A new Bayesian analysis of telecommuting data supports one of the oldest theories in social networking, with new implications for the future of work environments.

Weak ties describe the rare connections we have with acquaintances, casual colleagues, and occasional friends in our social networks.1. Despite their ‘weakness’, these connections often create new ideas, opportunities and advice in organizational settings1,2,3,4. While this finding has largely stood the test of time, much remains unknown about its causal mechanisms4. What dynamically generates these links? What keeps them going? Can they withstand ‘exogenous’ shocks from outside, such as changes in location motivated by the COVID-19 pandemic? This evolving landscape gives new urgency to calls for “particular insights [social] processes at certain times” through fresh data, methods and analytical rigor5. Now, an article on Computer science of nature answers this call, demonstrating that new computational methods can investigate difficult questions about organizational networks. For their part, Daniel Carmody and colleagues6 use Bayesian time-series analysis to provide evidence supporting an important, understudied theory in social networks called “matching”—which states that spatial proximity increases the chances of forming new ties and strengthening existing ones—in context of COVID-19.

Weak attachment formation often begins when family, community, or organizational activities bring people together. Simply being in physical proximity to someone increases the likelihood of strange interaction. In turn, these interactions give people the opportunity to explore common qualities, interests, and behaviors with each other, and thus form bonds. These steps describe the social process of closeness—the closer we are physically to another person, the more likely we are to form a new bond or rebuild an existing bond with them.7 (Fig. 1a). Existing work has found that sharing socially important qualities can amplify proximity effects8 and this proximity extends to virtual proximity9,10. But the concept is often taken for granted even though it has important implications for how we design organizations and social gatherings. This leaves a surprising lack of evidence showing this process as it happens, and thus we lack knowledge of how the process can improve everything from technology distribution to inequality.4.

Fig. 1: Loss of physical proximity due to remote work caused atrophy of weak ties with nearby researchers.
figure 1

or, Proximity relates the distance between two people (horizontal axis) to the likelihood that those individuals will form a relationship (vertical axis). People who are physically closer to each other are more likely to interact, and therefore form bonds with each other (curve shown). The referenced study provided empirical evidence supporting this social science theory. bThe central finding from the referenced study, showing the change in the number of weak ties between researchers as a function of the distance between their labs from March 2020 to July 2021. Data was collected from the original study6. Statistically significant increases in weak ties are shown in blue; significant declines are shown in orange; and non-significant changes are shown in gray. Error bars represent 95% confidence intervals, and *** indicates a statistically significant finding with p < 0.001 (all other bars had p > 0.1). The chart shows that researchers who once worked next to each other interacted less with each other throughout the COVID-19 pandemic, causing the weak ties between those individuals to disappear. Meanwhile, researchers who worked in the same (remote) lab group strengthened their existing relationships with those individuals and created weaker ties than they would have if they shared physical lab space.

Carmody et al. provided an important empirical demonstration of weak ties as they are formed and degraded by proximity. Most social network studies compare several snapshots of social networks over a time interval, because collecting temporal network data is often quite difficult, both logistically and ethically. Ultimately, this prevents us from witnessing when and how most bonds form. The authors overcome this obstacle by estimating the number of weak ties among researchers at the Massachusetts Institute of Technology (MIT). Their email data includes two dramatic changes in researchers’ workplaces over a year and a half during the COVID-19 pandemic. The first transition took place on March 23, 2020, when MIT stopped most in-person research activities. The researchers began working from home, hypothetically preventing the creation of weak ties through proximity. The second transition took place on July 15, 2021, when scholars began returning to campus, hypothetically increasing the formation of weak ties through proximity. The authors examined the e-mail network that includes these real-world location transitions through a synthetic counterfactual e-mail network with these transitions absent to assess how proximity affected weak link formation.

Methodologically, the authors constructed their synthetic counterfactual through a Bayesian Structural Time Series (BSTS) approach that separates the effect of a treatment (here, telecommuting) from qualities unaffected by the treatment (such as linear trends and cyclical variations). This enabled them to construct a reliable range for the expected number of weak ties with and without telecommuting. Their analysis showed that telecommuting may have cost about 5,100 weak ties during the telecommuting period—about 1.8 ties per person—in exchange, of course, for important public health goals because of the pandemic. Furthermore, researchers were more likely to lose weak ties with people working in nearby laboratories than with those working in the same or distant laboratories (Fig. 1b). Consequently, researchers were ‘stuck’ strengthening their existing connections. As a validation of this finding, they designed a generative network simulation to replicate several link formation mechanisms (such as sharing a lab, mutual friends, and co-location). In doing so, they show qualitatively that a proximity factor replicates the finding of their BSTS analysis.

Carmody and colleagues demonstrated the potential of modern computational techniques to support old social science theories and identify new phenomena. Their research may provide rigorous tools for testing elusive causal hypotheses of social networks and information.4,5. Future studies should consider how other confounders may affect these results. For example, the authors note that insufficient data from before the pandemic limited their ability to predict cyclical effects. This echoes the importance of properly constructing controls for quasi-experiments5. Size and location of the building, student demographics, university-required activities, and type of information exchanged (professional vs. friendly11) can confound the results, but it can also reveal unexplored questions. A word of caution, however: the benefits of computational power become unclear without proper framing from qualitative and theoretical social science5. Naive computational studies run the risk of drawing completely incorrect conclusions. Interdisciplinary collaborations between researchers with computational and theoretical perspectives can begin to answer difficult and long-standing questions, albeit—with patience and curiosity on all sides—in ways that can benefit how we create opportunities for social interaction in the years following.


  1. Granovetter, MS Am. J. Social. 781360–1380 (1973).

    Google Scholar Article

  2. Constant, D., Sproull, L. & Kiesler, S . Organ. Science. 7119-135 (1996).

    Google Scholar Article

  3. Hansen, MT Admin. Science. P. 4482–111 (1999).

    Google Scholar Article

  4. Aral, S. Am. J. Social. 1211931–1939 (2016).

    Google Scholar Article

  5. Davis, GF Organ. Res. Meth. 13690–709 (2010).

    Google Scholar Article

  6. Carmody, D. et al. Nat. Computer. Science. https://doi.org/10.1038/s43588-022-00296-z (2022).

    Google Scholar Article

  7. McPherson, M., Smith-Lovin, L. & Cook, JM the year. Rev. sociol. 27415-444 (2001).

    Google Scholar Article

  8. Reagans, R. Organ. Science. 22835–849 (2011).

    Google Scholar Article

  9. Walther, JB & Bazarova, NN Usual. Res. 35622–645 (2008).

    Google Scholar Article

  10. Lewis, K. & Kaufman, J . Am. J. Social. 1231684–1742 (2018).

    Google Scholar Article

  11. Balkundi, P. & Harrison, DA Acad. Manager. J. 4949–68 (2006).

    Google Scholar Article

Download references

Information about the author

Authors and affiliations

Corresponding author

Correspondence with John Meluso.

Ethics statement

Competing interests

The author declares no competing interests.

About this article

Verify coin and authenticity through CrossMark

Cite this article

Meluso, J. Model supports the social network history theory.
Nat Comput Sci 2, 471–472 (2022). https://doi.org/10.1038/s43588-022-00302-4

Download quote

By admin

Leave a Reply

Your email address will not be published. Required fields are marked *