Comparing methods for creating a national random sample of twitter users

Area of Science:

• Social Sciences
• Computer Science
• Data Science
• Computational Social Science

Background:

• Twitter data is extensively utilized in social and computer science research.
• A frequent research objective involves creating representative samples of users from specific countries.
• Existing methods for sampling Twitter users lack comprehensive comparative analysis.

Purpose of the Study:

• To implement and compare four common methods for generating random Twitter user samples in the US: 1% Stream, Bounding Box, Location Query, and Language Query.
• To evaluate these methods based on tweet- and user-level metrics and their accuracy in estimating the US population.

Main Methods:

• Implementation of four distinct sampling strategies: 1% Stream, Bounding Box, Location Query, and Language Query.
• Comparative analysis of collected data using tweet-level metrics (e.g., tweet frequency) and user-level metrics (e.g., followers, account age).
• Assessment of population estimation accuracy for each sampling method.

Main Results:

• The 1% Stream method yielded users with higher tweet volume, engagement metrics, and younger accounts, while also showing the lowest population estimation error.
• Users sampled via 1% Stream were disproportionately male compared to other methods.
• The 1% Stream method is time-intensive, limited to real-time data, and unsuitable for studies focusing on user engagement.

Conclusions:

• The 1% Stream method is superior for obtaining a statistically accurate US Twitter user sample for population estimation.
• The Bounding Box method emerges as a viable alternative when the limitations of the 1% Stream (time, historical data, engagement focus) are critical research constraints.