Cambridge Analytica’s (CA) mining and retaining of Facebook data during the election has become a hot topic of discussion. The claim is that they paid people, via Mturk (and maybe other avenues) to install a Facebook personality test app, which then mined not only their Facebook data but also the Facebook data of others. There’s some confusion about what they were doing, for what purposes, and some psychologists have questioned how effective it could have been, really.
I have some first-hand experience with related methods, and while my experience is from the pre-Facebook Stone Ages (mid 2000s), I have some thoughts. I should emphasize that everything I’ll say is speculation based on public news reports and extrapolation from my own experience in this area, and not first-hand knowledge of any kind. I would be curious about the perspective of people who have more recent experience with the modern iterations of these methods.
What is psychological profiling?
One of the more sensational claims is that CA used some highly effective personality profiling or psychographic modeling: CA CEO Alexander Nix touted their “model to predict the personality of every single adult in the United States of America.” What was this model? It looks like it was basically the standard workhorse of personality psychology, the Big-5. Unlike lots of so-called “personality tests”, the Big-5 is a real thing — it was developed via factor analysis: measuring many, many different self-report questions and then identifying those questions that best distinguish between people, and the underlying constructs that those questions seem to capture. Not only does the basic structure replicate well in other studies, but there has been a large literature testing what behaviors and attitudes are predicted by the Big 5, including political views and behaviors, although the relationships appear to be complex.
So, there was no breakthrough here in understanding personality. Instead, the purported innovation was in using Facebook data to predict people’s Big 5 personality, rebranded as OCEAN, from their online behavior, radically changing the scale at which personality data could be collected. This approach was pioneered by researchers at Cambridge in published academic research, and seemingly adapted, replicated and applied by Kogan and CA.
Does predicting personality matter?
The basic idea is an extension of an old database marketing tactic: purchase or compile a large-scale database with many variables, survey a much smaller sample of people to get their self-reported measures for key variables, and train a model to predict those key variables, to then apply that model to the large database. If the model predicts well, then you have an approximation of what people would say without having to get their participation. So, for example, you could survey 1000 people to ask if they support tax cuts or more funding for public schools, figure out which variables available in both the survey and in a million-person database predict these attitudes, and then score each of the million people in the database on their likelihood to support tax cuts or school funding, and use this to either customize communications (mailings and door-hangers) or for prioritizing turnout efforts.
The catch is that political attitudes are hard to predict. The irony is that, in my personal opinion, these kinds of models (based largely on demographics and location) are not very accurate, but still tend to be a lot more accurate than the subjective judgments of political consultants, even those who are supposed “experts” in a geographic area. That said, the promise of the Facebook data is that it is much richer and could yield models that predict much better than old-fashioned demographics-based models.
The other odd thing about the storyline that CA developed some kind of personality super-test is that personality, even the self-reported personality that their model was supposed to predict, is not all that useful in practice. The idea of the Big 5 is that it measures something very broad that is at least a little bit relevant to most human behavior, but that’s at the cost of not relating to behavior in any particular context that well. Political scientists study personality only a little, primarily as a bridge to psychology and out of curiosity whether personality relates to political behavior at all. Their primary focus is instead on the attitudes and beliefs that most directly relate to political decisions. By the same token, political pollsters (including those doing strategic work for campaigns) focus on attitudes and beliefs about candidates and issues, which are much more relevant to voting than is personality.
Predicting receptivity to political messages?
For all these reasons, I suspect that the personality profile aspect of this story is a red herring. My guess is that personality was useful because people like taking personality tests, and because the various stakeholders CA was interacting with would find personality easy to understand, and perhaps be impressed by their ability to predict it. But I’m guessing the main point was never to send neurotic messages to neurotic people and conscientious messages to conscientious people.
