45

Part III

STATISTICS IN RESEARCH
Much of a researcher’s time is spent accumulating and analysing data: data about
temperatures, occurrences, concentrations, durations, etc. The field of statistics deals
with the analysis of data.

Almost always one has imperfect data about a small subset (the sample) of the
intended population. The data is imperfect because of experimental, recording and
other errors. One has data only about a subset of the population because it is impossible
to examine the entire population. A botanist who proposes a law of genetics can only
test this law on a tiny fraction of the plants on the planet (to say nothing about those
plants not yet growing); sooner or later one has to stop and conclude that the patterns
found in the sample are also found in the whole.

We discussed earlier how to find a suitable sample. In the chapters on statistics

we examine how to analyse the sample data and draw conclusions. This is known as
statistical inference.

These chapters proceed as follows. We start with data and how to organize it. We
then look at probability and probability distributions especially the normal distribution.
This brings us to the main purpose of this part: explaining estimation and statistical
tests. These are used to answer questions such as: how confidently can one extrapolate
from the sample data, and does the sample data agree with a preconceived theory.

In this book we examine only the situations a researcher is most likely to encounter.
If you need more information then you should consider a statistics book (e.g. Net82,
Hog93, Fre90, Moo85, Mul87, Wad90) or a friendly statistician. A computer statistics
package can also be useful, both for performing the necessary calculations and for
representing the data in various ways. An example of a full statistics package is the SAS
program; more modest packages include SPSS and Minitab. A spreadsheet program
(such as QuattroPro, Excel or Lotus 1-2-3) can also be useful.
 8 Basic Statistics: Data Organization

In this chapter we introduce common statistical measures and discuss the represen-
tation and organisation of data. The chapter concludes with a discussion of regression,
which is a technique for fitting curves to data.

8.1 STATISTICAL MEASURES

Data can be qualitative or quantitative. Quantitative data has numerical values, e.g.
in the range 0 to 100. Qualitative data has values that fall into categories, e.g. animal,
vegetable or mineral. For example, whether it rained or not yesterday is a qualitative
question (the categories are rained and not rained); how much rain was measured is a
quantitative question. Quantitative data can be discrete or continuous: discrete if it
takes on only whole values and continuous if it takes on any real value in some interval.
The amount of rain yesterday is continuous data; on how many days it rained last year
is discrete data.

A statistical measure, or simply a statistic, is a summary of the data. This data can
be for the entire population or for the chosen sample. We now discuss some population
statistics.

8.1.1 Measures of central tendency

Most statistics of quantitative data are either measures of central tendency (where the
data is centred) or measures of dispersion (how spread out the data is). In the former
category are the mean, the median and the mode.

The population mean is the average and is denoted by the Greek letter µ. If the
population data values are x1 , x2 , . . . , xn , then:

n
1X
µ= xi
n i=1

The symbol Σ (pronouned ‘sigma’) means ‘the sum of’, so the above formula says that

x1 + x2 + x3 + . . . + xn
µ=
n
47

The median is the middle value were the data to be sorted from smallest to biggest. If
there is an even number of values, then the median is usually taken to be the average
of the middle two values. The mode is defined as the most common value(s) in the
data set.

Example: For the data 5, 3, 6, 7, 5, 5, 9, 1, 17, 9,

the mean is 6.7,
the median is 5.5, and
and the mode is 5.

Aside: In this book we use the decimal point in real numbers instead of the comma
(e.g. 5.42 rather than 5,42).

While the mean is more commonly used, the median can be a better summary of
the data if there are extreme values. For example, in statistics on the price of houses
in an area, the median is often given, since the mean is very much influenced by a large
mansion being sold. Say selling prices in thousands of rands were 100, 100, 110, 120,
130, 140, and 980. The mean here is 240 and the median is 120. The median is a better
summary as most people paid around 120 thousand rand for their house.

For qualitative variables (e.g. colour) the mode is the only possible measure. For
ranked data (e.g. 1st, 2nd, 3rd) both the median and the mode are possible, but the
median is preferred as it considers the rankings rather than just frequencies.

8.1.2 Measures of dispersion

In summarising quantitative data, one is also interested in measures of dispersion, that

is, how spread out the data is. Examples of these are the range, percentiles, the variance
and the standard deviation.

One simple statistic is the range: the gap between the largest and smallest values.
This is seldom useful, as there can be eccentric data. For example, the annual incomes of
South Africans range from zero to many million rands—however, the spread of incomes
has very few people at the top extreme and very many at the bottom, and most incomes
lie in a much narrower range.

One is more likely to use percentiles. For example, the 90th percentile is the value
below which lies 90% of the data. In this terminology, the median is the 50th percentile.

Another measure of dispersion is the variance; the variance of the population is

denoted by σ 2 . This is defined as the average squared deviation: if the data values are
x1 , x2 , . . . , xn , then
48

n
1X
2
σ = (xi − µ)2
n i=1

The (positive) square-root of the data is called the standard deviation of the
data; the population standard deviation is denoted by σ. The standard deviation is the
most used measure of dispersion.

Example: For the data 5, 3, 6, 7, 5, 5, 9, 1, 17, 9 (with µ = 6.7),

the range is 16,
the variance is 17.21, and
the standard deviation is 4.15.

One property of the mean and the standard deviation is that if the data comes from
a ‘reasonable’ population (to be defined later), then about 68% of the data lies between
µ − σ and µ + σ (that is, within one standard deviation of the mean), while 95% of the
data lies within two standard deviations of the mean.

8.2 GRAPHICAL REPRESENTATION OF DATA

Sometimes a picture explains the situation much more clearly than a jumble of numbers.
However, do not include a picture just to impress people; pictures should be relevant
and useful. There are several common types of pictures. Spreadsheets and many other
computer packages contain inbuilt graphing functions.

8.2.1 Depiction of qualitative data

A common device for displaying qualitative data is a pie chart. For example, if 60% of
a country’s music exports is kwaito, 25% rock and 15% rap, then this could be depicted
by the pie chart in Figure 1.

Figure 1. An example of a pie-chart.

49

Kwaito

Rap
Rock

8.2.2 Depiction of frequency data

In frequency data, one divides the range of values into intervals and counts the
number of data items that lie in each interval. For example, examination scores are
given as a mark out of 100 and a schoolteacher records the number of A-plus scores (90
– 99), As (80 –89), Bs (70 – 79), Cs (60 – 69), etc. The result is known as a frequency
distribution. An example of a frequency distribution is given in Table 1.

