SAMPLING METHODS

FOR GEOGRAPHICAL RESEARCH

C.Dixon & B.Leach

ISBN 0 902246 96 8

© C. Dixon and B. Leach 1977

 CONCEPTS AND TECHNIQUES IN MODERN GEOGRAPHY No. 17
CATMOG
SAMPLING METHODS FOR GEOGRAPHICAL RESEARCH
(Concepts and Techniques in Modern Geography)
by

CATMOG has been created to fill a teaching need in the field of quantitative C. Dixon and B. Leach
methods in undergraduate geography courses. These texts are admirable guides
for the teachers, yet cheap enough for student purchase as the basis of class- ( City of London Polytechnic and Thames Polytechnic)
work. Each book is written by an author currently working with the technique
or concept he describes.

1. An introduction to Markov chain analysis - L. Collins CONTENTS

Page
2. Distance decay in spatial interactions - P.J. Taylor
I BASIC CONSIDERATIONS
3. Understanding canonical correlation analysis - D. Clark
(i) Sampling in research design 3
4. Some theoretical and applied aspects of spatial interaction
shopping models - S. Openshaw (ii) Sampling and probability 3

5. An introduction to trend surface analysis - D. Unwin (iii) What is a good sample? 4

6. Classification in geography - R.J. Johnston (iv) Sample size 7

7. An introduction to factor analytical techniques - J.B. Goddard & A. Kirby (v) Target population and sampling frame 11

8. Principal components analysis - S. Daultrey II PROBABILITY SAMPLES

9. Causal inferences from dichotomous variables - N. Davidson (i) Simple random samples 13
10. Introduction to the use of logit models in geography - N. Wrigley (ii) Systematic samples 14
11. Linear programming: elementary geographical applications of the (iii) 'Systematic random' samples 15
transportation problem - A. Hay
(iv) Stratification 16
12. An introduction to quadrat analysis - R.W. Thomas
(v) The procedure for stratifying 17
13. An introduction to time-geography - N.J. Thrift
(vi) The choice of sampling fraction 18
14. An introduction to graph theoretical methods in geography - K.J. Tinkler
(vii) Weighting 19
15. Linear regression in geography - R. Ferguson
(viii) Stratification after selection 20
16. Probability surface mapping. An introduction with examples and
Fortran programs - N. Wrigley III CLUSTERING
17. Sampling methods for geographical research - C. Dixon & B. Leach (i) Reducing costs over a large area 20
18. Questionnaires and interviews in geographical research - (ii) Creating clusters 23
C. Dixon & B. Leach
(iii) Selecting clusters 23
Other titles in preparation (iv) Sampling with probability proportional to size 24

IV SAMPLING FRAMES
This series, Concepts and Techniques in Modern Geography
(i) Convenient sampling frames 25
is produced by the Study Group in Quantitative Methods, of
the Institute of British Geographers. (ii) Inconvenient sampling frames 26
For details of membership of the Study Group, write to
(iii) Sampling from lists: the example of the British 27
the Institute of British Geographers, 1 Kensington Gore,
London, S.W.7. Electoral Register
The series is published by Geo Abstracts, University of
East Anglia, Norwich, NR4 7TJ, to whom all other enquiries
should be addressed. 1
 (iv) Selecting within small clusters: the example of households 28
(v) Constructing frames from maps 30
(vi) Constructing frames by mapping 30
(vii) Longitudinal studies 31
(viii) Sampling continuous flows 32
I BASIC CONSIDERATIONS
V AREAL SAMPLING METHODS
(i) Sampling in research design
(i) Frames and methods 33
(ii) Problems of spatial frames We might want to design a research project to investigate the livestock
38
of farms in an area to be visited on a field course, or to find out where the
(iii) Sampling in space and time 39 households on a particular estate do their shopping, or to measure an island's
soil properties. In each of these cases, we want to make generalisations
VI NON-PROBABILITY SAMPLES about a large number of units (elements) in a population consisting of all
(i) Purposive samples the elements (farms, households, possible measuring points) whose character-
40
istics we want to describe.
(ii) Quota samples 40
(iii) Random routes In some cases, we can undertake a complete study of these elements, as
41
we might with shops in a small centre, and on other occasions we might
(iv) Combining quota sampling with a random route 42 deliberately pick out for study a few items, for example, several towns of
(v) Snowball samples the British Isles. But generally our resources do not stretch to collecting
42
data about all the elements in the population, and we have to make inferences
(vi) Justifying non-probability samples 43 about the whole population by studying only some of them.

VII RESEARCH DESIGN AND THE CHOICE OF SAMPLING METHOD 43 Even if our resources were adequate, a more economical way of carrying
out our study would be to draw a sample from the population. Clearly a sample
BIBLIOGRAPHY 46 will only approximately represent the characteristics of the parent population,
since it only contains part of it, but sampling theory enables us to estimate
the likelihood of our sample being a good representation of the population,
provided we have followed certain rules in the choice of elements for the
sample.

(ii) Sampling and probability

This book is mainly concerned with probability sampling, which requires

that each element in the population has a nown c ance of se ection for the
sample. Its simplest form is the 'simple random sample' in which all elements
have an equal chance of inclusion. This means that any particular sample, any
combination of elements, drawn from a population would have the same chance
of occurring as any other sample of the same size.

If a researcher only looks at the fields he can see from the road, or
interviews the people who happen to be in the street, his sample is clearly
biased. It does not surprise us that students sent to interview shoppers in
a market each return with a sample consisting of a different sort of people,
elderly ladies, young men, or whoever the student found it easiest to
approach. However, it is harder to realise that we cannot choose elements 'at
random', without any selection bias. Researchers sampling from a list of names
may tend to select more from one part than another; a biogeographer throwing
a stick to find a random point may subconsciously direct it towards some
interesting vegetation; if we place a small quadrat frame over a part of the
crop in a field, can we really do so without considering if the location
chosen is above, below or close to the average for the field? Only a method

2 3
in which all the subjective choice is removed ensures that each sample, each
We need to know this standard error if we are to have an indication of
possible combination of elements, has exactly the same chance of selection.
the likelihood of having drawn a 'good' sample, one which closely corresponds
Only if we have strict rules which prevent the researcher, or his agents such
to the parent population.
as field-workers, using 'judgement' at any stage can a sample be subjected
to statistical testing on the basis that it is truly a probability sample. Although we cannot actually calculate the standard error without knowing
One of the theoretical texts such as those in section A of the bibliography the true population mean - the very thing we want our sample to estimate -
should be consulted by those wanting a detailed discussion of sampling theory. the distribution of values in our sample can give us some indication of the
In the sections which follow, we summarise the main conclusions which have probable standard error. The more variable the parent population, the more
the greatest relevance for the practical design of samples. variable the samples from it will tend to be. If we draw a very small sample,
it might not reflect all the variability in the population, because sometimes,
(iii) What is a 'good sample'? by chance, all the elements in the sample will be rather similar. However,
the larger our sample, the more likely it is to be as variable as the popu-
The major paradox of sampling is that, since we are trying to find out lation itself.
about a population whose characteristics we do not know, we can rarely measure
the success with which our sample represents that population. Sampling theory We therefore calculate the standard deviation of the values in our sample,
enables us to estimate the range around each sample statistic within which we using formula 1. Then we estimate the standard error from it:
expect the true population value or parameter to lie, with a calculable margin
of error. (2a)
If a number of samples were drawn from a population they would provide
a number of estimates of such population parameters as the mean. While each where SE is the standard error, s is the standard deviation of the sample
sample might give us a different estimate of the mean, the spread of all (formula 1), n is the sample size, and N is the size of the population.
possible samples of that size that we could conceivably select forms a dis-
tribution scattered around the true mean, the population parameter. We can The expression (1-n/N) is designed to take into account the sampling
calculate statistics to describe the characteristics of any distribution, the fraction, the proportion of the population elements included in the sample,
most important for our purpose being the variance, a measure of the spread (n/N), because we should expect the sample to resemble the population more
around the mean, or its square root, called the standard deviation. closely if it is a large proportion of it. Unless n is a very large proportion
of N, it will make little difference, and for practical purposes this factor
The standard deviation is found by calculating the deviation of each is frequently ignored. Thus the standard error of the mean is usually estim-
ated:
item together, dividing by the size of the sample, and finally taking the
square root, so that it will be in the same units as the mean. Where the (2b)
sample is small, dividing by (n-1) rather than n, the sample size, will give
a significantly larger answer. This compensates for the fact that the smaller Although realistically we would never be able to draw all the possible
the sample, the less likely it is to have as large a variability as the popu- samples from the population in which we were interested, we are able to use
lation itself. their theoretical distribution as the basis for estimating the standard error
The usual formula for the standard deviation of a sample is: of any individual estimate, the value found in one sample. There are further
assumptions that we can make about the theoretical sample distribution.
(1) If we drew all the possible samples from the population, we would expect
that as many of the samples would have means above the true mean as below.
The mean of the sample means would be the population mean.

In addition, the distribution of sample means would always form a normal

curve, a distribution whose main mathematical property is that, regardless of
In sampling theory, the standard deviation of the distribution of all the any particular mean and standard deviation, there will always be a constant
possible sample means around the true, population mean is called the standard proportion of observations (in this case, sample means) within each standard
error of the estimate, because each sample mean will be a more or less accur- deviation unit (in the case of a sampling distribution, standard error unit)
ate estimator of the true value. The more variable the population itself, the of the mean. However variable the population, and thus however large the var-
more likely we are to have a sample which is very different from the popula- iation in the possible samples, we would always expect 95 per cent of the
tion, and the greater the range of possible estimates. The standard error is sample means to be closer to the true population mean than 1.96 times the
therefore larger when we are measuring something which is very variable in standard error; within 2.58 SE there will always be 99 per cent of sample
the elements of the population. means. Statistical tables give the percentage at other standard error dis-
tances, but these are the most generally used, because we are usually inter-
ested in ensuring that we will be able to predict that our sample value is

4 5
within a stated range of the true value, with a 95 per cent or 99 per cent the confidence limits for variables with a more erratic distribution, our
chance of being right. assertions about the value for the population as a whole must be very qual-
ified indeed. In the case of the farms mentioned above, we might find that
This chance is called the confidence level, and is usually expressed as the number of cows on each farm varied very little, and that we could say
a 95 or 99 per cent chance of being right, or conversely as a probability of that the average herd was 28 with a confidence limit of 2.2 and 99 per cent
5 per cent or 1 per cent (.05 or .01) of making a statement which is wrong confidence. In other words, we would be 99 per cent sure that the true value
about the relationship between the estimate and the parameter. was between 25.8 and 30.2 cows - a very close estimate indeed.

