Learning from the curve

COVID-19 in Belgium - what to do now? A commentary from the GEES experts

2020-09-27T18:26:00+00:00

A new blog post on the COVID-19 pandemic from the “About COVID-19” blog.

Experts from the GEES express their view on “what to do now?” regarding the COVID-19 pandemic in Belgium.

You can read the article in French or Dutch

Une évaluation de la contribution du premier tour des élections municipales à l’épidémie de COVID-19

2020-07-31T08:57:53+00:00

Les résultats d’une étude rédigée par Guilhem Cassan et Marc Sangnier.

Référence de l’étude

« Liberté, Egalité, Fraternité... Contaminé? Estimating the impact of French municipal elections on COVID-19 spread in France »

AMSE Working Papers, 2020-24, 22 juin 2020.

Environ 20 millions d’électeurs se sont rendus aux urnes le dimanche 15 mars pour le premier tour des élections municipales alors que s’amorçait en France l’épidémie de COVID-19. L’étude évalue la contribution de cet événement à la diffusion du virus. Cette étude se distingue des études déjà existantes portant sur l’impact du premier tour des élections municipales sur l’épidémie de COVID-19 en France (Zeitoun et al. 2020 et Bertoli et al. 2020). La méthode d’estimation permet en effet de prendre en compte la dynamique spécifique de l’épidémie dans chaque département tout en couvrant a priori l’ensemble de la population française. Elle permet en outre de proposer une quantification de l’impact du premier tour des élections municipales sur l’épidémie et de développer une évaluation qualitative des risques liés à la tenue du second tour à la fin du mois de juin. Les résultats de l’étude suggèrent que la tenue du premier tour des élections municipales a contribué significativement au développement de l’épidémie dans les départements dans lesquels l’épidémie était déjà à un stade avancé le jour de l’élection. À l’inverse, la situation ne s’est pas aggravée du fait des élections dans les départements où l’épidémie était alors à des niveaux plus faibles. Les estimations suggèrent que le premier tour des élections municipales a contribué à environ 4,000 hospitalisations supplémentaires pour suspicion d’infection par le COVID-19, soit environ 15% du total des hospitalisations cumulées à la fin mars. La méthodologie sur laquelle repose l’étude permet d’évaluer comme faible le risque que la tenue du second tour des élections municipales le 28 juin déclenche une nouvelle vague importante de contaminations dans la mesure où l’épidémie se trouvait déjà début juin dans la majorité des départements à des niveaux inférieurs au seuil au-dessus duquel les effets négatifs de l’élection sont détectés. La méthodologie utilisée combine les approches développées autour des rendements anormaux des actifs financiers et de l’évaluation des politiques publiques. Elle repose sur deux étapes. Tour d’abord, un modèle logistique standard est calibré sur les données antérieures au 15 mars pour chacun des départements de France métropolitaine. Il permet de prédire l’évolution attendue de l’épidémie en termes de nombre d’hospitalisations pour suspicion d’infection par le COVID-19 en l’absence des élections et des mesures de confinement mises en place dans les jours qui ont suivi. Les différences en matière de taux de participation entre départements sont ensuite utilisées pour identifier l’effet des élections sur les erreurs de prédiction issues du modèle de première étape et l’isoler des effets potentiels des autres chocs et mesures.

Les auteurs

Guilhem Cassan est chargé de cours à l’Université de Namur.

Marc Sangnier est chercheur à l’Université de Namur, en disponibilité d’Aix-Marseille Université (Aix-Marseille School of Economics).

References

Jean David Zeitoun, Matthieu Faron, Sylvain Manternach, Jerome Fourquet, Marc Lavielle et Jeremie Lefevre, « Reciprocal association between participation to a national election and the epidemic spread of COVID-19 in France: Nationwide observational and dynamic modeling study », medRxiv 2020.05.14.2009010, 2020.

Simone Bertoli, Lucas Guichard et Francesca Marchetta, « Turnout in the Municipal Elections of March 2020 and Excess Mortality during the COVID-19 Epidemic in France », IZA Discussion Papers 13335, Institute of Labor Economics (IZA), 2020.

