Q: Why was there such a large bump in the number of Hillary and Trump followers at 10:21 PM EDT during the second Presidential Debate?

A: I’m not sure, but maybe it has something to do with abortion or marriage equality?

During the last two presidential debates MITH has been collecting Twitter data using the words trump and hillary. Neil Fraistat’s motivation for organizing our Night Against Hate this week was to identify social media accounts for hate groups that are in the Southern Poverty Law Center’s Extremist Files, and then use this information to help see how these groups were using Trump’s own words in their tweets. So it was super to see How Trump Took Hate Groups Mainstream in Mother Jones just yesterday. Sarah Posner and David Neiwert have been on this trail for a while, and their write up is a must read.

We’re still talking about what we want to look for in the debate datasets and how we want to use the spreadsheet that we collaboratively built. Look for another blog post about the Night Against Hate soon.

One limitation of looking for patterns in collected Twitter data is that the Presidential Debates are high volume events. We know that the Twitter streaming API only gives us a portion of all the available tweets when there is a spike in traffic. For example our dataset for the first debate contains 1,303,084 tweets that mentioned “hillary” and the Twitter API let us know that at least 730,512 tweets were not delivered. Very little is known about what kind of sample Twitter provides, and without the full picture it is difficult to draw inferences from the numbers.


But one interesting metric that is available is the number of followers a person has. This number can be particularly interesting to look at over time as it rises and falls. Every tweet that you get from the Twitter API includes a User Object which in turn contains the followers_count for that user at the time that the request was made of the Twitter API. So if you collect data in real time from the Twitter Streaming API you get a moving picture of a user’s followers. What’s also super handy is that in every retweet you get the user object for the retweeter, and also for the user who sent the original tweet.

During the debates you can count on many people retweeting the candidates tweets. This means you can get smooth record of how many followers they have by the minute … or even second. The fact that you’re not getting all the tweets at a given point in time doesn’t really matter.

So I wrote a very simple utility that reads through Twitter data I collected with twarc looking for particular users and collects how many followers they have at particular times and writes it out as a CSV file, for analysis in a spreadsheet:

New Followers

I thought it could be interesting to look at the number of new followers each candidate received during the first debate. In theory these would be “people” that had decided to follow Clinton or Trump on Twitter as they were watching the debate. I put people in scare quotes there because it is possible, and perhaps even likely, that there are bot armies creating these accounts. You’d have to track who these new accounts were and sample them to know for sure.

So with that caveat in mind here’s what the graph for the first debate looked like:

It was encouraging (for me) to see that Hillary was gaining more followers than Trump during the debate. There were a few interesting changes in the graph, but I was mostly distracted by the fact that Hillary was outpacing Trump so much in new followers. We collected data for the second debate so I decided to look again with the same code:

I was pleased to see Clinton is gaining more supporters than Trump again; but this graph looks noticeably different. Do you see the bump in followers at 10:21 PM EDT (02:21 GMT). What happened there? Well, thanks to the Web you can see exactly what happened there: it was the question from Beth Miller about the Supreme Court that begins at 10:19:54:

If you watch the time tick by you can see in the exact minute of 10:21 PM Hillary is saying this:

I want a Supreme Court that will stick with Roe v. Wade and a woman’s right to choose, and I want a Supreme Court that will stick with marriage equality. Now, Donald has put forth the names of some people that he would consider. And among the ones that he has suggested are people who would reverse Roe v. Wade and reverse marriage equality. I think that would be a terrible mistake and would take us backwards.

What is truly bizarre is that there is a bump at that point for both candidates. Maybe this points at a problem in my code, or my spreadsheet data (I’m happy to share the Tweet ID datasets if you would like to look for yourself). Or perhaps the bump is part of some kind of backlog in Twitter’s infrastructure? Maybe there’s a bot army that’s working for both candidates? It would be necessary to inspect the users to get a sense of that.

But maybe, just maybe, this data points at the fact that a Woman’s right to choose and marriage equality are still polarizing and hot button issues, more so than any others, for folks in this years election? At least for folks who use Twitter…which is definitely not all voters. That’s probably the biggest caveat there is.

If you have ideas, questions, criticisms about any of this I’d love to hear from you.


I’m not sure my rate of change calculation was the best. I simply subtracted the new subscribers in the previous minute from new subscribers in the current minute. It does show that Trump got a bigger bump in this minute, and really highlights how much of a change it was. If you have an idea for calculating the rate of change that is better take a look at the data.