Table 1. Example of frequency data.

Interval Frequency Cumulative Cumulative

frequency percentage
30-39 1 1 1.56
40-49 6 7 10.94
50-59 14 21 32.81
60-69 17 38 59.38
70-79 16 54 84.38
80-89 8 62 96.88
90-99 2 64 100

A histogram is used to depict a frequency distribution in which the intervals have

equal size. In a histogram, one draws for each interval a rectangular bar with the base
of the bar the range of that interval and the height equal to the numbers recorded.
Figure 2a shows a histogram for the discrete data given in Table 1.

Figure 2a. An example of a histogram.

50

Frequency

15

10

Grade
34.5 44.5 54.5 64.5 74.5 84.5 94.5

One can also depict the data with a frequency polygon. In this, one plots for
each interval a point above the midpoint of each interval giving the number of data
items recorded, and then one joins up consecutive plots (see Figure 2b). Note that the
frequency polygon starts and ends at zero: one assumes that the intervals immediately
before and after those given both have 0 data items.

Figure 2b. A frequency polygon.

Frequency

15

10

Grade
20 40 60 80 100

In an ogive one looks at the data cumulatively. One calculates for each interval
the cumulative count (or percentage)—the number of data items up to and including
that interval. Then one plots for each interval a point at the maximum of the interval
giving the cumulative count (or percentage). These points are then joined up to form
the ogive. (If the data is discrete, then the end of an interval is halfway between the
maximum of that interval and the minimum of the next interval. So in our example
the ends are 39.5, 49.5, etc.) The ogive increases from 0 to total count (or 100 percent)
and is often S-shaped. An example is presented in Figure 2c.
51

Figure 2c. An ogive.

Cumulative Percentage Frequency

100%

75%

50%

25%

Grade
20 40 60 80 100

An ogive is also useful in calculating percentiles. For example, if one wishes to know
the 70th percentile, one draws a horizontal line from 70% on the Y-axis and determines
where it intersects the ogive. The corresponding value on the X-axis is the desired
percentile.

Another simple device is a box-and-whiskers diagram. This gives five statistics

of the data: the minimum, the maximum, the median, the 25th percentile and the 75th
percentile. An example is depicted in Figure 3. For this example, the minimum is 3.4,
the 25th percentile is 4.7, the median is 5.7, the 75th percentile is 8.1 and the maximum
is 12.

Figure 3. A box-and-whiskers diagram.

3.4 4.7 5.7 8.1 12

8.2.3 Depicting paired data

In paired data there is a series of observations, each with two values: say (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ).
For example, a chemical engineer measures the arsenic levels in a stream at various dis-
tances from a factory. He obtains a set of paired (distance, level) data. This data can
be represented by a scattergram on X-Y axes in which there is one plot for each data
point (see Figure 4). Such paired data often results from experimental research—e.g.
52

a physicist suspends a spring from the roof, attaches different weights to it and records
the stretch for each weight.

Figure 4. A scattergram.

Based on the data, one might propose a curve which ‘fits’ the data. The data
depicted in the example scattergram suggests a straight line fit. We discuss this type
of problem next.

8.3 REGRESSION

Suppose a meteorologist accumulates data on temperatures at different altitudes. The

question he would like answered is: is there a mathematical relationship between the
two quantities (altitude and temperature); if so, what is this relationship; how strong
is this relationship; and can one use the calculated relationship from the data for other
temperatures?

Typically one starts with a scattergram. Then one looks at it to see whether there
seems to be any pattern. This simplest case is whether there is a straight-line rela-
tionship and what the straight-line involved is. The process one uses to find the best
estimate of a straight-line fit for data is known as linear regression. We deal with
the question of how good the fit is in Section 11.1 (linear correlation).

In regression, one has a hypothesised function y = f (x) which is supposed to give

the relationship between x and y. The function f (x) starts with some undetermined
parameters: in the case of looking for the best line fit, the function is f (x) = A + Bx
53

with parameters A and B. (Recall that B is the slope of the line and A the y-intercept.)
The task is to find the best values of the parameters.

In linear regression the best values of the parameters are given by the following
rather intimidating formulas:

xi yi ) − ( xi ) ( yi ) yi ) − B (
P P P P P
n( ( xi )
B= and A=
xi − ( xi )2
P 2

n
P
n

The key is to calculate, for each pair (xi , yi ) of data, the value of x2i and xi yi and
then calculate the sums. Fortunately, many calculators have a statistical mode for
performing such a calculation.

Example: A study looked at the arsenic levels in a stream as a function of

the distance downstream from a factory.
x (distance in km) 2 4 6 8 10 12 14
y (arsenic in mg/kl) 24.5 18.7 16.3 12.0 9.3 6.2 2.5
A scattergram for this data is given in Figure 5. This suggests a negative
slope straight-line fit.

Figure 5. Scattergram for arsenic example.

25

20

15

10

2 4 6 8 10 12 14

Table 2. Results for the arsenic example

54

xi yi x2i xi yi
2 24.5 4 49.0
4 18.7 16 74.8
6 16.3 36 97.8
8 12.0 64 96.0
10 9.3 100 93.0
12 6.2 144 74.4
14 2.5 196 35.0
56 89.5 560 520.0

Calculations give n = 7, Σ xi = 56, Σ yi = 89.5, Σ xi yi = 520.0, and

Σ x2i = 560. The formulas for linear regression give
7×520−56×89.5 89.5−(−1.75)×56
B= 7×560−562
= −1.75 and A = 7 = 26.79

This means that the best straight-line fit is y = 26.79 − 1.75x. Figure 6
shows this fit.

Figure 6. Line fit on scattergram.

25

20

15

10

2 4 6 8 10 12 14

While linear relationships are the most common, there are many other possible
relationships. For example, the data in the scattergram in Figure 7 suggests a parabolic
fit. To find the best fit, one can uses regression but the actual formulas are beyond the
scope of the book.

Figure 7. Another scattergram.

55

The process of regression is guaranteed to give a fit. For example, one can fit a line
to the data in Figure 7, but this is clearly inappropriate. The best fit is not necessarily
a good fit. Testing whether a fit does in fact reveal a pattern in the data is discussed
in Section 11.1.

