Sexy, Labour-leaning polls get more traction on Twitter

(at least based on RTs of Britain Elects tweets)

Later today I'll be giving evidence to the House of Lords Select Committee on Political Polling and Digital Media.

One of the questions the committee is interested in is the role of social media both as a means of learning about public opinion, and as a way of disseminating findings about public opinion.

It's well known that Twitter is not representative of the broader population. In particular, "social media users are younger and better educated than non-users, and they are more liberal and pay more attention to politics".

It's also reasonably understood that one of the reasons we share or re-tweet information is to signal our belonging to a tribe, rather than because we generally believe other people need this information.

From these two statements it follows that some poll findings might well be shared not because they are particularly informative, but because they favour parties which people on Twitter like.

I did a quick test of this by using information on the number of re-tweets of tweets by Britain Elects. For each tweet prior to the 2017 election which reported a voting intention poll, I calculated:

• the mean absolute change in the poll reported in the tweet. Thus, if the tweet reads Lab… (+1), Con… (-1), LD (0) and UKIP (-1) the mean absolute change is 3/4.
• The change for Labour (in percentage points, using the same information as above).

I was able to find 73 tweets with this information. The number of re-tweets is almost always very high, but varies considerably.

First let's look at whether "sexier" polls get more re-tweets. Here's a plot of the mean absolute change against the number of re-tweets. Note that the y-axis is on a square-root-scale.

Now let's look at whether good polls for Labour get more re-tweets. Here's the change for Labour against the same information:

These effects remain when you include both variables plus time of day in a single regression. Here are the plotted coefficients from a regression on the log of the number of re-tweets:

Those of you who like that sort of thing can look at the raw regression output below:

What does this mean? It means that for every extra percentage point gained by Labour (lab_chg), the number of re-tweets increases by a little over thirteen percent. For every extra increase in the mean absolute change (here, mad), the number of re-tweets increases by more than twenty percent.

This means that those of you on Twitter are almost certainly getting a misleading impression of the state of the polls. In this case, a disproportionate presence for pro-Labour polls may have made the-polls-on-Twitter seem more accurate. In the case of last year's referendum, a similar effect may have made the-polls-on-Twitter seem less accurate (since (human) Twitter users are more pro-Remain).

Code for this analysis is on GitHub.