Looking through SSRN this morning, I came across a paper by David Nickerson (Notre Dame) and Todd Rogers (Harvard), "Political Campaigns and Big Data" (February 2014). It's a nice follow-up to yesterday's post about the software supporting new approaches to data analysis in Washington, DC.
In the paper, Nickerson and Rogers get into the math behind the statistical methods and supervised machine learning employed by political campaign analysts. They discuss the various types of predictive scores assigned to voters—responsiveness, behavior, and support—and the variety of data that analysts pull together to model and then target supporters and potential voters.
In the following excerpt, the authors explain how predictive scores are applied to maximize the value and efficiency of phone bank fundraising calls:
Campaigns use predictive scores to increase the efficiency of efforts to communicate with citizens. For example, professional fundraising phone banks typically charge $4 per completed call (often defined as reaching someone and getting through the entire script), regardless of how much is donated in the end. Suppose a campaign does not use predictive scores and finds that upon completion 17 of the call 60 percent give nothing, 20 percent give $10, 10 percent give $20, and 10 percent give $60. This works out to an average of $10 per completed call. Now assuming the campaign sampled a diverse pool of citizens for a wave of initial calls. It can then look through the voter database that includes all citizens it solicited for donations and all the donations it actually generated, along with other variables in the database such as past donation behavior, past volunteer activity, candidate support score, predicted household wealth, and Census-based neighborhood characteristics (Tam Cho and Gimpel 2007). It can then develop a fundraising behavior score that predicts the expected return for a call to a particular citizen. These scores are probabilistic, and of course it would be impossible to only call citizens who would donate $60, but large gains can quickly be realized. For instance, if a fundraising score eliminated half of the calls to citizens who would donate nothing, so that in the resulting distribution would be 30 percent donate $0, 35 percent donate $10, 17.5 percent donate $20, and 17.5 percent donate $60. The expected revenue from each call would increase from $10 to $17.50. Fundraising scores that increase the proportion of big donor prospects relative to small donor prospects would further improve on these efficiency gains.
If you've ever wanted to know more about how campaigns use data analysis tools and techniques, this paper is a great primer.