Content uploaded by Giorgia Michelini
All content in this area was uploaded by Giorgia Michelini on Mar 22, 2020
Content may be subject to copyright.
Changes in risk perception and protective behavior during the first week
of the COVID-19 pandemic in the United States
Toby Wise1,2,3, Tomislav Zbozinek1, Giorgia Michelini4, Cindy C Hagan1 & Dean Mobbs1,5
1Division of Humanities and Social Sciences, California Institute of Technology, Pasadena, CA
2Max Planck UCL Centre for Computational Psychiatry and Ageing Research, University College London, London, UK
3Wellcome Centre for Human Neuroimaging, University College London, London, UK
4Semel Institute for Neuroscience and Human Behavior, University of California Los Angeles, Los Angeles, CA
5Computational Neural Systems Program, California Institute of Technology, Pasadena, CA
By mid-March 2020, the COVID-19 pandemic spread to over 100 countries and all 50 states in the
US. Government efforts to minimize the spread of disease emphasized behavioral interventions,
including raising awareness of the disease and encouraging protective behaviors such as social
distancing and hand washing, and seeking medical attention if experiencing symptoms.
However, it is unclear to what extent individuals are aware of the risks associated with the
disease, how they are altering their behavior, factors which could influence the spread of the
virus to vulnerable populations. We characterized risk perception and engagement in
preventative measures in 1591 United States based individuals over the first week of the
pandemic (March 11th-16th 2020) and examined the extent to which protective behaviors are
predicted by individuals’ perception of risk. Over 5 days, subjects demonstrated growing
awareness of the risk posed by the virus, and largely reported engaging in protective behaviors
with increasing frequency. However, they underestimated their personal risk of infection relative
to the average person in the country. We found that engagement in social distancing and
handwashing was most strongly predicted by the perceived likelihood of personally being
infected, rather than likelihood of transmission or severity of potential transmitted infections.
However, substantial variability emerged among individuals, and using data-driven methods we
found a subgroup of subjects who are largely disengaged, unaware, and not practicing protective
behaviors. Our results have implications for our understanding of how risk perception and
protective behaviors can facilitate early interventions during large-scale pandemics.
The genesis of the novel coronavirus epidemic (spread of COVID-19 disease) has been tied to the Hubei
province of China and rapidly progressed to the level of a global pandemic, with multiple countries
across the globe reporting exponentially increasing numbers of cases (1). The first case in the US was
reported in January 14 2020 (2), followed by government interventions in travel restrictions. On March
11, however, COVID-19 officially become an global pandemic (3) and the introduction of a series of
governmental decisions to restrict social and economic behavior began. By March 17, all 50 states
reported at least one person with the virus (2). Like most developed countries, a major focus of the US
has been minimizing transmission of the virus in order to flatten the epidemic peak and lessen the
impact on healthcare services (4,5), enabling the most severe cases to be treated successfully and reduce
overall mortality. The success of these measures is particularly critical in the case of COVID-19 due to
its high transmissibility, even in the absence of symptoms (6,7), severity (4), and mortality rate, particular
among older individuals (5). However, these protective measures rely largely on rapid changes in
population behavior, which are dependent on individuals’ ability to perceive risks associated with the
virus and adapt their behavior accordingly (8).
Given the importance of human psychological and behavioral factors in managing pandemics, it is
crucial to assess psychological and behavioral responses to the situation and determine how perceived
risk is linked to engagement in protective behaviors (9). There is limited evidence on reactions to prior
pandemics in the early stages when preventative measures are most necessary (9). While some studies
have emphasized the role of risk perception, predominantly the personal effects of the disease (in terms
of likelihood and severity if infection for the individual), on preventative behaviors, these often take
place either in anticipation of an outbreak or long after its emergence (9). In addition, lab-based research
has suggested that increased perceived effects of disease spread on others may increase engagement
in social distancing (10). The few studies that have surveyed individuals during the early stages of a
pandemic have also suggested that perceived personal risk of infection and health effects are linked to
engagement in protective behaviors (11). However, it is also well established that individuals typically
tend to underestimate their likelihood of experiencing adverse life events (such as cancer) relative to
the average person, an effect known as optimism bias (12). Together, it is apparent that perceived risk
is likely to affect individuals’ behavior during a pandemic, but that individuals are often poor at
perceiving risk. However, it is unknown how perceived risk relates to protective behaviors in the early
stages of a pandemic on the scale of COVID-19. Additionally, we are unaware of any data of this kind
for the current COVID-19 pandemic; given COVID-19’s ongoing rampant nature, this data may have
global value to the medical community, government leaders, and society more broadly.