Discussion questions and exercises

1. How does one summarize data? Why does one do so?

2. For each of the measures defined in the text, find a situation (a) where the measure
is a good summary of the data and (b) one where it is bad.

3. Find some examples of data depiction in newspapers and scientific magazines.

Are the pictures informative? Are they misleading?

4. A geographer measures the relationship between the population size of a city and
the number of cinemas. Data is
population (ten thousands) 1 5 22 23 40 67
cinemas 1 1 10 10 17 28
(a) Plot the data on a scattergram.
(b) Find the best straight-line fit.
(c) Use the line to estimate the number of cinemas for a city of 500 000 inhabitants.

5. A bacteriologist counts the number of yeast cells in a test-tube at various time

intervals.
yeast 25 60 140 320 700 1605 3781 7017
time 1 2 3 4 5 6 7 8
56

(a) Plot the data on a scattergram.

(b) Describe in words a curve that would fit the data.
(c) (For the mathematically advanced.) Find a fit of the form y = AeBx . (Use
the fact that if y = AeBx then ln y = ln A + Bx and then use linear regression.)
 9 Normal Distribution and Estimation

The concept of probability is the foundation of statistical tests and processes. We

present a brief introduction to probability here. We then introduce the normal distri-
bution, and explain the process of estimation.

9.1 PROBABILITY

The probability of an event is the proportion of time the event can be expected to
occur over the long run. For a simple example, consider tossing a coin. If one repeats
this 1 000 times, and heads comes up 300 times, then it appears that the probability
of heads is 3/10. (Could a coin act this way? Think about it.)

The probability of an event can also be thought of as the likelihood or chance of it

occurring. One classic example is rolling a die. If the die is fair then there is an equal
chance that any of the six sides comes up. So the probability of a fair die coming up
with number 5 is 1 out of 6, or, in symbols:

1
Pr(D = 5) = 6

where D stands for the value on the die.

We need some definitions: The rolling of the die is an experiment. (In statistics,
an experiment is any action where the answer is not predetermined.) The experiment
has a number of outcomes, and associated with each outcome is a probability. For a
repeatable experiment, the probability of a specific outcome amounts to the proportion
of time the outcome is likely to occur. Proportions always lie between 0 and 1, so that
if something occurs 40% of the time, then its probability is 0.4. A probability of 0
means that the event never occurs, a probability of 1 means that it always occurs.

A numerical value resulting from the experiment is known as a random variable.

A discrete random variable takes on only discrete values and often results from counting
something; a continuous random variable may take on a continuous range of values and
often results from measuring something. The quantity D above is a discrete random
variable. For example, a geologist measures the masses of gold nuggets. This is a
continuous random variable: a random variable as it is the unknown value resulting
from an experiment, and continuous as it can have any real number as its outcome.

If one rolls a die 600 times, it is very very unlikely that one will get each number
exactly 100 times. However, if some numbers seem to come up a lot more than others,
58

then one would suspect that the die was crooked (not fair). A simulation the authors
ran produced the following frequencies of the numbers from 1 to 6: 86, 112, 122, 95,
98, 87. What do you think about the die?

The question of when to conclude that a die is crooked is precisely the type of
question that statistics is designed for. And the short answer is: if the probability of a
result as extreme is sufficiently small, then one may conclude that the die is crooked.

9.2 PROBABILITY DISTRIBUTIONS

Often one has an experiment with just two outcomes, such as tossing a coin. These
outcomes are often referred to as success and failure, and the experiment as a trial.
To obtain useful data one repeats the experiment several times and counts the number
of successes.

Suppose we have an experiment with a 60% chance of success. If we perform it 9

times and get 5 successes, how likely is that? The following formula gives the answer
to such questions:

n!
Pr(S = k) = pk (1 − p)n−k
k! (n − k)!

where:

g! = g × (g − 1) × (g − 2) × . . . × 2 × 1, for g nonzero, and 1 for g zero;

S is the random variable counting the successes;
n is the number of trials;
p is the probability of success is a single trial; and
k is the desired number of successes.

In the example we need the case n = 9, p = 0.6 and k = 5.

The calculation says 126 × 0.65 × 0.44 = 25.1%.

The random variable S is known as a binomial random variable. The probabil-

ity distribution of a random variable depicts the probability of each outcome for the
variable. Figure 8 shows the probability distribution for the binomial random variable
with n = 14 and p = 1/2.

Figure 8. A binomial distribution.

59

probability

0.20

0.15

0.10

0.05

outcome
2 4 6 8 10 12 14

We have made one assumption which one must be explicit about: the success or
failure of one trial does not affect the (probability of ) success or failure of another
trial. Statisticians would say the trials must be independent events. If the trials are
dependent, i.e. not independent, then the above formula is not valid. If I toss one
coin and then another coin, the results are independent. If I measure the height and
mass of one child, the results are dependent.

There are other useful probability distributions, including the normal, exponential,
geometric, and Poisson distributions. Of these, the normal distribution is the most
important.

9.3 THE NORMAL DISTRIBUTION

The most common probability distribution for a continuous random variable is the
normal distribution. Figure 9 shows the smooth bell-shaped curve of the normal
distribution.

Figure 9. The normal distribution.

60

68%

µ−σ µ µ+σ

A normal distribution has two parameters: its mean and standard deviation. The
distribution with µ = 0 and σ = 1 is called the standard normal distribution, and the
associated random variable is often denoted by Z. Many natural occurring things have
a normal distribution, including heights of people, speeds of ostriches, and birth-mass
of babies.

9.3.1 Using normal tables

A table of the standard normal distribution is given in Appendix B. This gives the
proportion of the population that will be less than or equal to z for various values of z.

For example, suppose that the heights of South African men had a mean of 1.77
m and a standard deviation of 4 cm, and one wanted to know what proportion is less
than or equal to 1.87 m tall. The method comprises two steps:

1) Calculate the z-value by the formula:

x−µ
z=
σ
2) Look up the proportion in the table in Appendix B.

187−177
For this example, z = 4 = 2.5. Now Appendix B is consulted. The proportion is
0.994.

If you need to find what proportion is greater than z, simply use 1 minus the
proportion less than z. For example, the proportion of men that are more than 1.87m
tall is 1 − 0.994 = 0.006 or 0.6%.
61

9.4 STATISTICAL ESTIMATION

A fundamental problem of statistics is the estimation of measures of populations such

as the population mean and standard deviation. A point estimate is a single-value
estimate of the population measure. A confidence interval gives a range of values in
which the measure probably lies—typically the confidence interval is such that there is
either a 95% or 99% chance that the actual measure lies in the interval.