If 95 per cent of the sample means lie within 1.96 SE of the population
mean, 95 per cent of the time we expect the population mean to lie within the It would be rare for research reports to qualify all their assertions
same distance of our sample mean. Therefore, in order to be 95 per cent con- about the population by the addition of confidence limits and a probability
fident of our prediction of the population value, we can add or subtract 1.96 of accuracy to each estimate, because their fluency would be greatly impeded.
times the standard error to our estimate, and assert that the true mean will However, the calculation should be performed for several variables reported,
lie within this range. and the results presented in a footnote or appendix, to indicate the confid-
ence with which the reader should approach the statements made in the text.
The distance from the sample mean (in either direction) within which we It is also important to give the standard deviation of variables in the study.
expect to find the population value is the confidence limit, calculated by: This will serve as an indicator of the standard error; a reader who sees a
large standard deviation and knows that the sample was small will approach
(3) statements made about the population with caution, knowing that the confidence
where c is the confidence limit and z is the standard-error-unit measure for limits will be large.
the desired confidence level. The sign ± indicates that the figure given is
the range on either side of the mean, since we do not know whether our sample (iv) Sample size
mean is above or below the true value.
In general, a sample of 30 is the smallest that can be expected to con-
For the 95 per cent confidence level, the confidence limit is: form to the normal distribution on which sampling theory, as outlined above,
is based. However, the larger the sample, the more accurate it will be, that
is, the nearer values calculated from it will be to the true population values.
If the mean land holding in an area were found from a sample of 100 farms to
be 53 hectares, and the standard deviation to be 26, the confidence limits Intuitively, we can see that if we want to estimate what proportion of
could be calculated for the 95 per cent probability level: all students study geography, a sample of ten could not be expected to reflect
the whole student body as closely as a sample of a hundred would.
It is primarily sample size, and not the percentage of the population
included in the sample, which determines the accuracy of a sample. If we deal
out four hands of playing cards, each will tend to contain a mixture of cards,
of all suits, and of both high and low values. If we increase the number of
packs of cards to twenty, and the size of the hand to 52, the spread of suits
and values will be more even, although the sampling fraction will have de-
creased from 1 in 4 to 1 in 20.
We could therefore say that we were 95 per cent confident that the true mean
lies in the range 53 ± 5.1, that is, from 47.9 to 58.1 hectares. If we wanted Thus the larger our sample, the more confident we can be in our predic-
to be 99 per cent confident in our prediction of the population value, we tion, or the narrower the range in which the predicted value lies. That is
would calculate c using 2.58 instead of 1.96 in formula (3), and this would why in the formula (3) for the confidence limit, the sample size, n, was in-
give a confidence limit of 6.7; our population mean could then be said to lie cluded. However, we should expect diminishing returns. Increasing the number
within the range 46.3 to 59.7 hectares with 99 per cent confidence. of cards in each hand from 13 to 18 would have a greater effect than adding
5 cards to a hand of 52. For this reason sampling theory indicates that it is
When making assertions about population values on the basis of an esti- the square root of n, rather than n, that should be used in these calculations.
mate from a sample, we can choose a high confidence level, a low chance of
being wrong, but then we have to make a less definite statement, that is, a The second factor which, as has already been mentioned, will affect the
statement with wider confidence limits, a bigger range within which we expect accuracy of a sample is the variability of the population. The more different
to find the population value. the elements are from each other, the larger the sample needed to represent
them accurately.
The standard error, and thus the size of the confidence limits, depends
primarily on the variability of each characteristic studied. Samples which In a pack of playing cards, there are two colours, four suits, and 13
are quite small may give very good estimates of phenomena which are fairly different values. We should need a comparatively small sample from the 20
uniform in the population as a whole, but we might find that when we calculate packs to have a balance of the red and black cards, but a larger one would be

6
needed before the suits would be found in equal numbers. A still larger hand
would have to be dealt before the hand contained all 13 values in their Some small scale studies will draw samples from small populations,
correct proportions. where the sampling fraction, the percentage included in the sample, will be
so large as to increase the accuracy of the sample significantly. The answer
The researcher will not usually know much about the things he sets out obtained by formula (4) or (6) can be corrected to take account of the samp-
to measure, but he will need some indication of the variability before he can ling fraction:
determine the sample size he needs for accuracy.
Equation (3) related the confidence limit to the standard-error-unit (7)
measure for the desired confidence level and the standard error: where n' is the corrected sample size, n is the sample size calculated with
(3a) formula (4b) or (6), and N is the population size.

If there were 10,000 households in the area, we could estimate car

(3b) ownership to within 2 per cent with a sample of 1936 instead of the previously
estimated sample of 2401. However, if there were 100,000 households, the
It therefore follows that we can determine the sample size needed for a given corrected sample size from formula (7) would be 2345, not a very great re-
confidence level and confidence limit by turning this information round: duction from the original estimate.
(4a) Normally only professional studies with considerable resources can con-
or, more conveniently: duct pilot exploratory work to find out the likely variability in the popu-
lation, but it is possible to draw on the experience of other research. For
(4b) human populations, sources such as population censuses can be used to make
an informed guess of the variability in the population to be sampled.
In the playing card example, we were concerned with attributes like colour
and suit which a certain proportion of the cards have. In the case of a pro- A more serious problem is that we are rarely concerned with only one
portion rather than a mean, we cannot calculate a standard deviation. The feature of the population, and the sample size needed for each will depend
measure of variability which we use instead, for example in calculating the on its variability. It is therefore the case that only an approximation to the
standard error of a proportion, is v, measured by: desired sample size can be made, and for this purpose Table 1 may be adequate.

(5) Table 1 gives the sample size needed to estimate population values to
within a chosen percentage (the confidence limit) with a desired probability
where p is the percentage with the characteristic. of being right (the confidence level) if the variability of the population is
Therefore, instead of formula (4b), we estimate the necessary sample size 50 per cent. This was the case in the car ownership example, where half the
using v instead of s: households had the characteristic. For a proportion, this is the maximum
possible variability, because v is the square root of 50 times 50, 2500, (49
(6) times 51 is 2499), and the table therefore gives a conservative estimate of
sample size.
Using these equations, if we need a sample to estimate the proportion of
households with cars to within 2 per cent (the confidence limit) with 95 per For a continuous variable, 50 per cent would not be the maximum possible
cent confidence (the confidence level), in an area where we expect that only value. Not infrequently, the standard deviation is more than half the mean,
half will have cars, we substitute in equation (6): giving a coefficient of variability (100 s / R) over 50 per cent. The table
may therefore give an underestimate of the sample size needed for continuous
variables.

It can be seen that in order to obtain answers to within a narrow range

it is necessary to have a larger sample size if the population is variable.
In choosing a research project for a small scale study, it might be possible
to scale down the original ambitions to reduce the sample size needed.

Reducing the variability of the population could reduce the sample size
needed; to generalise about shopping patterns, a study of all households would
This means that we would need a sample of 2401 households to ensure the almost certainly need a larger sample than one concentrating on one particular
precision specified. type of household, such as families with small children.

8 9
 The size of the target population can also be reduced, so that instead Redefinition of the target population is not something which a client
of drawing the sample from a large area, a smaller one with fewer elements or sponsor of a commercial study would usually accept, but in academic re-
is chosen. The sampling fraction may then have some effect on the sample size search the choice may be between an inadequate sample of a highly diverse
needed, which can be adjusted with equation (7). This could not be done in population, or a less wide-ranging one about which it is possible to infer
cases where the population was of unlimited size, for example consisted of characteristics with greater precision.
all possible sampling points in a river basin; in such a case, unless reducing
the area reduced the variability, there would be no increase in accuracy. In considering sample size, it is also necessary to decide the minimum
numbers needed in particular sub-groups of the population which are to be
compared. Statistical tests will require a minimum in each category for any
TABLE 1 meaningful conclusions to be reached.
Sample sizes needed to estimate population
A final consideration is the size of the data body that can be handled
values with given levels of confidence ,
at the processing stage. If the measurement to be undertaken on each element
assuming a variability of 50% 1
is very lengthy, then the sample size may be limited accordingly.
and a very large population.'
Efficiency and resources of time and money have generally to be weighed
(from equations 4b or 6)
against each other. To reduce the sample error by 30 per cent, the sample
size will have to be doubled. This can be seen in Table 1, where the reduction
of the confidence limit from ± 6% to ± 4% necessitates an increase in sample
size from 267 to 600 at the 95% confidence level.

1 In the case of sampling from maps or lists, this increase in size may
16587 9604
2 not lead to a doubling of time and effort, but in the case of a questionnaire
4147 2401
3 survey, one of the leading research agencies estimates that doubling the
1843 1067
sample size would increase costs by 80%. Other studies involving field work
4 1 037 600
would expect similarly large increases.
5 663 384
6 461 267
In practice, in student work, sample size is usually determined by time
7 339 196
and resources. It may be that, having calculated the confidence limits for
8 259 150
the sample size that is feasible, it will be clear that statements could not
9 205 119
be made about the population with any degree of reliability. If we are unable
10 166 96
to conduct a study with a sample size sufficient to improve our population
15 74 43 estimates above the level of guesswork, then we might as well not conduct it
20 41 24 at all. Small scale work, even if it is only a very exploratory study, in-
tended to generate ideas rather than to draw conclusions about a population,
must recognise its limitations. Before embarking on any sampling procedure
the researcher should calculate the likely confidence limits for his estimates.
It will be in many cases a discouraging exercise, but is better done before-
hand than after collecting some meaningless data.

