Gordon S. Black is Chairman and CEO of
Harris Black International, Ltd., where George Terhanian is Director of Internet Research.
This article appeared in the Oct. 26, 1998,
edition of The
Polling Report.
Using the
Internet for
Election Forecasting
by Gordon S. Black and George
Terhanian
For more than 20 years, election polling in the United States and many other countries has
been conducted by telephone. And why not? Telephone polls, on the whole, have proved to be
remarkably accurate predictors of voter behavior -- the gold standard of all polling
research. Telephone polls are also far less expensive to conduct than door-to-door
interviews, the method they have largely replaced.
Although telephone polls possess many virtues,
they are not without their imperfections -- ones we do not like to mention too often or
too loudly to our clients. For instance, refusal rates today routinely exceed 40% of all
households. Unreachable respondents (due to traveling, working, answering machines, the
absence of telephones in college dorms, etc.) can regularly run another 30% of the sample.
The cost of telephone research can also be prohibitive -- making it virtually impossible
for one survey firm to cover every statewide race during an election year, not to mention
congressional races.
A 16-State Test
Drive
We at Harris Black International believe, however, that the strengths of the Internet as a
platform for conducting research more than compensate for the documented limitations of
the telephone, as well as the purported limitations of the Internet. This is one major
reason why we are now testing a new methodology of forecasting statewide elections via the
Internet.
Using an Internet database of more than
one million cooperative respondents nationwide, we are inviting individuals from 16 states
to complete three polls. The first is under way as this is being written, the week of Oct.
18. The second will be completed on the Sunday before the election. The third will be an
exit poll, the day of the election.
This effort is a combined venture with Excite,
Inc., one of the most popular Internet portals, and the political science faculty at the
University of Rochester. Excite is driving Internet traffic to our web site, where
respondents are invited to register to participate in the Harris Poll Online(sm) at
www.harrispollonline.com. Excite will also
display the results, as they become available.
The political science team from the University
of Rochester, led by Dr. Richard Niemi, will receive the entire database after the
experiment is complete. They will be able to use it for scientific evaluation and
publication.
Accurate
Election Forecasts
Conducting research on the Internet is by no means a simple or simple-minded exercise. In
the past six months, for example, we have had to overcome a variety of complex problems,
including software breakdowns and bugs, security problems, hardware failures, and so
forth. We feel a little like Chilton must have felt when they introduced the first
computerized telephone interviewing system back in the 1970s.
Our experience has convinced us, however, that
properly weighted findings from Internet studies compare favorably to findings from
telephone research. Nevertheless, we realize the validity of any polling method hinges on
its ability to reliably and accurately forecast voter behavior.
We also realize that many of our colleagues in
the polling community are rightfully skeptical of our efforts. To do what we are doing we
have had to set aside the staple of our industry -- the simple random sample. Virtually
every social scientist in America was educated on the power of random sampling.
Random sampling is a very powerful tool in
every avenue of science and industry for increasing the accuracy of estimates while
decreasing the cost of the process. It has been canonized by academics, and it is the
standard by which we judge research in most scientific fields. To describe something as a
"convenience sample" is to assign the research to the bottom of the ladder of
scientific quality and value.
We are not challenging the validity of random
sampling. In fact, we employ random sampling daily in our telephone polling and believe
that it is a wonderful statistical tool for those applications for which it is designed.
We are instead investigating whether findings from huge samples of Internet respondents,
coupled with sophisticated weighting processes, are as accurate as anything done on the
telephone or door-to-door.
We have decided to mount this investigation in
public for myriad reasons, including our desire to advance the collective understanding of
our colleagues in the polling and scientific communities who, at times, tend to dismiss
this methodology without any empirical evidence.
Campaign
2000
We are also preparing for the future. In the year 2000 election, the first presidential
election of the new millennium, we will possess a database of millions of cooperative
respondents, and we plan to use that database to cover the entire campaign, from the
primaries to the general election, beginning in January. Our coverage will include the
presidency, state offices, the U.S. Senate, and all congressional districts for which we
have adequate data. On election day 2000, we will conduct the largest -- by a factor of 15
-- "exit poll" ever conducted.
In a classic example of the impact of
scientific advancement, we expect to improve the performance while dropping the prices.
That is the vision that motivates this experiment in polling. |