Phone Calls and Formulas

How call-center phone calls were wrangled into data models and micro-targeting tools...

In the early years of doing call-center-based voter contact in Canada, clients often asked, “How am I doing?” It’s not always an easy question to answer. But in seeking a solution the data model we developed eventually evolved into a micro-targeting platform.

Strict spending limits mean few of our district-level campaigns have access to professional polling based on valid statistical samples of the population. Campaigns might commission a pre-election poll but rarely had district-specific nightly tracking during the election. And while we conducted voter identification with a large number of observations—up to 100,000 per night—our sample was not truly random, as required by polling. Our challenge was to create a model from the most granular data, one that would allow us to make an inference about the whole electorate—and ultimately, to predict whether our clients would win or lose their elections.

The First Model
Our experience showed that if a certain percentage of voters indicated they supported the candidate and the Liberal Party (we work exclusively for Liberals), the candidate was very likely to win. Though this heuristic worked well, it could be confounded by three-way races or electoral districts with unusually high or low voter turnout. Clients could bias the sample of voters we were calling, too: they ranked the order that we called their precincts and could prioritize calling past supporters or even reserve these past supporters for contact by their own volunteers. These factors impacted the sample and were confounding in close races.

We started working on a solution in advance of the 2003 Ontario General Election, when we would be working on behalf of 60 candidates. We developed a model that assigned every voter—about 7.5 million total—a probability, from 0 – 1.0, that we expected them to vote Liberal. These were the expected values. We then took the range of possible answers to the VoterID script and converted these to more probabilities. These were the actual values. At the end of each day we took the sample and compared the expected to the actual values, making a determination of whether we were doing better (actual > expected) or worse (actual < expected). We then reviewed the previous election results and made predictions about whether we would win or lose the district.

This first iteration of the model worked. We immediately noticed sharp spikes in support in some unexpected electoral districts and the model ultimately predicted the result of the election quite closely.

Refining the Model
Despite this early success, the model needed further refinement. I had made up the table converting voters’ answers (“I am undecided at this time,” for example) to numerical probabilities based on experience rather than statistics. The model allowed us to tell clients they were doing better or worse than expected but not how much better or worse. We also felt that we could create two attributes to describe voters: a probability of voting and a probability of voting Liberal. Finally, we wanted to validate our results to see how closely they tracked with actual voting patterns on Election Day.

We turned to Matt Lebo, a Canadian teaching political science at Stony Brook University on Long Island, N.Y., and his colleague Andrew Sidman, a recent Ph.D. graduate embarking on a career at City College in New York. Lebo has published a volume of work on electoral modeling and has studied elections in the United States and the United Kingdom.

Lebo and Sidman first validated the ten years of data we had collected, which showed a high degree of correlation between our results and actual election results both on the precinct- and on electoral-district levels. The pair then re-created the coefficients for voters’ answers using more advanced statistical techniques than we had. Lastly, they described every Canadian voter in terms of a probability of voting and a probability of voting Liberal. We then put the model to work in the 2007 Ontario and 2008 federal general elections, each day comparing expected and actual results. This more complex model proved even more successful in predicting outcomes.

The model has also become a micro-targeting tool. We can display a graph that plots every voter in an electoral district according to their voting behavior. That graphs helps identify core supporters, supporters who vote infrequently, swing voters, supporters of other parties and non-participants. In the 2008 election we targeted infrequent voters with GOTV calls during the advance polls and targeted swing voters with persuasion calls.

We can also display a matrix showing the number of voters in each of these groups, helping our clients see how to build a winning plurality: Do I need to activate my core vote, or do I need to recover voters who moved to other parties? We expect increased interest in this type of targeting following the success of the Obama campaign in “re-shaping the electorate” and convincing traditional non-participants to vote in 2008.


Mike O’Neill is the president of First Contact, a Canadian company that specializes in voter-contact management for Liberal clients and First Voter Contact, a U.S. company serving Democrats. O’Neill has been active in politics for 20 years as a staff member, campaign manager and service provider and has been a frequent speaker at campaign training events.