The value α (the Greek letter alpha) denotes the probability of error (that is, the
measure lying outside the interval). So the probability of being correct is 1 − α. For
example, for a 95% interval α is 0.05 (5%) and for a 99% interval α = 0.01 (1%).
Confidence intervals are (usually) constructed such that the probability of error is
shared equally between the two sides: the probability that the actual measure is above
the interval is α2 and the probability that it is below the interval is α2 .

In this book we discuss the problem of estimating the population mean. The sample
mean is denoted by x̄ and is defined as the average of the n sample values:

n
1X
x̄ = xi
n i=1

The sample mean is a point estimate of the population mean µ.

To determine a confidence interval for the population mean we need, in addition to

the sample mean, the sample standard deviation. The sample standard deviation
s is defined by:
sP
n
− x̄)2
i=1 (xi
s=
n−1

This is similar to the formula for the population standard deviation σ except that there
is the value n − 1 in the denominator rather than n. (There is a deep statistical reason
for this difference which we do not explore here.) There is an alternative formula which
is easier for calculations:
sP
n 2
i=1 xi − n × x̄2
s=
n−1

Example: For the sample data 8, 12, 10, 16, 7, 13,

the sample mean is 11;
and the sample standard deviation is:
62

r r
782 − 6 × 112 56
= = 3.35
5 5

In this section we discuss two common situations where one can provide confidence
intervals for the mean.

9.4.1 Large sample method for estimating the mean

The large sample method can be reliably applied if there are at least 30 data values.
Given some choice of α, a 1 − α confidence interval for µ is given by:

s s
x̄ − z × √ < µ < x̄ + z × √
n n

In this expression, the value z is obtained from the table in Appendix C by looking up
the entry in the two-tailed column corresponding to the chosen value of the error α.
(The two-tailed values are always used for confidence intervals.)

For example, suppose one wanted to be 95% sure that the value of µ lay within the
interval. Then choose α to be 5% (i.e. 0.05), and the value 1.96 is obtained for z in the
column corresponding to 95% and two tailed. (Recall that one divides the chance of
error equally on both sides or tails, so that there is a 2.5% chance of the actual value
of µ being smaller than the calculated lower bound, and a 2.5% chance of the actual
value of µ being larger than the calculated upper bound.)

Example: A biologist measures the salinity of the leaves of 60 white man-

grove trees in an estuary. He wants to estimate the average salinity with a
99% confidence interval.
Procedure: He calculates the sample mean and sample standard deviation.
Say he obtains x̄ = 5.39 and s = 2.04. He looks up the critical value for 99%,
two-tailed in Appendix C. The value found is 2.58. So a 99% confidence
interval for the salinity is
2.04 2.04
[5.39 − 2.58 × √ , 5.39 + 2.58 × √ ] = [4.71, 6.07]
60 60

9.4.2 The small sample method

One can estimate the population mean from a small sample if one knows that the
underlying population has a normal distribution, or one has good reason to believe
63

this. In this case the sample mean is related to Student’s t distribution (so-called
because the discoverer Gosset published the work under the pseudonym ‘Student’).
The process of estimation is the same as the large sample method, except that one
looks up values in the t-table given in Appendix D rather than the z-table.

The t distribution has a bell-shaped curve. It has one parameter df which is known
as its degrees of freedom. For df small the curve is a bit more peaked than the
normal distribution, but for df large the two curves are indistinguishable.

A 1 − α confidence interval (the value α denotes the probability of error) is given

by:

s s
x̄ − t × √ < µ < x̄ + t × √
n n

where the t-value is that for df =n − 1 degrees of freedom. The value t is found in
the t-table on the row for n − 1 degrees of freedom in the appropriate column for the
α-value required (again, the two-tailed values are used).

Example revisited : In the previous example we used the large sample method
to determine a 99% confidence interval for mangrove salinity. Say our biolo-
gist could only obtain 20 leaves, and again obtained a sample mean of 5.39,
but a sample standard deviation of 1.84. Since he believes that the salinity
has a normal distribution, he uses the small sample method. Degrees of
freedom df = 20 − 1 = 19. α = 0.01 (1% chance of error). Value in table is
2.861. So the confidence interval is
1.84 1.84
[5.39 − 2.861 × √ , 5.39 + 2.861 × √ ] = [4.21, 6.57]
20 20

The above example shows that a confidence interval can be narrowed by taking n
larger (think about why). This means: the larger the sample, the better the estimate.

Discussion questions and exercises

1. If you roll two honest dice, what is the probability of getting double-sixes? Any
double? An odd number for the total?

2. From the population of the digits 0 through 9 take a random sample of 3 digits.
Calculate µ, σ, and s.

3. Say the size of king protea flowers is normally distributed with mean 127 mm and
standard deviation 13 mm. What proportion is more than 140 mm?
64

4. Find a 95% confidence interval for the population mean from the sample data

7.2, 9.3, 10.2, 11.4, 14.8, 16, 10.3, 11.4, 12.3, 9.9, 8

stating any assumptions you make.

5. Find a 99% confidence interval for the population mean if the sample size is 50,
the sample mean is 43.41, and the sample standard deviation is 2.61.
 10 Testing Hypotheses

A fundamental principle in the scientific method is the formulation of hypotheses.

Once one has formulated a hypothesis, and accumulated the data, one analyses the data
and then accepts or rejects the hypothesis. In this chapter we describe this process,
from the formulation of the hypothesis to the test of the data.

The purpose of a hypothesis is to predict a relationship between variables that can

be tested. Hypotheses direct one’s research by indicating what procedures should be
followed. This means that hypotheses may be reformulated before empirical studies
start, but not afterwards.

10.1 HYPOTHESES, ERRORS AND TAILS

10.1.1 Hypotheses

Any statistical test revolves around the choice between two hypotheses. These are
labelled H0 and H1 . H0 is often called the null hypothesis.

We asked in a previous chapter: how would one decide if a coin was fair? In this
case the null hypothesis H0 would be that the coin is fair. The alternative H1 is that
it is biased.