Figure 1. Timeline of events early in the United States COVID-19 pandemic. Days of current study data acquisition shown in gray.
News events in green are most relevant for United States. COVID-19 data acquired from European Centre for Disease Prevention
and Control. Major news events retrieved from National Broadcasting Company (NBC) News, Columbia Broadcasting System (CBS)
News, and Cable News Network (CNN).
We conducted an online survey of 1591 individuals in the USA during the early stages of the country’s
outbreak in March 2020, asking about their perceptions of risk and behavioral responses to the
pandemic (see Figure 2 for demographic information). Subjects were recruited through Prolific (13)
between 3/11/20, the day when the WHO declared COVID-19 a pandemic, and 3/16/20. The study was
approved by the Institutional Review Board (IRB) at California Institute of Technology. We focused on
how perceived risk from the virus and propensity to engage in protective behaviors developed as the
pandemic progressed. We also sought to quantify the extent to which engagement in protective
behaviors was dependent on perceived risk. 375 subjects of the 495 who participated on the first day
were followed up after 5 days to provide a picture of within-person changes in perceptions and
behaviors. We also replicated cross-sectional results in a subset of our data to ensure robustness.
Figure 2. Demographics of survey respondents. The red dotted line on the lower panel represents the median age (30 years).
Perceptions of risk from COVID-19
While at the time of submission it remains unclear exactly how widespread the pandemic will be, current
estimates suggest that up to 80% of the population may contract the disease (4). We sought to
characterize perceptions of infection likelihood and severity, for both the study participants themselves
and others. All responses were recorded on a visual analogue scale coded between 0 and 100. We also
examined changes in behavior over time by sampling independent groups of subjects over five days
and retesting subjects who participated on the first day after a five-day period. Despite being a short
period of time, multiple significant political events occurred during this time period, including travel
bans and restrictions on public gatherings (Figure 1).
Figure 3. Distributions of responses to items regarding risk perception (n=1591). All responses were recorded on a visual analogue
scale ranging from 0 to 100. Bar plots indicate mean responses to these items over the two timepoints where a subgroup of
subjects was retested (n=375).
As shown in Figure 3, subjects assessed their risk of being infected as relatively high (mean = 43.06, SD
= 26.62). Additionally, they reported perceiving the disease as being a threat to their health (mean =
44.70, SD=26.93). They also indicated that they would be personally affected economically, such as
through loss of work (mean = 45.68, SD = 34.35), and that they would be affected by the global
economic consequences, such as through economic recession and effects on healthcare provision
(mean = 64.38, SD = 24.02), although responses to this question were not unimodally distributed..
Subjects were also aware of the potential for contagion, indicating that if they became infected, they
would be likely to pass it to someone else (mean = 66.18, SD = 27.39, Figure 3). As with perceptions of
infection likelihood, subjects believed that if they did infect another person, they would be worse
affected than themselves, both in terms of health and of economic effects (mean difference = 14.82, SD
difference = 26.67). Linear regression indicated that the difference between perceived effects on another
person and reported personal health risk was partially dependent on age (t(1550) = -8.33, p < .001),
suggesting that this may be explained, in part, by the relatively young age of the participants (median
age=30 years) and knowledge of the worse health effects in older individuals. However, the intercept in
this model remained positive and significant (β=29.68, p < .001), indicating the presence of such a bias
even after accounting for age.
Perceived likelihood of infection however differed according to who participants were rating (F(3, 4737)
= 579.00, p < .001, 𝜂𝑝
2 = 0.27), with participants rating the average person in the US to have the highest
risk of infection, but themselves to have the lowest risk, in line with work on optimism bias (12) (Figure
4C). Perceived likelihood of infection differed across samples tested on different days, demonstrating a
higher rate over time (F(6, 1579) = 6.48, p < .001, 𝜂𝑝
2 = 0.024, Figure 4C). An increase in perceived
likelihood was found within-subjects in a subsample followed up after 5 days (F(1, 374) = 69.19, p <
2 = 0.16, Figure 4D). There was an interaction between time and subject of rating (F(3, 1122) =
7.56, p < .001, 𝜂𝑝
2 = 0.02), representing the greatest changes in risk perception for the self, however this
was weak and likely influenced by ceiling effects.