(v) Target population and sampling frame

Where the sample size given by the table is a substantial proportion of the
population size, the figure on the table should be corrected by formula (7)
in the text. The results of a sample study are only applicable to the population from
which the sample was drawn, although they may stimulate other research, refute
or support generalisations made elsewhere, or contribute to the formulation
For a complex sample design which is expected to increase the standard
of ideas about a wider universe.
error by 1.5, the confidence limit should be set to 2/3 of the desired
limit, for example, for a desired confidence limit of ± 6%, read the sample
The researcher will normally define a population of a size and variabil-
size for ± 4% (Section III, i).
ity that he can study with the resources he has available. However, he may
also have to re-define this target population to enable it to be drawn from
The confidence limit is measured as a percentage. For an attribute, it is
a convenient source, such as a list or a map, known as the sampling frame.
expressed as the percentage having the characteristic. For a continuous
variable, it is given as a percentage of the mean. Establishing a sampling frame, which contains all the elements in the
target population, is the first stage in sampling. If the frame is imperfect,
the researcher has to decide in the circumstances of his particular study
whether to accept its deficiencies by re-definition of the study population,
or to amend the frame. It is only possible to compensate for obvious
10
11
inaccuracies; often one can only evaluate the frame after completing the
research. If an old frame is being used, it may need up-dating. II PROBABILITY SAMPLES

There are two possibilities if the frame excludes part of the original (i) Simple random samples
target population. For example, if the land use in a country is to be studied
from a map, but this does not cover the north east corner, we could obtain The most straightforward sort of sample is the simple random sample
another map, or prepare an additional map in some way, or alternatively we already mentioned, where elements from a list or other sampling frame are
could re-define the population in terms of the existing map. This would mean selected randomly, usually by using random numbers.
stating that the study only concerned the part of the country shown. Provided
that this was made clear in any discussion of findings, the results for the Each element in the sampling frame is allocated a unique identifying
rest of the area would be by no means invalidated. number. Numbers are then read from a random number table, and elements in the
frame with those numbers are included in the sample. Random numbers, published
Similarly, in dealing with human populations, it is quite usual to ex- in every set of statistical tables, are simply lists of numbers in which each
clude those living in institutions such as army barracks, halls of residence digit has an equal chance of occurring in each position. As many can be used
or hospitals, and the report will point this out. at once as are necessary to correspond with the size of the identifying num-
Where an ineligible element is drawn in a sample, it cannot be replaced bers on the frame. For example, if the list were numbered 1 to 1000, three
by the next one on the list, as this would give the one following a second digits could be read at a time, with the number 000 corresponding to 1000.
chance of selection. If a list contains a large number of ineligible entries, If the frame included 1001, four digits would be needed. If numbers are read
or substantial parts of a map are to be excluded, enough elements must be which do not correspond to numbers on the sampling frame, then these in-
chosen to enable the final total of eligible elements to correspond to the eligibles are ignored, and the reading of numbers continues until enough
desired sample size. valid numbers have been read. It is usual to start from a randomly selected
point on the table.
Sometimes an element will occur twice in a frame. In special cases,
duplication will be so rare that it can be conveniently ignored. For example, Sampling theory assumes that if the same element is picked twice, it
so few individuals would qualify as members of more than one household that, will be included twice in the sample. In practice, this would not generally
although they would actually have had two chances of selection, they would be done, since it would tend to reduce the variability in the sample, because
hardly ever receive special treatment. two elements included would be identical. The technical term for not allowing
the same item to occur twice in the sample is sampling without replacement;
More usually, though, if an element occurs more than once, one occurrence, once an element has been selected, it ceases to have another chance of being
normally the first, is taken as the one qualifying for inclusion in the sample. drawn, and those remaining have a slightly increased chance - although this
If any of the others are selected, they are rejected as ineligible. is only significant if the sample is a very large proportion of the population.

It is not always possible to find a comprehensive source to serve as a The procedure of the simple random sample will produce a sample which
sampling frame, but there may be two or more sources, which, with some over- can be of exactly the desired size, from any part of the sampling frame. It
lap, include all the elements in he population . If all the sources were used,
t
might happen, since each digit has an equal chance of occurrence and there-
elements which occurred in more than one would have an increased chance of fore all combinations of elements in the frame could theoretically be en-
selection. The various sources are therefore placed in order, starting with countered, that the chosen sample is concentrated in one part of the popu-
the most comprehensive or best organised. The first entry of each element is lation. If the list contained males and females, the chosen sample might be
then taken to be the decisive one. The initial selection f elements is made
o
of one sex only, and in theory it would be sometimes. If the sampling frame
from all the sources, using the same sampling fraction. All those in the covered different rock types, only one might be found in a sample drawn at
first list or on the first map are automatically included, but those chosen random.
in the others are checked to see if they had an entry in an earlier list,
or were on an earlier map; they would then be rejected as ineligible. Although simple random sampling is in many cases reasonably quick to use,
provided that the sampling frame can be readily numbered (as lists and maps
The form of the sampling frame will to a large extent determine the with grids usually are), better coverage could be ensured by other methods.
possible sampling methods, but before looking in detail at the different
types of frame (chapters IV and V), it is helpful to consider the alternative Deviations from the simple random sample will be used to increase pre-
methods of selecting a sample. cision, that is, to reduce the spread of sample estimates around the popu-
lation value, and eliminate the sort of 'freak' samples just mentioned. They
will also be used for practical reasons, or to reduce the cost of drawing the
sample. The alternative methods are discussed in the succeeding sections.

12 13
 A systematic sample is not equivalent to a random sample, but it may be
(ii) Systematic samples
treated as one if the pitfalls discussed above are considered, and the re-
searcher feels confident that regularities in the sampling frame will not
One simple and convenient method of ensuring even coverage throughout
a sampling frame is the systematic sample. Instead of using random numbers lead to bias in the sample.
which could indicate elements in any part of the population, a regular spacing
is used, taking every kth individual, or the intersections of a regular grid (iii) 'Systematic random' samples
laid over a map.
Some of the problems of systematic samples can be minimised, while re-
The sampling interval, k, is the reciprocal of the desired sampling taining the advantages of even coverage, by combining systematic sampling
fraction, that is, k is N/n. If the sample is to contain 50 elements and the with a greater random component. There are a number of ways of doing this.
sampling frame lists 5000, one element in 100 would be chosen by setting k The simplest method where periodicity is suspected is to use a series
at 100. of random starts scattered throughout the frame, and to select with the
The sample is selected systematically from a random starting point. This intercal k between them, as if each part constituted a separate list. For
is usually a number between 1 and k, in order to ensure that all the items example, in order to select 100 from a population of 1000, k would be 10.
in the list before the first one chosen have a chance of selection. This is We might choose 5 random numbers, starting with a random number between 1 and
the only random number which needs to be drawn. Thereafter the selection of 1 0. If these happened to be 4, 26, 364, 505 and 787, we would start by samp-
the components of the sample proceeds simply, with every kth one chosen. ling 4, 14, 24, and then 26, 36, 46 .... 364, 374, 384 ....

Alternatively, a series of low starts could be made and a systematic

It is subsequently very easy to check that the sampling method has been
sample drawn with the same interval from each. If we wanted to take 100 from
correctly applied, because the sample will be regularly spaced; the sample
1000 with 5 starts, we would use k = 50, and make 5 starts between 1 and 50.
will come from every part of the original frame in the correct proportions.
Using an interval of 50 from each of these starts would have the same effect
However, there are two major reasons why the systematic sample, for all its
as using an interval of 10 throughout the whole sequence.
attractions and its frequent use, is not strictly a random sample.
A spaced random sample has the advantage of even coverage, but lacks the
Firstly, the systematic method does not enable all possible different
advantage of quick selection. Random numbers are read for each member of the
samples of the given size to be selected. The number of different samples is
sample, but the choice of an individual element excludes all its neighbours
limited by the sampling interval k. The selection of members of the sample
up to a chosen distance away. This distance must be less than k/2, and the
is not independent, because once the first individual has been chosen, all
closer to k/2 it is, the more evenly spaced this sample will be. Sampling
the others are fixed. This violates one of the assumptions on which sampling
from maps might lend itself to this method, where the selection of a point
theory is based.
could exclude all the points in a circle around.
Secondly, if the sample produced by this method is to be analysed as if
it were a simple random sample, the frame needs to be in a random order. A randomly varying interval can be used instead of a fixed k. The inter-
val is varied each time by a small random number, r, added or subtracted, so
Many lists can be assumed to be randomly ordered, but many lists and certainly
that the elements are chosen at intervals of k ± r, from a start between 1
most spatial arrangements, whether plants, landscape features or the built
and k. Whether r is to be added or subtracted each time should be randomly
environment, contain order. In many cases there is a periodicity which may
determined. One way of doing this would be to read an extra random digit, an
not be immediately apparent. If a list of households is arranged in order
odd number indicating plus, an even number (or 0) minus. Small random numbers
corresponding to their position along a street, taking a sampling interval
can be drawn quickly using playing cards; if r was to have a maximum value
with an even number might produce a sample consisting entirely of the left-
of ± 7, the higher cards (and the picture cards) would be removed. Red cards
hand pair of semi-detached houses, which might be different in size or other
could indicate positive values, black cards negative values of r.
i mportant features. Terrain or crops, too, might have periodic features which
could be picked up by a sampling method which used a regular interval.
The desired k and the limits of r would have to be decided first, and
the random start made between 1 and k. With k = 50 and r between 1 and 7, a
An additional problem can occur when a list is arranged in an order,
starting point of 48 might be made. The next element would be chosen by
for example, in ascending order of magnitude of some variable, (either one
drawing a card. Red four would mean that the interval was 50 + 4, and the
being studied or one which correlates with variables being studied). In such
element chosen would be 102. The next card might be black 1, and the interval
cases, the initial random start can affect the estimate of the population
therefore 49, giving element number 151.
mean, for a high starting point will bias the sample upwards, a low start
downwards. The effect may be reduced by treating the list as circular. After
The final sample size using the k ± r method would be very close to the
drawing every kth element, those left at the end are added to those at the
sample size using a fixed k.
beginning, and if there are more than k, another one is drawn; this is more
likely to give equal representation to high and low values but will not Many sampling frames will lend themselves to division into systematic
completely solve the problem. blocks, where the number of blocks is the same as the desired sample size.
One element is then chosen from each at random. The advantage of this method
is that only small random numbers will have to be drawn.

14 15
 If the sample size is not an exact division of the population, the final
block will be smaller than the others. 'Blanks' are therefore added to it to
make it the same size, and if one of the blanks is chosen, no element from
the last section will be included in the sample. Thus each element in the (8a)
smaller section will have the same chance of inclusion as those in the other
blocks.