10.1.2 Errors

Consider the following example: A biotechnology firm develops a new drug against
tuberculosis. To test the drug, they administer the drug to some patients and a placebo
to others. They then compare the results to determine whether the drug is effective or
not. Suppose H0 is that the drug has no effect, H1 that the drug is effective.

There are two possible ways the answer could be wrong. One possibility is that one
concludes that the drug is effective when in fact it has no effect. The other possibility
is that one concludes that the drug has no effect when in fact it does. These are known
as Type I and Type II errors respectively.

Ideally, one would like neither error to occur, but this is impossible. The smaller the
risk of one type of error, the greater the risk of the other type. If one is not prepared to
accept any risk of a Type I error, then one must accept a high risk of a Type II error,
and vice versa. (Think about why this should be so.)
66

In a hypothesis test, one takes as H0 that situation which one doesn’t mind as
much being accepted as true when it is in fact false, and takes an alternative as H1 .
The benefit of the doubt goes to the hypothesis H0 . In other words, one focuses on the
probability of a Type I error. This probability is denoted by α. The probability of a
Type II error is denoted by β (beta).

Table 3: Types of errors

H0 true H1 true
H0 accepted okay Type II error
H1 accepted Type I error okay

10.1.3 Hypotheses and tails

An agricultural researcher wishes to test a new fertiliser, and see if it increases maize
yield. There are four possible hypotheses:

H0 : The fertiliser does not affect maize yield

H1a : The fertiliser does affect maize yield
H1b : The fertiliser increases maize yield
H1c : The fertilizer decreases maize yield

Hypothesis H1a allows for either an increase in yield or a decrease in yield—it is a

combination of hypotheses H1b and H1c . Since there are two ways in which hypothesis
H1a can be true, it is termed a two-tailed test. Hypothesis H1b and H1c , on the
other hand, only test for a difference in one direction (at one extreme only) and are
termed one-tailed tests. H1b is an example of an upper-tailed test whereas H1c
is an example of a lower-tailed test. Whenever hypothesis H1 uses a relative term
(e.g. heavier, less than) it is a one-tailed test; whenever H1 states a difference without
specifying the direction of the difference (e.g. unequal, affects) then the test is two-
tailed.

When the test is one-tailed, the null hypothesis must be adjusted accordingly. In
our example above, is H1b is used then H0 is that the fertiliser does not increase maize
yield. Similarly if H1c is used then H0 is be that the fertiliser does not decrease maize
yield. In all cases H0 is the opposite of H1 —that is, any situation not covered by H1 is
covered by H0 . Figures 10a, 10b and 10c show this diagrammatically.

Figure 10a. A two-tailed test.

67

H0

H1a H1a

Figure 10b. An upper-tail test.

H0

H1b

Figure 10c. A lower-tail test.

H0

H1c

10.1.4 Level of significance

A hypothesis test is based on the probability of sample data being as extreme as the
data encountered, assuming that H0 is true. This is α, the probability of Type I error.
The researcher must choose a level of significance—an upper bound for α. Common
levels are 5% and 1%; for a test at the 5% level of significance there is at most a 5%
chance of Type I error. In reporting the results of a hypothesis test one must state the
level of significance.

10.2 THE STEPS IN A HYPOTHESIS TEST

In every statistical test, the procedure is:

1. Identify which test is appropriate.

2. Choose a level of significance α.
68

3. Formulate H0 and H1 .
4. Determine whether the test is one-tailed or two-tailed. If it is one-tailed,
determine whether it is an upper- or a lower-tail test.
5. Calculate the test statistic.
6. Based on the chosen level for α, compare the test statistic with a value
from a table (the critical value).
7. Conclude: Either accept or reject H1 .

In particular, suppose we have a test statistic u and a table value v. Then H1 is

accepted:

(a) in a 2-tailed test, if u > v or u < −v.

(b) in a 1-tailed upper-tailed test, if u > v.
(c) in a 1-tailed lower-tailed test, if u < −v.

If H1 is accepted, then H0 must be rejected. However, since tests are structured so

that the benefit of the doubt goes to H0 , the converse is not true—the fact that one
cannot be 95% (say) sure of H1 does not mean that one is 95% sure of H0 (think about
it). So if H1 is rejected one simply states that H0 is not rejected.

10.3 TESTS OF THE MEAN OF ONE SAMPLE

One takes a sample and has a value in mind for the population mean µ. Then the
question is, does the sample x̄ contradict this value significantly? This is similar to the
estimation of the mean from the sample mean described in the previous chapter.

We describe two versions:

 the large sample (z) test, and

 the small sample (t) test.

The large sample test can be used if the sample size is greater than 30, the small
sample test can be used if the sample comes from a normal distribution (regardless of
sample size).

In a test, the H1 hypothesis is that the population mean is either different from
(two-tailed), greater than (one-tailed upper tail) or less than (one-tailed lower tail) a
particular value. H0 is the complement of this. That is,
69

H0 : µ = c, H1 : µ 6= c;
H0 : µ ≤ c, H1 : µ > c; or
H0 : µ ≥ c, H1 : µ < c.

The sample statistics are denoted by n, x̄ and s. The large sample test uses the
following test statistic:

x̄ − µ √
z= n
s

The test statistic calculated is compared to a value from the z-table. The small sample
test uses the following test statistic:

x̄ − µ √
t= n
s

The test statistic calculated is compared to a value from the t-table with df =n − 1.

Example 1 : A climatologist measures the freezing points of water taken from

42 different lakes. She wants to test whether the average freezing point is
0 degrees Celsius, and be 95% sure of her conclusion. The sample statistics
are x̄ = −0.49 and s = 1.23.
Procedure:
1. The test is the large sample test.
2. The significance level is 0.05.
3. H0 : the population mean is 0.
H1 : the population mean is not 0.
4. This is a two-tailed test.
5. The test statistic is
−0.49 − 0 √
z= 42 = −2.58
1.23
6. The table value for 5% level of significance is 1.96.
7. Accept H1 , as −2.58 < −1.96. There is sufficient evidence to conclude
that the population mean is not 0.

