Three primary means of getting online data: Curated data is ideal

• Curated datasets
  • Wikipedia dumps, COVID tweets, etc.
  • Often flat files
  • Often cleaned

APIs are the next best thing

• Application Programming Interfaces
  • Programmatic access to a server’s database
  • Often built for external developers, not researchers
  • Outages, code changes, etc. are often invisible
  • Typically in JSON
  • Some wrappers to make this easier (e.g., rtweet in R)

Screen scraping is often the only option

• Easy for static pages - just download them
• Can also be used on complicated pages
  • Program pretends to be a browser and extracts data from the HTML page
• Can be against ToS
• Programs break often
• The only way to get data not made available via APIs
• requests + BeautifulSoup

Where are the good data sources?

• Places where the conversation is already public
  • Twitter
    • Very easy in Python and R
    • Incredibly generous Academic API for retrieving tweets
  • Reddit
    • API is OK but rate limited
    • Pushshift is amazing!
  • Wikipedia, Github
• Kaggle

Benefits and drawbacks to online data

• So much data!
• Unobtrusively collected
• Amenable to machine learning approaches
• These are all benefits and drawbacks!

Dealing with lots of data

• Can find even small effects
• Can do analyses of often neglected subgroups
• Requires IT + cluster computing for many questions

Unobtrusively collected data

• Subjects don’t modify their behavior
• But they also often don’t know that their data is accessible to researchers!
• Informed consent is typically impossible
• There are emerging best practices around the ethics of using online data
  • e.g., aggregating, not quoting directly or including usernames, etc.

Amenable to machine learning

• Great advances in NLP and ML
• Machine-only approaches have biases and blind spots

Some other downsides

• Users != individuals
  • Multiple accounts, bots, etc.
• Very little demographic information
• Ethically and practically difficult to tie people to offline behavior
  • Aggregate behavior can still be useful, though (e.g., Google Flu Trends)

Project 1 - Google Maps and OpenStreetMap

“Zooming in” on the individual-level causes

• Big question is how competition influences an open source mapping project community
  • It hurts it!
• Having digital data lets them ask why
  • New members stop contributing but established members increase their dedication

Figure 3 from Nagaraj and Piezunka, 2020

Project 2 - How do protests spread on Twitter?

How do people decide when to post a supportive hashtag?

• People have different thresholds before joining in (Granovetter, 1978).
• Gonzáles-Bailón et al. measured this threshold empirically

Project 3 - Wikipedia’s declining userbase

• Wikipedia study found a “rise and decline” pattern.
• Caused by Wikipedia’s quality control

Comparing many communities

Looking at many communities helps to identify patterns

Project 4 - Combining computational and qualitative approaches

How do different communities characterize risk?

• Project led by Tiwalade Adekunle
• Get comment data from pro- and anti-mask subreddits
• Use NLP to identify salient topics
• Gain qualitative understanding of topics and identify other insights
• Develop and test hypotheses

Further reading

Thanks!