(iv) Stratification

Systematic samples effectively choose elements from even-sized blocks

of the population. Strata, sub-divisions of the frame within which elements
are chosen, can be s imply even divisions of the population (or approximately
even, as in the case of the randomly varying interval systematic method),
but they can be sub-groups of the population of any size that are expected Unless the sampling fraction were very large, we could ignore the ex-
to differ from each other in ways important to our study. pression (1-n/N), and calculate the standard error:

If we want a sample to contain both men and women, it would be sensible

to sub-divide the frame on the basis of sex before sampling (if we had the (8b)
information to do this), rather than run the risk of obtaining a sample
unrepresentative of the population. If we expect different areas to have dif- Many researchers make the assumption that the variability in all the
ferent characteristics, it will be as well to consider them as separate strata will be very similar and continue to use the formulas given in section
strata, and choose within each of them. I, which are applicable to simple random samples. This has the advantage of
greatly easing the calculations, but each variable studied may have a differ-
There are two benefits from stratifying in this way. Firstly, it will ent variability in the strata, and a different degree of correlation with the
ensure that each part of the population which we might want to compare will variable used to create the strata. For most variables the standard error
be represented with a sample of sufficient size for the comparison to be made. calculated by the simple random sample formula would be an over-estimate of
If it is known in advance what groups are to be tested, and they can be ident- the sampling error.
i fied in the frame, even approximately, it will be beneficial to stratify.
(v) The procedure for stratifying
The second major advantage of stratifying is that it will tend to improve
population estimates, provided that the variable used for the stratification In some cases, the sampling frame will be already divided into suitable
is related to the subject of the study. For example, we might expect soil strata, for example, a trade directory may be arranged in different categories
types to be important determinants of aspects of plant communities. The more of businesses, but in others the researcher will have to decide how to divide
different the plants on each soil type are, and the more internally homo- up the frame.
geneous the communities are, the more the estimates of the population will
gain from the use of the strata. If the sampling frame is conveniently divided, it is possible to draw
the part of the sample in each stratum separately, either randomly or systemat-
If the frame were randomly ordered in terms of the variables being ically. Alternatively, a random sample with stratum quotas can be drawn by
studied, then there would be no gain from stratification (or from systematic dividing the sampling frame into 'strata ', which are either blocks of equal
sampling) and a simple random sample would be as good. However, this complete- size, or which correspond to differentiated parts of the frame, for example
ly random arrangement will be unusual, and although we will not normally be soil types or areas of a town. Each part is allocated a proportion of the
able to calculate how much we have gained from stratifying, we can expect sample, and random numbers are read until the quota in each stratum is full;
that in almost all circumstances our population estimates will be more precise numbers are discarded if they fall in a part whose quota is completed.
as a result. Greater precision means a reduction in the spread of sample
estimates around the population parameter. Each individual sample may not be There is no need for the strata to be of identical size, and there is
better than each simple random sample, but freak samples will be less likely no rule about the appropriate number of strata, although Kish (1965, p.102)
to occur. This means that, at a given confidence level, the confidence limits suggests that between three and ten will normally be best.
will be narrower for a given sample size - but the same effect might be
achieved more cheaply by drawing a larger simple random sample. If the sampling frame provides little information with which to perform
the stratification, supplementary information from another source may be used,
Even if the simple random sample will not produce great savings, it will provided that it correlates with the desired stratifying variable.
facilitate calculations, for even when the same sampling fraction has been
used in each stratum, the standard error is estimated by summing the variance Thus, if we wanted to include even coverage of areas with small and large
percentages of pensioners or particular house types, we could use a frame
in each stratum:
such as the Electoral Register (section IV, iii) which did not contain any
information on these variables, but distinguish our strata by Census,

16 17
 portion of working mothers is very small, it might be necessary to sample
airectory or map information for the same areas. This would probably be valid them at a higher rate even if we expect that they will prove more homogeneous.
even if the source was out of date, but the researcher would have to be sen- The effect of this would be to increase the overall sampling error, making the
sitive to any changes (new buildings, demolition and so on) which might sample a less efficient way of determining the characteristics of the whole
radically have altered the areas. population of mothers, but this disadvantage might be over-ridden by the
desire to be able to make comparisons between the two groups.
This suggests another important point: the researcher should always know
his study area as well as possible before commencing the formal data collec- (vii) Weighting
tion stage. He may in fact know the characteristics of the frame sufficiently
to make an approximate division on a subjective basis. The categorisation of If the strata are sampled with different fractions, in order to generalise
towns, house types, central and peripheral parts of a settlement or remote about the population as a whole, the sample values for the individual elements
and accessible parts of a nature reserve might be used to create strata which must be weighted by the reciprocal of the sampling fraction to represent their
would improve the population estimates. Provided that, having created the true proportion in the population.
strata, sampling proceeds rigorously, there is no obligation to use only in-
formation internal to the sampling frame. Table 2 illustrates a stratified sample for a study of three villages;
they have 200, 300 and 500 households respectively, and these are therefore
(vi) The choice of sampling fraction

The simplest form of stratification involves applying the same sampling less variable than the first. The Table shows the calculations for the
fraction, that is, taking the same proportion of elements, in each stratum.
There are occasions, though, when the use of a different sampling fraction
in particular strata can be beneficial.

The first case is when the variability of the population is very differ- owning households in the total population of the three villages. We can esti-
ent from one stratum to another, and is known or can be guessed in advance. mate the proportion of car-owning households by dividing u by N.
Thesma leh the variability, the smaller the sample size needed to provide
an estimate of its characteristics (section I, iv), and the same is true of
the part of the population in a stratum. A smaller number can be taken from TABLE 2
a more homogeneous stratum, saving resources which can profitably be applied
to the more variable strata. Estimating population values with a weighted sample
If our sampling frame consists of mothers of children attending a partic- Number of
ular nursery, for example, we might expect the mothers with jobs to differ Number of Sample car-owning
from those without. If the topic under investigation is perception of the households size in households Weight for
immediate vicinity, we might expect those not working to be more alike than Strata
in stratum stratum in sample stratum i
those who travel away from the area to work. If , on the other hand, we were from
studying the behaviour patterns during the time between leaving and collecting stratum i
the children, the working mothers would almost certainly be less variable.

If we feel sufficiently confident in our decision about the differential

variability of the two strata, we might produce a more efficient sample (one Village one 200 35 18 200/35 = 5.7 102.9
which gives a better population estimate than a simple random sample of the
same size) by using a larger sampling fraction in the more variable stratum. Village two 300 30 21 300/30 = 10.0 210.0
However, it might happen that the working mothers are more uniform in their Village three 500 35 14 500/35 = 14.3 200.0
perceptions, since they all work on a single industrial estate, and we would
then discover that we had made the wrong decision if we had over-sampled them.
Our allocation of the samples to the strata would then produce a less effic- Total N = 1000 n = 100
ient sample than a simple, unstratified, random sample. Only if w e
ave strong
h

evidence for differential variability should we use different sampling frac-

tions, both because of the possibility of producing a worse sample than a
simple random sample, and because unequal allocation will necessitate weight- u, the estimated number with the characteristic, car-ownership, is 512.9
ing, which, as the next section shows, complicates calculations. p, the estimated proportion with cars, is calculated:

The second case in which we might want to use different sampling frac- p = u / N
tions is when one stratum is very small, and we need to have a sufficient = 512.9 / 1000
sample in that part to enable it to be compared with the rest. If the pro-
= .5129 or 51.29%
19
18
 Kish (1965, p.427) makes the point that there is no need for the weights If the costs of including each element in the sample vary between the
to be precise to several places of decimals; in most situations where weights strata, this factor can also be taken into account in determining the alloc-
vary from 10 to 99, rounding the weights to 2 digits will be sufficiently ation of effort to the strata. The best sample is obtained when the sampling
precise. Using 6, 10 and 14 as weights in the example above gives u = 514,
not very different from the precise answer. Weights do not have to equal the
reciprocal of the sampling fraction; more workable numbers may be obtained
by using numbers which are in proportion. Instead of 6, 10 and 14, we could
use 3, 5 and 7, but then instead of N in the calculations we would use
It would be rare indeed to have sufficient information to obtain a
sample which conformed to this optimum, but given the constraint about the
The formula for estimating the standard error of a weighted sample is: complexity of processing samples with very large weights, some such alloc-
ation can be attempted. In addition, only specialised sample designs of very
unusual populations would have strata where variability and costs differed
(9a) substantially.

(viii) Stratification after selection

Occasionally, we may want to stratify on some variable which is not clear

on the frame, but which is measured in the course of the study. In these
situations, the original sample is sometimes weighted to correspond with the
known distribution of the population. This method should be used with care,
for if the final sample does not correspond closely to the characteristics of
the population in one or two important respects, it may be that the elements
that are included are not truly representative.

For example, we may want to weight a sample of people to correspond with

the age groups in the population as at the most recent Census. However, there
may be reasons why our sample does not have the correct proportions; our
sampling frame may have excluded mobile sections of the population, for
(9b) example, young people, or we may, by excluding those in institutions, have
reduced the number of elderly people. Simply to weight those young people or
those in the oldest age group whom we have included in the sample may not
improve our sample's estimates for the whole population, and may well make it
worse (Moser and Kalton, 1971, pp 99-100).

(9c) III CLUSTERING

(i) Reducing costs over a large area

Weighting will counteract to some extent the beneficial effects of
stratifying because if some elements in the sample have very large weights, Sampling over a large area, or from a very dispersed population, can
it will reduce the variability of the sample. As a general rule, though, for produce a sample which is very precise, but enormously costly in time and
most variables we would expect our stratification to show appreciable gains effort of data collection. Examples occur when field-work is necessary, and
over a simple random sample, even with some weighting. a widely scattered sample, or one which is evenly spread as a result of using
systematic sampling, is difficult to survey. Similar problems occur when data
The most efficient sample will be obtained when the sampling fraction are from maps or other sources, and not all can be bought or sorted through
in each stratum is proportional to the variability in that stratum, measured for the sample. In physical geographical studies, problems of access over
by the standard deviation. In other words, if we consider only one variable, large areas might make a scattered sample very hard to use.
and the standard deviations in three strata are 5, 6 and 9, the best sample
is obtained when the sampling fractions are in these proportions. A common solution is to use a sample which is deliberately grouped
in a convenient number of small areas. In other words, we divide the frame
into parts, and not all of them are sampled. Small scale projects rarely
need to use cluster samples because they do not cover large areas.

20 21
 Yates (1960, p.19) writes that 'it will be more accurate to take 10 per The clusters should therefore be internally as heterogeneous as possible,
cent of all farms in each parish than to take all the farms in 10 per cent /so that the omission of whole clusters from the sample will not remove un-
of the parishes', in other words, a clustered sample will be less good at usual parts of the population and thus bias the final sample. In stratified
estimating the population values than a sample of similar size drawn from sampling, the object is to create strata which are as different from each
throughout the sampling frame. other as possible, since all will be included. The opposite principle is
applied in creating clusters; each should be a microcosm of the population.
Exactly how detrimental will be the effect of clustering on the pre-
cision of the sample is hard to determine. It will not be the same for all In many cases, clusters will be geographical sub-areas since that is why
variables, having the most adverse effects on the variables for which the clustering is more economical and has to be used. If it is thought that these
clusters are relatively homogeneous. will be internally homogeneous, and different from each other, cluster samp-
ling is not appropriate. Maximising heterogeneity and concentrating data
The ratio of the standard error of any sample to the standard error of collection in small areas are mutually contradictory for many of the features
a simple random sample of the same size is the square root of what is usually of interest to geographers.
called the design effect (the ratio of the variance of a sample to that of
a simple random same of the same size). Since cluster sampling will almost (ii) Creating clusters
always increase the standard error, the ratio will almost always be more than
one while for stratified samples it will usually be less than one. Experience Clusters tend to be more homogeneous than the population as a whole, in
shows that with clustered samples it usually varies from 1 to 2, depending other words, elements within them tend to exhibit high intra-cluster cor-
on the variable, and it is often around 1.5. relation. The procedures for creating clusters are designed to minimise the
effects of this.
The calculation of the standard error of a clustered sample is complex,
Larger clusters, from each of which a small number of elements are taken,
particularly when, as is usually the case, clustering has been combined with
will be best, because they are more likely to be heterogeneous.
stratification. Rather than simply using the formulas for a simple random
sample, it will be better to correct them by the 'rule of thumb' ratio of 1.5 The number of sampled elements in the cluster will be determined by the
to take account of the design effect. Instead of formula (2), we might there- savings that can be made, since clustering is undertaken for economic reasons.
fore estimate the standard error by: If sample size can be increased by reducing travel time between elements, the
clusters should contain a reasonable assignment for one field-worker or one
(10) day's work; a larger sample in one cluster would be unnecessary.