Figure 4. Changes in protective behaviors and risk perception over time. A) Reported likelihood of attending events with a given
number of other people in separate samples tested on 5 days in the early stages of the outbreak in the United States. B) Reported
likelihood of attending events of different sizes in a subset of subjects followed up 5 days after initially completing the survey. C)
Perceived likelihood of becoming infected for participants themselves and average people at different geographic scales in
separate samples tested over 5 days. D) Perceived likelihoods of infection in a subset of subjects followed up after 5 days.
Engagement in protective behaviors
We next assessed the extent to which subjects reported engaging in protective behaviors, such as social
distancing and hand washing, in addition to superficially helpful behaviors such as buying more food
and water. On average, subjects indicated that they were engaging in such behaviors more than usual,
although response distributions included peaks at the extremes (Figure 5). Five out of six protective
behaviours had a peak for not engaging in the protective behaviour more than normal, and three out
of six had a peak for engaging in the protective behaviour more than normal. In particular, subjects
reported washing their hands more than normal (median = 77, IQR = 38) and staying home more than
normal (median = 62, IQR = 69), representing high engagement with sanitization and social distancing
measures. In subjects who completed the survey a second time point 5 days after first completion
(3/11/2020), responses had changed for both hand-washing (Wilcoxon W(375) = 25027.5, p < .001) and
social distancing (W(375) = 12269, p < .001) to reflect increased engagement in these behaviors.
We also asked people how likely they would be to attend events with varying numbers of people (10 to
1000) to assess how they were adapting their behavior according to transmission risk. As expected, we
observed a main effect of group size (F(4, 6316) = 1311.68, p < .001, 𝜂𝑝
2 = 0.45, Figure 4A), whereby
individuals were less likely to attend an event with more people. We also saw markedly lower likelihood
ratings over time in separate samples collected across multiple days (F(6, 1579) = 22.84, p < .001, 𝜂𝑝
0.08, Figure 4A). Congruently, a decrease over time emerged in our within-subject analysis (F(1, 374) =
279.02, p < .001, 𝜂𝑝
2 = 0.43, Figure 4B), providing evidence that individuals reported dramatically
changing their intended behavior within the space of only a few days. Notably, this occurred before and
after the CDC’s recommendation of avoiding gatherings of 50+ people (3/15/20) and on the same day
of President Trump’s announcement to avoid gatherings of 10+ people (3/16/20).
Figure 5. Distributions of responses to items regarding behavior (n=1591). All responses were recorded on a visual analogue scale
ranging from 0 to 100. Bar plots indicate mean responses to these items over the two timepoints where a subgroup of subjects
was retested (n=375).
Influence of risk perception on protective behaviors
We next investigated the extent to which risk perception was predictive of engagement in protective
behaviors. We used multiple linear regression to assess the extent to which of our 10 items assessing
risk perception (shown in Figure 6) were associated with engagement in two primary protective
behaviors, hand washing and social distancing (assessed through asking subjects whether they were
staying home more than normal), controlling for age. We performed this analysis in a subset consisting
of 75% of participants and repeated it in the remaining 25% to ensure reproducibility of our results.
Results reported here are from the larger dataset but were consistent across both subsets (Figure 6). All
data were scaled to zero mean and unit variance prior to analysis to allow comparability of regression
The clearest effect common to both behaviors was a significant effect of perceived likelihood of
personally becoming infected (hand washing β = 0.17, p < .001, social distancing β = 0.20, p < .001,
Figure 6), while perceived severity of illness was not a significant predictor (hand washing β = -0.03, p =
.37, social distancing β = 0.002, p = .95, Figure 6). Perceived impact from global consequences (e.g.
economic recession, healthcare overcapacity) also significantly predicted engagement in both behaviors
to a lesser extent (β=0.08, p = .01, social distancing β=0.14, p < .001, Figure 6). Notably, likelihood of
passing the virus on to others and perceived negative effects for another individual who contracted the
virus did not significantly predict behavior (Figure 6). Age did not have a significant effect in either
Figure 6. Results of linear regression predicting engagement in hand washing and social distancing (represented by responses to
an item regarding staying home) from measures of risk perception, with validation in a subsample of 25% of subjects. A represents
the discovery dataset and B represents results from the validation dataset.
