On the junk science and excellent PR of Cambridge Analytica

Cambridge Analytica claimed that it could sway elections thanks to the devastating power of psychometric profiling, and they may have even believed it, but those claims should be read with a critical eye, because they're marketing hype aimed at people whom Cambridge Analytica was pitching as client; and because Cambridge Analytica is not a scientific enterprise, but a secretive corporation whose researchers never had to subject their experiments and results to critical, peer-reviewed scrutiny, opening up endless possibilities for self-deception and truth-shading.

Cambridge Analytica was undoubtedly very good at some kinds of targeted persuasion: as Bloomberg's Faye Flam writes, they were incredibly talented at targeting marketing messages to rich, corrupt politicians and their backers. But that marketing pitch relied on exploiting a common statistical shell-game in which effect sizes are expressed in relative, not absolute terms

Think of how the Daily Mail has spent decades dividing all the universe into cures and causes for cancer by misreporting scientific studies that show that very rare forms of cancer are only slightly less rare (or more rare) in people who eat a certain food, as in, "One in ten million people get fingernail cancer, but among Chardonnay drinkers, it's 12 in one hundred million!"

Cambridge Analytica — and psychographics advocates in general — are masters of this kind of framing, pointing out that "the big five traits" are a better predictor of someone's preferences than a survey of their close friend. What they don't mention is that close friends really suck at predicting certain preferences, and the "Big Five" only suck a very tiny bit less. Omit that important fact and psychographics sounds very exciting indeed.

A close examination of the literature on psychographics for analysis, prediction and persuasion reveals a very "scant science," as Elizabeth Gibney explains in Nature. While targeting has some major benefits over nontargeted advertising (putting pet food ads in magazines targeted at pet owners more product-per-exposure that putting the ads in magazines targeted at, say, pharmaceutical reps), the literature-backed additional benefit of psychographic targeting is pretty modest, even according to its biggest proponents.

The 2016 election was won by a tiny number of votes, a statistical rounding error. Even very small nudges could have pushed things into victory for Trump. Maybe Cambridge Analytica secured enough votes — either through normal targeting or through the tiny (and possibly imaginary) incremental gains from psychographics to carry the vote for Trump, but so did everything else that moved the needle even a tiny amount. The real question is, how did the needle get so close to the red zone, when the red zone had a bozo like Trump in it?

The danger of focusing on the self-proclaimed sorcerous powers of Cambridge Analytica is it offers people who (for example) don't want to reform the Democratic Party to advance an agenda that puts people over finance, mass surveillance and endless war a way out: don't offer real alternatives to neoliberalism, they can say, but instead fight the sorcerous interventions of Cambridge Analytica and all will be fine again.

(via Naked Capitalism)