Heterogeneity can also be increased by skilful division of the sampling

where SE is the standard error of a sample where we expect the standard error
frame. In urban areas, for example, gains may be made by producing elongated
to be 1.5 times that of a simple random sample of the same size, s is the
clusters rather than compact ones, perhaps by aggregating wards; interviewing
standard deviation, and n is the sample size. The expression (1-n/N) will be
would still be facilitated, because the areas would be contiguous and de-
ignored unless the sampling fraction is very large.
limited, but the possibility of a cluster containing a single house type,
household type or 'social area' would be reduced.
The formula for calculating the sample size needed would be altered to:
If a list which is not spatially arranged is to be used as a sampling
(11)
frame, and it will be prohibitively expensive to spread data collection
throughout the area covered by the list, it can be arranged into clusters.
where n is the sample size needed, s is the standard deviation, z is the
The quickest way of doing this is to draw a sample several times the size of
standard-error-unit measure for the desired confidence level, and c is the
the required sample. These elements are then grouped into geographical areas
confidence limit.
each of which contains the same number of elements. These clusters can then
be selected randomly, and the result will be a sample arranged in a more eco-
Table 1 can still be used to give an approximation to the sample size
nomical way, although clusters will clearly not be as compact as might be de-
needed if the result for the confidence limit 2/3 of the one desired is used.
It can be seen that if we make the assumption that the ratio will be about sired.
1.5, our sample size needs to be twice as large as a simple random sample to
(iii) Selecting clusters
achieve the same precision. For e a pl , at the 95% confidence level, if we
x m e

want to be within 9 per cent of the true value, we would have to have a sample
The simplest form of cluster sampling is when the population is divided
of 119 for a simple random sample, but 267 for a clustered sample.
into equal sized groups, and the resultant clusters randomly selected. All
elements within these, or a fixed proportion within the clusters drawn randomly
Although the actual number used is arbitrary, multiplication by 1.5 will
or systematically, are then included in the sample.
generally be better than ignoring the adverse effects of clustering on the
ability of the sample to estimate the population values. The researcher should,
however, endeavour to minimise the design effect.

22 23
 If clusters of different size are randomly selected, and a fixed samp- IV SAMPLING FRAMES
ling fraction is used, individuals in smaller clusters will have had a greater
chance of selection than those in larger ones. Elements will have to be (i) Convenient sampling frames
weighted by the size of the cluster from which they were drawn.
Where resources are limited or time is short , the rapid application of
Where cluster sizes and sampling fractions both vary, two weights will a sampling procedure to a readily obtainable sampling frame will greatly
be needed for each element. Firstly the variable sampling fraction is correct- assist the project.
ed by multiplying each element by the reciprocal of its cluster's sampling
fraction, as in Table 2, and then a weight proportional to the size of the At the expense of reducing generality, it might be reasonable to limit
cluster is needed. a study to a sampling frame which can be easily obtained, and to devote pro-
portionally more resources to other aspects of the data collection.
(iv) Sampling with probability proportional to size
For example, pilot or exploratory work on time geography (Thrift,
Clusters are usually created to economise at the data collection stage, CATMOG 13 1977) might reasonably be limited to an institutional population,
and the work-load in each is the smallest possible number of elements that such as nurses in a home, if the co-operation of the appropriate authorities
will make these savings. It is also preferable for each field-worker to be could be obtained. This might have the added advantage of reducing the expected
allocated the same size of sample (except in circumstances where the work is variability in the sample (there are only a certain number of shifts) and
expected to be substantially more difficult in some areas, because of terrain thus reducing the necessary sample size.
or greater difficulty in contacting respondents, perhaps). If the clusters
vary greatly in size, the equalisation of the sample size in each can be Similarly, it might be possible to discover something useful about a
achieved without weighting by selecting the clusters with a probability which small area, or about the users of a particular facility. Even an inexperienced
depends on their size. researcher who had designed a sensible project might manage to gain access
to membership records of a small organisation which felt its work might
The need for weighting results from the fact that selecting uneven-sized benefit from his study. Clearly, for example, a study of young people based
clusters equally gives those in the smallest clusters a higher chance of on those who used youth clubs would be in no way representative of young
selection. If the initial selection of clusters is dependent on the cluster people in general, but case studies are certainly not invalid because of
size, this effect is compensated for. their lack of generality.

The simplest method in practice is to allocate sequential numbers to the (ii) Inconvenient sampling frames
elements in each cluster. In other words, the clusters are listed with a
cumulative total of elements within them. If a researcher is interested in a population for which no listing or
source is available, it will be necessary to investigate the possibility of
Random or systematic sampling can be used to select from the list of
constructing a sampling frame.
cumulative numbers; the clusters within which numbers fall are used in the
second stage of the sampling procedure. This is demonstrated in section IV,
The compilation of large-scale sampling frames is complex and time-
vi (Figure 1). consuming. Where the population to be examined consists of a rare sub-group
In a multi-stage sampling design which is designed to avoid weighting, of a population some form of screening might be used. For example, to find
every stage of selection except the last is made with probability proportional a sample of freezer-owners to compare with other shoppers, either a short
to size. For example, if constituencies are used as the first stage, and postal questionnaire or a brief interview might be necessary.
wards as the second, both should be chosen with probability proportional to
their size (number of electors). An equal number of elements should be taken Screening is more economical when resources are concentrated in clusters,
from each of the chosen wards. rather than spread out to create a frame over the whole area. An area of
sufficient size should be screened or investigated to yield a sample of the
The method requires that the population of each unit is known, or that desired size without the rejection of any, so that all the resources employed
any errors will be constant for all the clusters (for example, if the inform- to identify the sampling frame would be used to the full. When the percentage
ation on size is out of date, the error might be consistent if all the units with the rare characteristic is not known, clearly it must be estimated con-
have increased their population in the same proportion). servatively, with the aim of finding at least as many as are needed for the
sample.
Sampling with .probability proportional to size is the most convenient
form of cluster sampling. If at the preliminary stage there is doubt about whether an element is
eligible or not, it should be included, because false inclusions can be re-
moved at the data collection stage, but false exclusions will not be sub-
sequently checked.

Screening involves a great deal of time and cost to find each element in
the sample. Where each member of a research team wants to investigate different

24 25
 (possibly selected from the Electoral Register, as explained above) where we
sub-groups or features of a population, and could pool resources to identify
were really interested in individuals.
the sampling frame, perhaps screening could yield a more justifiable return.
Often we would find that attitudes and behaviour were highly correlated
A similar process might be used where it is necessary to obtain a large
for members of the same household, and it would be detrimental to the sample
amount of information from a small number of individuals, and a small amount
to include all the members of the household in the sample (section III, i).
from a large number. This might take the form of conducting short interviews
There is also an additional problem with interviewing more than one person
and asking a sub-sample to complete diaries or undergo longer interviews.
at the same address where the first interview may influence the others, either
In other studies, it would be economical to utilise the differential vari-
by suggesting responses, or by taking up an unreasonable amount of time.
ability of the information being collected by, in effect, having samples of
different sizes, with the more uniform variables collected from only a small
Therefore one member of the household is usually chosen at random. The
part of the larger sample from which the others would be collected. Farm size
problem is that we have to establish all the members of the household before
might be highly variable, crop yields much less so; this factor could be used
making the selection; as well as being tiresome for the field-worker this
in the sample design to speed up data collection by not determining yields
procedure can be discouraging to the respondent and may prevent him from co-
on every farm. However, if one aim of the study were to discover if more
operating altogether.
extensive farms had higher or lower yields than smaller ones, the method would
be inappropriate.
If a rare part of a population, for example a particular crop, rock or
group of people, is unevenly distributed, the reduction of generality to
The arrangement of the list will determine the ease with which the
areas where the concentration is greatest might be essential in order not to
'first entry' method can be used. In rural areas, names may be listed in
dissipate energy searching out needles in the rest of the haystack - but the
alphabetical order making it hard to work out the households; in urban areas,
isolated part of the population might well have distinctive characteristics,
the listing is alphabetical within addresses, street by street. Problems
arise when there are several surnames at the same address, who could and its exclusion would have to be explicit.
be boarders, relatives, or separate single person households in bed-sitters.
(iii) Sampling from lists: the example of the British Electoral Register
If a name is chosen and there is any doubt about whether others listed are
members of his household, all the names at the address would have to be listed
The ideal sampling frame would be a list which contained all the target
and checked by a field-worker, so that if the voter listed is not the first
population with no elements excluded and none included more than once. This
named in a household, he can be crossed off.
is rarely encountered, but in many cases an imperfect frame can be adapted to
serve. The British Electoral Register, the most commonly used sampling frame
The definition of what constitutes a household will have to be rigorous;
it is usual to include people who regularly live together and are catered for studies of people in Britain, is by no means a perfect source, and the
for by the same person for at least one meal a day, but specialised surveys ways of using it illustrate the techniques adopted to deal with problems en-
may need their own definition. The researcher will have to decide how to de- countered more widely.
fine any institutional households he wishes to exclude.
The Electoral Register lists all the adults aged eighteen and over
eligible to vote who have complied with the legal obligation to complete
A supplementary frame can be used where the main frame is incomplete;
registration forms on the basis of residence in October each year. Registers
an area of new buildin might be sampled from an address list. However, the
g

sampling fraction would have to take into account the fact that the Register are available in public libraries or can be purchased from local authorities.
gives individuals, the list dwellings.
Although compiled in October, the Register is not published until the
following spring, so that it will always be at least three months, and
Another way of dealing with omissions is to use the half-open interval
possibly nearly fifteen months, out of date. Towards the end of the life of
method. With each element drawn for the sample a note is made of the next
element in the list. Any subsequently found to lie between them and not on the list (that is, in the early part of the year) alternative frames might
be more attractive. It is possible to substitute a new occupant of the same
the original frame are included. This works well for small amounts of new
address for someone who has died or moved, although it would have to be some-
building, or where several households are found at one address. The method
one not already included on the Register, and randomly selected within the
assumes that each element omitted is linked unambiguously to one included,
new household (see section IV, iv). If it is only the selected individual who
and thus compensates for exclusions from the original frame. Problems arise
is no longer there, no substitution can be made, because the others in the
when the spatial arrangement makes it difficult to ascertain to which included
house would already have had a chance of selection.
element an omitted ohe should be attached, and when the intermediate elements
are very numerous, when an unacceptable level of clustering would arise. It
Lists frequently contain a number of omissions: in the case of the
also depends on field-work to identify all the exclusions.
Electoral Register, foreigners, immigrants or others not sure of their en-
titlement to vote, as well as people who have recently moved. The simplest
(iv) Selecting within small clusters: the example of households
solution is to ignore these deficiencies, and note in the research report any
Sometimes our sampling frame constitutes, in effect, a list of small qualifications which result.
relatively homogeneous clusters. This would be the case with households

