Media Continues to Struggle With the Whole Correlation/Causation Thing


Wikimedia Commons

It seems hardly a week goes by without media outlets leaping at a study that purports to explain the real reason Donald Trump won the 2016 election. Such studies never focus on Trump's strengths or Hillary Clinton's weakness, but coincidentally often end up indicting the cultural forces that mainstream journalists worry about the most. Two such studies caught my eye in the past week because they both seem to fall for the same fallacy.

The first thing they teach you in a high school statistics class is that correlation does not imply causation. Some things correlate because one causes the other: say, suntanning and skin cancer. But other things correlate because they're both caused by another factor: like suntanning and ice cream sales. And sometimes when two factors correlate, you might think one causes the other when it's actually the other way around: perhaps violent teenagers gravitate towards violent video games.

Over the weekend, Politico published a story indicating that "the decline of local media sources by itself may have played a role in the election results." They reached this conclusion by examining subscription data and discovering that there is a correlation between "news deserts"—counties in the bottom-10 percent of newspaper and digital subscriptions—and counties where Donald Trump outperformed Mitt Romney. Their explanation for this phenomenon is that Trump did "worse overall in places where independent media could check his claims."

I believe wholeheartedly that counties with low newspaper subscriptions moved toward Trump, but the notion that they voted that way because of low subscriptions is taken as an article of faith, despite plenty of other explanations coming to mind. For one, newspapers likely struggle to maintain circulation in the sparsely populated rural counties that voted for Trump, but thrive in densely populated urban areas. As one of my followers snarked on Twitter, "Trump did well in places with grain silos. Grain silos are why we have Trump."

For another, conservatives tend to have a healthy skepticism of the media while liberals tend to lionize it. One poll showed that when asked why they subscribe to their local newspaper, a majority of liberals say it made them "feel good." Conservatives and independents are more likely to say they just want to clip coupons. It stands to reason that counties with many conservatives—particularly those who strongly support a candidate who rails against "the fake news media"—are more likely to have low newspaper subscription rates.

Both explanations strike me as far more likely than the notion that central Indiana went for Trump because The Kokomo Warbler wasn't around to factcheck him.

Politico cites the example of Price County in Wisconsin, which went from a dead heat in 2012 to supporting Trump over Clinton by 25 points in 2016. In Politico‘s reckoning, this was because Price "had virtually no paid local-news circulation."

Or it might be because Price County has a heavy manufacturing base and Trump's protectionist policies were more attractive than Romney's. It may have been because Barack Obama spent millions in Wisconsin depicting Romney as the sort of man who shut down factories, while Clinton ran ads on Trump's naughty words. It could be because the Trump campaign made a conscious effort to target historically neglected Northern Wisconsin. It might be because residents there were energized by Trump's speeches in Green Bay, Waukesha, West Bend, and Eau Claire, but didn't attend [Note to editor: I can't find any example of Hillary Clinton speeches in Wisconsin, please find and insert before publication].

In short, Politico fails to explain why we shouldn't chalk up Trump's strong showing in "news deserts" to a third factor. Trump almost single-mindedly targeted white, rural, working class voters—including labor voters who traditionally supported Democrats—and Clinton chose instead to focus on turning out minority voters. That led to Trump outperforming Romney in the kinds of communities that also struggle to maintain local newspapers and do not trust the coastal newspapers.

The piece does cite two outliers in an attempt to dispel the notion that their results can all be chalked up to rural voters breaking for Trump. Arizona's Yuma County, a poor rural county that nonetheless has high subscription rates, voted heavily for Clinton four years after being evenly split between Obama and Romney. North Carolina's Robeson County has a similar economic profile to Yuma County, but exists in a "local news desert." Trump outperformed Romney's performance in the county by 20 points.

But there are mundane explanations for that disparity. Yuma County rests on the Mexican border and is around 60 percent Hispanic, not exactly Trump country. Robeson County is a bit of a political oddity; around 40 percent of residents are Native-Americans and members of the Lumbee Tribe. The Washington Examiner reports that Trump's strong showing was attributable to local Republicans successfully courting the tribe.

In a similar vein, The Washington Post ran a story last week headlined, "A new study suggests fake news might have won Donald Trump the 2016 election." That piece was based on a paper by three Ohio State researchers that is still awaiting peer review.

"Our analysis leads us to the conclusion that fake news most likely did have a substantial impact on the voting decisions of a strategically important set of voters…" the paper reads. The researchers reached that conclusion by polling 2012 Obama voters and asking them about three false stories: that Clinton was seriously ill, that the Pope endorsed Trump, and that Clinton gave weapons to ISIS. Voters who defected to Trump in 2016 were more likely to believe those stories were true.

So what they've established is … people who didn't vote for Hillary Clinton are more likely to believe bad things about Hillary Clinton. This is not remotely surprising. I'm not a professor, but it seems to me that an obvious causal explanation presents itself: not supporting a politician leads one to more readily believe negative stories about them.

A January Gallup poll showed that more than three-fourths of Democrats and Republicans alike said that "accurate news stories casting a politician or political group in a negative light" is sometimes or always "fake news." A huge swath of Americans will look at any negative story about a politician they support and declare it fake. By asking about fake stories that portray Trump in a positive light and Clinton in a negative light, the researchers basically guaranteed they'd get results showing Trump voters were more susceptible to fake news.

Even the authors admit that "data from a single-wave survey cannot ‘prove' that these fake news items caused former Obama voters to defect from the Democratic ticket." That admission didn't slow their analysis based on just that assumption, nor did it temper WaPo's appetite for the story.

Professors at both Dartmouth and Princeton were skeptical of the study on Twitter, noting that fake news in 2016 was mostly concentrated to highly partisan voters and made up only a fraction of shared news stories. "When authors include a perfunctory ‘to be sure' note that ‘‘causality’ cannot be proven on the basis of a single-wave survey like ours' [and] then proceed to discuss the results using causal language … beware!" tweeted one.

This is of course why WaPo says the study "suggests" that fake news led to Trump's win, and Politico says the decline of local news "may have" led to the same. It's possible that either phenomena was a determining factor, but neither outlet give sufficient weight to eminently reasonable and far more likely alternative explanations. Which "suggests" these outlets are less concerned about what's actually true, and more concerned with what they want to be true.