As David Cameron once joked, Twitter isn’t Britain. We know, for example, that polls that look good for Labour get more retweets than those that don’t, because Chris Hanretty ran the numbers. And we know from Chris Prosser and Jon Mellon’s work using the 2015 British Election Study, that social media users in general and Twitter users in particular, are to the left of centre and more socially liberal compared with the general population.
But we now know from their 2017 BES data that the left liberal lean of the Twittersphere increased substantially over the previous two years. In 2015, Twitter users voted Labour 37-31, while non-Tweeters voted Conservative by 38-31. But in 2017, there were two Labour voters on Twitter for each Conservative (54-27), but off Twitter, the Conservatives won 45-39. And in the EU referendum, Twitter voted 68-32 to Remain in the EU, while everyone else voted 54-46 to Leave.
What about Facebook? Well, it’s not representative either, but not to anywhere near the same extent (46-36 Labour in 2017 and 54-46 Remain in 2016):
Given that (per Mellon & Prosser) the explanation for the Twitter “skew” in the first place was its demographic profile, combined with what we know from elsewhere in the BES about how different demographics have changed their votes recently, it shouldn’t come as a surprise that this too explains the increase in the skew since 2015. Chris Prosser has kindly rerun his earlier analysis using the latest data, and found that demographics (especially age) still explain the disparity. In other words, it seems that younger-than-average people swinging disproportionately to Labour explains the change, rather than there being any evidence that social media has attracted liberals or repelled others in recent years.
What does all of this mean? Well, anyone pretending that Twitter was an impartial gauge of reaction to news events already needed unfollowing… Now they deserve to be ratioed. That applies, to a lesser extent, to Facebook, and possibly to other platforms (though the BES only asked about these two). And if these biases go beyond what can be explained by demographics, pollsters probably need to look carefully at where they’re recruiting panelists from.
The penultimate paragraph of this article was updated on 9th February to reflect Chris Prosser’s updated analysis. Our thanks to him for that.