Example 2 : It is known that the I.Q. of humans has a normal distribution

with mean 100. It is believed that rugby players’ I.Q.s are higher. Twenty
members of rugby teams are selected using cluster sampling (each team
70

forming a cluster) and tested for I.Q. The sample mean is 102.7 and the
sample standard deviation 14.8. Are rugby players’ I.Q.s higher?
Procedure: The test is the small sample t-test. H0 is that rugby players
have normal or below normal I.Q. H1 is that rugby players have higher
than normal I.Q. This is a one-tailed upper tail test. The calculated test
statistic is t = 0.816. The degrees of freedom df = 19. Value in the table
is 1.73 for 5% level of significance. As 0.816 < 1.73, at 5% significance level
one cannot conclude that rugby players have higher I.Q.s.

10.3.1 Estimation and hypothesis testing

These hypothesis tests of the population mean are linked to the estimation of population
mean discussed in Section 9.4. For example, in a two-sided test, H0 is rejected if and
only if is outside the confidence interval.

10.4 A TEST FOR EQUIVALENCE OF RELATED SAMPLES

Related sample pairs can be formed in two ways. One is where two sets of measure-
ments are compared (for example, the results of ‘before and after’ tests on the weights of
mice). The other is where the data is gathered from measurements of matched samples
(see Section 5.4).

To perform the test on pairs (ai , bi ), one calculates the difference data di = ai − bi .
This difference data is itself a sample. For a two-tailed test, the null hypothesis H0
is that the two population means are equal. This is the same as testing whether the
population mean µD of the difference data is zero, i.e. H0 is that µD = 0. Hypothesis
H1 is that µD is non-zero. We use the difference data for a t-test. The sample statistics
for the difference data are denoted n, x̄d and sd . Note that for matched data n denotes
the number of data pairs.

The t-test for equivalence of related samples uses the following test statistic:

x̄d √
t= n
sd

Degrees of freedom: df =n − 1.

Example: An educator proposes that using multimedia would improve the

teaching of calculus to first-year students. So she picks a sample group of 20
71

students at random. Students are matched in terms of matric mathematics

symbols into 10 pairs, with one member of each pair being placed in the
experimental group and the other in the control group. The measurements
she uses as data are the final calculus exam marks of the students.
Procedure: The test is the t-test for equivalence of related samples. It is
decided to do the test at both the 1% and the 5% levels of significance. Hy-
pothesis H0 is that the multimedia has no effect on or worsens the calculus
results. Hypothesis H1 is that the multimedia improves the calculus results.
The test is one-tailed upper tail. The measured data is tabulated.

Table 4: Results for the calculus example

Pair Marks Group Marks Group Difference d2i
Experiment Control
a 18 16 2 4
b 16 15 1 1
c 14 15 −1 1
d 16 16 0 0
e 17 10 7 49
f 14 10 4 16
g 12 15 −3 9
h 13 8 5 25
i 11 11 0 0
j 9 5 4 16
Totals 19 121

√
10
Calculations: x̄D = 1.9 and sD = 3.07 so that test statistic is t = 1.9 3.07 =
1.96. The number of degrees of freedom is df =9, so the critical values are
1.83 and 2.82. So one can conclude an improvement at the 5% level of
significance but not at the 1% level of significance. (The educator can be
95% certain that H1 is true, but not 99% certain.)

Notes:

 The necessary assumptions for hypothesis tests must be checked. If not known
beforehand to be true, this may involve further statistics tests!

 It is sometimes advisable to estimate β, the chance of a Type II error.

72

Discussion questions and exercises

1. Explain the meaning of: hypothesis, null hypothesis, tail, level of significance,
hypothesis test, test statistic, critical value.

2. In drug testing the null hypothesis is normally taken that the drug has no effect
and the alternative is that the drug is effective.
(a) What are Type I and Type II errors?
(b) Which is preferable?
(c) What about a disease such as HIV/Aids which is rampant and for which no
cure has been found yet? Should one deny people the drug just because one is
not totally convinced that it is effective? Discuss.

3. The sample data of Question 4 from Chapter 9 is believed to come from a popu-
lation with mean 12.8. Test the hypothesis that:
(a) this is not true; and
(b) the actual mean is smaller than 12.8.

4. If we always use a 0.05 level of significance, does this mean that on average 1 out
of 20 conclusions will be wrong?

5. A quantity surveyor calculates the amount of material needed for a project as

60 units. For a sample of 40 projects she obtains a mean of 65.32 units and a
standard deviation of 3.17 units. Test to see whether the sample data conforms
with her calculations.

6. A student makes a series of measurements of the same compound on two scales

in milligrams and obtains the following results:
Scale 1: 473 321 555 678 299 407
Scale 2: 472 320 556 682 307 415
(a) Find a 95% confidence interval for the average difference between the two
scales.
(b) Test the belief that scale 2 gives higher readings than scale 1.
(c) What does this say, if anything, about the validity and reliability of the two
scales?
 11 More Hypothesis Testing

In this chapter we describe some frequently-used tests. First we describe a test

for linear correlation, and then introduce the concept of partial correlation. Then we
describe two more tests: the chi-square test for patterns in frequency data and the test
for the equality of means for independent samples.

11.1 A TEST FOR LINEAR CORRELATION

In Section 8.3 we described how to find the best line fit for paired data. Determining
whether there is in fact a linear relationship requires another hypothesis test. Pearson’s
product-moment coefficient of linear correlation is calculated by the formula:

x i yi − (
P P P
n x i ) ( yi )
r=q P q P
2
n x2i − ( xi ) 2
yi − ( yi )2
P P
n

(Wow!) This parameter lies between −1 and 1. A value of 1 indicates a perfect linear
dependence with positive slope. (An increase in the value of variable X is associated
with a proportionate increase in the value of variable Y .) A value of −1 indicates a
perfect linear dependence with negative slope. (An increase in the value of variable X
is associated with a proportionate decrease in the value of variable Y .) A value of 0 or
thereabouts says very little.

The hypothesis test: H1 can be that either

(a) there is a linear correlation (two-tailed),

(b) there is a positive slope correlation (one-tailed upper tail), or
(c) there is a negative slope correlation (one-tailed lower tail).

The hypothesis H0 is, as usual, the complement of H1 . The test statistic is the Pearson
coefficient of linear correlation r. The test statistic is compared with the critical values
obtained from the table in Appendix E, with n − 2 degrees of freedom (where n is the
number of data pairs), and value desired of α.

