Comp. by: VPugazhenthi Stage : Revises1 ChapterID: 9781405132879_4_S Date:1/4/
09 Time:15:23:34 Filepath:H:/00_Blackwell/00_3B2/Gregory-9781405132879/appln/3B2/
revises/9781405132879_4_S.3d
. Coherence: the degree that the measured
entity conforms to the theoretical con-
struct being studied. This is particularly
important in qualitative research when
often necessarily small samples need to be
theoretically relevant (Mitchell, 1983);
consequently, for example, we need to de-
fine and sample the different mechanisms
that place people in poverty (exclusion
from the job market, divorce and separ-
ation, etc.) and make sure that each type
is recognized and studied.
. Bias: the degree to which the parameter is
accurately estimated, without systematic
error; for example aninternet-based sam-
pling strategy may seriously underestimate
the extent of poverty.
. Precision: the degree to which the parameter
is reliably measured and random, stochastic,
innumerable small errors are controlled.
. Efficiency and cost: efficiency is the relative
precision of an unbiased sample compared
to others of the same size. Concern with
precision can often be over-ridden by con-
venience, practicality and cost; moreover,
it is generally much better to do a well-
designed small-scale study than a botched
large-scale one.
A highly convenient sample design (often used
by commercial organizations) is the quota
method, in which, for example, an interviewer
approaches people on a shopping street until
the desired quota of 25 each of young and old
men and women is reached. This approach is
simple but also open to bias, as certain types of
people are not generally found in shopping
streets; there can also be interviewer bias, as
the interviewer may be resistant to approach-
ing certain groups of individuals. Non-
response is inevitably ignored in this approach.
Snowball samplingis when having found a key
informant (e.g. a drug user) they are asked to
recommend others. While useful with hard-
to-reach populations, there is again the poten-
tial for bias as isolated drug users may be
missed and one circle of users may not inter-
mix with others. Asystematic sampleis when
respondents are chosen in an ordered way,
such as every fourth house on a street. Such
a design is highly convenient in an the field
and when the total population is not known,
but can produce biased results when the entity
being studied has a corresponding systematic
patterning, so that all even-numbered houses
on one side of a street are social housing,
but all odd-numbered houses are owner
occupiers.All the designs discussed so far are non-
probabilistic. Inprobabilistic designsthe key
feature is that neither the interviewer nor
the interviewed can affect the selection mech-
anism, which is done at random. With such
samples, the likelihood of being sampled is
knowable and non-zero; consequently we can
use statistical theory (based on the central
limit theorem) to guarantee unbiased, repre-
sentative estimates and to estimate the degree
of precision in those estimates. Thus we can
say that the proportion in poverty is 25 per
cent and that we can be 95 per cent confident,
given our sample of 7,200 households, that
the true underlying rate lies between 24 and
26 per cent. Non-probabilistic sampling is not
necessarily biased and unrepresentative, but
we lack the necessary formal framework for
making any judgement.
There are three basic types of probabilistic
sample:. Simple random sample(SRS) requires a
complete listing of the population (the
sampling frame) from which a sample is
chosen at random so that each and every
unit has an equal chance of being selected.
With such EPSEM sampling (equal prob-
ability of selection method) the standard
error, which defines the precision of the
sample estimate, is proportional to the
square root of the absolute sample size.
Consequently, the larger the sample (in
absolute terms, not as a percentage of the
population) the greater the precision but
there are diminishing returns, as the sam-
ple size must be quadrupled to halve the
standard error. This can be an expensive
design as in a national survey the inter-
viewers will be required to travel the entire
country. In practice, quasi-random sam-
ples are often used; for example, the British
national birthcohortsof 1946, 1958 and
1970 were based on babies born in a par-
ticular week, while the ONS Longitudinal
Study uses record linkage of individuals
born on four days of a year, which equates
to some 1 per cent of the national popula-
tion. Indeed, providing there is no period-
icity in the sampled variable, systematic
sampling can be treated as quasi-random.
With probabilistic sampling, an effective
sample size of 10,000 respondents is
needed in order to be 95 per cent confident
of being+1 per cent of the true value,
when that is 0.5. Typically, scientifically
credible national opinion polls contain
around 1,500 respondents. In student and
Gregory / The Dictionary of Human Geography 9781405132879_4_S Final Proof page 662 1.4.2009 3:23pmSAMPLING