Validation of Dunbars number in Twitter conversations

Modern society’s increasing dependency on online tools for both work and recreation opens up unique opportunities for the study of social interactions. A large survey of online exchanges or conversations on Twitter, collected across six months involving 1.7 million individuals is presented here. We test the theoretical cognitive limit on the number of stable social relationships known as Dunbar’s number. We find that users can entertain a maximum of 100-200 stable relationships in support for Dunbar’s prediction. The “economy of attention” is limited in the online world by cognitive and biological constraints as predicted by Dunbar’s theory. Inspired by this empirical evidence we propose a simple dynamical mechanism, based on finite priority queuing and time resources, that reproduces the observed social behavior.

💡 Research Summary

The paper investigates whether the well‑known cognitive limit on stable social relationships—Dunbar’s number, traditionally estimated at 100–200 individuals—holds in the modern, highly connected environment of Twitter. Using a massive dataset collected from Twitter’s firehose over a six‑month period (November 2008 to May 2009), the authors retrieved the complete tweet histories of three million accounts, amassing more than 380 million tweets spanning roughly four years. By focusing on the “reply‑to” metadata, they reconstructed conversation trees (a forest of >25 million trees) and projected these onto a directed, weighted user‑user network comprising 1.7 million unique users and 68 million weighted edges. In this network, the out‑degree (k_out) of a node counts the number of distinct users to whom a person has replied, while the in‑degree (k_in) counts distinct users who have replied to them.

To differentiate casual contacts from stable relationships, the authors introduced a weight ω_ij for each directed edge, defined as the total number of replies sent from user i to user j. They then computed for each user an average interaction strength per outgoing connection: ω_out(i) = Σ_j ω_ij / k_out(i). Plotting ω_out against k_out revealed a clear saturation: ω_out rises with the number of contacts up to roughly 150–200 contacts, after which it plateaus or even declines. This pattern mirrors the hypothesized cognitive ceiling: beyond a certain number of acquaintances, individuals cannot allocate sufficient attention and time to maintain the same level of interaction per contact. A similar saturation was observed for the number of reciprocated connections (ρ) as a function of k_in, which flattens once incoming contacts exceed about 200–300, indicating that additional inbound demands are not met with proportional replies.

To explain these empirical findings, the authors proposed a minimalist dynamical model. They start with a static scale‑free network (degree distribution P(k) ∝ k^‑γ, γ≈2) of N=10⁵ nodes. Each node possesses a finite‑capacity priority queue that can hold at most q_max,i messages at any time step. Incoming messages are queued with a priority proportional to the sender’s total degree (a proxy for social prominence). At each discrete time step, a node processes a random number S_t of messages, bounded by the current queue length and q_max,i, replying preferentially to higher‑priority messages. Each reply increments the corresponding edge weight ω_ij by one. After processing, replied messages are removed, new incoming messages are added up to the queue limit, and any excess is discarded. With a small broadcast probability p, nodes may also generate a status‑type message to all contacts, which does not affect edge weights.

Simulations were run for T=2×10⁴ steps, averaging results over at least 10³ independent runs. By varying q_max (from 50 to 300) the model reproduced the empirical ω_out vs. k_out curve: the peak of ω_out shifted linearly with q_max, confirming that the “attention queue capacity” is the primary driver of the observed saturation. Sensitivity analyses showed that changes in broadcast probability p, total simulation time T, or heterogeneity in q_max (Gaussian distribution with σ=10) did not qualitatively alter the outcome.

The authors discuss several implications. First, the persistence of Dunbar‑type limits in an online micro‑blogging platform suggests that digital tools do not fundamentally expand human cognitive bandwidth; they merely change the medium of interaction. Second, the model’s reliance on a simple queue‑capacity constraint and degree‑based priority captures essential features of human attention allocation without invoking complex psychological mechanisms. Third, the study highlights the importance of distinguishing between superficial network ties (e.g., follower counts) and genuine interactive relationships when analyzing online social structures.

Limitations are acknowledged. The dataset reflects Twitter usage patterns from 2008‑2009, before algorithmic timelines, widespread retweet culture, and the introduction of direct messages, which may affect contemporary interaction dynamics. Moreover, the analysis excludes private communications and offline interactions, potentially underestimating the true size of stable social circles. The model also assumes homogeneous decision rules across users, whereas real individuals employ diverse strategies for time management, emotional regulation, and content prioritization.

In conclusion, the paper provides robust empirical evidence that the cognitive ceiling on stable social relationships manifests in Twitter conversations, with a maximum of roughly 100–200 active contacts. The proposed priority‑queue model offers a parsimonious mechanistic explanation rooted in limited attention and time resources. These findings have relevance for the design of social media platforms, the study of information diffusion, and broader theories of human social cognition in the digital age.