QUALITATIVE AND QUANTITATIVE SAMPLINGEXAMPLE BOX 6
Example Sample, the 2006 General Social SurveySampling has many terms for the different types
of samples. A complex sample illustrates how
researchers use them. We can look at the 2006
sample for the best-known national U.S. survey in
sociology, the General Social Survey (GSS). It has
been conducted since 1972. Its sampling has been
updated several times over the years based on the
most sophisticated social science sampling tech-
niques to produce a representative population within
practical cost limits. The populationconsists of all res-
ident adults (18 years of age or older) in the United
States for the universeof all Americans. The target
populationconsists of all English- or Spanish-speak-
ing mentally competent adults who live in house-
holds but excludes people living in institutional
settings. The researchers used a complex multistage
area probability sample to the block or segment level.
At the block level, they used quota samplingwith
quotas based on gender, age, and employment sta-
tus. They selected equal numbers of men and
women as well as persons over and under 35 years
of age.
The sample design combined a cluster sample
and a stratified sample. U.S. territory was divided into
standard metropolitan statistical areas (SMSAs, a U.S.
Census Bureau classification) and nonmetropolitan
counties. The SMSAs and counties were stratified by
region, age, and race before selection. Researchers
adjusted clusters using probability proportionate to
size (PPS)based on the number of housing units in
each county or SMSA.
The sampling design had three basic stages.
Stage 1: Randomly select a “primary sampling unit”
(a U.S. census tract, a part of a SMSA, or a county)
from among the stratified “primary sampling units.”
Researchers also classified units by whether there
were stable mailing addresses in a geographic area
or others. Stage 2: Randomly select smaller geo-
graphic units (e.g., a census tract, parts of a county),
and Stage 3: Randomly select housing units on
blocks or similar geographic units. As a final stage,
researchers used the household as the sampling ele-
ment and randomly selected households from the
addresses in the block. After selecting an address, an
interviewer contacted the household and chose an
eligible respondent from it. The interviewer looked
at a quota selection table for possible respondents
and interviewed a type of respondent (e.g., second
oldest) based on the table. Interviewers used
computer-assisted personal interviewing (CAPI).
In the 2006 sample, researchers first identified
9,535 possible household addresses or locations.
However, this number dropped to 7,987 after they
eliminated vacant addresses and ones where no
one who spoke either English or Spanish lived. After
taking into account people who refused to partici-
pate, were too ill, were ineligible, or did not finish an
interview (23.3%), the final sample included 4,510
persons (for details, see http://publicdata.norc.org:
41000/gss/Documents/Codebook/A.pdf)size grows, the returns in accuracy for sample size
decrease.
In practical terms, this means for small popula-
tions (under 500), we need a large sampling ratio
(about 30 percent) or 150 people, while for large
populations (over 150,000), we can obtain equally
good accuracy with a smaller sampling ratio (1 per-
cent), and samples of about 1,500 can be equally
accurate, all things being the same. Notice that the
population of 150,000 is 30 times larger but the
sample is just 10 times larger. Turning to very large
populations (more than 10 million), we can achieve
accuracy with tiny sampling ratios (0.025 percent),or samples of about 2,500. The size of the population
ceases to be relevant once the sampling ratio is very
small, and samples of about 2,500 are as accurate for
populations of 200 million as for 10 million. These
are approximate sizes, and practical limitations (e.g.,
cost) also play a role.
A related principle is that for small samples, a
small increase in sample size produces a big gain in
accuracy. Equal increases in sample size produce
an increase in accuracy more for small than for large
samples. For example, an increase in sample size
from 50 to 100 reduces errors from 7.1 percent to
2.1 percent, but an increase from 1,000 to 2,000