The Three Sources of @realDonaldTrump
There’s a lot more information buried inside a tweet than you see on the screen. But anyone can access this “meta-data” through the twitter API. One of the pieces of hidden data is what program/platform was used to send the tweet. Some examples are predictable, like “Twitter for Websites,” “Twitter for iPhone,” “Twitter for Android.” Some are not, like “TwitLonger” or “Periscope.”
Back in mid-2016, people started proposing that there were two different people tweeting from the @realDonaldTrump account: Trump and his staffers. Data-scientist David Robinson wrote a very extensive article called “Text analysis of Trump’s tweets confirms he writes only the (angrier) Android half,” in which he shows that there are undeniable differences between the tweets sent from “Twitter for Android” and all the other sources (mostly iPhone and from a web browser).
For example, the ones from the Android phone are a lot angrier and negative. Robinson used a very sophisticated method of testing for 10 different moods in addition to sentiment. One of his conclusions is that the Android tweets used “about 40-80% more words related to disgust, sadness, fear, anger, and other ‘negative’ sentiments than the iPhone account does.”
The Atlantic went a little further into the weeds after Robinson’s article came out, showing that as the campaign went into high gear, fewer tweets came from the Android phone:
This agreed with reports that his campaign staff were clamping down on how often he sent out tweets himself.
But is there any evidence that only Trump uses the Android phone?
Trump has been using some variant of an Android phone for years. We’ve seen photos in the media and he has acknowledged it himself. So in this project, I implicitly assumed that any tweet from an Android phone was made by Trump himself. The fact that my model correctly identified 98% of the tweets coming from an Android in the months following the inauguration strongly argues that this assumption is correct. But is there actual evidence?
Maybe. My is argument based on four pieces of evidence:
The extensive linguistic and sentiment analyses by David Robinson (here and here) show that tweets made from an Android phone are clearly written by different authors than tweets made from other devices.
2. The phone.
Trump has been repeatedly photographed using a Samsung Galaxy S3, which was released in mid-2012 (9 months before Android tweets started showing up).
3. A successful classification model
This model was trained on a small subset of all the tweets from @realDonaldTrump. However, it is remarkably accurate at classifying a tweet as coming from an Android device (over 98% correct for the tweets made after Trump took office). If there weren’t different author(s) for Android and non-Android tweets, then classification models could never do better than 50% accuracy (the same as random).
4. Timestamp analysis
Tweets from an Android device didn’t start showing up until Feb 5, 2013, even though his account had tweets dating back to May 4, 2009. Trump himself admitted back in 2011 that he farms out some of his tweets (this was before the Android device shows up (by the way, my model predicts a 66.5% probability that these are his exact words).
I dictate my tweets to my executive assistant and she posts them. "Time is money"—The Art of the Deal.
— Donald J. Trump (@realDonaldTrump) November 23, 2011
If he dictated most of his tweets, they should have been coming out mostly during working hours (9 AM to 6 PM). Here are the distribution of tweets made throughout the day based on device.
The blue curve shows when tweets were made before the Android phone started tweeting: mostly during working hours (92%), as suspected. Once tweets started coming from an Android device (green curve), they were mostly outside of working hours (85%), while those from non-Android devices (yellow curve) were made during working hours (91%). That certainly suggests that someone different (i.e. Trump himself) was using the Android phone.
What does the machine-learning model say?
I trained my machine-learning model only on tweets made during the campaign, but as you can see by perusing the entire Trump twitter database, I ran the classifier on all of his tweets (except retweets, of course). Now I never included the time-of-day (or twitter platform) as features for training the prediction model, so we can look at when the tweets were made and how they were classified.
The following chart show how many tweets were made throughout the day for those classified as Written by Trump.
Note that many tweets have been classified as Written by Trump before the Android showed up (blue curve). I assume he was either dictating (as he claimed) or writing them himself, and as we saw above, these were mostly during working hours (i.e., when an assistant would have been present). But as soon as the Android showed up (green curve), tweets classified as his came out overwhelmingly outside of working hours (85%). There are two notes to make here. First, as shown below, during the time period before the campaign started, the distribution of tweets made on the Android, and those identified as Written by Trump, are indistinguishable. Conclusion: the same author(s) was writing all those Android tweets. Second note is that the prevalence of early-morning tweets continues for all time periods after the Android phone came on-line. Conclusion: the same author(s) continued to use the Android phone after the campaign started.
Now let’s look at the tweets classified as Written by Staff.
As expected (since Trump was apparently writing all the Android tweets), the distribution of tweeting times up until the campaign was launched hardly changed before and after the Android device showed up. Even during and after the campaign, most tweets classified as Written by Staff come out during work hours.
To summarize, Android-based tweets are undeniably different than all the others and we know that Trump used an Android phone. Now either many different people had Android phones that they were using almost exclusively outside of work to write @realDonaldTrump tweets, or only Mr. Trump did. Occam’s razor suggests the latter.