Internet Polls

THE USE OF the internet to conduct surveys (including pre-election and other political polls) has grown rapidly over the past decade. The advantages of this mode of information exchange have been extensively documented: cost-effectiveness, large samples, the ability to administer intricate graphics and multimedia to respondents, and efficiency of data management. However, as web surveys have become more popular, concerns regarding sample representativeness and data quality have increased in importance.

When the cost of in-person interviewing became prohibitively expensive in the 1970s, and the cost of long distance telephone calls became cost-effective, a great deal of survey research shifted to random digit dialing (RDD) and telephone interviewing. The rising difficulty of reaching respondents due to answering machines and cell phones increased the costs of telephone interviewing significantly, and response rates and sample representativeness dropped. With the emergence of the internet, survey researchers have been intrigued with the possibility that this medium might permit cost-effective and quick data collection.

Willem Saris, professor at the University of Amsterdam, pioneered the possibility of using computers for data collection. He created the Dutch “Telepanel” in the 1980s, by recruiting a representative sample of Dutch citizens and giving them computers and modems. Each week, these panel members completed an electronic questionnaire and sent their responses in through their modems. A similar approach was adopted by the U.S. firm Knowledge Networks, and by firms in other countries (such as TNS Forsa in Germany). These data collection organizations initially recruit panelists using random sampling techniques such as RDD. Panelists then continue to complete questionnaires over an extended period of time in exchange for some form of compensation, such as free internet access.

Internet survey researchers generally eschew representative sampling, and collect data from volunteer samples. Respondents can be recruited into a panel through various means, such as banner ads on web pages and email address lists. For each survey, the demographics of the sample can be fine-tuned by imposing quotas at the time of data collection (for example, only half of the respondents can be female) and/or at the time of sampling (invitees can be chosen in numbers designed to produce a sample that resembles the population with regard to some demographic variable). In addition, post-stratification weighting can be implemented to yield the appearance of demographic representativeness. These methodological strategies can help ensure that the results of the survey approximate a random sample.

There are several potential sources of error in internet polls, particularly those that employ volunteer respondents. Because a substantial number of households do not have internet access, it is difficult to conduct general public surveys over the web, that tap the attitudes of the U.S. population as a whole. This produces what is known as coverage error. Coverage error can affect poll results because internet users are more likely white, wealthy, younger, and educated than those who do not have access to the internet. Firms can provide free internet access and computer hardware to respondents to mitigate coverage error. Additionally, there is no comprehensive list of email addresses that [Page 326]comprises the population (as there is with phone numbers and home addresses) and people may have multiple email addresses, complicating random sampling. Accordingly, it is impossible to calculate sampling error, or what is commonly referred to as the margin of error. Finally, as with telephone polls, internet surveys may suffer from non-response error, as many people who are asked to complete questionnaires refuse to do so, and these individuals may be different from those who choose to participate.

...