The Spread of COVID-19 in Belgium: a Municipality-Level Analysis

2020-07-17T13:31:32+00:00

Table of Contents:

Introduction
Description
Analysis
Results
Conclusion
References

Philip Verwimp, ECARES, ULB

Email: Philip.verwimp@ulb.ac.be

July 8, 2020

Abstract

In this contribution I analyse socio-economic and demographic correlates of the spread of the COVID-19 epidemic across Belgian municipalities. I am interested in the onset of the epidemic, its intensity early on as well as the growth of contaminations in April. The paper uses contamination data from Sciensano, the Belgian health agency in charge of epidemiological information. In the period under investigation, March and April 2020, Belgium used a uniform and restrictive test policy for COVID-19, which changed on May 4th. The data are completed with socio-economic and demographic data published by governmental agencies. Employing linear and log-linear models I find that COVID-19 spread faster in larger, more densely populated, higher income municipalities with more elderly people and a larger share of the elderly population residing in care homes. Richer municipalities managed to slow down the epidemic in April more compared to poorer ones. Municipalities which were more exposed to migration, foreign travel for business, leisure or family affairs were affected earlier on in the epidemic. Income correlates with the contamination rate in particular in the Flemish Region whereas the share of foreign nationalities correlates with the contamination rate in particular in the Walloon Region.

Key words: COVID-19, Belgium, municipality, regression analysis

Note: Ilaria Natali, Francois Ryckx and Jan Van Bavel provided useful comments to an early draft of the paper. All responsibility for remaining errors rest with the author only.

Introduction

Back to Table of Contents

COVID -19 does not stop at national of municipality bounderies. Nevertheless, once a country goes into lockdown, the spread of the virus across national or municipality bounderies decreases and a large part of the new cases occurs from a positive case within the municipality. In this paper I want to investigate how the spread of COVID-19 correlates with characteristics of a municipality such as the wealth/poverty of its citizens, population density, the age structure of the population as well as the exposure of the municipality to international migration and business relationships. I am interested in the onset of the epidemic, its intensity early on as well as its evolution towards the end of the lockdown.

Desmet and Wacziarg (2020) find substantial spatial heterogeneity across US counties. They use population density, modes of transportation, housing arrangements, the age distribution, health conditions, among other variables. At any point in time, they write, locations will continue to differ according to these characteristics. They will differ no matter the number of days since onset of the epidemic, and the dfferences will persist, perhaps even increase over time. This provides a foundation for policies that are sensitive to local specicities, where less affected places can have less stringent lockdowns or earlier reopenings. That same rationale applies to study the Belgian case.

I am taking the period March 31 till May 4 as the period under investigation for the following two reasons: First, Sciensano, the data provider, has released contamination data on a municipality level in March only when, on a given day, the number is at least five. In practice, only larger municipalities and cities make that threshold in the course of March. For smaller municipalities we have to wait till March 31 as the first day at which Sciensano realised cumulative contamination data. Of the 581 Belgium municipalities only 205 reached the threshold in March of at least 5 cases on a given day. This paper proposes a method to date the onset of the epidemic in the other muncipalities in the absence of Sciensano start date data. And second, Belgium went into lockdown from March 13 to May 4th, meaning that during this period, the same set of rules applied to the entire territory, including police enforcement as well as testing for potential cases. The latter is important for this paper as local-level discretion on testing would mean that one would find more cases (mostly mild cases) in municipalities with very broad testing and very few cases in municipalities with very low testing. That did not occur, because the testing policy during this period was nationwide the same and it was very conservative due to the absence of testing reagentia.

Recently, the team of Piet Maes (Rega Institute, KU Leuven) released >250 sequences of the virus deposited in the GISAID database. This data set represents a unique opportunity to investigate the dispersal history and dynamic of SARS-CoV-2 in Belgium: origin of introductions into the Belgian territory, relative importance of external introductions in establishing Belgian clusters of transmission, spatio-temporal distribution of these clusters, etc. Two main conclusions arise from his work so far: (i) the importance of external introduction in a municipality, (ii) the clusters resulting from these introductions are widely distributed across the country. In future work, they want to assess if this pattern evolves with the inclusion of more sequences sampled during the lockdown. These analyses are based on sequences available on the 7th April, but in the future, they will update these analyses with newly available sequences. I refer to https://spell.ulb.be/news/covid19_analyses/ for his work.

Belgium comes forward as one of the countries with the highest spatial density of sequenced SARS-CoV-2 genomes. At the global scale, his analysis confirms the importance of external introduction events in establishing transmission chains in the country. At the country scale, the spatially-explicit phylogeographic analyses highlight a global impact of the national lockdown on the dispersal velocity of viral lineages. The dispersal velocity of viral lineages was 5.4 km a day before the lockdown and 1.2 km a day in the first few weeks of the lockdown (see Dellicour et al, 2020).¹

In this paper, I use linear and log-linear models. The latter capture the exponential growth of contaminations very well, but results are often similar to the linear model, in particular when we take the lagged dependent variable into account. The contribution of this paper is to obtain a better understanding of the effect of the socio-economic and demographic variables at the municipality level, apart from the inclusion of the lagged dependent variable.

I will use this research to situate the date at which the epidemic entered a municipality in the absence of Sciensano data. I will do that in the next section. Afterwards I explain the hypotheses that I wish to test in this paper, present the estimation strategy and then the results. I use several graphs to illustrate the findings.

