Ana I. BentopersName, Thuy NguyenpersName, Coady WingpersName, Felipe Lozano-RojaspersName, Yong-Yeol AhnpersName, Kosali SimonpersName (2020date). “Evidence from internet search data shows information-seeking responses to news of local COVID-19 cases.” Proceedings of the National Academy of Sciencestitle·j.
Object of study — collective information-seeking after state announcements of first local COVID-19 cases
Research question — whether and for how long a focal local announcement raises public attention to an emerging crisis
Contribution — rapid event-study evidence using state-level internet searches to trace the onset, content, and decay of collective attentionterm·dom
phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | local official news creates a brief burst of basic information seeking but not broad engagement with coordinated policy responses — “The third row suggests that first case announcements do not induce search for community-level policies (quarantines, school closures, and coronavirus testing) or more-elaborate personal health strategies (face masks, grocery delivery, over-the-counter medications).” | |
| claim-type | mechanism — “Overall, our analysis of internet search data suggests that government information disclosure does help focus public attention on the crisis.” | |
| method — evidence | — “where Announcement i , k is an indicator variable set to 1 if the first case of state i was announced k days ago. The reference category was 1 d prior to the first case announcement, and we created single indicators Announcement i , 31 , for any k ≥ 30 , and Announcement i , − 41 , for any k ≤ − 40 . We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends. SEs allowed for heteroskedasticity and clustering at the state level. Poisson coefficients and CIs are shown for the 40 d leading up to the first case announcement, and up to 20 d following it. The vertical axis measures the differential percentage change in search frequency as states approach the first case announcement ( Figs. 1 and 2 ). We performed regression analysis using Stata, version 16.0 (StataCorp).” | |
| object of study — evidence | — “To provide rapid evidence to inform policy making, we use internet search data in an event study design to examine how collective attention and information-seeking behaviors respond to state government announcements of first COVID-19 cases.” | |
| research question — evidence | — “In settings where restrictions on the free movement of residents are less strictly enforced, it is especially critical for public officials to know whether and for how long statements from leaders motivate individuals to efficiently seek and absorb information, as officials put their communication strategies in place.” | |
| contribution — evidence | — “Using an event study framework for all US states, we show that such news increases collective attention to the crisis right away.” | |
| mechanism | A locally relevant first case makes the pandemic salient and prompts immediate crisis searching. — “There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attention to the pandemic.” | mechanisms[0] |
| mechanism | The attention pulse decays even as stronger policies are announced. — “However, the increased level of information seeking faded within 2 wk, even though many announcements of school closures or other mitigation strategies followed, suggesting that increased attention is only short-lived.” | mechanisms[1] |
| mechanism | Early public discourse limits which mitigation strategies people know to search for. — “This may be because first cases are reported relatively early in the epidemic when many of these more serious mitigation strategies are not yet part of public discourse.” | mechanisms[2] |
| mechanism | Official disclosure focuses public attention toward actions people can take. — “Our evidence also indicates that people seem to mainly react by seeking information on what they can and should do in response to the epidemic.” | mechanisms[3] |
| mechanism | Focal events create a short window in which policymakers can supply actionable guidance. — “One conjecture is that policy makers may be able to take advantage of a brief period of active information seeking following focal events (first cases and first deaths) to provide clear advice regarding the actions people can undertake to avoid risk.” | mechanisms[4] |
| sense | attention - aggregate crisis-focused search behavior following locally salient news — “There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attention to the pandemic.” | senses[0] |
| sense | salience - local relevance conferred by a state's first confirmed case — “There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attention to the pandemic.” | senses[1] |
| sense | control - fixed effects and announcement-timing adjustments that isolate an event response — “where Announcement i , k is an indicator variable set to 1 if the first case of state i was announced k days ago. The reference category was 1 d prior to the first case announcement, and we created single indicators Announcement i , 31 , for any k ≥ 30 , and Announcement i , − 41 , for any k ≤ − 40 . We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends. SEs allowed for heteroskedasticity and clustering at the state level. Poisson coefficients and CIs are shown for the 40 d leading up to the first case announcement, and up to 20 d following it. The vertical axis measures the differential percentage change in search frequency as states approach the first case announcement ( Figs. 1 and 2 ). We performed regression analysis using Stata, version 16.0 (StataCorp).” | senses[2] |
