Over the past year, a lot has been made of how President Barack Obama's 2012 re-election campaign was able to successfully utilize data analytics to help turn out voters. One key part of the campaign's analytics efforts involved a cutting-edge predictive analytics technique that is reshaping the way marketing department, healthcare organizations and other users are reaching out to their target audiences.
In a presentation at the 2013 Predictive Analytics World conference in Boston, Daniel Porter, director of statistical modeling for the Obama for America 2012 campaign, described how he and his eight-person team used an uplift modeling program to precisely identify voters who were leaning toward Republican Mitt Romney but were likely to be receptive to the Obama campaign's message.
People tend to really like [President Obama] or really oppose his agenda, and we were still able to identify people who were likely to be persuadable.
Daniel Porter, director of statistical modeling, Obama for America 2012
Uplift modeling is all about finding the "persuadables." In any group of voters, or potential customers, some people have already decided; in marketing terms, they plan to buy your product even if you don't advertise to them, or they're dead set against doing so. Either way, it's a waste of resources to contact them, Porter said. Another set of people are likely to be turned off if they're contacted; they should also be left alone. But then there are those who need some convincing -- and are open to being convinced.
Identifying them is no small feat. Porter said political campaigns have known for decades that contacting certain voters pays off more than interacting with others does. But in the past, campaign managers typically relied on generalizations about voters based on limited demographic data. Porter's team of data analysts had the advantage of being able to use a database created by the Democratic National Committee that was a trove of public record voter files and the DNC's own voter data. During the course of 2012, the analysts built, tested and refined a set of predictive models based on demographic, geographic and political data in an effort to statistically identify the characteristics of persuadable voters in swing states. Last October, the Obama campaign started using the models to optimize online and TV ad buys and to determine which doors canvassers should knock on, which voters should be called and which ones should receive mailings.
Porter, now a partner at analytics services provider BlueLabs in Washington, D.C., said the success of the re-election campaign is proof that uplift modeling works.
"People tend to really like [President Obama] or really oppose his agenda, and we were still able to identify people who were likely to be persuadable," he said. "So I'd imagine in other organizations, where attitudes aren't as crystalized, there'd be an even greater opportunity for uplift modeling."
Uplifting experiences beyond politics
One of the places outside of politics where uplift modeling is gaining ground is in marketing and advertising. Peter Amstutz, an analytics strategist at Minneapolis-based ad agency Carmichael Lynch, said at Predictive Analytics World that the uplift approach enables marketers to spend their advertising budgets more efficiently -- and potentially increase sales in the process -- by targeting ads at persuadable consumers.
Carmichael Lynch is helping advertising client Subaru of America use uplift modeling to identify Internet users who need, and are likely to respond to, a marketing push for the automaker's cars. After developing an initial uplift model in-house, the agency signed up to use an automated analytics service offered by Rocket Fuel Inc., which operates an online ad-buying platform. Rocket Fuel's model scores potential Subaru buyers against 300,000 data variables that are updated on a daily basis.
The agency tested the model by showing people deemed to be persuadable online ads for either Subaru or a charity. Amstutz said the first run, in June, generated an incremental lift of about 10% in car purchases by the group shown the Subaru ad compared with the control group that saw the public service announcement. By September, the lift was up to 35%. "We're hopeful we'll continue to see that go up," he said. "We don't know what the ceiling is, or whether there is a ceiling."
READ MORE ABOUT PREDICTIVE ANALYTICS
Learn how much data you need for effective predictive analytics
See how businesses are organizing predictive analytics programs
Read how predictive analytics software can strain IT systems
There are some tradeoffs and complications to be aware of. Amstutz acknowledged that the cost per conversion, a standard measurement of marketing success, typically increases with uplift modeling because sure-thing customers are no longer counted toward the conversion number. There's also the cost of the data science work required to determine the characteristics of persuadable customers and identify people who fit that mold. And an uplift modeling campaign requires some patience while models are built, tested, refined and updated. But if it's done right, Amstutz said, businesses can do more with fewer, more targeted ads. "We're getting our [ad] media in front of the people who we think really need to see it," he said.
Because of the tradeoffs, uplift modeling initiatives need to be supported by an organization's top executives, Porter said; otherwise, they might be tempted to pull the plug if a program doesn't show a quick return on investment. He noted that his efforts with the Obama campaign benefited from the support of campaign manager Jim Messina, who was very interested in analytical modeling. Without Messina's backing, the uplift modeling project likely would have collapsed early on, Porter said.
Measuring the right things
Eric Siegel, president of consultancy Prediction Impact Inc. and founding conference chair of Predictive Analytics World, said most marketing managers measure the response rate to ads, mailings and other campaigns and then use that to calculate the cost per conversion. But, he pointed out, that largely measures people who would have bought a product anyway.
What businesses really should be thinking about is influence, not response, Siegel said: Instead of asking whether consumers will buy something if they're contacted, ask whether the customer will buy only if contacted.
With the amount of data that many organizations are looking to analyze growing rapidly in the big data era, Siegel thinks companies that leverage applications like uplift modeling will be the ones that differentiate themselves in their markets. "Data always speaks," he said. "If you can learn how to interpret it appropriately, it will tell you something."