Social distancing in networks: A web-based interactive experiment
GAICIL Insights
Every crisis policy is an experiment with no control group.1 Governments pick an instrument, impose it on everyone, and never learn what the alternative would have done. Social distancing in networks builds the missing control group. During the first wave of the COVID-19 pandemic, the authors ran a framed web-based interactive experiment on a near-representative sample of the US population and pitted two favorite instruments against each other: fines and informational messages, or nudges. Fines promote distancing. Nudges have a smaller and less robust impact, with an estimated effect about half in magnitude and not robust to all specifications.
The design is the hidden star. Participants interacted in networks, so each choice to distance changed the spread of infection to others. That captures what mobility records and surveys struggle to isolate: behavior that depends on what everyone else does. One finding stands out. Individuals did more social distancing when they knew they were a superspreader. Telling people where they sit in the network changed what they did, a hint that targeted, network-aware information can accomplish what generic messaging cannot.
Swap the virus for an algorithm and the template fits artificial intelligence almost perfectly. The hard questions about AI deployment are networked questions under uncertainty. How do users react to warnings about synthetic content? How do firms respond to disclosure duties? How do platforms respond to penalties? The authors state that the methodology can play an important role in informing policymakers and can complement data obtained from sociomobility and survey-based studies. Framed interactive experiments, online, at scale and in real time, let regulators try an instrument on a representative sample before imposing it on a whole population. Call it a wind tunnel for rules.
The fines-versus-nudges result carries an uncomfortable message for soft governance. Nudges are popular because they look cheap and respectful of choice. Yet a single informational video emphasizing harm to others moved behavior less, and less reliably, than a monetary penalty. The authors are careful: they test one specific nudge, and nudge efficacy is sensitive to a multitude of factors. They cite evidence from India in which a similar nudge, delivered through 2.5-minute information videos, was effective. Their related study on a Chinese sample found a significant and robust effect of the same informational nudge. The same message, three countries, different results.
That is the real warning for AI policy: instruments do not travel well. AI rules are drafted across very different legal and cultural environments, often by copying what worked elsewhere. The paper suggests that an instrument's effectiveness may depend on national and cultural context, and that demographic characteristics such as age shape responsiveness. Importing a labeling mandate or disclosure regime untested risks importing its costs while leaving its benefits behind at the border.
Politics enters too. Using an instrumental variable approach, the authors find suggestive evidence that progressives are more likely to practice distancing and are somewhat more responsive to fines. Identity shapes compliance alongside incentives. Any AI policy that relies on voluntary cooperation, from content provenance standards to safety reporting, will meet the same split. An instrument that works only on the already willing protects least where protection matters most.
The broader case is for evidence before mandates. Experiments like this are fast, inexpensive and adaptable. This one was run while the crisis unfolded, and such studies can be rerun as conditions shift. For emerging technologies, where regulators face genuine uncertainty about behavioral responses, routine testing on representative samples would cut the risk of rules that look reassuring and deliver little. The gentle-sounding instrument may be the one that fails. Only measurement reveals which.
There is an innovation story buried here as well. Web-based interactive experiments are a research technology, and their value climbs as digital tools make large, representative, real-time samples cheaper to assemble. AI itself can lower those costs further, making policy testing faster still. Better measurement of behavior is a public good: it lets policy target real problems with instruments that demonstrably work, and it leaves space for the experimentation that drives progress.
The takeaway for anyone writing technology rules is blunt. Before choosing between the stick and the message, run the trial. A few weeks of testing can save years of policy that looks good on paper and changes nothing in practice.
Introduction
The outbreak of the COVID-19 pandemic led to an unprecedented rise in the use of social distancing as a means to curb the spread of infection. We examine the effectiveness of fines and informational messages (nudges) in promoting social distancing in a framed web-based interactive experiment conducted during the first wave of the pandemic on a near-representative sample of the US population. Fines promote distancing, while nudges have a smaller and less robust impact. Individuals do more social distancing when they are aware they are a superspreader. Using an instrumental variable approach, we show suggestive evidence that progressives are more likely to practice distancing, and they are somewhat more responsive to fines.
Takeaway
The sudden disruption brought about by the COVID-19 pandemic demanded a quick response from policymakers who had to implement novel social distancing policies whose success determined the fate of tens of thousands of lives. Our work provides experimental evidence on the efficacy of these policies and how it varies across social and individual characteristics. Additionally, it shows how the methodology of framed interactive web-based experiments can play an important role to inform policymakers, and complement data obtained using sociomobility and survey-based studies. Our main finding is that fines are effective at increasing social distancing, while the estimated impact of nudges is about half in magnitude and not robust to all specifications. A limitation of our design is that we examine one specific form of nudge – an informational video that emphasizes the harms to others of failing to practice social distancing. Moreover, the ineffectiveness of nudges in our experiment seems to be at odds with some other evidence in the literature. For instance, Banerjee et al. (2020) show that a similar nudge – in the form of 2.5-minute information videos – was effective in encouraging social distancing in the Indian context.2 However, the efficacy of nudges is clearly sensitive to a multitude of factors, and whatever works in one context may not perform well in another. For example, as we show in Section 4, demographic characteristics, such as age, could contribute to the effectiveness of nudges.1
Further, in a related study conducted on a Chinese sample, we actually find a significant and robust effect of the same informational nudge at increasing compliance with social distancing (Barak et al., 2022).3 This suggests that the effectiveness of nudges may crucially depend on national and/or cultural contexts. There is also a growing literature on the effectiveness of nudges in a range of other contexts. For example, de Ridder et al. (2022) show the importance of personal preferences in moderating the effects of nudges.
Notes
- Edoardo Gallo, Darija Barak & Alastair Langtry, Social distancing in networks: A web-based interactive experiment, 107 J. Behav. & Experimental Econ. 102090 (2023), https://doi.org/10.1016/j.socec.2023.102090. Takeaway source: Concluding remarks, reproduced from the openly licensed version (CC BY 4.0). ↩
- Banerjee, A., M. Alsan, E. Breza, A. G. Chandrasekhar, A. Chowdhury, E. Duflo, P. GoldsmithPinkham, and B. A. Olken (2020). Messages on COVID-19 prevention in India increased symptoms reporting and adherence to preventive behaviors among 25 million recipients with similar effects on non-recipient members of their communities. Working Paper 27496, National Bureau of Economic Research. Baradaran Motie, G. and C. Biolsi (2020). County-level determinants of social distancing (or lack thereof) during the COVID-19 pandemic. Contemporary Economic Policy 39, 264–279.. ↩
- Barak, D., E. Gallo, K. Rong, K. Tang, and W. Du (2022). Experience of the COVID-19 pandemic in Wuhan leads to a lasting increase in social distancing. Scientific Reports 12(1), 18457.. ↩
Cite as: Edoardo Gallo, Darija Barak & Alastair Langtry, Social distancing in networks: A web-based interactive experiment, 107 J. Behav. & Experimental Econ. 102090 (2023), https://doi.org/10.1016/j.socec.2023.102090
Header image: California highway 85 near highway 17 interchange.jpg by Evan0512, CC BY 4.0, via Wikimedia Commons.