| sense | representation - limits on population inference from one segment of online information seekers — “First, our state-level data from Google Trends database do not allow us to adequately understand intrastate variation of search volumes following the announcements or to generalize conclusions about population search behavior (as Google Trends only captures the search behavior of a particular segment of information seekers).” | senses[3] |
| control | pre-announcement trend inspection — held — “There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attention to the pandemic.” — rules out attributing an existing search trend to the announcement — “There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attention to the pandemic.” | controls[0] |
| control | state fixed effects — held — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” — rules out time-invariant differences in state search patterns — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” | controls[1] |
| control | date fixed effects — held — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” — rules out national search trends shared across states — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” | controls[2] |
| control | other-state and neighboring-state announcement timing — held — “To investigate information spillovers in which states respond to epidemics in other states, we examined augmented models that controlled for the timing of the announcements of Washington, Illinois, California, and Arizona, as well as for the timing of the earliest announcement among a state’s neighbor states [section A of additional results at https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” — rules out information spillovers from earlier outbreaks — “To investigate information spillovers in which states respond to epidemics in other states, we examined augmented models that controlled for the timing of the announcements of Washington, Illinois, California, and Arizona, as well as for the timing of the earliest announcement among a state’s neighbor states [section A of additional results at https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” | controls[3] |
| control | updated Google Trends extraction — held — “We checked whether the findings were driven by accessed date of the data by rerunning the analysis using the latest search output from Google Trends, which we retrieved on April 18, 2020 and compared to the current results using the data retrieved on March 20, 2020 [ https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” — rules out dependence on the original data-access date — “We checked whether the findings were driven by accessed date of the data by rerunning the analysis using the latest search output from Google Trends, which we retrieved on April 18, 2020 and compared to the current results using the data retrieved on March 20, 2020 [ https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” | controls[4] |
| outcome | control — held — “There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attention to the pandemic.” | controls[0] |
| outcome | control — held — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” | controls[1] |
| outcome | control — held — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” | controls[2] |
| outcome | control — held — “To investigate information spillovers in which states respond to epidemics in other states, we examined augmented models that controlled for the timing of the announcements of Washington, Illinois, California, and Arizona, as well as for the timing of the earliest announcement among a state’s neighbor states [section A of additional results at https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” | controls[3] |
| outcome | control — held — “We checked whether the findings were driven by accessed date of the data by rerunning the analysis using the latest search output from Google Trends, which we retrieved on April 18, 2020 and compared to the current results using the data retrieved on March 20, 2020 [ https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” | controls[4] |
| limitation | State aggregates cannot identify intrastate variation in search response. — “First, our state-level data from Google Trends database do not allow us to adequately understand intrastate variation of search volumes following the announcements or to generalize conclusions about population search behavior (as Google Trends only captures the search behavior of a particular segment of information seekers).” | limitations[0] |
| limitation | Google Trends represents only a segment of information seekers. — “First, our state-level data from Google Trends database do not allow us to adequately understand intrastate variation of search volumes following the announcements or to generalize conclusions about population search behavior (as Google Trends only captures the search behavior of a particular segment of information seekers).” | limitations[1] |