Example: In Section 8.3 we considered a chemical engineer who obtained

the following measurements for arsenic levels at different distances from a
factory. The engineer wishes to test whether the arsenic distances decrease
linearly the further one is from the factory.
74

x (distance in km) 2 4 6 8 10 12 14
y (arsenic in mg/kl) 24.5 18.7 16.3 12.0 9.3 6.2 2.5
Procedure: This is a one-tailed lower tail test with H0 : arsenic levels do
not decrease linearly with distance from factory, and H1 : arsenic levels do
decrease linearly with distance from the factory.
Calculations give n = 7, Σ xi = 56, Σ yi = 89.5, Σ xi yi = 520.0, Σ x2i = 560
and Σ yi2 = 1490.81. Hence
7 × 520 − 56 × 89.5
r=√ √ = −0.995
7 × 560 − 562 × 7 × 1490.81 − 89.52
The df =7 − 2 = 5. Table values are 0.669 (5% level) and 0.833 (1% level).
So the engineer can accept H1 at the 1% level (he can be 99% certain that
H1 is true).

11.2 PARTIAL CORRELATION

It is important to note that the existence of high linear correlation between two variables
does not necessarily imply a cause-and-effect relationship between the two variables. It
is likely that ownership of motor vehicles and ownership of television sets have a high
correlation. This does not mean that owning a car causes someone to own a television
set. Some other factor could be at work, e.g. wealth.

Partial correlation tries to remove the effect of the third variable and thereby find
the ‘true’ relationship between the first and second variables. The formula for partial
correlation is:

r12 − r13 r23

r12.3 = q q
2
1 − r13 2
1 − r23

where:

r12.3 is the partial correlation between variables 1 and 2 with the effect of
variable 3 removed;
r12 is the correlation between variables 1 and 2;
r13 is the correlation between variables 1 and 3;
r23 is the correlation between variables 2 and 3;

and correlation is measured by the Pearson coefficient of linear correlation (see previous
section).
75

To continue our example, say a researcher found that owning TVs (variable
1) had a correlation of 0.7 with owning cars (variable 2). It was also found
that the correlation of TV ownership with wealth (variable 3, which the
researcher defined as assets over R50 000) was 0.8, and the correlation of
car ownership with wealth was 0.9. The formula gives:
0.7 − 0.8 × 0.9
r12.3 = √ √ = −0.08
1 − 0.82 1 − 0.92
In other words, the researcher finds almost no correlation at all (−0.08 is
a tiny negative correlation) between TV ownership and car ownership once
the effect of wealth has been removed.

11.3 THE CHI-SQUARE TEST

In this section we describe the chi-square test for dependence of two qualitative vari-
ables. For example, the colour of a particular chemical solution can be red, orange,
purple or blue and its acidity can be categorised as high, medium or low. We wish to
know if colour and acidity are related.

Say the first qualitative variable divides the entire population up into R classes and
the second divides the entire population up into C classes. In our example above R = 4
and C = 3. To apply the chi-square test, one takes a sample of the population, and
categorises each item according to both variables. The sample items are divided into
R classes according to the first variable and into C classes according to the second.

The data is represented in a contingency table. Each row corresponds to a

different class of the first variable and each column to a different class of the second
variable. The entry oij in row i and column j gives the number of sample items which
are classified as class i for the first variable and class j for the second. Table 5 shows the
observed frequencies for our example. Totals for each row and column are calculated,
as well as the overall total.

Table 5: Observed frequencies

High Medium Low row total

Red 15 8 5 28
Orange 12 9 9 30
Purple 25 24 31 80
Blue 6 14 20 40
Column Tot 58 55 65 178
76

Alongside that table, one constructs the expected contingency table which
would result if the two variables were independent: To calculate the entry eij in row i
and column j of the expected contingency table, one multiplies the total in row i by the
total in column j and divides the result by the overall total. The expected contingency
table for our example is given in Table 6.

Table 6: Expected frequencies.

High Medium Low

Red 9.12 8.65 10.22
Orange 9.78 9.27 10.96
Purple 26.07 24.72 29.21
Blue 13.03 12.36 14.61

In the test, the hypothesis H0 is that the two variables are independent and H1 is
that they are dependent. The test statistic is:

X (oij − eij )2
χ2 =
eij

where oij denotes the observed frequency in row i, column j, and eij the expected
frequency in row i, column j, and the summation is over all entries in the contingency
table.

For the above data, the contribution to the test statistic χ2 is as follows:

Table 7: The contribution to the test statistic

3.785 0.049 2.670

0.506 0.008 0.349
0.044 0.021 0.109
3.796 0.218 1.991

(5−10.22)2
For example: the entry 2.670 is calculated by 10.22 .

One needs to compare this with the chi-square tables. The degrees of freedom for a
χ2 test are given by (r − 1)(c − 1). In our example above, which is a 4-by-3 table, this
is df = (4 − 1)(3 − 1) = 6. The concept of a lower tail does not exist for the χ2 test
statistic as the statistic is always positive. As all the error is thus in the upper tail, a
two-tail test becomes an upper-tailed test.
77

The null hypothesis H0 is that colour and acidity are independent, and H1 is that
colour and acidity are dependent. The test statistic χ2 = 13.546 with df = 2 × 3 = 6.
So critical values are 12.59 (5%) and 16.81 (1%). Accept at the 5% level of significance
but not at the 1% level of significance.

There are limitations on the use of this test. An important rule of thumb is that
every entry in the expected contingency table must be at least 5.

The chi-square test is an example of a non-parametric test. Such tests can be

used when the data is qualitative or discrete as they do not assume anything about the
underlying distribution of the population. Such tests are especially useful if the data
is ordinal rather than cardinal (i.e. the data items are ranked from 1 to n rather than
having values showing quantity).

11.3.1 Comparing one qualitative variable with a theory

A version of the chi-square test can be used where one has frequency data and wishes
to check whether this conforms with a theorized distribution. For example, one might
wish to test whether data comes from a random distribution or not.

There is one qualitative variable. Say this divides the population into M classes.
The data gives the observed frequencies oi and the theorized distribution gives the
expected frequencies ei . Then the test statistic is given by a similar formula to the one
above:

X (oi − ei )2
χ2 =
ei

There are M − 1 degrees of freedom.

