In class we have discussed that online polls carry the most bias and are not the best way to conduct the most accurate poll. In this past presidential election, however, the results of the online polls were closer to the actual results than the results of phone polls, mail polls, etc. according to the New York Times. The Gallup poll consistently showed Romney ahead by six percentage points in late October. In their final poll they showed Romney ahead by one point. The final poll conducted by the Google Consumer Survey, however, showed Obama with a lead of 2.3 percentage points in the popular vote. His actual lead was 2.6 percentage points based on ballots counted through the first Saturday after election day.
The writers looked at nine polling firms that conducted their polls online, either wholly or partially. Their average error was 2.1 percentage points, while the average error for live telephone interviews was 3.5 points and 5 points for "robopolls."
The phone polls showed more of a republican bias than the online polls. One reason why this could have happened is the people the phone polls reach. If cell phones were not called as much as landlines, a whole group of people were ignored. People who depend on cell phones for communication are usually younger, more urban, make less financially, etc. Many democrats could have been missed by these polls.
Although online polls may present more bias, this election shows that the way people communicate has changed.
http://fivethirtyeight.blogs.nytimes.com/2012/11/10/which-polls-fared-best-and-worst-in-the-2012-presidential-race/
This is really interesting. Just to clarify, though, it is the sampling method (self-selection) used by many internet polls that makes them biased. So I was curious how these internet polls, such as Google Consumer Surveys, performs sampling. It took a little bit of digging, but I found a paper from Google that explains that they actually do use probability sampling. Calculating the probabilities requires inferring the demographics of respondents and re-weighting the responses. The reliability of this approach is not guaranteed, so their paper experimentally compares their results to other methods.
ReplyDeleteOther internet polls may have different methods for probability sampling.
http://www.google.com/insights/consumersurveys/static/consumer_surveys_whitepaper_v2.pdf