| limitation | Anticipatory behavior and unobserved local events may confound event-study estimates. — “The third limitation comes from potential confounding factors in our event study framework, including anticipatory behavior and unobservable local events ( 15 ).” | limitations[2] |
| limitation | The early study period preceded mature policy discourse and later public responses. — “These results are representative of the study period being relatively early in the epidemic, and more-elaborate policy responses were not yet part of the public discourse.” | limitations[3] |
| limitation | The underlying search data are restricted and require an application for access. — “State news data and code are available at https://github.com/anabento/GoogleBehaviorCovid ( 9 ). The search data for this study are available from Google Trends but are restricted in use; researchers may apply to Google Trends API for access.” | limitations[4] |
finding · quantity · variable · definition · key-term — grounded in the full text| finding | Coronavirus searches rose sharply immediately after a state's first-case announcement. — “Searches for “coronavirus” increased by about 36% (95% CI: 27 to 44%) on the day immediately after the announcement but quickly decreased back to the baseline level in less than a week or two ( Fig. 1 ).” | content_tags.findings[0] |
| finding | The announcement-associated increase in information seeking faded within two weeks. — “However, the increased level of information seeking faded within 2 wk, even though many announcements of school closures or other mitigation strategies followed, suggesting that increased attention is only short-lived.” | content_tags.findings[1] |
| finding | Pre-announcement searches showed no observable trend. — “There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attention to the pandemic.” | content_tags.findings[2] |
| finding | Searches for coronavirus symptoms increased after first-case news. — “News of the first COVID-19 case in a state leads to a 52% increase in searches for “coronavirus symptoms” but does not increase searches for coronavirus treatment options.” | content_tags.findings[3] |
| finding | Hand-sanitizer searches increased and remained elevated over the observed period. — “The second row shows searches for “hand sanitizer” increasing 35% immediately after first case announcements, and, unlike in the previous two cases, the search activity remains high for the remainder of the observable period.” | content_tags.findings[4] |
| finding | First-case announcements did not immediately raise searches for community policies or elaborate personal strategies. — “The third row suggests that first case announcements do not induce search for community-level policies (quarantines, school closures, and coronavirus testing) or more-elaborate personal health strategies (face masks, grocery delivery, over-the-counter medications).” | content_tags.findings[5] |
| finding | Later-reporting states showed larger and faster information-seeking responses. — “The results suggest that information-seeking responses were larger and faster in late states compared to the early states.” | content_tags.findings[6] |
| finding | The principal results persisted when the Google Trends extraction date changed. — “These core results were not sensitive to this replication.” | content_tags.findings[7] |
| quantity | 36% — increase in coronavirus searches on the day after the first local-case announcement — “Searches for “coronavirus” increased by about 36% (95% CI: 27 to 44%) on the day immediately after the announcement but quickly decreased back to the baseline level in less than a week or two ( Fig. 1 ).” | content_tags.quantities[0] |
| quantity | 52% — increase in searches for coronavirus symptoms — “News of the first COVID-19 case in a state leads to a 52% increase in searches for “coronavirus symptoms” but does not increase searches for coronavirus treatment options.” | content_tags.quantities[1] |
| quantity | 35% — immediate increase in hand-sanitizer searches — “The second row shows searches for “hand sanitizer” increasing 35% immediately after first case announcements, and, unlike in the previous two cases, the search activity remains high for the remainder of the observable period.” | content_tags.quantities[2] |
| quantity | 51 regions × 78 days — balanced state-day panel from January 1 through March 18, 2020 — “Our analysis is based on a balanced daily panel of COVID-19−related search intensity data in 50 US states and Washington, DC between January 1, 2020 and March 18, 2020 ( n = 51 regions × 78 d). We collected the data using a restricted-access Google Health Trends Application Program Interface (API) account ( 10 ).” | content_tags.quantities[3] |
| quantity | 40 days before and 20 days after — reported event-study window around the first-case announcement — “Poisson coefficients and CIs are shown for the 40 d leading up to the first case announcement, and up to 20 d following it.” | content_tags.quantities[4] |
| variable | first-case announcement (event exposure indexed by state and relative day) — “where Announcement i , k is an indicator variable set to 1 if the first case of state i was announced k days ago.” | content_tags.variables[0] |
| variable | search intensity (outcome measured as query shares scaled per ten million searches) — “The outcome variables measure the daily share of all Google queries in a state that correspond to a particular term during our study period.” | content_tags.variables[1] |
