Our Inaugural Cycle Deconstructed: Many Great Calls and a Few Blunders

The Phillips Academy Poll—America’s first high-school-led public opinion poll—concluded its inaugural year with many great calls, as well as a few blunders. This article will deconstruct what went right and what went wrong.

In the three weeks preceding the November 8th election, The Phillips Academy Poll surveyed likely voters in Nevada, Arizona, and New Hampshire. Prior to the December 6th Georgia Senate runoff, we also polled likely voters in Georgia. All four polls were conducted with text messages to mobile phones, while the polls in Arizona, New Hampshire, and Georgia also featured interactive voice response calls to landline phones. All four polls were subsequently weighted via iterative proportional fitting.

Phillips Academy achieved a 100% record for polling the Senate on Election Day, though we later incorrectly projected a narrow win for Herschel Walker in the Georgia Senate runoff. Nevertheless, Phillips Academy’s Senate polls were among the most accurate of the cycle.

Nevada, our most accurate Senate poll, found Catherine Cortez Masto leading by 2%—she actually won by just under 1%. Election Day results demonstrated that our most accurate prediction was also the most unexpected; most other polls, polling averages, and experts predicted that Republicans would flip the state. We attribute the accuracy in Nevada to the use of text messaging, which allowed us to better sample younger voters.

Georgia’s senate runoff, our only senate poll that misidentified the eventual winner, was 3.6% off the mark. However, our Georgia poll allowed respondents to indicate that they were undecided—an option that was chosen by 5% of respondents. With the razor-thin margins, it is likely that these undecided voters tipped the balance, as both Walker and Warnock reached the percentages our poll indicated based on decided voters.

Phillips Academy maintained a less accurate record in races for governor. In Nevada, we incorrectly found Steve Sisolak winning by 5%, when in actuality Joe Lombardo eked out a narrow victory. Most problematic, however, was the Arizona poll that found Kari Lake winning by 11%, when, on Election Day, she lost a close race.

These blunders in Nevada and Arizona can be attributed to our weighing process, which did not account for income or urbanity for these specific two polls, as was indicated by those polls’ methodology statements. For these two polls, we attempted to impute data on income and urbanity rather than asking explicitly. However, due to missing data for some respondents, the imputed data were often unreliable, and as a result, we did not incorporate this information into our raking process or likely voter model.

This was not initially thought to be a major issue, as many pollsters do not weigh by income and urbanity. However, as a result of having fewer weighting parameters, the remaining parameters—including party affiliation—became greater in prominence. In races for governor, party affiliation is less correlated with voting patterns. Since these races have no national implications, crossing party lines becomes more common. As a result, the underemphasis on income and urbanicity, and consequential overemphasis on party affiliation, were detrimental to our accuracy in the Nevada and Arizona gubernatorial polls.

All in all, Phillips Academy performed respectably, particularly with regard to our Senate predictions. For Senate polls, our average error was 3.6%—slightly higher than the 3.4% average error of the FiveThirtyEight polling average for those same races, though less than the 4.5% average error of RealClearPolitics. When also factoring in gubernatorial races, between Phillips Academy, FiveThirtyEight, and RealClearPolitics, the average error became 5.5%, 2.6%, and 3.6% respectively.

The Phillips Academy Poll is honored to have informed 2022 election discussions and analyses across a range of media outlets, including MSNBC, Fox News, BBC News, Politico, Newsweek, and The Guardian. We intend to address the aforementioned imperfections in our polling practices in future polls throughout the 2023 and 2024 election cycles.

More Articles