Identification of subgroups demonstrating low engagement in protective behavior
The distributions shown in Figure 5 clearly indicate that the pattern of responses to questions on
protective behaviors was not Gaussian, and was not consistently unimodal, suggesting that there are
likely to be subgroups of individuals responding to the outbreak in qualitatively different ways. To
explore this further, we used a Bayesian Guassian mixture model (GMM) to decompose the distribution
of responses to four primary questions (avoiding social interaction, hand-washing, staying home, and
travelling less) into latent components. The Bayesian GMM approach assigns weights to these
components and we rejected any with a weight below 0.01 as these had a negligible contribution to the
model, leaving 16 components as the final solution (Figure 7B). Based on the mean response scores of
the components, two components (components 4 and 6) were characterised by high and very low
reported engagement with the four protective behaviours respectively (Figure 7A). Others indicated that
there were clusters of individuals selectively engaging in certain protective behaviours but not others
(components 3, 10, and 16 for example).
The model allowed us to assign a probability of each subject being described by each component, which
we used to select individuals most likely to belong to the low or high engagement cluster. Having
labelled individuals according to their behaviour, we then assessed Z-scored responses to other items
to examine how these individuals compared to the group average in terms of percevied risk, information
seeking, and personal effects of the pandemic (Figure 7C). This revealed a broad pattern of below
average perceived risk for both themselves (mean Z = -0.68) and others (mean Z = -0.38), perceived
likelihood of transmission (mean Z = -0.28), low engagement with information sources (mean Z = -0.89),
and low perceived personal effects (mean Z = -0.66), while the opposite pattern was observed in the
high engagement group. Significant differences from the group average are shown in Figure 3C.
Together, this indicates that there exists a subgroup in the population who are generally disengaged in
terms of information seeking, feel unaffected by the situation, and perceive the risk of COVID-19 as
being low for themselves and others, and who do not engage in protective behaviours.
Figure 6. Results of Bayesian Gaussian Mixture Model (GMM) decomposing response distributions for protective behavior items
into clusters. A) Mean scores for each component in the GMM model on the four items used to generate clusters. B) Weights of
retained components. Four components were rejected due to having negligible weights (< .01). C) Z-scored responses on other
questionnaire items for the low engagement and high engagement clusters, demonstrating how they compare to the average
individual. Asterisks represent significant differences from the group average (one sample t-test on the Z scores versus zero, FDR
corrected for 18 comparisons).
Understanding how psychological factors influence behavior in severe, global pandemics such as that
COVID-19 is key to facilitating disease minimization strategies. Our analyses indicate that although most
individuals are aware of the risk caused by the pandemic to some extent, they typically underestimate
their personal risk relative to that of others, an example of optimism bias (12). In turn, higher perceived
personal risk predicts engagement in protective behaviors such as hand washing and social distancing,
as shown in studies of prior pandemics (9). Notably, we identified and characterized a non-negligible
subset of subjects reporting little to no engagement in protective behaviors, who rated overall likelihood
of infection as low and reported being generally disengaged in information seeking and being
personally unaffected. Overall, the presence a subgroup is concerning given the threat posed by COVID-
19 and the beneficial effects of widespread behavioural changes.
One explanation for our results is the optimism bias (12). This bias is associated with the belief that we
are less likely to acquire a disease than others, and has been shown across a variety of diseases including
lung cancer (14). Indeed, those who show the optimism bias are less likely to be vaccinated against
disease (15). Recent evidence suggests that this may also be the case for COVID-19 and could result in
a failure to engage in behaviors that contribute to the spread this highly contagious disease. Our results
extend on these findings by showing that behavior changes over the first week of the COVID-19
pandemic such that as individuals perceive an increase in personal risk they increasingly engage in risk-
prevention behaviors. Notably, we observed rapid increases in risk perception over a 5-day period,
indicating that public health messages spread through government and the media can be effective in
raising awareness of the risk. This effect was strongest for perceptions of subjects’ own risk, diminishing
the optimism bias. The speed at which perceptions changed is such that this could have a meaningful
effect in terms of reducing disease transmission.
Our results point to candidate targets for intervention in public information campaigns during
pandemics on this scale. Clear communication of risk could aid the development of accurate risk
perception, in turn facilitating engagement in protective behaviors. It would be particularly important
to target the subset of individuals who remain disengaged and are not themselves seeking information
on the pandemic. This suggests the need to expand outreach methods to individuals who do not seek
information themselves (e.g., emergency alerts on phones). Furthermore, such disengagement should
be considered in epidemiological models used to forecast the effects of behavioral interventions on
disease spread. Additionally, education on the beneficial effects of such behaviors for others may
improve engagement, particularly in those at low perceived personal risk; it is possible that links
between protective behavior and perceived personal risk minimization are merely easier to appreciate.