| variable | state fixed effects (adjustments for time-invariant state differences and national trends) — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” | content_tags.variables[5] |
| variable | date fixed effects (adjustments for time-invariant state differences and national trends) — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” | content_tags.variables[6] |
| definition | collective attention — “Because the extent of the epidemic is likely to be underestimated at this stage, it is crucial to know how much the early case announcements induce heightened collective attention and information-seeking behavior.” | content_tags.definitions[0] |
| definition | information seeking — “We find that people respond to the first report of COVID-19 in their state by immediately seeking information about COVID-19, as measured by searches for coronavirus, coronavirus symptoms, and hand sanitizer.” | content_tags.definitions[1] |
| definition | search intensity — “The outcome variables measure the daily share of all Google queries in a state that correspond to a particular term during our study period.” | content_tags.definitions[2] |
| definition | information spillover — “To investigate information spillovers in which states respond to epidemics in other states, we examined augmented models that controlled for the timing of the announcements of Washington, Illinois, California, and Arizona, as well as for the timing of the earliest announcement among a state’s neighbor states [section A of additional results at https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” | content_tags.definitions[4] |
| concept | collective attention — not defined in this paper's own words; population-level concentration of searches on a public crisis — “Because the extent of the epidemic is likely to be underestimated at this stage, it is crucial to know how much the early case announcements induce heightened collective attention and information-seeking behavior.” | concepts[0] |
| concept | information-seeking behavior — not defined in this paper's own words; observable effort to obtain crisis-relevant information — “To provide rapid evidence to inform policy making, we use internet search data in an event study design to examine how collective attention and information-seeking behaviors respond to state government announcements of first COVID-19 cases.” | concepts[1] |
| concept | first case announcement — not defined in this paper's own words; state disclosure used as a focal information event — “We collected data on the timing of the first COVID-19 case announcements from media reports in each state; see Dataset S1 .” | concepts[2] |
| concept | Google Health Trends — not defined in this paper's own words; restricted API providing state-level query shares — “Our analysis is based on a balanced daily panel of COVID-19−related search intensity data in 50 US states and Washington, DC between January 1, 2020 and March 18, 2020 ( n = 51 regions × 78 d). We collected the data using a restricted-access Google Health Trends Application Program Interface (API) account ( 10 ).” | concepts[3] |
| concept | event study — not defined in this paper's own words; lead-lag comparison around a first-case announcement — “We estimated Poisson models in an event study ( 11 ) framework to examine the effect of the first COVID-19 case announcements on searches. In the model we use, the expected number of searches per 10 million is” | concepts[4] |
| concept | state fixed effects — not defined in this paper's own words; adjustment for stable interstate search differences — “We controlled for state fixed effects to allow for time-invariant differences in search patterns by state, and for date fixed effects to account for national trends.” | concepts[5] |
| concept | nonpharmaceutical interventions — not defined in this paper's own words; behavioral and institutional measures used to reduce transmission — “The first confirmed case of novel coronavirus disease 2019 (COVID-19) in the United States occurred in Washington State on January 21, 2020. Since then, the virus has spread across the country ( 1 , 2 ). There are confirmed cases in every state, and the vast majority are not connected to international travel ( 3 ), indicating that the virus had been circulating for several weeks before the first positive test ( 4 , 5 ). Slowing the transmission of the virus will help reduce the burden of the disease, save lives, and reduce strain on the health care system ( 4 , 6 ). In the absence of a vaccine, the main strategies for reducing transmission involve sanitation and handwashing, disciplined social distancing, quarantines, and school and workplace closures ( 6 ). These nonpharmaceutical interventions (NPIs) are effective and powerful ways to control transmission ( 7 ), as shown by earlier epidemics ( 8 ) and the steady decline of COVID-19 cases in China ( 2 ). To be successful, NPIs require people to undertake behavioral changes that may be personally costly. This is particularly true in countries with deeply rooted norms about personal freedom and reluctance to impose mandatory policies.” | concepts[6] |
