First, we explored people's COVID-19 concerns by analyzing search queries over different time periods. We initially established a baseline of common search queries prior to and during the pandemic by selecting the search queries of Yahoo! JAPAN's users from April to May 2019 and from April to May 2020 (we chose to examine queries during April and May due to the Japanese government’s declaration of a state of emergency in April 2020 [18]). Additionally, since older adults (those aged >65 years) are at a significantly greater risk of adverse COVID-19 outcomes [19], we speculated that there might be more COVID-19–related search queries from this group. Therefore, we started the analysis by targeting the search queries of people over 65 years of age. We extracted 100,000 search queries of this older population for the two aforementioned time periods and ranked them in reverse order according to the search counts. A change in time period results in a change in search query ranking; thus, we quantified the difference in ranking by defining the following rank change index:

The larger the index, the fewer the search counts for this query in 2019 or the greater the number of search counts in 2020, and vice versa. A constant of two was added to both the numerator and denominator, to avoid instances of zeros (uncountable) in the search counts. Table 1 shows the top five rapidly increasing search queries in April 2020 compared to April 2019. As expected, these terms appeared to be COVID-19–related.

Consistently, of the top 100 queries in the ascending query list, 76 queries contained COVID-19–related keywords (eg, コロナ [corona], マスク [mask]); of these, 33 queries contained prefecture names combined with コロナ感染者 (corona cases). A query pattern, such as the prefecture name plus “coronavirus cases,” clearly displays the prefecture’s information and it also reflects the user's concern about COVID-19 to some extent. Therefore, we chose this query pattern as our target query and sorted the queries by region according to the prefecture names mentioned in the query. Table 2 lists some query samples used in this study.

However, when we calculated the search counts of target queries that comprised the prefecture name and “coronavirus cases” from January to September 2020, we found that the search counts in Tokyo were much higher than in the other prefectures, as shown in Figure 1. It is possible that Tokyo’s larger population meant that the frequencies of the general search counts were higher than in less populated prefectures. We speculated that this excessive disparity would inevitably have an impact on our subsequent calculations. To mitigate this effect, we introduced baseline queries for each prefecture that were frequently searched for in that prefecture and that had a relatively stable search count for a certain period. The baseline queries contained the prefecture name and “X,” where “X” referred to any keywords as long as the sample variance of their monthly search counts was as small as possible. For example, for Tokyo, 東京 23区 (Tokyo 23 wards) and 東京 天気 過去 (past weather in Tokyo) were some of the baseline queries, and their sample variances of search counts in these nine months were 499.6 and 683, respectively. In summary, we compared the search counts of possible queries with the above pattern from January to September 2020, and then identified the top three queries with the smallest sample variance during those nine months to form the baseline queries. This was done to balance the impact of excessive disparities by using the quotient of the baseline query and our target query, as the search counts of the target queries were high for urbanized prefectures (such as Tokyo), where the search count frequencies of the baseline queries were also high.

Next, our query-based equation to quantify the level of concern about COVID-19—the LCI equation—was defined for each prefecture pref as follows:

where tq_pref and bq_pref are the target query and the baseline query, respectively, and Count(.) is a function that counts the occurrences of a query. We used the logarithmic result for the LCI calculation, which suggests that a higher LCI means a higher frequency of baseline-controlled, prefecture-specified COVID-19–related queries, which in turn reflects a greater level of concern about COVID-19 in that prefecture.

However, the LCI appeared to be greatly influenced by the severity of the COVID-19 situation in each prefecture. For example, urbanized areas, such as Tokyo and Osaka, which are the first and second most populous cities in Japan and the hardest hit, naturally had the highest search counts for target queries and a higher LCI. The COVID-19 pandemic has raised people's anxiety levels [20], and the LCI appears to reflect this attitude in hard-hit, largely urban areas. However, heightened concern about COVID-19 in rural areas [21] does not seem to be reflected in the LCI. Rather, given the lower number of cases in rural areas, it is likely that concern about COVID-19 in rural areas may be related to socioecological variables unrelated to the risk of infection.

For this reason, we attempted to improve our LCI equation to be able to examine COVID-19 concerns beyond the direct influence of actual COVID-19 cases. We argue that this would quantify the prevalence of concern about COVID-19 beyond the risk of actual infection. To do so, we modified the LCI to account for the number of new cases per month in each prefecture by the population of that prefecture to calculate the percentage of infected patients; that is, Number of monthly new cases/population.

In summary, the LCIPP for each prefecture pref was calculated as follows:

Similarly, we used the logarithmic result for the LCIPP calculation.

To examine whether the differentiated patterns of associations for the LCI and LCIPP reflected the COVID-19 concerns in rural and urban areas, we examined prefecture-level correlations with both indices. Accordingly, we included the corresponding measures and examined the following aspects, with the corresponding indicators in parentheses: (1) the prevalence of farming (number of farming households, percentage of farmland, rice production), (2) population change (rate of population change), (3) ease of accessibility (reachable areas within one hour from major stations of each prefecture, travel time from Tokyo to major stations in each prefecture), and (4) public health outcomes (reported symptoms, daily outpatients, COVID-19 cases per million residents).