26 27
 a list of households, to each of which is attached one row of the Table.
However, an alternative is available which may be preferred, particularly
The rows are allocated in the proportions set out in the first column, so
in areas of high mobility or when the list is out of date. It is more likely
that if an interviewer went to twelve addresses, two would have the first row,
that one elector will have been left off than that all those in a household
one the second, one the third, and so on. When the field-worker has listed
will have been. It is therefore a better source for households than for
all the eligible individuals in the house in a pre-determined order (usually
individuals. by age), he selects for interview the one with the number indicated for that
We do not know how many voters are listed at each household. We therefore household size by the row of the Table given to that address. For example,
decide that the first entry on the list of any household will have to be if there were four people eligible in a household, and the address had been
chosen if that household is to be included. This gives each household only allocated row V, individual number three would be interviewed; using the same
one chance of selection irrespective of the number of voters in it. row, if there were five in the house, number four would be interviewed, if
three, number two.
In order to know what sampling fraction to use when many of the names
selected will be treated as ineligible because they are not the first listed,
we need to know that on average there are 2.2 in each household, and we (v) Constructing frames from maps
therefore start out by selecting 2.2 times the number of households we actually
Published maps or aerial photographs can be used to create lists of a
want in the sample.
variety of phenomena for which no suitable sampling frame already exists,
It will be necessary to weight the answers for each individual by the such as vegetation, water courses, shops, factories and dwellings. Clearly
reciprocal of the sampling fraction, within his cluster, that is, by the there will be errors due to inaccuracy of the source map, changes such as
number of eligibles in the household divided by the number chosen. Someone demolition and construction or disturbance of vegetation, and misinterpret-
who is the only eligible member of a household will therefore be weighted by ation. Comparison of the map with the area might indicate the extent of these
1/1, where there are two people eligible, the one chosen will be weighted by problems.
2/1, where there are 3, 3/1, and so on. To avoid weighting, instead of using
the 'first entry' selection method from the Electoral Register we include a Some phenomena can be identified readily from maps, but care would have
household when any of its members is selected. This gives, in effect, selec- to be taken over multiple-occupied premises, moveable stalls, mobile shops,
tion with probability proportional to size (number of voters). Taking one and so on.
from each household prevents the need for weighting (III, iv). Specialised maps may name all the elements to be included. An example
of this type of map is the series of Goad shopping centre maps, on which each
TABLE 3 shop is named and its main functions indicated (Rowley and Shepherd, 1976).
Such maps might be preferable to directories if their coverage were more com-
Kish's tables for selecting one individual from a household plete, and if a sample were desired with a spatial spread. One way would be
to number all the units shown by a 'postman's walk' route, and then to use a
form of systematic sampling.

Rows are Where such a map is out of date, it might be possible to up-date it
allocated Number of eligible individuals found: without re-surveying the area. If, for example, a shop had been sub-divided
in the into smaller units, all should be included in the sample if the original was
proportions: 1 2 3 4 5 6 included. The half-open interval method can be used for new units inserted,
or more provided that they can be unambiguously linked to their neighbour on one side.
select individual number: The removal of a unit, such as a closure, would mean that element would be
considered as an ineligible blank. A sufficient size of sample would have to
Row I 1/6 1 1 1 1 1 1 be drawn to allow for some deletions in this way, because it would be quite
1 1 2 2 incorrect to substitute a neighbouring unit which would then have an extra
Row II 1/12 1 1
chance of selection.
Row III 1/12 1 1 1 2 2 2
1 2 2 3 3 (vi) Constructing frames by mapping
Row IV 1/6 1
Row V 1/6 1 2 2 3 4 4 Where there are no obtainable maps which mark the population to be
3 3 3 5 sampled and are up to date, a researcher may have to construct a sampling
Row VI 1/12 1 2
frame on the ground, or at least amend the existing maps in order to construct
Row VII 1/12 1 2 3 4 5 5 such a frame.
Row VIII 1/6 1 2 3 4 5 6 Buildings such as industrial or commercial premises in a small area,
dwellings in rural areas, new buildings, or in countries where there is no
suitable sampling frame for households, a listing may be prepared by mapping.
An impartial selection within the household can be made by using
Clear procedures and forms for recording information will be needed,
Table 3, devised by Kish (1965, pp. 398-400). The interviewer sets out with

28 29
and Kish (1965, pp.322-351) discusses the instructions needed to list build-
ings for such a frame. The area is usually divided into segments delimited by
natural features or streets, clarified by sketch maps where necessary.

Complete enumeration of every element in the population may be possible,

but will often be too costly and time-consuming. The segments may be used
as clusters, and only those included in the sample need be mapped fully.
Clusters are most conveniently sampled with probability proportional to
size, to avoid weighting (section III, iv). If population sizes for the
segments are known, this method can be used, and then the same number of
elements would be sampled from each one.

Where segments contain fewer than 6 or 8 elements, it may be more eco-

nomical not to use probability proportional to size sampling, but to include
all the elements in the selected clusters. The value for each element would
have to be weighted by the size of the cluster from which it was drawn, but
this may be acceptable if the segments vary little in size. The advantage is
that there is no need to list all the elements first and then select within
them.

If the selected clusters were subsequently found to contain only slightly

more or slightly fewer elements than the enumerator first thought, there will
be little difficulty. If more were to be found than expected, all should be
included in the sample.

Figure 1 illustrates the kind of sheet that could be used to record the
segments, in a convenient form for the selection of clusters with probability
proportional to size. Random numbers would be drawn once the listing was
complete, and whichever clusters they were in would be the ones included in
the sample. In the example shown, if the numbers 9 and 21 were drawn, the
clusters selected would be segments 2 and 5. If the same number of elements
was chosen from each, the sample would be self-weighting (section III, iv).
In this example, two elements would probably be included from each. The dis-
advantage of the method is that the chosen segments have to be visited again,
and the elements within them listed.

It would be possible for the identifying numbers for the clusters to be

drawn in advance, so that the field-worker could proceed directly to data
collection after the listing of segments. In this case, to avoid conscious or Fig. 1 Listing Segments: An Example
unconscious 'cheating', he should not have the random numbers to hand when
listing the segments, but should carry them with him in an envelope to be
opened only after all the first stage had been completed. In some situations, may become more conscious of their actions, or be conditioned to respond in
particularly where the area to be covered is compact and the data to be a particular way. They may therefore be very different from the population
collected readily available (as might be the case with observations rather they are supposed to represent at the subsequent interview.
than interviews), field-work could be conducted rapidly by this method.
With human populations, unless the time period between the interviews
(vii) Longitudinal studies is very short, there will be substantial 'attrition' of the sample as people
move out of contact or die. Substituting its replacement for a household which
Where studies are designed to measure changes over time, it may be has moved is sometimes used, although it may not be the case that the new
necessary to re-visit measuring points or a panel of respondents. This assumes inhabitants resemble the old, and the method is anyway inapplicable to in-
that the process of measurement does not destroy or affect the variables being dividual respondents. The response rate on the second interview is also likely
studied, as trampling vegetation or disturbing the soil does. A different to be lower than on the first, unless the interview is very enjoyable or the
sort of change takes place in people who have been interviewed once, for they sample 'captive', as in the case of a school class.

30 31
 To avoid repeating the measurement on the same individual or at exactly
the same point, some studies select two samples with the second sample ad- V AREAL SAMPLING METHODS
jacent to the first. In the case of physical features, this method seems
legitimate, but it may not be safe to assume that people will be like their (i) Frames and methods
next-door neighbours.
It might be expected that geographers would make considerable use of
Another method of studying change, therefore, is to sample afresh. The areal sampling techniques, but especially in human geography samples are
problem is that each sample may have such large sampling error that it is generally drawn from frames which seldom have more than an incidental spatial
i mpossible to distinguish differences due to that from real trends in the aspect. Land use, vegetation, soils and other continuous phenomena are, how-
data. It is more reliable, although more costly, with large-scale longitudinal ever, sometimes sampled from maps or aerial photographs, and similar sampling
studies, to draw an initial master sample several times larger than the sample methods may be used on a small scale in the field.
needed each time. After the initial observation, sub-samples are drawn from
it for subsequent phases. Each individual is only examined twice, which Many of the spatial sampling methods are directly analogous to the a-
reduces the effects of 'contamination' due to observation. Moser and spatial methods already discussed.
Kalton (1971, pp 137-145) discuss panel studies and longitudinal surveys Normally we use a grid system with co-ordinates, preferably one already
for human populations; their suggestions have applications for other present on the map as in the case of the National Grid on British Ordnance
studies. Survey maps. If the area is of irregular shape, the grid must overlap the
study area, as shown in Figures 2 to 7. The grid reference for each point
serves the same function as the identifying number in a list.
(viii) Sampling continuous flows
Figure 2 shows the spatial version of the simple random sample, using
When records are constantly being added to a card index, or people are points as elements. Random number tables are used to provide co-ordinates on
passing a point, the researcher cannot start off with a complete sampling both axes. Points selected which fall outside the study area, in this case in
frame and then make his selection. The only feasible sample will be systematic, the sea, are rejected, just as any blanks in a sampling frame would be.
taking every kth element, or every (k ± r)th, where r is a small random num-
ber. The systematic sample incorporates the advantages of coverage through This form of sampling is used to estimate land use on the assumption that
time, representing early and late arrivals equally. the points will fall on each land use type in its correct proportions; the
percentage of points on waste ground, for example, will give the percentage
It is not possible, however, to pre-determine the sample size, unless of the whole area covered by waste land.
the rate of flow can be precisely predicted. The interval k will have to be
guessed on the best available evidence. It is likely that waste ground will not be distributed across the whole
area; a small random sample might miss it altogether. Most work on an areal
The processing of individuals in the sample may hinder the counting of basis requires even coverage, and a systematic point sample (Figure 3) is
others. Counting and interviewing every kth shopper to pass through a super- therefore often used. This method is easier to use in the field than random
market checkout is a difficult task for this reason, and there will certainly sampling, because the field-worker can move easily from one point where he
have to be someone counting as well as sufficient people interviewing. This has to examine vegetation or soil to the next, perhaps using chains to measure
will inevitably mean that field-workers will be standing around for a large the distance. Just as in sampling from lists, though, there is a danger that
proportion of the time, and what seems superficially a straightforward and the interval between the points will correspond to some periodicity in the
economical way of sampling is in fact less practical. It may be necessary to data, as it might with land forms, or man-made landscapes, or, within a field,
interview half as many in the busiest time periods, for example, and then the spacing of crops (Zarkovich, 1966).
weight them accordingly.
Many of the methods of combining random and systematic sampling can be
Another problem with sampling sequential flows is that what is being adapted for the spatial situation. If the area is divided into parts, a quota
sampled is passage past a point or entry onto a list. If a researcher was can be set for each which, when exceeded , leads to rejection and the selection
really interested in users of a particular facility, rather than usage, the of other points. Alternatively, a point can be selected at random in each
visits might have to be weighted to represent the users. The chance of selec- square of a regular grid (Figure 4).
tion of an individual would depend on the frequency of his use of the facility,
so the results should be weighted by the reciprocal of the frequency of use Berry and Baker (1968) advocate a stratified systematic unaligned samp-
to produce an unbiased estimate of the users. This means that the results for ling method (Figure 5), which utilises grid squares each with its own internal
someone who uses the facility fortnightly should be given double the weight grid co-ordinates. We start by drawing two random co-ordinates, a and b,
of those of a weekly user. within the first square. The marginal grid squares along the top row are then
allocated points by drawing a random y co-ordinate within each square and
There is also the general problem that the survey may cover an un- using the x co-ordinate, a, from the first square. The y co-ordinate, b, is
representative time and may need to be repeated (section V, iii); different used with a random x co-ordinate in each grid square of the marginal column.
time periods might be treated as strata (section II, iv). Each subsequent square's sample point is determined by the y co-ordinate of
its column and the x co-ordinate of its row.