| concept | information spillover — not defined in this paper's own words; cross-border attention response to another state's news — “To investigate information spillovers in which states respond to epidemics in other states, we examined augmented models that controlled for the timing of the announcements of Washington, Illinois, California, and Arizona, as well as for the timing of the earliest announcement among a state’s neighbor states [section A of additional results at https://github.com/anabento/GoogleBehaviorCovid ( 9 )].” | concepts[7] |
Full text — tagged
Evidence from internet search data shows information-seeking responses to news of local COVID-19 cases Abstract The COVID-19 outbreak is a global pandemic with community circulation in many countries, including the United States, with confirmed cases in all states. The course of this pandemic will be shaped by how governments enact timely policies and disseminate information and by how the public reacts to policies and information. Here, we examine information-seeking responses to the first COVID-19 case public announcement in a state. Using an event studyterm·mth framework for all US states, we show that such news increases collective attentionterm·dom to the crisis right away. However, the elevated level of attention is short-lived, even though the initial announcements are followed by increasingly strong policy measures. Specifically, searches for “coronavirus” increased by about 36% (95% CI: 27 to 44%) on the day immediately after the first case announcementterm·dom but decreased back to the baseline level in less than a week or two. We find that people respond to the first report of COVID-19 in their state by immediately seeking information about COVID-19, as measured by searches for coronavirus, coronavirus symptoms, and hand sanitizer. On the other hand, searches for information regarding community-level policies (e.g., quarantine, school closures, testing) or personal health strategies (e.g., masks, grocery delivery, over-the-counter medications) do not appear to be immediately triggered by first reports. These results are representative of the study period being relatively early in the epidemic, and more-elaborate policy responses were not yet part of the public discourse. Further analysis should track evolving patterns of responses to subsequent flows of public information.
Keywords: COVID-19, Google Trends, information The first confirmed case of novel coronavirus disease 2019 (COVID-19) in the United States occurred in Washington State on January 21, 2020. Since then, the virus has spread across the country ( 1 , 2 ). There are confirmed cases in every state, and the vast majority are not connected to international travel ( 3 ), indicating that the virus had been circulating for several weeks before the first positive test ( 4 , 5 ). Slowing the transmission of the virus will help reduce the burden of the disease, save lives, and reduce strain on the health care system ( 4 , 6 ). In the absence of a vaccine, the main strategies for reducing transmission involve sanitation and handwashing, disciplined social distancing, quarantines, and school and workplace closures ( 6 ). These nonpharmaceutical interventionsterm·dom (NPIs) are effective and powerful ways to control transmission ( 7 ), as shown by earlier epidemics ( 8 ) and the steady decline of COVID-19 cases in China ( 2 ). To be successful, NPIs require people to undertake behavioral changes that may be personally costly. This is particularly true in countries with deeply rooted norms about personal freedom and reluctance to impose mandatory policies.
In settings where restrictions on the free movement of residents are less strictly enforced, it is especially critical for public officials to know whether and for how long statements from leaders motivate individuals to efficiently seek and absorb information, as officials put their communication strategies in place. Such coordinated actions are needed at the first signs of community spreading and cannot be guided through traditional polling methods, which take too long. Because the extent of the epidemic is likely to be underestimated at this stage, it is crucial to know how much the early case announcements induce heightened collective attentionterm·dom and information-seeking behaviorterm·dom.
To provide rapid evidence to inform policy making, we use internet search data in an event studyterm·mth design to examine how collective attentionterm·dom and information-seeking behaviorsterm·dom respond to state government announcements of first COVID-19 cases. We highlight changes in search patterns occurring in the days leading up to and following first case announcementsterm·dom in a state ( 1 , 2 ).
Results Fig. 1 depicts the event studyterm·mth estimates of the effects of initial announcements on overall coronavirus search. Searches for “coronavirus” increased by about 36% (95% CI: 27 to 44%) on the day immediately after the announcement but quickly decreased back to the baseline level in less than a week or two ( Fig. 1 ). There was no observable trend in the search behavior in the days leading up to the announcements, suggesting the first “local” case indeed heightened the collective attentionterm·dom to the pandemic. However, the increased level of information seekingterm·def faded within 2 wk, even though many announcements of school closures or other mitigation strategies followed, suggesting that increased attention is only short-lived. This is consistent with no observable trend in the sense of urgency prior to the local announcement.
