Giuliano and Rasul, 2020), as well as forecast disease spread (see e.g. Bengtsson et al., 2015;
Wesolowski et al., 2012, 2015; Peixoto et al., 2020). A third strand uses crowdsourced infor-
mation, including surveys, to monitor disease symptoms and detect potential outbreaks (see
Facebook Symptom Survey; Smolinski et al., 2015; Paolotti et al., 2014).
In comparison to this literature, our stable network-based measure is less likely to suffer
from changes in internet behavior or seasonality, both of which have hampered Google Flu
Trends (Olson et al., 2013). In addition, our measures do not require individuals to have
experienced symptoms, which potentially allows us to identify at-risk localities before disease
transmission.
1
Finally, because our measures are based only on aggregated connections
(instead of individual movement), they are easily accessible to researchers and consistently
available for a large number of geographies around the world.
More generally, our results add to a literature that has applied aspects of network the-
ory to build spatial epidemiological models (for overviews, see Keeling and Eames, 2005;
Keeling and Rohani, 2011; Danon et al., 2011). Works in this literature move beyond the
basic assumption that individuals within a population are “fully mixed”, or equally likely
to interact; instead, they better represent the dynamics of real-world connections (see e.g.
Newman, 2002; Klovdahl, 1985; Klovdahl et al., 1994; Mossong et al., 2008; Yang et al.,
2020). While some of these studies parameterize models with information on local networks,
we are unaware of any that introduces a measure with comparably high levels of coverage
and granularity.
2
Our hope is that our unique measure of social connectedness can help pa-
rameterize future epidemiological work. In addition, we hope that the Social Connectedness
Index can advance the literature on the determinants and effects of urban and regional social
networks (see Bailey et al., 2020a; Kim et al., 2017; B¨uchel and von Ehrlich, 2016; Mossay
and Picard, 2011; Brueckner and Largey, 2008; Glaeser et al., 1992).
It is important to note that our objective in this paper is not to incorporate social
connectedness data into a state-of-the-art epidemiological model. Instead, we provide a
unique measure to assess regions’ outbreak risk, answering the call of Avery et al. (2020),
among others, who highlight the “urgent need” for “creative and entrepreneurial methods”
of interpreting and sharing data to model coronavirus spread. To that end, the data in this
paper, as well as similar data for a wide range of other geographies, are accessible by emailing
1
However, it suggests that our data might partner well with these measures. For example, if one can
detect an early outbreak using surveys, they could then predict (and potentially prevent) the next outbreak
using information on social connectedness.
2
For example, the Social Connectedness Index is available at the ZCTA level in the U.S., the NUTS3
level in Europe, the GADM2 level in the Indian Subcontinent, and the GADM1 level throughout much of
the rest of the world.