considering
independent samples, so choose FALSE.

exact a logical indicating whether an exact p-value should be computed. If this

argument is not specified, an exact p-value is computed by default if the samples
contain less than 50 finite values and there are no ties. Otherwise, a normal
approximation is used.

To conduct Kruskal-Wallis test, we will use the function [Link]( ).

Ø [Link](outcome ~ grouping variable, data = name of data

frame)

Here are the meanings of these arguments:

outcome The outcome variable

Grouping variable Categorical variable that defines the groups
data Name of your dataframe that contains the variables for the test.

3. Exercises

Exercise 1. Two fuel additives are being tested to determine their effects on gasoline mileage. Ten (10)
cars were tested with additive 1 and 10 cars were tested with additive 2 (the cars are randomly assigned
to each group). The following data show the miles per gallon obtained with the two additives. Using α =
0.05, conduct a hypothesis test to see whether there is a significant difference between gasoline mileage
for the two additives. Data for this exercise are hypothetical.

Additive 1 Additive 2

1|Page
 BES - LAB 6

17.3 18.7
18.4 17.8
19.1 21.3
16.7 21
18.2 22.1
18.6 18.7
17.5 19.8
15.4 20.7
16.2 20.2
15.1 26

Set up H! , H" : We want to see if significant differences exist between 2 additives:

H# : The two population distributions of gasoline mileage are the same

H" : The two population distributions are different

Check the assumptions to see whether the MWW test is appropriate.

Because the samples sizes are quite small (10), we produce a stem and leaf display for each sample. Note
that when sample sizes are small, it is difficult to determine whether the populations are normal or not.

Remember to create 2 data vectors of Additive 1 and Additive 2, called add1 and add2:

Ø add1<-c(17.3,18.4,19.1,16.7,18.2,18.6,17.5, 15.4, 16.2, 15.1)

Ø add2<-c(18.7,17.8,21.3,21,22.1,18.7,19.8,20.7,20.2, 26)

Stem and leaf display for Additive 1:

The decimal point is at the |

15 | 14
16 | 27
17 | 35
18 | 246
19 | 1

Stem and leaf display for Additive 2:

The decimal point is at the |

16 | 8
18 | 778
20 | 2703
22 | 1
24 |
26 | 0
These figures show that normality should not be assumed here. Also, the samples are independent and
random samples (why?).

2|Page
 BES - LAB 6

Then if you write the [Link]() function:

Ø ex1<-[Link](add1, add2, alternative="[Link]",

paired=FALSE)

you will have the following ouput:

Wilcoxon rank sum test with continuity correction

data: add1 and add2

W = 6, p-value = 0.001004
alternative hypothesis: true location shift is not equal to 0

Question 1. What is the difference between the choice of exact = TRUE and exact = FALSE?
Question 2. Let’s calculate the test statistic by hand and compare with the value given in R output. Why
are they different?

Type in ?[Link] to see the note about test statistic given by R. It states that R produces the
same value as Mann-Whitney test statistic, which equals the sum of ranks of the first sample with the
minimum value subtracted of 𝐦(𝐦 + 𝟏)/𝟐 (m is sample size of the first one).

Then if you want to express test statistic as sum of ranks, you must add 𝒎(𝒎 + 𝟏)/𝟐 to the value given
in R output. If you want this sum of ranks to appear in the output at the end, try the following
commands:

Ø n1<-length(add1)
Ø ex1$statistic<-ex1$statistic+n1*(n1+1)/2
Ø names(ex1$statistic)<- "T.W"
Ø ex1

The new output will appear:

Wilcoxon rank sum test with continuity correction

data: add1 and add2

T.W = 61, p-value = 0.001004
alternative hypothesis: true location shift is not equal to 0

Important note: We should not use R to run Mann-Whitney-Wilcoxon test in case a sample size is
smaller than 7 and there are ties.

Exercise 2. In the consumer preference study, one group of consumers was “primed” to images that
would appear on product labels, while the other group of consumers was not “primed.” All consumers
were then asked their attitude toward a product on a seven-point scale (from 1 = dislike very much to 7
= like very much). Carry out the Mann-Whitney-Wilcoxon procedure to test if “primed” consumers are
associated with systematically higher preference scores. What do you conclude?

What type of hypothesis is this? What is the type of the data we wish to compare?

3|Page
 BES - LAB 6

At first, we must import file [Link] into R:

Ø brand<-[Link]("[Link]",header=TRUE,sep = ",")
Ø brand

The second command is to check how the data are presented. You will see 2 groups of primed and not
primed customers presented by index number 0 and 1. So you must separate these 2 groups first, using
subset() function:

Ø primed<-subset(brand, Primed == 1)
Ø notprimed<-subset(brand, Primed == 0)

Note that the samples are independent and we do not need to consider normality assumption (why?).

Then, you’re expected to produce the following output:

Wilcoxon rank sum test with continuity correction

data: primed$Preference and notprimed$Preference

T.W = 600.5, p-value = 0.00042
alternative hypothesis: true location shift is greater than 0
95 percent confidence interval:
0.99994 Inf
sample estimates:
difference in location
1.00004

The code to produce the above output is:

Ø ex2 <- [Link](primed$Preference, notprimed$Preference,

alternative="greater", paired=FALSE, exact=FALSE)
Ø ex2

Question 3: Try setting exact = TRUE in the above code and see what happens.

ADDITIONAL EXERCISES

Exercise 3. “Conservationists have despaired over destruction of tropical rainforest by logging, clearing,
and burning”. These words begin a report on a statistical study of the effects of logging in [Link].
Does logging significantly reduce the number of species in a plot after 8 years?

The R output for the test is provided below. Try to demonstrate that the populations are not normal and
that the assumptions of the Mann-Whitney-Wilcoxon test are satisfied.

Wilcoxon rank sum test with continuity correction

data: logged$Species and unlogged$Species

T.W = 72, p-value = 0.029
alternative hypothesis: true location shift is less than 0

4|Page
 BES - LAB 6

Exercise 4. A sample of 20 engineers employed with a company for three years has been rank ordered
with respect to managerial potential. Some of the engineers attended the company’s management-
development course, others attended an off-site management-development program at a local university,
and the remainder did not attend any program. Use the following rankings and α = 0.025 to test for a
significant difference in the managerial potential of the three groups. Remember to state the hypotheses,
and carry out other necessary steps for the test.

Firstly, you must import file [Link] as program data frame into R. Then just apply the
function [Link]. You’re expected to produce the following output:

Kruskal-Wallis rank sum test

data: RankOfManagerialPotential by Program

Kruskal-Wallis chi-squared = 12.611, df = 2, p-value = 0.001826

Exercise 5. In Exercise 3 you compared the number of tree species in plots of land in a tropical rain
forest that had never been logged (Group 1) with similar plots nearby that had been logged 8 years
earlier (Group 3). The researchers also counted species in plots that had been logged just 1 year earlier
(Group 2). Check the normal assumptions, are there features that might prevent use of ANOVA? Use
the Kruskal-Wallis test to compare the distributions of tree counts. State hypotheses, the test statistic and
its p-value, and your conclusions.

Import file [Link] as borneo3 data frame into R.

The R output for the test is provided below:

Kruskal-Wallis rank sum test

data: species by groups

Kruskal-Wallis chi-squared = 9.4382, df = 2, p-value = 0.008923

5|Page