Time-varying effects of announcements of the first COVID-19 case in a state on searches for coronavirus. The period prior to the treatment (first confirmed case) is set as a reference: gray vertical bar. In red are the estimated coefficients (95% CI, gray band) in the Poisson model (differences in log-expected counts of search relative to the period prior to the event). The average search frequency of this term is 97,023.9 per state per day.
Fig. 2 shows the event studies of internet search for 1) symptoms and treatments, 2) hand sanitizer and diagnostic tests, 3) coordinated responses, and 4) narratives that undermine public responses. News of the first COVID-19 case in a state leads to a 52% increase in searches for “coronavirus symptoms” but does not increase searches for coronavirus treatment options. The second row shows searches for “hand sanitizer” increasing 35% immediately after first case announcementsterm·dom, and, unlike in the previous two cases, the search activity remains high for the remainder of the observable period. However, the announcements do not induce searches for nearby coronavirus testing opportunities, at least in the period under study. The third row suggests that first case announcementsterm·dom do not induce search for community-level policies (quarantines, school closures, and coronavirus testing) or more-elaborate personal health strategies (face masks, grocery delivery, over-the-counter medications). The final row examines how first case reports affect searches about the credibility of the epidemic. There is no indication that government confirmation of the first case increases or reduces searches for coronavirus hoaxes and overreactions; one might have expected official news to reduce concerns about false news.
Time-varying effects of announcements of the first COVID-19 case in a state on searches for 1) symptoms and treatments (red), 2) hand sanitizer and diagnostic tests (green), 3) coordinated responses (orange), and 4) narratives that undermine public responses (blue). Point estimated coefficients are the dotted lines (95% CI, gray band) in the Poisson models (differences in log-expected counts of search relative to the period prior to the event). Average search frequency per state per day for the terms: coronavirus symptoms (8,522.6), coronavirus treatment (448.2), hand sanitizer (3,214.7), testing near me (381.6), quarantine (2,158.2), isolation (631.2), coronavirus conspiracy (205.1), and coronavirus hoax (123.0).
Robustness Checks.
To investigate information spilloversterm·dom in which states respond to epidemics in other states, we examined augmented models that controlled for the timing of the announcements of Washington, Illinois, California, and Arizona, as well as for the timing of the earliest announcement among a state’s neighbor states [section A of additional results at https://github.com/anabento/GoogleBehaviorCovid ( 9 )]. The results show that COVID-19 searches responded to Washington’s first case ( Dataset S1 , column 2), to first case announcementsterm·dom in other states that occurred in January ( Dataset S1 , column 3), and to announcements in neighboring states ( Dataset S1 , column 4). These results indicate some spillovers, but the main conclusions regarding the effect of a state’s own first case remains. Each state may experience different treatment effects of the first COVID-19 cases on online searches. Therefore, we present single-state time series of search volumes in figure B-2 in additional results available at https://github.com/anabento/GoogleBehaviorCovid ( 9 ), to descriptively depict the treatment effect heterogeneity. The results suggest that information-seeking responses were larger and faster in late states compared to the early states. We checked whether the findings were driven by accessed date of the data by rerunning the analysis using the latest search output from Google Trends, which we retrieved on April 18, 2020 and compared to the current results using the data retrieved on March 20, 2020 [ https://github.com/anabento/GoogleBehaviorCovid ( 9 )]. These core results were not sensitive to this replication.
