The Floating Sample method substantially reduces the cost and complexity of conducting systematic election observation based on statistical principles. It allows observer groups to provide representative data on the election process (the conduct and administration of the voting process at the polling stations), as well as draw accurate projection of electoral results with extremely narrow margins of error.
The Floating Sample approach builds upon the classic method of Quick Count (also known as Parallel Vote Tabulation, or PVT) first pioneered by the Philippine National Citizen Movement for Free Elections (NAMFREL) in 1986 and later adopted by the National Democratic Institute (NDI). The Quick count method is grounded mainly in the central limit theorem and the law of large numbers. By drawing a representative sample from an official list of polling stations, observer groups can then deploy observers who would monitor the process of the election and the posted vote count at each polling station. The data collected from the sample observers can then be used to make statistically grounded projections of voter turnout, election results per candidate or list, and the quality of election procedures’ implementation at the polling stations level.
The Quick Count method is tried and true. When done well, it produces extremely accurate results. However, it requires mobilizing a large number of observers and asking those observers to travel (sometimes great distances) to randomly selected polling stations. This not only risks disenfranchising a large number of (mostly young) citizens, but also costs more money in transportation, per diem, lodging, etc. Finally, it increases the risk of observer dropout.
The Floating Sample method imagines solving these challenges by drawing a sample of polling stations that are nearest to the district in which existing, accredited observers are registered and/or reside. Integrating artificial intelligence and statistical principles, IWATCH’s in-house expert Manel Ben Achour developed a unique algorithm to draw just such a sample. The algorithm works by taking the official polling stations database “A,” and the existing observers’ whereabouts database “B”, then creating a random sample of polling stations “C” and testing that against database B. If the match is not at least 12%, the algorithm would then create a new random sample “C” and test it again against database B. The process would automatically repeat until database B and C are as closely matched as possible or until 2 million iterations are completed. The final sample is floating above two databases A and B, hence the name the “Floating Sample”.
This method doesn’t guarantee a sample that is a 100% match to the locations of existing observers, but it substantially minimizes the number of observers who would need to be relocated according to the random sample, thus increasing the number of observers in the field and decreasing the costs to deploy them. This method, like the conventional Quick Count, does require reliable data-reporting tools to ensure timely data collection from the observers.
In order to ensure that the statistical underpinnings of the methodology were sound, IWATCH had to test if the integrity of random sampling principles would be jeopardized when only using the “random” sample that is most convenient for observers. Luckily, IWATCH had the opportunity to test this method during the legislative elections that were carried out one week before the presidential runoff of October 13. During the legislative elections, IWATCH tracked the voter turnout rate at two points throughout E-day: at 1pm and 6pm (the closure of polling stations). Comparing the official voter turnout numbers to the numbers produced by its observers, IWATCH was confident that going forward it would able to release election results projections with high degrees of statistical confidence.
Tunisia’s presidential runoff presented unique challenges that were perfect to showcase both the need for and the efficacy of the Floating Sample. It was the third electoral event for the country in less than 28 days, and election fatigue affected observers as well as voters. Many observers, most of whom are students, had to return to their universities, leaving many areas uncovered. Moreover, the deadline for submitting accreditation for new observers had already closed. This left most observer groups in the country limited to the observers they already had accredited, many of whom were not available for redeployment during the runoff. While other observer groups fell short due to their insufficient numbers of observers, the Floating Sample allowed IWATCH to use all of its existing observers, and more conveniently deploy them according to representative sampling. The algorithm had to run around 1,440,000 iterations to find the most convenient sample.
The difference between the projections published by IWATCH six hours after the closure of polling stations on e-day, and the official results announced the following day by the Independent High Authority for Elections (ISIE), was just 0.07 percent. This gap was the narrowest of all the projections made by exit poll agencies, and certainly of all observer groups who attempted the Quick Count that day.
In a style reminiscent of James Bond, à la 007, relying on the hard work and expertise of a dynamic young team, IWATCH was able yet again to innovate tools and methods to ensure a high quality of citizen oversight over the political process.
By Mustapha Ben Zine
Co-founding member, technical advisor.