There are limitations to our work that should be considered. First, the median age (30 years) of our
sample is relatively young. However, many of our results do not appear to be highly dependent on age,
for example age was not a significant predictor of hand-washing or social distancing. In addition, young
people are typically the primary target of efforts to encourage social distancing, having on average
larger social networks (16) and therefore a higher likelihood of engaging in social contact. This is
particularly important in the context of COVID-19, where there is evidence that the spread of the virus
has been facilitated by the movement of young people with limited to no symptoms (5,6). Second, our
data only reflects views of those in the United States and may not be applicable to other cultures. It will
be important to characterize psychological and behavioral responses across the globe during
pandemics in order to recommend and implement the most optimal strategies for effecting behavioural
change, which often are culturally specific.
Adaptation of behavior will be fundamental to the management of a pandemic on the scale of COVID-
19. Our results provide insights into key psychological and behavioral states during a crucial time in the
This work was supported by US National Institute of Mental Health grant 2P50MH094258 and a Chen
Institute Award (P2026052) (support to D.M.). TW is supported by a Wellcome Trust Sir Henry
Wellcome Fellowship. TZ is supported by the National Science Foundation (#1911441).
1. World Health Organisation. Coronavirus [Internet]. 2020 [cited 2020 Mar 17]. Available from:
2. CDC. Coronavirus Disease 2019 (COVID-19) in the U.S. [Internet]. Centers for Disease Control
and Prevention. 2020 [cited 2020 Mar 19]. Available from:
3. WHO Director-General’s opening remarks at the media briefing on COVID-19 - 11 March 2020
[Internet]. [cited 2020 Mar 19]. Available from: https://www.who.int/dg/speeches/detail/who-
4. COVID-19 reports | Faculty of Medicine | Imperial College London [Internet]. [cited 2020 Mar
17]. Available from: https://www.imperial.ac.uk/mrc-global-infectious-disease-analysis/news--
5. Lai C-C, Shih T-P, Ko W-C, Tang H-J, Hsueh P-R. Severe acute respiratory syndrome coronavirus
2 (SARS-CoV-2) and coronavirus disease-2019 (COVID-19): The epidemic and the challenges.
International Journal of Antimicrobial Agents. 2020 Mar 1;55(3):105924.
6. Bai Y, Yao L, Wei T, Tian F, Jin D-Y, Chen L, et al. Presumed Asymptomatic Carrier Transmission
of COVID-19. JAMA [Internet]. 2020 Feb 21 [cited 2020 Mar 17]; Available from:
7. Li R, Pei S, Chen B, Song Y, Zhang T, Yang W, et al. Substantial undocumented infection
facilitates the rapid dissemination of novel coronavirus (SARS-CoV2). Science [Internet]. 2020
Mar 16 [cited 2020 Mar 17]; Available from:
8. Xu J, Peng Z. People at Risk of Influenza Pandemics: The Evolution of Perception and Behavior.
PLoS ONE. 2015;10(12):e0144868.
9. Bish A, Michie S. Demographic and attitudinal determinants of protective behaviours during a
pandemic: A review. British Journal of Health Psychology. 2010;15(4):797–824.
10. Kappes A, Nussberger A-M, Faber NS, Kahane G, Savulescu J, Crockett MJ. Uncertainty about
the impact of social decisions increases prosocial behaviour. Nat Hum Behav. 2018
11. Liao Q, Wu P, Lam WWT, Cowling BJ, Fielding R. Trajectories of public psycho-behavioural
responses relating to influenza A(H7N9) over the winter of 2014-15 in Hong Kong. Psychology
& Health. 2019 Feb 1;34(2):162–80.
12. Sharot T. The optimism bias. Current Biology. 2011 Dec 6;21(23):R941–5.
13. Palan S, Schitter C. Prolific.ac—A subject pool for online experiments. Journal of Behavioral and
Experimental Finance. 2018 Mar 1;17:22–7.
14. Brnstrm R, Brandberg Y. Health Risk Perception, Optimistic Bias, and Personal Satisfaction
[Internet]. 2010 [cited 2020 Mar 19]. Available from:
15. Brewer NT, Chapman GB, Gibbons FX, Gerrard M, McCaul KD, Weinstein ND. Meta-analysis of
the relationship between risk perception and health behavior: The example of vaccination.
Health Psychology. 2007;26(2):136–45.
16. Wrzus C, Hänel M, Wagner J, Neyer FJ. Social network changes and life events across the life
span: A meta-analysis. Psychological Bulletin. 2013;139(1):53–80.