Human societies around the world have been facing epidemic spreading since their birth. Nevertheless, the COVID-19 emergency is an unprecedented event because human societies have never been so intertwined in the previous history. In particular, this is true at both the institutional and the informational level. On the one hand, worldwide governments have strong political and economic relations and bindings to each other (e.g., United Nations, European Union, etc.). On the other hand, the extreme accessibility of information of our age makes the population of each nation more aware than ever about what happens in other countries.
The high level of interconnection among countries raises important questions about the reactions to the emergency. How did information needs and the attention of the citizens evolve in time? How did governments take actions? How are these two levels connected? How did different countries influence each other? Did collective phenomena, in terms of synchronized multi-state interventions, emerge? In this preliminary report, we provide insights and interpretations that could be useful to formulate the answers to these questions.
Thanks to the Searches index extrapolated from the activity on the Google search engine (source: Google Trends ), we have a powerful proxy for the attention of the general public to the Covid-19 emergency. In this report, we used data about searches for "Coronavirus" topic for more than 200 countries and territories. It is worth noting that not all the countries use "coronavirus" as a term of the queries, as some use "corona" or "covid", but worldwide, "coronavirus" is used by the vast majority of countries, and even when it is not the most used term in a country it is still significantly used.
Data about the actions taken by every government are much harder to find and to deal with, due to the lack of standardized codification of the action, and for many other reasons. We base our analysis on the dataset aggregated by the Vienna Complexity Science Hub. Interventions data are available and categorized for more than 50 countries. Unfortunately, most of the countries from Africa and Oceania are missing, so for some analysis these two continents were not included (updates will be performed as soon as more data become available).
In the following set of graphs, we show for Asia, Europe and the Americas the time series of the Searches index together with the crucial days of interventions at the international level. More precisely, we highlight the days in which at least 50% of the countries of the continent took any intervention (the threshold is 25% for the single categories of interventions).
For Asia, we can find synchronized interventions of different nations already at the beginning of February in a crescendo that culminates in mid-March when many interventions concentrated in a couple of weeks. Though it is not surprising that this peak occurred earlier compared to the other continents, what is interesting is that these international interventions days were happening before the peak of interest shown in the Searches index, that is reached at the very end of March. This evidence is valid for the average (dominated by India), but it is also true for the majority of the countries considered. The most significant exception is Singapore. It is worth noting that presently, China is currently not represented. It will be added as soon as intervention data become available.
Europe shows the maximum concentration of international intervention days in mid March as Asia, but they appear to be much more focused. Also, no synchronized international actions (i.e. more than 50% of the considered countries) were taken before the beginning of March. In this case, the peak of the average Searches index for the continent is the same days, or slightly before. The more significant exception is Italy, that shows it highest peak in the Searches at the end of February.
Americas' considered countries show a high concentration of international interven tion days, also starting in mid-March and continuing until the beginning of April. The peak of the Searches index happens to be in mid-March, like in the other continents. So, differently from the others, the international intervention days seem to follow the peak of the interest of the population mainly.
It is quite hard to establish causal relations for what we observed; still, the three continents have different timing in interventions compared to the peak of interest. So while Asiatic countries interventions happened synchronously even before the population attention to the issue reached the apex, in Europe, these two events happened at the same time. In contrast, for the American countries, action slightly followed the attention of the population.
This fact is subject to different possible interpretations. It could be that the European and American country governments needed people awareness to be able to take the critical and unprecedented actions needed while Asiatic country did not. As many suggested, many Asiatic countries were more prompt to face the emergency due to previous epidemic experiences (e.g. 2002 SARS outbreak). Another possible interpretation is that the interest in Asiatic countries was triggered by the government's interventions, while in Europe and the Americas was raised by media. This reasoning would also explain why the interest in the Asiatic countries seems to occur later compared to the other continents. At the same time, since they were involved in the emergency before due to geographical proximity, the interest should have arisen earlier. Further investigation will be needed to understand which interpretation is correct.
So far, we split countries according to their continental belonging, but is this a meaningful distinction? The high level of interconnection of the contemporary world makes this point questionable. Also, inside a given continent, it is interesting to spot the similarities and the differences between the reactions of different countries. In the following part, we explore the similarities in the interest and the temporal intervention development for all the countries included in the datasets. We aim to spot and represent similarities between countries, and to analyze whether this similarity does cluster meaningfully at the continental level. We will also check the specific coherence of the group constituted by the European Union( EU27 in the following).
The following graphs show of the strength of intra and inter continent similarities taking into the account 190 countries and territories. Our first target was to show similarities in the Searches index temporal curve. To this aim we filtered those with a small population, and we trimmed all the temporal series to synchronize their starts. We did this because the start of the interest is obviously linked to when the virus arrived in a given nation. We are more interested in what happens to the interest later, since we assume the interest is a sort of reaction. From the trimmed Searches timeseries we calculate similarities as Spearman coorrelation coefficients between all the possible couples of time series.
In the previous graph, each dot represents a country and each link a similarity relation with another country. To make the representation more intelligible, we connected each country only to the three most similar countries. It is still possible that a country has more than three links if this country belongs to the top 3 of many nations. While similarity is symmetric, ranks in similarity are not (if country A is the first for country B, country B might not be the first for country A). Also, in this representation:
Moving the mouse over each dot will show the name of the country and the top 3 more similar countries. The similarity is the correlation coefficient normalized between the maximum an the minimum observed in the whole similarity dataset. So a value of 100 will mean that the two countries are at the maximum of the observed similarity, while 0 will mean they are at the minimum.
