Zika is an important emerging illness that has been linked to neurological syndromes in adult patients, and birth defects in patients infected in-utero. Here, we use Twitter to explore the discourse of individuals tweeting about Zika. We labeled a sample of 500 English tweets to identify common themes, used keywords to track two themes, reproduction and travel, and identify spatial patterns in tweets based on the language of the tweet. We also observed tweets made in the first person that might be indicative of reflections of personal behavior on Twitter. Future work will delve further into first person tweets and continue to examine spatial and temporal trends in the themes temporally observed.
Tweets were collected from Gnip, based on the keywords “zika”, “zica” (a common Portuguese spelling of the virus1, or ‘zikv’ (a common abbreviation of ‘Zika Virus’). Tweets were collected from March 1, 2015 until October 31, 2016. This resulted in just under 15.5 million tweets, 7 million of which are in English. For this initial work, we only coded English tweets. We identified the likely gender of person tweeting using Demographer12, and identified the likely location of the tweets using Carmen.13
This initial work has focused on the relationship between Zika and behavioral decisions (e.g., those relating to reproduction, travel, mosquito interventions etc.), as well as to important global events (e.g., 2016 Olympics). We identified a list of keywords related to these behaviors (see Table 1). We then filtered the dataset for tweets that included the keywords. Where applicable we included different plural forms (-ies rather than -s endings). For words that might match longer words or strings (e.g., ‘birth’ would match ‘birthday’), we included \b, a Python regular expression marker for a word boundary. microcephalies \bsti\b barre pope cancel paralysis brain std repellent catholic protect terminate babies sex\b spray ultrasound paralyze The resulting subset included 3,572,320 tweets. Of these, we randomly sampled 500 for coding. Labels were decided iteratively as a group, with the final labeling schema available in Table 2. Tweets were coded independently by two team members. Multiple labels per tweet were allowed. Labels were assigned based solely on the content of the tweet; we did not attempt to look at the context of any URLs linked, or find additional context from that user’s other tweets. Inter-rater reliability is presented in Table 2 (see Cohen’s ).
CDC says pregnant women should consider delaying travel Men, if you’re partner is pregnant, abstain or use condoms to prevent zika Delay pregnancy if you live in a place with zika Americans weigh in on late-term abortions in cases of zika-linked defects Baby born with zika-related birth defect Guillain-Barre´ cases associated with #Zika infection rise Country to spray for mosquitos to prevent zika Olympic athlete decides to not compete over zika fears Brazilian government admits microcephaly not caused by zika
Freq
(%)
34
(6.8%)
13
(2.6%)
17
(5%)
14
(3.4%)
56
(11.2%)
17
(3.4%)
37
(7.4%)
68
(13.6%)
8
(1.6%)
0.47
0.64
0.65
0.41
0.65
1.0
0.56
Lastly, we performed initial explorations of two specific types of tweets. Because our initial labeled corpus of 500 is
too small to build a robust classifier system, we identified a few keywords that related (
Within the behavior Twitter corpus, we have a somewhat higher proportion of male tweeters (61%) compared to female tweeters (37%) (the remaining 2% could not be identified). This is in contrast to surveys that suggest that men and women use Twitter at roughly the same frequencies.14 We were able to predict a user’s location for 61% of tweets. The majority came from the United States (20%). Venezuela, United Kingdom, Brazil, Canada, and India contributed 1-3% each. 179 additional countries had at least 1 tweet present in the dataset. Additional spatial analyses are available in Figure 2.
Labels, descriptions and example tweets are listed in Table 2. The most common label in our random sample was ‘birth defects’, followed by discussions of the Olympics, travel, ‘intervention strategies’, safe sex behaviors and ‘reproductive rights’.
We also noticed a rare, but interesting handful of tweets that were written in first person. These are an interesting subset because they provide insight into individual’s opinions on policy decisions, or reactions to guidelines on health behaviors, beyond what might be inferred by noticing what kinds of informational things an individual retweets or links to. A paraphrased example is ‘Canceling our baby-moon trip due to Zika concerns’.
These keywords are not meant to be exhaustive representations of discourse, but are useful to identify general patterns to explore more in future work. Several spikes within the dataset seem to correspond to policy discourse in the US and have been added as annotations onto Figure 1. Overall, travel tweets are much more common than those about reproduction. Given the intensely personal and political nature of abortion and reproductive rights, this is unsurprising. However, further investigation into the content of these tweets is warranted.
This work describes common themes in Zika-related tweets, as well as temporal and spatial trends in those tweets. Future work will employ other techniques, like topic modeling and named entity recognition to identify topics discussed within each theme identified. Additional focus will be on first person tweets because they provide insights into people’s individual decisions to change behavior. Although first person tweets were rare, we found tweets describing the decision to change travel plans based on advisories from public health organizations. Analysis of these tweets will allow us to understand how public health agencies can better use Twitter to communicate important public health information.