Here at Civis, we’re very interested in studying how events shape public opinion. Survey research is one great way to track public discourse, but ambient data from social media provides an excellent look at the organic conversation around an issue or event. We’ve been using Twitter to track US discussion on access to healthcare for months now, and the overall conversation seemed to be shifting in a productive direction, toward jobs and implementation and away from anti-administration partisan bickering.
Last Thursday our Twitter feeds went into overdrive over the Supreme Court ruling on the Affordable Care Act (ACA), which upheld ACA authorized federal tax credits for Americans in all 50 states, not just those living in states with their own healthcare exchanges. It inspired us to dig in and see what this meant for the trends we’d observed.
What did the conversation look like before June 25?
We’ve been tracking the conversation around healthcare for almost a year now, with a focus on distinct tweets in English and only in the US. Our analysis focused on two main points: what conversations are happening about healthcare, and which communities are participating in each discussion.
We found that there were several distinct conversations happening around healthcare. Here are the topics we found:
- Obamacare/ACA Perceptions
- Healthcare System Perceptions
- Healthcare Statistics
- Personal Healthcare Issues
- Healthcare Jobs
- Supreme Court Ruling
To understand the conversation, we sampled millions of tweets from the Twitter firehose that met basic search criteria (keywords and phrases) we considered part of the healthcare discussion. We excluded retweets from the set, and retained only independently-authored material. In order to make that giant set of unstructured data more manageable, we used a technique called Latent Dirichlet Allocation (LDA) to learn what topics are discussed most frequently. LDA learns a set of human-interpretable topics (technically, frequency distributions over words) that can be used for high-level analysis and visualization. We still need to assign human-readable labels for each topic once they are learned, but this is typically easy to do by looking at the words and tweets most strongly associated with a topic.
In early June, prior to the announcement of the Supreme Court’s ACA announcement, the partisan discussion had increasingly moved towards a focus on the growth of healthcare-related jobs. The Healthcare Jobs topic had grown to about a quarter of all Twitter activity on the issue, up from around 10% at the beginning of the year. It looked like we might be on our way to recognizing that ACA was not such a job-killer after all!
The second thing we wanted to understand was which distinct communities were participating in the conversation. To do this, we built communities using the Louvain method for community detection. In simple terms, this approach looks at the follower relationships across the full spectrum of Twitter handles and clusters them into distinct groups. We found that there are four groups participating in this conversation: Centrists, Healthcare Thought Leaders, Progressives, and Conservatives. Conservatives are ‘farther away’ and more isolated from the rest of the conversation than all of the other groups. Here’s what the communities look like:
How has the conversation changed?
Well, one significant way the Twitter conversation around healthcare changed after the ACA decision was that it became a lot more active — no surprise there! On the average day in the first half of 2015, we saw about 18,000 tweets per day on healthcare; on June 25th, we saw a spike of 130,000 tweets on healthcare.
But the increase in volume itself is not particularly remarkable or interesting. We also wanted to know how the topics that were being discussed in the wake of the decision had changed. To that end, we constructed a daily time series of the activity on our LDA topics. This graph shows the relative activity of each of our seven topics over the previous ten days.
To complement this, we looked at the relative share of the conversation each topic occupied and saw that the share of the conversation about healthcare jobs had gone way down. It was replaced by partisan commentary about the ACA and the Supreme Court ruling.
It certainly looks like something is going on here, given the tails on many of these graphs after the June 25 announcement. We can validate this by using a method called Bayesian change point analysis to pick out those words and topics showing a marked change in distribution during the time in question. Our initial suspicions were confirmed, with big losses for words like ‘#healthcare’ and ‘#hiring’ and big gains for ‘Obamacare’ and ‘SCOTUS.’ We also see clear change points in the conversation for the following discussion topics:
So it seems that the big shift in the conversation following the ruling was a significant drop in the share of postings about healthcare jobs and personal healthcare issues. These were replaced by an increase in tweets about the Supreme Court ruling (obviously) and a rehashing of vitriolic rhetoric around ‘Obamacare.’
So have we slipped back into partisan bickering?
The next question we wanted to answer was whether or not this change in the conversation has led to a regression to partisan bickering. While it’s too early to tell if any change is going to last, we can certainly look at the difference between today and what had been going on recently.
To do this, we decided to take a look at sentiment around healthcare tweets. What we found was disheartening, albeit unsurprising. The overall sentiment of healthcare tweets began plummeting on the day of the announcement.
This drop in sentiment is largely due to a change in the dialogue from talk about implementation and employment opportunities to political partisans complaining about the ACA.
To further hammer this point home, we can look at the standard deviation on the sentiment, which skyrocketed. This means that the sentiment for the average tweet got further away from the mean. For non-statisticians, that means that people weren’t converging around good ideas, but moving farther away from each other as time went on.
Who is driving this change on Twitter?
Finally, we wanted to take a deeper look into who was participating in the discussion so we returned to the communities of users that we identified using the Louvain algorithm. Similar to what we saw on the content side, we found that participation increased among political partisans and decreased among healthcare thought leaders. The table below describes the change by group:
On Twitter, while the bulk of the discussion continues to be driven by healthcare domain specialists, political factions have ramped up their activity.
For fun, here is a list of the terms most strongly associated with each community’s tweets in the day following the ACA announcement. Deciphering the favorite hashtags of each group is left as an exercise for the reader”¦
In many ways, the conversation last week represented a regression, with political partisans drowning out a healthy discussion about economic opportunity under the new ACA. Before the SCOTUS ruling, the discussion had been focused on job prospects for nurses and other health-care professionals. Thought leaders were engaged in discussions about where the health-care market was headed. That moment waned, and for now, we’re witnessing the same rhetoric we heard in the summer of town-hall meetings in 2009.
*This post is co-authored with Derrick Higgins.