Example: The crooked die story completed. Recall that we rolled a (simu-
lated) die 600 times and got 86, 112, 122, 95, 98, 87. Is this die crooked?
Procedure: Hypothesis H0 is that the die is fair, hypothesis H1 that it is
crooked. Expected values are 100 each. The test has 5 degrees of freedom.
2 (−5)2
122 222
The test statistic χ2 = 10.22 (calculation is (−14)
100 + 100 + 100 + 100 +
(−2)2 2
100+ (−3)
100 ). Testing at the 5% level, we find that the table value is 11.07.
So we cannot reject the hypothesis that the die is fair.
78

11.3.2 Tests with one degree of freedom

In general, the chi-square test is a test of dependence. However, in the special case where
df = 1, one can test for a one-tailed hypothesis. One case is a 2-by-2 contingency table.
For example, one may be testing a new drug with two categories: drug and placebo, and
with two results recover or not recover. One can use a chi-square test to test whether
the drug is effective or not. Another case is illustrated in the next example.

Example: An economist wrote a computer program to forecast whether

tomorrow’s JSE industrial average would be higher than today’s. She ran
it for three months. It is known that an increase occurs on average 3 days
out of 5. As the program’s forecast was right only 66 days out of 90, a cynic
suggests that one might as well toss a coin which comes up heads 60% of
the time. Is he correct?
Procedure: This test is the chi-test for one qualitative variable. H0 is that
µ = 0.6 and H1 is that µ > 0.6. This is a one-tailed test. The observed
frequencies are 66 (correct) and 24 (incorrect). The expected frequencies
are (3/5) × 90 = 54 (correct) and (2/5) × 90 = 36 (incorrect). The test
2 2
statistic χ2 = (66−54)
54 + (24−36)
36 = 6.67. The value in the table is 2.71 (5%
level) and 5.51 (1% level). Therefore, one can conclude with 99% confidence
that the program is worthwhile.

11.4 TESTS FOR EQUALITY OF MEANS (INDEPENDENT SAM-

PLES)

Another situation one may encounter is where one has two independent samples from
different populations but wonders whether the means are the same.

We describe two tests. These are the large sample (z) test and the small sample
(t) test. The t-test can only be applied if both samples come from normal populations,
and the standard deviations of the normal populations are similar. (The equality of
standard deviations can be checked by using what is known as the F -test.)

In both tests there are two samples which we denote by A and B. The null hypoth-
esis H0 in a two-tailed test is that the two population means (which we denote by µA
and µB ) are equal. The sample statistics for A are denoted by nA , x̄A and sA , and for
B are denoted by nB , x̄B and sB .

The z-test for equality of means (large samples) uses the test statistic:
79

x̄A − x̄B
z=r
s2A s2B
nA + nB

The t-test for equality of means (normal populations) uses the test statistic:
s
x̄A − x̄B nA nB (nA + nB − 2)
t= q
(nA − 1)s2A + (nB − 1)s2B nA + nB

with df = nA + nB − 2 degrees of freedom.

Large sample example: A large software house uses the languages C and
Natural for programming. They wish to compare the development time
as they believe that Natural is faster to program in. They know that re-
cently they have completed 42 projects in Natural and 30 in C. The sample
statistics in person-months are x̄N = 3.3, sN = 1.1, x̄C = 4.0 and sC = 0.8.
Procedure: The test is the two-sample z-test. H0 is that Natural is the same
or worse than C, and H1 that Natural is better. This is a one-tailed lower
tail test. Test statistic: z = −3.13. Critical values −1.64 (5%) and −2.33
(1%). As the test statistic is less than both critical values, one can conclude
with 99% certainty that Natural is quicker to develop.

Small sample example: A zoologist has developed a new nutrient which

when taken as a supplement by rats is believed to increase weight by at least
4 grams. He tested the nutrient on rats using an experimental group of 10
rats and a control group of 10 rats. He has good reason to believe weight
gain to be normally distributed. In the control group the gain statistics in
grams were x̄C = 37.3 and sC = 4.2. In the experimental group they were
x̄E = 44.6 and sE = 4.7.
Procedure: The test is the two-sample t-test. H0 is that the nutrient in-
creases mass by 4 g or less, and H1 that the nutrient increases mass by more
than 4 g. This is a one-tailed test (upper tail). Test statistic: t = 1.66. De-
grees of freedom: df =18. Critical values from the table are 1.73 (5% level)
and 2.55 (1% level). Therefore one can be 95% certain that the nutrient
increases weight by more than 4 grams, but not 99% certain.

There is also an extension of these tests to a null hypothesis H0 in a two-tailed

test of the form µA − µB = C. The test statistics are the same as those in the above
formulas, except that the numerator x̄A − x̄B is replaced by x̄A − x̄B − C.
80

Discussion questions and exercises

1. What is linear correlation? What is negative linear correlation? Is it possible to

have a negative linear correlation without a linear correlation? A linear corre-
lation without a negative one? A negative correlation without a negative linear
correlation?

2. Test the data of Question 4 of Chapter 8 for linear correlation.

3. Pick 10 number pairs at random. Each number in each pair should be a whole
number in the range 1 to 10. Is there a linear correlation among the number pairs
chosen?

4. The drug company finally does the trial on a new treatment for tuberculosis.
There are 32 patients who complete the trial. Of the 18 patients on placebo, 8
recover and 10 do not. Of the 15 patients on the drug, 12 recover and 3 do not.
(a) Is there any difference between the effectiveness of the drug and that of the
placebo?
(b) What does your result in (a) tell you about the drug?

5. According to Mendel’s laws of genetics, the population of plants produced from

a certain crop of seedlings should be red flower green stem, orange flower yellow
stem, red flower yellow stem, orange flower green stem in the proportion 9:3:3:1.
A random sample of 100 seeds are sprouted and the observed frequencies are
52:19:24:5. Are these results compatible with the laws?

6. In a case-study, a researcher into poverty measures the correlation between income

and IQ and obtains r = 0.6. She also measures the correlation between income
and years of schooling, obtaining 0.8, and between years of schooling and I.Q.,
obtaining 0.55. Use partial correlation to compute the correlation of income and
I.Q. without the influence of schooling. What does your result mean?

7. An economist measures the prices of samples of 40 widgets in both Ankara and

Bisho in rands. In Ankara he obtains a mean of 14.31 and standard deviation 3.12
and in Bisho a mean of 14.64 and standard deviation 3.03. Is there a significant
difference in the price?

8. Two samples from different normal populations yield the following sample statis-
tics: nA = 6, x̄A = 100, sA = 10, nB = 8, x̄B = 80, and sB = 9. Test for
inequality of population means.