If you want your samples to be unbiased, you need to be aware of how
you generate your samples. Biases can be introduced because of the interaction
between the survey participants or units and the person conducting the
survey. Possible sources of error in surveying include:
response errors (e.g. people might lie about their age, their weight,
how many cigarettes they smoke, how much alcohol they drank last week,
and so on);
missing data (e.g. you might be unable to contact a subject in your
study);
the effect that the wording of questions has on responses; and
the effect that the interviewer might have on participant responses.
Simple random sampling (SRS) is an attempt to ensure that the samples
are more representative. However, mistakes can be made with simple random
sampling, as the following story shows [4].
SCENARIO
The Literary Digest, a weekly newsmagazine in the US, sent mail surveys to 10 million households to develop its opinion
poll for the 1936 presidential election. The addresses were obtained from databases of car owners and households with
telephones. As a result of its surveys, the Digest predicted the election would be an overwhelming victory for the Republican
candidate, Alfred Landon. However, Democrat candidate, Franklin Roosevelt won in a landslide.
SCENARIO
Below is an account of the process that another newspaper
used to obtain a representative sample for a poll it conducted
[5]. As you
read this account, think about the effort involved .
The poll was based on telephone interviews conducted over
3 days in all parts of the city. A sample of 900 adults was
achieved. The interviews were conducted in the two most common
languages spoken in households. The sample of telephone exchanges
called was selected by computer from a complete list of city
exchanges. The exchanges were chosen to ensure that each area
in the city was represented in proportion to its population.
For each exchange, the telephone numbers were formed by random
digits, which ensured that both listed and unlisted numbers
were selected. These numbers were screened to ensure that
only households (not businesses) were chosen. The results
were weighted to take into account household size and the
number of telephones in the household, and also to adjust
for variations relating to factors such as suburb, sex, age,
education and race.