Description

Back to Table of Contents

Onset of the epidemic

Let us first consider the start of the epidemic in each municipality. The data obtained from Sciensano inform us about the first date a municipality reaches the threshold of at least 5 positive cases, these are cases where the contamination is confirmed in the laboratory. Sciensano does not reveal the number of cases on a given day in a give municipality when this figure is below 5. Correspondance between the author and Sciensano reveals that they refuse to do this for privacy reasons, stating that they want to avoid that contaminated persons can be identified in the community. From the 581 municipalities we have the exact date of the start of the epidemic, defined here as at least 5 confirmed contaminations for 309 municipalities of which 205 reached that threshold in March and 104 in April. For the remaining 272 municipalities Sciensano published the accumulated number of contaminated cases on March 31. If example given, a municipality had 9 contaminated cases by March 31, but never reached the threshold of 5 on a given day (eg. two on March 23; three on March 25 and four on March 30), then we only see in the data that this municipality has accumulated 9 cases by March 31.

On Graph 1, in order to use the figures for all 581 municipalities, I use two methods to inpute the earliest date the epidemic started in those 272 which remained under the 5 cases threshold during the month of March. The first method, resulting in the solid black line in the graph, depicts the start date for the 309 municipalities with known start date. And it uses the average date per district among the 205 municipalities with a start date in March in the following way: a district in Belgium is known as the “arrondissement” and is the administrative level between a municipality and a province. There are 43 districts in Belgium with an average of 13 municipalities each. In this first method I use the municipalities with known dates in March and average these dates per district. I then assign that average date to each of the municipalities in that district that remained under the 5 cases threshold but was listed by Sceiensano with its cumulative cases on March 31. This makes sense for two reasons: (i) we know these 272 municapilities registered their first case somewhere in March, thereby remaining under the Sciensano publication and privacy threshold of 5 cases per day; (ii) contaminations spread through proximity between persons. Maes and Dellicour (2020), who researched the genoom of the virus in Belgium, found that it travelled 5.4 km per day before the lockdown in Belgium (March 13) and 1.2 km per day during the lockdown. Hence it is not farfetched to assume that people living in a municipality are first contaminated by infected persons from their own municipality but in a matter of days also from infected persons from a neighbouring municipality (municipalities in Belgium rarely are more than 10km wide). The results of this method is that it moves the onset of the epidemic in these 272 municipalities to March 24 on average, rather than March 31. ²

The data reporting strategy of Sciensano (protecting privacy) allows municipalities with a smaller population to remain under the radar as larger municipalities reach the 5 cases per day threshold much easier. This can be seen from the population size in the 205 municipalities with known start day, which is on average 37,000 people. For the 272 municipalities with accumulated contamination count available on March 31 only, the population size is on average only 13,000. The method above can thus be regarded as a correction to put smaller municipalities also on the radar.

The second method, resulting in the red dashed line in Graph 1, deals with one shortcoming of the first method, to wit that in a few districts the average can be based on only one or a few municipalities with more than 5 registered cases before March 31. In an extreme example of only 1 municipality with more than 5 registered cases before March 31, method one assigns its date to all other municipalities in that district (provided of course these other municipalities have more than 5 cases by March 31, otherwise they will turn up only in April).³ To account for that I use a second method in which I subtract the number of municipalities with known dates from 31. Thus, example given, if in a district of 15 municipalities 6 reached the 5 cases threshold in the course of March, 4 reached that threshold in April and 5 have accumulated more than 5 cases by March 31, then I inpute 31-6=25 as the date in March at which the epidemic started in these 5 municipalities which feature under “March 31” in the Sciensano data . The logic is akin to method 1 above and follows the findings of Maes and Dellicour (2020): if you are surrounded by many municpalities in your district that reached the 5 cases threshold in the course of March, then you are more likely to register cases yourself prior to March 31. And, in contrast, if you are not surrounded or only a few municipalities in your district have reached the threshold, then your start date will not be far from March 31. Using this second method I arrive at March 25 on average as the start date for these 272 cases, only 1 day more compared to method 1.

To summarize, Graph 1 shows the onset of the epidemic for all belgian municipalities, in a kernel density estimate, whereby the exact Sciensano startdate is used for 309 municipalities, because they reached the threshold of at least 5 cases on a given day. For the remaining 272 municipalities we know that they accumulated at least 5 cases by March 31. Their start date is thus earlier and is advanced by 5 to 6 days on average depending on method 1 or 2 above.