So, if the point was not some amazingly effective personality test, what was the point? The NYT framed the issue as “the company that can perfect psychological targeting could offer far more potent tools: the ability to manipulate behavior by understanding how someone thinks and what he or she fears“. Mere personality targeting would not achieve that. Kogan bragged about having “a sample of 50+ million individuals about whom we have the capacity to predict virtually any trait.” I think this is the real strategy — to identify specific issues that motivate individual voters.
So, the goal is not to target generically neurotic voters with neurotic messages, it’s to adapt a much more standard strategy in political campaigns to the social media environment: find the pro or anti-gun-control people, find the pro or anti-immigration people, find the pro or anti-gay-marriage people, etc.. Then send them targeted messages, arguing that your candidate is on their side, and/or that the other candidate will bring about the end of civilization based on their evil position on the issue that the person you’re messaging cares the most about.
It seems that personality was incorporated into this strategy to some degree, but my guess is that the real key to the strategy was access to extensive Facebook behavioral data and using targeted communications to hit the right issue buttons for each person. When David Carroll sued CA in the UK to get the data they had on him, what he got back was exactly this — their predictions of his positions on various issues, rather than personality predictions. This is not new — it’s an old strategy, but on steroids. In the 1980’s, campaigns would target their messages geographically — using different messages in different zip codes. In the 2000’s, messages could be targeted to the household, so that next-door neighbors might get different mail or different phone calls. Now, the targeting is even more granular, by device or account, so that different people in the same household can be targeted with different messages.
Large scale A/B testing.
There is another key aspect of the strategy that has largely been overlooked. According to Theresa Hong, a member of Trump’s digital team, “It wasn’t uncommon to have about 35 to 45 thousand iterations of these types of [Facebook] ads everyday.” In the context of the emphasis on personality, this sounds like micro-targeting, with the campaign creating ads customized to the unique personality of each voter. But I doubt it — personality testing is about understanding big differences, splitting people up into large groups that are different from each other, rather than nuanced differences that would yield thousands of different micro-targeted ads. In fact, this volume of advertising can only be created algorithmically, by assembling ads from separate components, like a political Mad-Lib.
So, this seems to be about micro-targeting via testing, rather than via understanding personality. According to the vice.com article, “On the day of the third presidential debate between Trump and Clinton, Trump’s team tested 175,000 different ad variations for his arguments, in order to find the right versions above all via Facebook. The messages differed for the most part only in microscopic details, in order to target the recipients in the optimal psychological way: different headings, colors, captions, with a photo or video.” The key here is that the targeting is not being done by understanding each voters’ personality. In fact, this kind of testing makes any such understanding completely unnecessary. Instead, the strategy is to push out a high volume of different ads, see which ones perform best in terms of what is observable on Facebook (clicks, likes, shares, etc..), and then use those outcomes in conjunction with Facebook data on the people who responded to determine a targeting strategy (i.e. to predict which ad version will work best for whom).
Such testing is not especially sinister — it underlies much of the behavior that people experience on the web, where much of the content we interact with is constantly being tested. In fact, the Obama campaign pioneered testing in online fundraising and recruiting volunteers. However, what’s potentially unprecedented is using testing at this scale for modifying political messages and ad copy/targeting. Traditionally, campaign messaging was something that high-level consultants within the campaign fought over the minutae of, either for ideological reasons or to maintain their own position within the campaign. While survey data would be leveraged in these debates, the degree to which control of the message and format was shifted to algorithmic testing in this case seems novel. It’s also unclear to what degree the Trump campaign took what they were learning about what worked on Facebook and applied it offline, such as at campaign rallies.
So, what’s the big deal?
I think a lot of the discussion about these issues misses the mark, overselling the impact of personality profiling. That said, I think the skepticism about whether personality profiling would impact an election also misses the impact that CA’s methods might have had. In the past, targeting would result in the voter receiving a paper mailer or email that was clearly from the campaign, or a volunteer from the campaign might ring the voter’s doorbell. Such approaches are not all that effective — people throw out junk mail and delete emails at a high rate. (There is some evidence that door-to-door campaigning is effective, but costly and not easily scaled up).