32 33
 While this method is not strictly one in which each point is selected
independently, for all practical purposes it can be assumed to be a stratified
sample with independent selection within the strata (the grid squares); Figure 7 shows the simplest form of random transect sample, in which
Berry and Baker have demonstrated the successful use of this form of sample lines are drawn across the area between pairs o random co-
f
ordinates. It is
for land use proportions. assumed that the area of phenomena will be in proportion to the length of
line crossing them. The percentage of each land use can therefore be simply
If part of the area, for example a soil or land use type, is very small assessed by measuring distances along the transects.
or variable, and is to be compared with a larger and more uniform area,
stratification with different sampling rates may be used. A larger quota of For a systematic transect sample, the transects are arranged across the
random numbers would be allocated to the smaller or more variable stratum. study area, usually as a set of parallel lines. Purposively located transects
are also commonly used; rather than using probability sampling the transects
The quickest way to draw the sample may be to take random numbers within are sited by the researcher across the contours, or to correspond with an
grid squares. Approximate demarcation between the strata may be adequate; 'environmental gradient'. If we were to position our transects at right angles
squares could be allocated to the more variable stratum if they contain more to the sea, at regular intervals, we would not have a probability sample, and
than 50 per cent of the more variable feature. This misallocation of the would have to make this clear in discussion of the techniques of our project.
fringes may not matter if the grid squares are small relative to the size However, along these transects we might use a random or systematic sampling
of the strata (Figure 6). procedure, treating the initial transects as sampling frames. Transects have
the advantage that they can be illustrated readily by a cross-sectional dia-
gram or a series of diagrams showing such features as topography, land use
or vegetation along their length.

34 35
 n

Fig. 8 Estimating Species Density By The Method of

Cottam et al (1953)

All these spatial samples are easily applied to maps, but are more
difficult in the field. Trying to identify points on the ground, or to follow
straight lines, may be rendered impossible even without such obstacles as
buildings, termite mounds or impassable boundaries. The samples normally used
in the field are systematic rather than random, because it is convenient to
relate each observation point to the last.

Biogeographers in particular make considerable use of quadrat sampling,

which is not strictly a sampling method, but a data collection method at
points chosen by one of the areal methods discussed. A quadrat is a rigid
frame of fixed area and shape placed on the ground, for example with one cor-
ner touching the chosen point; the vegetation, soil or crop within the frame
is studied.

The shape and size of the quadrat will to some extent determine the re-
sults. There seems no generally agreed improvement on the square quadrat,
but the size has to be chosen for the particular features being studied.
Examples and advice on choosing suitable sizes are given in a number of refer-
ences (Peltier, 1962; Haggett, 1965, pp.198-199; Shimwell, 1971; Kershaw,
1973). A large number of small quadrats seems better than a small number of
large ones, because they can be more spread out, but they have longer edges
and therefore create more problems of determining inclusion or exclusions.
Shimwell (1971, p.17) discusses the question of the 'minimal area needed for
vegetation studies, the smallest area in which all species in the area will
be found, and which is thus 'large enough to represent the characteristic
structure and floristics of a plant community'.

Extra problems arise in the use of quadrats for crops, because if a field
is planted in rows, it might be possible to place the quadrat in such a way
as to contain either x rows, or x + 1. The solution clearly is to take several
randomly aligned quadrats from a field. A decision has to be made about the
border of the field, for the crop may be sparser at the margin. If quadrats

37
Fig. 7 Random Transect Sample
which overlap the boundary are excluded, there will be a biased over-estimate One mistake is to use a spatial sample for a phenomenon which is not
of yield (Zarkovich, 1966, pp.331-365). continuous. If random points were located, and the sample were then to con-
sist of the nearest oak tree, or the closest farm building, then the method
Quadrat measurements can be taken at randomly or systematically located would be biased in favour of the more scattered, isolated elements.
points, or along transects. They cannot legitimately be located by throwing
objects at random', for this method will never approximate to a true random If we were to include elements if a point fell on them, buildings or
sample. Even if a field-worker is genuinely concerned not to direct his throw, farms with large ground areas might be more likely to be included. The best
the vegetation height may affect the point where the object falls, and points method for phenomena such as these is to list (or mark) all the elements.
around the border of the study area might also be under-represented (partic- For economy, we could do this in parts of the study area chosen by a probabil-
ularly as a long throw may make the object hard to find!). ity method, and sample within these.

Not all vegetation studies use quadrats; there are alternative methods Occasionally, where phenomena are relatively evenly spread, as in the
which have particular applications. For example, the contact between species case of shops in a shopping centre or factories on an industrial estate,
can be determined by sampling points by any of the usual methods, recording elements are allocated to a single dimensionless point. This might be their
the plant at each point, and any plants touching it; this produces an southwest corner, or their central point. The nearest one to each selected
'association matrix' between species (Yarranton, 1966) although it ignores sampling point in a given direction (and perhaps within a specified distance)
association in which plants do not 'touch'. Recording the distance from the might then be chosen. The larger units would have a higher chance of selection
sample point to the nearest plant of a particular species in each of four by this method, but if their size does not vary greatly, and a speedy selec-
'quarters' (Figure 8) gives an indication of the density of the species and tion is required, the method might sometimes be justifiable.
may be quicker than estimating coverage by a large random sample (Cottam
et al., 1953). (iii) Sampling in space and time

In the field, particularly when a quick sampling method is required in Sampling methods are sometimes used to locate points in time and space
open country, use has been made of the random walk to find sampling points, for traffic counts or stream flows, or measures of a variety of phenomena
often points at which to take quadrat measurements, or as an alternative to which are to be repeated.
transects. From a random start, the observer moves a set distance (or number Traffic counts are usually performed at purposively selected sites, for
of paces) in a direction randomly determined for a predetermined number of example on all major roads leading into a settlement, but occasionally they
'legs'. Such a method might give very poor coverage. The use of random walks may be evenly spaced every k kilometres. Yates (1960, pp.363-364) discusses
in urban areas is discussed in section VI, iii. this use of sampling.
(ii) Problems of spatial frames Occasionally it is necessary to divide the measurement period in order
to cover a larger number of sites than there are field-workers. This might
The greatest problems with spatial sampling are concerned with access. occur when studying atmospheric pollution at a variety of locations over a
There is a temptation to ignore areas which have the most difficult terrain, week (Haggett, 1975, pp.541-543).or when interviewing people using a number
or to place transects across accessible points. Coastal vegetation studies of public libraries during different opening shifts (SCPR, 1973, pp.78-80).
have sometimes been limited to points at which landings could feasibly be In both cases a grid would be drawn up which ensured that each point was
made. These are not then probability samples, and the researcher would have sampled the same number of times, at different time periods, for example,
to use his judgement to guess how different the inaccessible areas might be. in the library case, in both morning, afternoon and evening equally.

Similarly, a rigorous sample might be drawn from a very incomplete frame. Distinct time periods may be regarded as strata and sampled accordingly.
The target population might include all the areas with a particular soil type, However, when a researcher divides his sample between different times and
but the sample might be restricted to areas not built upon. The research re- places, he is in effect creating a clustered sample design. As was explained
port would have to indicate such deficiencies, and likely inadequacies would in chapter III, such a sample probably only includes part of the variability
need to be discussed. Rigorous sampling within the accessible areas could not of the population as a whole. In order to counteract adverse effect, the
compensate for the initial lack of a probability sample, but as a case study largest possible number of small clusters (sampling points and times) should
the work would not be invalidated. be used as will be economic.

Interaction between the researcher and the study area can be detrimental, Geographers might perhaps give more attention to choice of sampling
for example trampling across an area undoubtedly effects its vegetation. The time, because both time of day and season will affect certain kinds of data.
problem might not be too severe the first time an area is studied, but if the
field-work is to be repeated later, there could be problems analogous to the
'polluting', 'conditioning' or mortality of a human population. Placing
markers to indicate points to be returned to is particularly likely to affect
the vegetation, especially if the markers are posts on which birds may perch.
In such cases, sample points can be a set distance from their markers.

38 39
 In the interrelated form, the interviewer is given specific categories
VI NON-PROBABILITY SAMPLES of people to find, and told how many of each to interview. For example, a
quota sample of young people might control for occupation and sex. The quota
(i) Purposive samples would then consist of a specified number of young unemployed males, young
unemployed females, young workers of each sex, and young people continuing
When a geographer selects a 'typical' example, or set of examples, from their education of each sex, giving six separate categories with a target
the total population he purports to study, this sample cannot be analysed as number for each. This would prevent the interviewer confining his attention
if it were a probability sample. No matter how carefully chosen, it clearly to the most accessible young people, for example by interviewing a large
cannot conform to the requirement of probability sampling that each element number of unemployed or school pupils.
has a known probability of selection, and therefore estimates of standard
error and confidence limits cannot be produced. When the controls are independent, the interviewer would be told simply
how many of each sex, and how many in each occupation category to interview,
In fact, the purposive sample drawn in this way may well contain the but not how many in each of the six categories. It would therefore be possible
extreme cases, or those which best illustrate the point the researcher wishes for one category, for example, working girls, to be omitted altogether while
to make. fulfilling the marginal totals. The interrelated form is therefore often
Within purposively chosen samples it is possible to use probability preferable, but it depends on the researcher knowing how many of each category
methods. Case studies of particular places are of this type. Transects of there are in the population, and the particular variables he is using may not
vegetation or topography can be purposively located, but probability samples be cross-tabulated anywhere.
drawn within them; data collected in origin and destination surveys, taken
at deliberately selected points where the work can be undertaken, may be It may be possible to determine quotas appropriate to a particular study,
analysed quantitatively. In these cases, the researcher would make it clear and indeed a researcher may feel confident that in doing so nothing is lost
exactly what the limitations to the study were. by this method. It is also possible to set quotas which over-sample particular-
ly variable parts of the population, in the way that particular strata can be
(ii) Quota samples over-sampled in probability methods.

