Trump’s data gurus

x Sponsored content on Research Live and in Impact magazine is editorially independent.
Find out more about advertising and sponsorship.

At an event last year, to predict the trends for the coming 12 months, one of the presenter’s tips was to ‘look out for Cambridge Analytica – it is getting a lot right and we’re going to hear a lot more from it’. It was the start of people outside the niche world of big data hearing the name Cambridge Analytica (CA), and its profile has only increased since, though not always for the reasons it may have hoped.

When it comes to the influence, strategy, client list, behavioural-prediction accuracy and technical success of the company, there’s a counter claim for every claim. Its work for the Donald Trump presidential campaign has secured it publicity and press scrutiny, but other aspects of its business remain opaque – such as the billionaire US conservative political donor Robert Mercer reportedly being a shareholder, and its involvement with the Leave.EU campaign (CA had denied it worked for the campaign group, while Leave.EU communications director, Andy Wigmore, told The Observer that CA had worked for it, although not paid).

Impact sought to find out more about CA’s methodologies and practices. After lengthy negotiation, SCL Group and CA’s lead data scientist, Dr David Wilkinson, answered some questions. 

Q: You have garnered fame by working with Donald Trump in the US. Can you explain the methodology you used to help his campaign? 

Data science played a huge role in the 2016 US presidential election. Our data analytics were used to inform strategic planning and drive marketing and communication. We used: extensive research surveys; voter files; data from the campaign; and commercially available data. 

Q: I understand it included psychological profiling of voters – how did you manage to profile so many people? How many did you do? 

This is incorrect. We didn’t have the opportunity to dive deeply into our psychographic offering because we simply did not have the time. Building a presidential data programme often takes campaigns well over a year. We started working for the campaign in the summer, so focused on core elements of a political data-science programme. 

Q: What behaviour traits were most important to identify in terms of understanding voting intentions? 

As before. However, we did focus a lot on key political issues, finding that Trump supporters were more likely to support a strong stance on law and order, and immigration, whereas persuadable voters were much more motivated by issues relating to international trade – particularly taxing companies that send work overseas – and increasing wages. The cross-section of issues with persuadable voters was key to identifying messaging strategy. 

Q: How did you use these profiles with other, external data sources? 

Our models were constructed with a combination of campaign data, commercial data and research. The results of these models were used to segment the target audiences – mainly focused on the most persuadable voters – by messages that would resonate most with them. These audiences could then be made into custom lists that were used in digital marketing platforms and social media. 

Q: In particular, how do you use social media platforms? 

Any media platform we use to distribute content to our audiences is strategically chosen, based on the audience we’re trying to reach and the campaign objectives. Because social platforms are a thriving centre of time spent online and have the added benefit of peer-to-peer influence, we find a lot of success in delivering content to our targeted audiences across Facebook, Twitter and Snapchat. Additionally, designing tactics that are unique to each platform – such as Twitter’s conversational ads or Snapchat’s geofilters – is an important way of ensuring that all our advertising is data-led, but still exciting and creative. The highly accurate, people-based targeting methodologies of Facebook facilitate an unparalleled match rate with our offline universes, as well as being able to test results very precisely. 

Q: Other pollsters have found difficulties in managing online responses vs telephone responses; how important do you think the medium of response is, and what biases do you see between the two? 

What you describe is what we call ‘method bias’. We correct for this by using a blended methodology – we were surveying using a mix of online panels, telephone to landlines, and telephone to mobile. We chose an optimum combination of these methods, based on achieving as much demographic balance in representing the voting population as possible. 

Q: How much were you influencing Trump’s decisions and strategy choices with your findings? 

Senior campaign staff had round-the-clock access to our findings and results, including information on those voters who were persuadable on any given week, and those who should be messaged as part of a get-out-the-vote strategy. These were detailed for each battleground state, with demographic and geographic breakdowns within each state. We showed how this electoral map moved with time, and what issues were most relevant at any given time. This led to important tools, such as an optimum Electoral College path to victory, prioritisation of states with recommended spend in each state, and optimum locations for rallies and events. We don’t know for certain when and how much each of these pieces were used, but we know the results were read and used to some extent by the campaign team. 

Q: I’ve read about you personalising advertising messages and individual targeting; do you have evidence of how effective that was? 

We conducted a multitude of tests to measure lift (that is, increase in support of the candidate compared to a control group), which continually fed back into our research and data programme. Typical lift figures for generic advertising were around 3% per week, but we were able to get to 12% lift per week using our individual modelling methods, and 6-9% lift per week using data-informed geographic targeting. 

Q: Your website talks about 5,000 data points per person – how are these achieved? 

CA’s database is a combination of demographic, political and consumer data from commercial sources, and models we have constructed for previous projects. 

Q: What do you think is the difference between what you do and what other pollsters or data analytics firms do? 

We built a research, data and digital programme that was optimum for the candidate in question. The regularity with which we could update our models and segmentation was behind our analytics programme success. In general, we are able to extract the deeper reasons about why people behave and react in a certain way. The ability to take fundamental research about behaviours for a political or commercial client – and expand that into a full communications strategy via data science – produces the rigour, efficiency and completeness that gives us an edge over competitors. 

Q: Despite all the work you did, you still didn’t predict Trump’s win; do you think the days of being able to predict election outcomes have gone? 

We saw about a 25% chance of a Trump victory going into election day. But more importantly, we showed that, if he was to win, it would be a combination of rust-belt state support and a low turnout from African-Americans in southern battleground states. We are much more interested in maximising the chance of victory than being a crystal ball. 

Q:You nailed your colours to the mast of right-wing politics, but where else do you see your data science business going in the future? 

We’re a politically neutral company. We work with Republicans in the US, while – in the rest of the world – we have worked with political campaigns and groups from the centre-left and the centre-right, government departments and agencies, unions and charities. We help commercial brands and small companies. SCL Group has a long history in national and regional projects in conflict resolution, social, health, development and welfare, and that is still a core division within the company.