However, what happens online is different, in ways that may make a big difference. Social media content blurs the boundaries between advertising vs. reputable news vs. fringe (or even fake) news sources, and often comes with an implied or explicit endorsement from others in one’s social network. These factors may transform traditional campaign approaches into something far more effective.
Lastly, there is an entire other set of issues around how these traditional approaches were deployed online. The current attention is driven by revelations suggesting that CA harvested data from Facebook from people who were either told it was for academic research or from unknowing social contacts of those people, and kept the data even after Facebook required them to delete it. There are also major issues of accuracy and disclosure in online political advertising. If CA’s methods were accompanied by them or someone else using bots and fake accounts to reinforce the messages, based on the same person-level targeting, people’s skepticism about online information could have been very effectively overcome, by creating a targeted echo chamber.
I’m not an election lawyer, so I don’t know to what degree existing election laws prohibited the tactics used, were too vague to proscribe tactics that occurred online, or were actually intended to allow these tactics. Was the Clinton campaign flat-footed, and out-teched, in the same way that the Obama campaign had a tech advantage over the Romney campaign? Or did CA bend or break the rules, getting sole access to powerful data and using it for unethical messaging tactics, while the Clinton campaign did not because they followed the rules? I don’t know. But it seems very unlikely that the big difference between the campaigns was that one campaign knew which Facebook users were neurotic, and the other did not.
One last point: incentives and skepticism.
Going back to the controversy in the 1960s about “hidden persuaders” manipulating consumers for profit, marketers (including political marketers) have been shaped by two competing incentives. Regardless of the actual effectiveness of their marketing tactics, there is a strong incentive for practitioners (and potentially even for academics) to oversell effectiveness, for example, to convince potential clients that spending on marketing messages will pay off in desired changes in consumers’ (or voters’) behavior. However, consumers and voters usually don’t like the idea that their minds are being played with, exaggerated though it often is, and may push for sanctions or regulations that constrain or punish marketers, leading marketers to plead that they are not trying to persuade anyone, just to honestly inform them. This push and pull can even lead to societal panic about marketing tactics that may turn out to be more ineffective than harmful.
We can see this dynamic in the CA case. CA sends out a press release quoting Nix, “We are thrilled that our revolutionary approach to data-driven communication has played such an integral part in President-elect Trump’s extraordinary win“. Alexander Taylor of CA then says “It’s not about tricking people into voting for a candidate who they wouldn’t otherwise support. It’s just about making marketing more efficient.” Even without the legal and political scrutiny of the Trump campaign, CA has the incentive to portray themselves as both wizards who can manipulate minds and simple purveyors of objective information, and to both highlight and downplay the role of personality in what they do, to claim that they have massive data and that they would never use third-party data. This brings to mind another scandal, involving whether Facebook could manipulate emotions or not — in that case, a similar mix of incentives to exaggerate and downplay made it difficult to pinpoint reality.
All of which is to say, healthy skepticism is needed all around.
Update: This interview sheds a bit more light on the methods and scope, although the whistleblower, Chris Wylie, seems to draw a distinction between persuasion and manipulation that rests on whether or not messages are customized via personality profiling. That doesn’t make much sense to me.
Update # 2: Ch. 4 has just released a new report, with hidden camera video. Suffice it to say, I think this has even less to do with data, let alone personality prediction, than I did before. It seems possible that the personality profiling that people find objectionable was really just an academic-seeming cover for seriously dirty campaigning. In any case, I think the focus on personality is likely to be obscuring the much bigger problems we now face with anonymity of political actions and deliberate deception and outright propaganda.
Update #3: Latest Ch. 4 video. It was not about personality prediction — the goal was A/B testing and targeting untraceable ads and communications that seemed to be coming from independent advocacy groups or individuals.