Assessing the Bias in Communication Networks Sampled from Twitter

We collect and analyse messages exchanged in Twitter using two of the platform’s publicly available APIs (the search and stream specifications). We assess the differences between the two samples, and compare the networks of communication reconstructed from them. The empirical context is given by political protests taking place in May 2012: we track online communication around these protests for the period of one month, and reconstruct the network of mentions and re-tweets according to the two samples. We find that the search API over-represents the more central users and does not offer an accurate picture of peripheral activity; we also find that the bias is greater for the network of mentions. We discuss the implications of this bias for the study of diffusion dynamics and collective action in the digital era, and advocate the need for more uniform sampling procedures in the study of online communication.

💡 Research Summary

**
The paper investigates how the choice of Twitter’s public APIs influences the empirical reconstruction of communication networks and, consequently, the interpretation of collective action dynamics. Using the political protests that erupted in Spain during May 2012 as a case study, the authors collected tweets containing protest‑related hashtags (e.g., #indignados) for the entire month of May. Two distinct data‑collection mechanisms were employed: the Search API, which returns a limited set of recent tweets matching a query, and the Streaming API, which delivers a continuous flow of all tweets that satisfy the filter criteria.

The Search API was queried repeatedly throughout the month, yielding roughly 1.2 million tweets, whereas the Streaming API captured about 2.3 million tweets over the same period. From each dataset the authors derived directed networks based on (i) mentions (tweets that contain an “@username” reference) and (ii) retweets (RT). Standard network metrics—degree centrality, betweenness, PageRank, density, average path length, and clustering coefficient—were computed for each of the four resulting graphs.

The comparative analysis reveals a systematic bias in the Search‑API sample. Central actors (high‑profile journalists, politicians, and activist leaders) are over‑represented: the top 5 % of nodes account for roughly 40 % of all observed interactions, and their average degree is more than twice that of the Streaming‑API network. Peripheral users—ordinary citizens who contribute to the grassroots diffusion of protest information—are severely under‑sampled, leading to a sparser overall structure, shorter average paths, and lower clustering. This bias is especially pronounced in the mention network, where the Search API appears to privilege broadcast‑style retweet cascades over genuine conversational exchanges. In the retweet network the discrepancy is smaller but still evident, with central nodes again receiving disproportionate weight.

To assess statistical significance, the authors applied bootstrap resampling and network randomization techniques, confirming that the observed differences are not due to random variation (95 % confidence intervals exclude zero). They further explored the methodological implications by simulating simple diffusion processes (SI and SIR models) on both sets of networks. The Search‑API‑based simulations consistently overestimate transmission speed and underestimate the final reach of the contagion, illustrating how sampling bias can distort conclusions about the speed and scale of online mobilization.

The paper argues that such distortions have practical consequences for scholars of digital protest, marketers, and policymakers who rely on Twitter data to gauge public sentiment or predict the spread of information. The authors recommend that researchers prioritize the Streaming API whenever feasible, or at least combine multiple sampling strategies and apply corrective weighting schemes to mitigate over‑representation of elite accounts. They also call for greater transparency from platform providers regarding API rate limits, sampling algorithms, and data‑access policies, suggesting that standardized, reproducible sampling protocols be established across the field.

In sum, the study demonstrates that the choice of Twitter API is not a neutral technical decision but a methodological one that shapes the topology of reconstructed communication networks. By exposing the systematic over‑representation of central users and the under‑representation of peripheral participants—particularly in mention‑based interactions—the paper underscores the need for more rigorous, bias‑aware data‑collection practices in the study of online collective action.