It is immediately evident that continents do cluster somehow, but with different behaviors, and with exceptions. E.g., some European countries (blu) like Russia or Ukraine are far away from the European cluster, and deep inside the Asiatic one. Instead, middle-eastern countries do seem more close to Europe than to Asia. Also, African countries do not seem to cluster as much as others. Instead, they seem scattered in the Asian and in the American clusters. To give a more accurate account of these phenomena, we performed a very simple clustering analysis starting from the (trimmed) Searches correlation results. For this analysis, we considered each continent and the unique group EU27. We simply compare the similarity between couples of countries belonging to the same continent with the similarity of a couple of countries belonging to different continents. Once we have these sets of similarities, we compare them in terms of average and with a statistical test to measure the meaningfulness of the difference (with a two-sample T-test).
In the previous sets of histograms, we show internal similarities of continents (and EU27) compared with their similarity with other continents. E.g., Americas countries are, on average, 23% more similar among each other more than they are similar to non-American countries. This confrontation is proved to be statistically meaningful by a two-sample T-test index with a p-value threshold of 1e-3. If the differences between the averages (above the top right corner of each histogram) are not shown, it means that the test showed that the difference is not statistically meaningful.
The strength of the internal binding of the continents appears to be quite heterogeneous, ranging from +32% of Europe to -7% of Asia. This last value, in particular, seems to mean that Asiatic countries are (slightly but meaningfully) more similar to countries of other continents than to each other. Interestingly enough, Europe is the most clustered continent, with a difference in average that rise to 43% if we consider the EU27 group. To better understand the importance of this value, it should be compared with the corresponding measure made on the special group of the single states of the USA. Anyway, that is the federation of nations that as a different level of aggregation so for the present analysis USA have been considered as a single state, leaving more in-depth analysis for further studies.
It is worth noting that the reaction to the emergency in terms of informational needs of the population is probably more linked to the cultural and informational layers of societies than to the political and institutional ones. Thus, the high value of similarity between EU27 countries is probably related to their strong cultural bonds, rather than to similar institutional reactions. To study this hypothesis, we can perform the same clustering analysis with intervention patterns.
The following graphs show a representation of 50 countries and of the strength of continental bindings. In this case, our focus is on the intervention strategy of each country. The dataset aggregated by the Vienna Complexity Science Hub makes available categorized interventions for several nations. This categorization is hierarchically structured in different levels of details. For the sake of simplicity, and to avoid semantic complications (i.e. differences between the distance intra category and inter category), we are considering only the most coarse-grained distinction, named "L1" in the dataset, that contains the following categories:
These categories are also visible in the first three graphs of the report. To compare different countries, we had to define similarity in the sequence of the interventions. For each country, we translated the series of interventions in the corresponding series of category "L1" of intervention. Then taking any couple of countries, we can calculate the Levenshtein distance between this sequence of categories. This distance can be considered as the minimum number of single edits (insertions, deletions or substitutions) required to change one sequence into the other. E.g., if country X and country Y have the following sequence of categories of interventions, Country X (6 interventions): ABEEHA Country Y (7 interventions): ABEDHAC the distance between them will be 2, because there are just two edits: the second E has been turned into a D (substitution) and a C has been added to the end. The distance has been then normalized dividing it for the minimum length, taken as a reference. In the example, the final result would be 2/6=0.33, roughly meaning that we need to change 33% of the shorter sequence to get the longer one. In the following analysis, we use the reciprocal of this normalized distance between the sequence of the categories of intervention as a measure of similarity between couples of countries reaction strategies.
This graph follows the same rules of the previous network representation of countries (i.e., dots are countries, links are similarities, only top 3 are shown, etc.). The main differences are the following: reduced number of countries (50), an abbreviation of the country is shown inside each dot, and the value of similarity in the overlay is now the actual value of the similarity measure adopted (i.e. the reciprocal of the normalized distance explained previously).
The situation now seems less prone to straightforward interpretations, due to the scarcity of data for all continents but Europe. Still, it seems quite evident that countries do generally align forming an unclustered disposition. This result seems to suggest a higher level of general dissimilarity, compared to the previous situation. The only highly clustered area seem to involve southern-European countries. To better quantify the presence (or the absence) of a continental cluster, we do repeat the analysis done in the previous sets of histograms. Again, we split a couple of countries into different sets, and we compare the links inside a continent with the link from that continent to the others in the following graph.
This histogram follows the same rules of the previous histogram. The main differences are the following: reduced number of countries (50), fewer continents taken into account (Europe, Asia, Americas plus the special group EU27), and a different threshold for the p-value (5%) has been adopted. Threshold has been changed due to the vast differences in samples sizes.
As previously suggested, the analysis shows that the difference between intra and inter continent similarity is much lower than in the previous case. In other words, reactions in term of interest seem to be much more cluster at the continental level than the response in terms of interventions. The only meaningful cluster that does emerge is in Europe (+6% of internal similarity), remarking again the presence of a level of clustering unequaled in the other continents. Still, this self-similarity is much weaker than before and, contrarily to what we observed for the pattern of interest, the EU27 group has a self similarity equal to +5%.
The interpretation of preliminary and incomplete analysis can be useful but must be considered with due caution. The results we observe for these measures seem to point out that there has been a substantial heterogeneity of the intervention strategies, at least at the continental level. This dissimilarity in th e actions taken by the governments contrasts with the similarity of the people reaction in terms of the temporal evolution of the interest in the Coronavirus topic. This contrast might be interpreted in several ways. One possibility is that the contrast could be a consequence of a more substantial relationship between the different infospheres of the various countries compared to the international coordination in facing the emergency, that appeared to be somehow weaker. The lack of agreed global strategies and standards is actually one of the most discussed points about the COVID-19 response, so that, for example, even comparing the epidemics stats between different countries is a non trivial task. Of course, such a conclusion is just a possible interpretation, and further analysis will be needed to study this hypothesis better.