Conclusion Our results suggest that first state COVID-19 case announcements do lead to a widespread increase in the extent to which people seek out information about the epidemic. We find that announcements of first cases have the biggest effect on searches for basic information about the virus and its symptoms, suggesting that people attempt to educate themselves about COVID-19 and its effects. However, first case reports do not lead to a differential increase in searches involving large-scale NPIs (e.g., school and workplace closures) that will likely have important consequences for everyday life. This may be because first cases are reported relatively early in the epidemic when many of these more serious mitigation strategies are not yet part of public discourse. Finally, although one might expect official news to reduce concerns related to false messages, we do not find much indication that state announcements of first COVID-19 cases affect searches questioning whether the epidemic may be a hoax. Overall, our analysis of internet search data suggests that government information disclosure does help focus public attention on the crisis. Our evidence also indicates that people seem to mainly react by seeking information on what they can and should do in response to the epidemic. One conjecture is that policy makers may be able to take advantage of a brief period of active information seekingterm·def following focal events (first cases and first deaths) to provide clear advice regarding the actions people can undertake to avoid risk.
Materials and Methods Samples.
Our analysis is based on a balanced daily panel of COVID-19−related search intensityterm·var data in 50 US states and Washington, DC between January 1, 2020 and March 18, 2020 ( n = 51 regions × 78 d). We collected the data using a restricted-access Google Health Trendsterm Application Program Interface (API) account ( 10 ).
Measures.
The outcome variables measure the daily share of all Google queries in a state that correspond to a particular term during our study period. We multiply the shares by 10 million and round to the nearest integer to make the measure more interpretable. We collected data on the timing of the first COVID-19 case announcements from media reports in each state; see Dataset S1 . All states reported their first case by March 17, 2020.
Data Analysis.
We estimated Poisson models in an event studyterm·mth ( 11 ) framework to examine the effect of the first COVID-19 case announcements on searches. In the model we use, the expected number of searches per 10 million is where Announcement i , k is an indicator variable set to 1 if the first case of state i was announced k days ago. The reference category was 1 d prior to the first case announcementterm·dom, and we created single indicators Announcement i , 31 , for any k ≥ 30 , and Announcement i , − 41 , for any k ≤ − 40 . We controlled for state fixed effectsterm·mth to allow for time-invariant differences in search patterns by state, and for date fixed effectsterm·var to account for national trends. SEs allowed for heteroskedasticity and clustering at the state level. Poisson coefficients and CIs are shown for the 40 d leading up to the first case announcementterm·dom, and up to 20 d following it. The vertical axis measures the differential percentage change in search frequency as states approach the first case announcementterm·dom ( Figs. 1 and 2 ). We performed regression analysis using Stata, version 16.0 (StataCorp).
Data and Code Availability.
State news data and code are available at https://github.com/anabento/GoogleBehaviorCovid ( 9 ). The search data for this study are available from Google Trends but are restricted in use; researchers may apply to Google Trends API for access.
Limitations Although the eventual failure of Google Flu Trends suggests that building and maintaining a complex model for a long period of time is difficult ( 12 , 13 ), internet search volume has been shown to be a good proxy for many socioeconomic behavioral indicators, such as automobile sales ( 14 ) or dietary patterns ( 9 ). This study has, indeed, several limitations. First, our state-level data from Google Trends database do not allow us to adequately understand intrastate variation of search volumes following the announcements or to generalize conclusions about population search behavior (as Google Trends only captures the search behavior of a particular segment of information seekers). Second, the major issues with search query data—long-term drift and representation bias across geographic regions—does not come into play in our study because we focus on responses in a short time window surrounding the government announcements. The third limitation comes from potential confounding factors in our event studyterm·mth framework, including anticipatory behavior and unobservable local events ( 15 ). We mitigate these concerns by controlling for the timing of the announcements of Washington, Illinois, California, and Arizona as well as for the timing of the earliest announcement among a state’s neighbor states, in several sensitivity analyses.
Supplementary Material Acknowledgments We thank Raj Thakkar and Chrislin Priscilla for research assistance.
Footnotes References Associated Data Supplementary Materials Data Availability Statement State news data and code are available at https://github.com/anabento/GoogleBehaviorCovid ( 9 ). The search data for this study are available from Google Trends but are restricted in use; researchers may apply to Google Trends API for access.