If no convenient sampling frame exists, and costs or time preclude the When an interviewer has to approach individuals in the street, there is
creation of one, it may be permissible to include elements so long as they a great temptation to omit certain people altogether. It is difficult to
have certain specified characteristics of the target population. If the pro- approach someone and ask them the questions necessary to establish if they
portions in which some features occur in the target population are known, are in the quota; asking age and occupation does not usually start off an
these features are used as quota controls, and elements are selected by the interview well. Similarly, it requires a certain amount of tact to refuse to
most convenient method until the number in each quota is complete. interview someone who does not happen to fit into a required category, and
there may even be a temptation to squeeze a co-operative but ineligible person
The method is usually used in samples of people, and particularly in into an adjacent classification. Most interviewers are reluctant to approach
market research, but could be used for other phenomena. Quota samples depend someone and then reject them, and consequently will approach people who look
on accurate knowledge of the target population. It is known, for example, like the stereotypes of the set categories. For example, if age groups are
from the Census how the population of Britain (or of particular areas of the given, the interviewers are likely to produce a sample which clusters in the
country) divides into age and sex groups. If the variables in which the re- middle of the age ranges, although this is less likely when the quota con-
search project is interested are highly correlated with these variables whose trols are not linked. If interviewing is conducted door-to-door , the problem
distribution is known, a sample in which they are represented in the correct of excluding people not on the street is overcome, but the sample may instead
proportions can be expected to resemble the population. However, we are be biased towards those at home, such as housewives or the housebound. Bias
usually interested in variables that are not highly correlated with the quota ca never be removed from quota samples, but their greatest drawback is that
n

control variables, and we can rarely assume this correlation; therefore the we can never know how severe it is in any given sample.
method can lead to a very misleading sample.
(iii) Random routes
The bias stemming from interviewer or respondent selection could be
considerable. If a researcher was interested in users of a particular central When there is no sampling frame, and the researcher wants to sample
facility who came from a particular part of the town, he could approach people households or buildings of some kind, it may be quick and cheap to sample by
using the facility and ask them where they came from. If the resultant sample means of a set of rules governing a walk around the study area.
consisted wholly of people under 30 he would have no way of knowing if this
was typical of users of the place, or if he had chosen an unrepresentative A random start is chosen, and then instructions are followed, for example,
time of day to carry out the survey, or if he had failed to approach all sorts first right, first left, first right, calling at addresses at a fixed interval
of people. along the way.

Quotas are established in one of two forms, depending on whether the

variables used as controls are given in an independent or interrelated form.

40 41
 Instructions should be sufficiently unambiguous for two people following (vi) Justifying non-probability samples
the same route to make the same decisions about where to turn and which
addresses to count, but in practice seldom are. Decisions will be needed about If a probability sample can be used, in other words, if a reasonable
maisonettes, tower blocks, multiple-occupied dwellings and cul-de-sacs. The sampling frame exists or can be created, it is hard to justify using a non-
technique is attractive in that it dispenses with a sampling frame, but it probability method. It is impossible to establish the absence or extent of
is open to misapplication, and the creation of a sampling frame would be the bias in these methods, and the statistical tests applied to probability
only way to ensure that buildings obscured from the road, or houses containing samples are invalid.
more than one address, were included.
Their speed, cheapness and the ability to dispense with a sampling
The selected addresses should be treated as any other sample, and if frame may occasionally make their use essential, and they may be justified
there is no reply, they will have to be visited again. Random walks depend in pilot or exploratory work. Paradoxically, though, they rely on experience
very much on the integrity of the practitioner, and are at best quasi- and a certain amount of knowledge of the target population - which the re-
probability samples. searcher probably lacks when he is embarking on a pilot study. Whenever
possible, probability methods should be used.
(iv) Combining quota sampling with a random route

To ensure that the initial selection of respondents is relatively un-

biased, and that if a sample is to represent residents of a place, they are VII RESEARCH DESIGN AND THE CHOICE OF SAMPLING METHOD
not all contacted in one part of it, interviewers can be sent to randomly
located starting points and told to follow random route instructions, approach- The cost of using a sampling method is a combination of the cost of
ing households at a fixed interval, and interviewing until the quotas are collecting the measurements, and of the frame; the time taken is similarly
full. a combination of the time spent in the field and the time spent beforehand
obtaining the frame and drawing the sample. It is these costs and time limits,
The interviewer's honesty is greatly strained, particularly if told to as well as constraints at the analysis stage, which will largely determine
call at every kth house, because if the house next door seems more likely to the sample size and the methods which are feasible.
fulfil the quota, it will be very tempting to include it. Permitting the
interviewer to call at every house along the random route is preferable, and After fixing the relevant population and properties of it which are to
i f the quota is for a comparatively rare trait, then the level of clustering be studied, and the method of measuring these characteristics (interviews,
may not be too great; after a set number of eligibles have been found, the observations, measurements), we have to decide the approximate size of sample,
interviewer can be told to move to a new point. the number of units, which can be studied. This size can be increased by re-
ducing costs or time, usually by clustering, but this will increase the samp-
The method suffers from the disadvantage of quota sampling in general, ling error.
that those who are not in the right place at the right time never get in-
cluded. It might have applications for immobile populations such as morpholog- Conversely, stratification may be more costly, and lead to a reduced
ical features or buildings. sample size, but to an increase in precision which is effectively equivalent
to an increase in sample size. It is possible to cost ways of reducing the
(v) Snowball samples sampling error to see which are feasible.
One way of creating an approximation to a sampling frame for a rare Therefore, the initial decision about sample size may be modified as
trait in human populations is to ask a number of initial contacts for the the research design proceeds, but it will be a major consideration and a rough
names of any other eligible individuals, who are then approached and asked approximation will be needed from the outset.
the same question. This has been successfully used for deaf people, or players
of particular sports. Ideally the initial contacts should be randomly located, Figure 9 shows a flow diagram for choosing between probability and non-
and the trait should be one which leads to a high degree of contact. The probability methods. It is only after discovering the absence of a sampling
'snowball' should be allowed to grow until each respondent is producing a frame, and deciding that it is not feasible to amend any existing frame or
list which consists largely of those already identified, but even then there create a new one, that non-probability methods should be resorted to. Having
is no guarantee that there is not a substantial number of isolated indivi- established a sampling frame, decisions about the form of probability sample
duals, or a self-contained group not in touch with those found by the snowball. are set out in Figure 10. There will be applications for all five methods at
the bottom of the chart, and their use will be largely determined by the
Snowball sampling has limited applications, but where a researcher felt frame. A degree of clustering may be needed to reduce costs, and can be com-
confident in the enclosed nature of the rare attribute qualifying for in- bined with other methods. Complex sample designs are harder to use, and cal-
clusion in the sample, and no sampling frame was available, it might be the culations of standard errors are made more difficult. In small scale studies
only feasible method. elaborate sample designs will rarely be needed.

42 43
 Fig. 10 The Process of Sample Design: Choosing between probability methods

Fig. 9 The Process of Sample Design: probability and non-probability methods

 Shimwell, D. (1971) The description and classification of vegetation.
The greater the knowledge the researcher has of the population he is to
(London: Sidgwick and Jackson), 4-23 - particularly quadrats,
study, and of the frame from which he is to draw the sample, the better able
transects and 'plotless sampling'.
he is to take decisions about the research design. But having made the various
choices, it is necessary to set out rigorous rules of procedure and follow
E. Examples of areal sampling
them, and finally, in any research report, to state exactly what they were,
and if there were any problems with them. The simplest and smallest piece
Berry, B. and Baker, A. (1968) Geographic sampling. in: Spatial analysis,
of research needs to have its methods clearly stated to enable anyone reading
B. Berry and D. Marble, (New Jersey: Prentice-Hall).
it to understand its strengths and weaknesses.
Cottam, G., Curtis, J. and Hale, B. (1953) Some sampling characteristics of
a population of randomly dispersed individuals. Ecology, 34,
741-757.
BIBLIOGRAPHY
Peltier, L. (1962) Area sampling for terrain analysis. Professional
Geographer, 14(2), 24-28.
A. Basic texts on sampling
Yarranton, G. (1966) A plotless method of sampling vegetation. Journal of
Moser, C. and Kalton, G. (1971) Survey methods in social investigation. Ecology, 54, 229-237.
(London: Heinemann), 61-210.
Zarkovich, S. (1966) Quality of statistical data. (Rome: Food and
Social and Community Planning Research, SCPR (1973) Sample design and Agriculture Organisation of the United Nations).
selection. (Technical Manual No.2).
F. Sampling in space and time
Stuart, A. (1962) Basic ideas of scientific sampling. (London: Griffin).
Haggett, P. (1975) Geography: a modern synthesis. (New York: Harper
B. More Advanced texts
and Row) 2nd edn. 541-543.
Kish, L. (1965) survey sampling. (New York: Wiley).
G. Quota sampling
Yates, F. (1960) Sampling methods for censuses and surveys.
(London: Griffin). Moser, C. and Kalton, G. (1971) Survey methods in social investigation.
(London: Heinemann), 127-137.
Cochran, W. (1963) Sampling techniques. (New York: Wiley).
Stephan, F. and McCarthy, P. (1963) Sampling opinions. (New York: Wiley).
C. Sampling frames for human geography in Britain

Gray, S. (1971) The Electoral Register. Government Social Survey M151,

(London: HMSO).
Pickett, K. (1974) sources of official data. (London: Longmans),
115-131.
Rowley, G. and Shepherd, P. (1976) A source of elementary spatial data for
town centre research in Britain. Area, 8(3), 201-208.
Davies, W., Giggs, J. and Herbert, D. (1968) Directories, rate books and the
commercial structure of towns. Geography, 53(1), 41-54.
Oliver, J. (1964) Directories and their use in geographical enquiry.
Geography, 49(4).

D. Introductions to areal sampling

Chorley, R. (1966) The application of statistical methods. in: Essays in

Geomorphology, G. Dury, (London: Heinemann), 275-387,
especially 292-307.
Haggett, P. (1965) Locational analysis in human geography. (London:
Edward Arnold), chapter 7.
Kershaw, K. (1973) Quantitative and dynamic plant ecology. (London:
Edward Arnold), 28-37 - particularly deals with quadrats and
transects.
46 47