We present an RDF Data Cube { integrated from numerous sources on the Web { that describes countries in terms of general variables (e.g., GDP, population density) and COVID-19 variables. On top of this data cube, we develop a system that computes and visualises correlations between these variables, providing insights into the factors that correlate with COVID-19 cases, deaths, etc., on an international level. Demo link: https://c19.dcc.uchile.cl/
The among-country variation in the numbers of reported COVID-19 cases
and deaths per capita is not well-understood [
In this demo, we will discuss on-going work regarding the preparation of an RDF Data Cube that collects together a wide variety of both general and COVID-19-speci c variables at a country level. On top of this RDF Data Cube, we have built a system to visually explore the correlations that exist between such variables in order to gain insights on the among-country variations observed for COVID-19. We call this data cube and associated system \COVIDCube". Code is made available online at https://github.com/tmnvrd/COVIDCube.
We rst describe the data preparation for generating the RDF Data Cube. Thereafter we describe the system used to visualise correlations.
Copyright c 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). RDF Data Cube: There are a wide range of datasets available online that provide di erent types of indicators for countries. In order to narrow down the scope of the variables considered, we conducted an initial survey in search of hypotheses relating to the among-country variations observed for reported COVID-19 cases and deaths. From these, we broadly identi ed three main categories of indicators relating to economics (e.g., GDP, wages, unemployment), health (e.g., obesity and other comorbidities, blood type, vitamin-C de ciency, adult and child mortality rates), and climate (e.g., temperature, precipitation, pollution). We also identi ed hypotheses relating to miscellaneous factors, such as political ideologies, transportation networks, tourism, etc. Based on these factors, we began to identify potential sources of data online at the international level, extracting 525 variables from sources including Our World In Data, Wikipedia, The World Health Organization, The World Bank, among others. These datasets, mostly tabular in nature, were extracted as CSV les. We further extracted data for 4 variables pertaining to COVID-19 at the international level from Johns Hopkins University Center for Systems Science and Engineering (CSSE) and Our World in Data (OwID), namely: con rmed cases (CSSE), con rmed deaths (CSSE), con rmed recovered (CSSE), and stringency index (OWiD)1
The data were diverse: some datasets were broken down by temporal or
regional dimensions; measures were provided in di erent units; naming variations
were present for countries or regions; etc. We wished to integrate the data while
modelling their provenance. We thus chose to adopt the RDF Data Cube
vocabulary and model [
We provide a screenshot of part of the visualisation in Figure 1 referring to the top-10 health variables in terms of positive correlation to con rmed COVID-19 cases. The colours and values indicate the value of correlation, where the colours range from red (positive) to blue (negative). Results that reach a particular level of statistical signi cance (p < ) are noted with ( = 0:05) and ( = 0:01). We see that variables such as life expectancy, body mass index, prevalence of overweight children, obesity, blood types A and O, etc., correlate positively in terms of the number of con rmed COVID-19 cases. The reader may notice that Prevalence of overweight children appears twice in the results. This is because y-axis labels are sometimes hierarchical, where only the rst level of the label is shown to avoid overly-long variable names; dimension 5 refers to male children, while dimension 6 refers to female children, which can be seen by hovering over the respective result in the interface. Further such results for other variables can be explored in our online demo: https://c19.dcc.uchile.cl/. Evaluation: In order to initially understand users' opinions of the system, we created a quantitative survey that was published in a university forum and received 52 responses. On a Likert scale of 1{5, users were most positive with respect to the functionality (average: 4.25) and usefulness (average: 4.12) of the platform, but were less positive regarding how easy it was to understand the information provided (average: 2.98), where users require some base statistical knowledge of correlations, p-values, etc., in order to fully understand the data presented. 3
Correlation obviously does not imply causation, but correlations may lead to novel hypotheses and prompt further study regarding potential underlying factors for among-country variations. Though an in-depth analysis of the correlations found is out-of-scope, in summary, we did nd a number of variables spanning the di erent categories that had statistically signi cant positive or negative correlation with the COVID-19 variables: more than would be expected according to the chosen signi cance level (e.g., we nd more than the expected 1/20 variables satisfying the = 0:05 threshold). Examples of signi cant correlations can be seen in Figure 1, and also by visiting our demo.
Some correlations observed were to be expected, such as in the case of positive
correlations for obesity, which is associated with comorbidities for COVID-19.
However, other correlations were surprising. For example, one of the most
negatively correlated health-related variables was No access to handwashing facility,
which suggests, with statistically signi cant results, that countries where fewer
people can wash their hands tend to have fewer per-capita cases of COVID-19.
This seems counter-intuitive as handwashing is considered a way to reduce
transmission of the virus. Another (initially) counter-intuitive result is that increased
health expenditure per capita correlates positively with COVID-19 cases, where
one would expect that better health services would help to reduce transmission
and cases. However, by considering the overall matrices, a common
confounding factor begins to emerge: namely the level of development of the country,
with more developed countries tending to have more con rmed cases. Similar
observations can be found elsewhere, and possible explanations include a lack of
testing [
While it is not possible to draw rm conclusions about the among-country variance observed for COVID-19 from the currently available data, COVIDCube does provide some insights and clues as to potentially important factors. Deriving more de nitive conclusions will require integrating diverse data from further diverse sources, for which RDF Data Cubes provide a relevant solution.
Regarding the use of RDF Data Cubes, in the original plan for the project, we had intended to use a relational database as our back-end to load di erent variables by country. However, when accessing the raw data, we encountered a variety of issues relating to their diversity, including the use of di erent units and multipliers; di erent names being used; di erent countries being included/excluded from certain sources; the same demographics (such as gender, age, etc.) appearing in di erent measures; some countries having their results being presented by state or region; di erent measures having di erent temporal granularities; etc. Rather than cleaning and preprocessing the data to \ t" a clean relational schema, we rather chose to use RDF Data Cubes to represent the diverse underlying data in a more complete way, and thereafter use SPARQL queries to compute the tables from which correlations were extracted. This approach o ered a greater decoupling between the data preparation and the application design, where we could focus on representing the underlying data as completely as possible in the RDF Data Cube, and then later use SPARQL queries to extract the data needed for the application from multiple sources.
In terms of future directions, we wish to investigate correlations along
temporal dimensions; such data are available in the data cube but are not exploited
by the visualisation. Similarly, although the underlying dataset tracks
provenance information, this is not shown in the current interface; adding links to
the underlying sources used would help others to build upon and reproduce the
results shown. Incorporating other data sources { including integration with
existing RDF datasets (e.g., on COVID-19 [