We argue that rural areas can be identified by the higher prevalence of farming households, more farming area, and higher rice production. Conversely, urban areas can be identified by their higher population density, ease of accessibility (the proportion of reachable areas within one hour from major stations of each prefecture), and rate of population change (urban areas should exhibit population growth, while rural areas should show a population decline). Finally, prefectural public health was measured as the number of cumulative COVID-19 cases per 1 million residents (as of September 30, 2020), the proportion of the population that reported general symptoms (non–COVID-19–related), the number of ambulance dispatches, and the number of daily outpatients in 2019. Public health outcomes are able to reflect the current state of local health care to some extent. Since COVID-19 outbreaks would take over existing health care resources, we aimed to find the relationship between public concern about COVID-19 and the current state of local health care through the public health outcomes indicators.

Figure 2A and Figure 2B show the geographical results of the LCI and LCIPP from January to September 2020, respectively. Table 3 shows the top prefectures with the highest and lowest LCI and LCIPP. These are also shown as the darkest and lightest regions in Figure 2, respectively. In addition, we examined the LCI and LCIPP across three phases following the timing of the state of emergency: (1) before the state of emergency (January-March), (2) during the state of emergency (April-June), and (3) after the state of emergency (July-September). For a breakdown of the LCI and LCIPP by prefecture, please refer to our online supplementary material [22]. The LCI results show that before the declaration of the state of emergency, there was a low level of concern about COVID-19 nationwide. However, with the state of emergency declaration, some prefectures (including Tokyo and Ibaraki) began to show a high level of concern. At the end of the state of emergency, the overall level of concern in the country did not change much, though prefectures such as Hokkaido showed a slight decrease. The LCIPP results show a high level of concern for COVID-19 in Hokkaido even before the national state of emergency was declared. This is because Hokkaido activated COVID-19 response measures at an earlier stage (March 2020) [23], which resulted in a high level of concern for COVID-19 among local residents. With the declaration of a nationwide state of emergency in April 2020, most prefectures also showed a high level of concern, and with the end of the state of emergency, the level of concern declined accordingly.

Of all the prefectures, we selected four to be displayed in Figure 3: Tokyo, Osaka, Niigata, and Ibaraki. We used the number of farming households as a criterion for distinguishing between urban and rural areas. According to a survey by the Statistics Bureau of Japan [24], Tokyo and Osaka have the lowest numbers of farming households in Japan, while Ibaraki and Niigata have the highest numbers. This suggests that Niigata and Ibaraki have relatively lower urbanization rates than Tokyo and Osaka. Nevertheless, Figure 3 indicates that some urban areas (ie, Tokyo and Osaka) have a relatively lower LCIPP than some rural areas (ie, Ibaraki) in terms of the general trends.

By calculating the correlation coefficients of LCI and LCIPP on the aforementioned indicators, our results showed that the LCI was significantly correlated with the COVID-19 infection risk (cumulative cases) and urbanized prefectures. The latter was identified through population density and ease of accessibility. The LCI was higher in prefectures that have a high proportion of reachable areas within one hour from their major stations and in prefectures that had growing populations. In contrast, the LCIPP was higher in prefectures that were more rural, as determined by their higher number of farming households and rice production. Table 4 displays the correlation results.

We also investigated the LCIPP results for three different age groups: (1) age 25-44 years, (2) age 45-64 years, and (3) age >65 years. At this point, population in the LCIPP equation was replaced by the number of people in each age group. Figure 4 shows that older people over the age of 65 years had fewer concerns than those aged 25-44 years and 45-64 years, while those aged 25-44 years and 45-64 years had almost the same level of concern about COVID-19.

We note the limitations of our approach. Our method of extracting prefecture information from search queries relied on searches for the prefecture name + “coronavirus cases,” and not location-based information, such as IP addresses. This may not necessarily represent queries that are only from residents of these prefectures; it may also reflect queries from nonresidents who are interested in the COVID-19 situation in these prefectures (eg, a user who may have family in these prefectures). In addition, although Yahoo! JAPAN has a large user base in Japan, in the broader search engine market, Google still has a significant share. Search queries and results might vary between search engines due to user preferences and service provider settings. Therefore, from the perspective of data diversity, it would be a better choice to consider a combination of search queries from multiple search engines.

In summary, this study used search queries from Yahoo! JAPAN users to quantify the degree of concern about COVID-19 in rural and urban areas. We first established that Yahoo search queries could be used to quantify COVID-19–related concerns. We then defined the LCI and LCIPP as quantitative indicators of prefecture-level COVID-19 concern. The LCI was indicative of COVID-19 concern in urban areas, whereas the LCIPP appeared to be indicative of COVID-19 concern in rural prefectures. By investigating the relationships between these concern indices and prefecture-level information, we showed that the LCI and LCIPP have external validity. These results suggest that one potential application could be conducting differentiated public campaigns about COVID-19 prevention and misinformation. In light of the different sources of concern in rural and urban areas, such campaigns could adopt different risk communication strategies in these areas.