Graph 1

We derive from Graph 1 that most municipalities registered their first set of contaminations between March 15 and March 30, hence in the first two weeks of the lockdown in Belgium, 70% of them to be exact, with 4% of municipalities before March 15 and 26% after March 30. This does not mean that the persons testing positive have contracted the virus during the lockdown. Given that the incubation period is on average 10 days and that only persons with symptoms were tested in the period under investigation (March 1 to May 4), it may well be that these persons contracted it prior to the lockdown.

Intensity of the epidemic on March 31

Back to Table of Contents

Moving the description to the intensity of the epidemic, defined as the number of registered contaminations by a given date. The earliest date at which we have accumulated contamination data from Sciensano is March 31. This allows us to study the total number of registered contaminated cases for all belgian municipalities on March 31. By that date Belgium registered 12,300 cases. This figure is an undercount as there are 105 municipalities that will only reach the 5 cases threshold in the course of April, hence they may have registered a few cases by March 31. By May 4 - the end of our period under study - the total count for registered contaminations in Belgium will be 50,000. Thus the total count on March 31 gives us an idea of the intensity of the epidemic at the municipality level at a relatively early stage in the epidemic, in any case before the “high point” of the epidemic, defined at the date at which the number of new contaminations is lower then the day before (or lower then the average of the last few days). This turning point in the epidemic in Belgium is situated around mid-April. The number of contaminations registered by March 31 tells us how many people contracted the virus before the lockdown and in the first few days of the lockdown (give that the average incubation period is 10 days). Map 1 presents the contamination rate per 1000 inhabitants on March 31.

Map 1: Number of registered contaminations per 1000 inhabitants on March 31, 2020

Legend, quartiles: 0 to 0.5 0.5 to 1 1 to 1.5 , +1.5

The map shows a number of clusters with a high contamination rate, notably around two flemish cities, Sint-Truiden in southern Limburg and Kortrijk in the west of the country as well as two clusters in Wallonia, one around the city of Mons in the south-eastern part and in south of the country around the city of Arlon. In general the flemish region is more affected then the walloon region, with more municipalities in the north having darker color, meaning more contaminations per 1000 inhabitants compared to the south.

Growth of contaminations in the month of April

Back to Table of Contents

The number of registered contaminations in Belgium grew from 12,500 on March 31 to 50,000 on May 4. Graph 2 depicts poor and rich municipalities. It shows the growth of contaminations, expressed as the change in contaminations per 1000 inhabitants during the month of April, meaning between the first date that we have accumulated contamination data per municipality from Sciensano (March 31) and the end of the strict lockdown (May 4). It is clear from the graph that the richest municipalities started with a disadvantage, meaning they have more registered contaminations per 1000 inhabitants compared to the poorest municipalities, 1/1000 versus 1.2/1000 to be precise on March 31.

Following the evolution of the epidemic in the month of April, on a weekly basis, which is possible with the Sciensano data, we see that by April 14th, the richest municipalities are doing better than the poorest ones, meaning they have turned the disadvantage of the early and intensive hit into an advantage, i.e less contaminations per 1000 inhabitants. Most likely this is because the population in these rich municipalities is better able to isolate itselve from fellow citizens, in the sense that they have jobs where they can work at home, a house with a garden that they do not have to share, a car which allows them to avoid public transport and so on. We come back to this in the analysis part of this paper.

The difference between 31/3 and 4/5 on the one hand, and between rich and poor municipalities on the other hand, is statistically significant at the 5% level, as can be seen in table 1. At the end of the observation period (May 4), this difference amounts to .65 cases per 1000 inhabitants. For a municipality in the poorest decile, of, on average 35,000 inhabitants, this accounts for a difference of 22 more persons contaminated compared to a municipality in the richest decile. Multiply that figure 58 times for all municipalities in this poorest decile and one obtains a difference of 1,276 more contaminated persons compared with the richest decile.

Table 1: Difference-in-Differences, #registered contaminations per 1000 inhabitants in the course of the month of April, by poorest and richest decile, N==117

Week	March 31	May 4	Difference
Income decile
Poorest	0.99	4.24	3.25^***
Richest	1.18	3.78	2.60^***

Difference	0.19^*	0.46	0.65^**

Graph 2

This finding resonates with Bartsher et al (2020) who exploit within-country variation in the spread of COVID-19 and in social capital (proxied by voter turnout) in seven european countries and they find that find fewer accumulated cases of the virus on areas of high social capital.

We can also look at the entire income distribution over all municipalities and visualise the effect of income on the growth trajectory of the epidemic by comparing the contamination rate on March 31 with that at the end of the observation period (May 4). We notice that the initial disadvantage (early contaminations) and later advantage (slower growth of the rate of contaminations) kicks in at around 20,000 euro per capita per year. Around 30% of Belgian municipalities have more than 20,000 euro per capita income in 2018 according to fiscal data.

Graph 3

Analysis

Back to Table of Contents

Several factors can influence the onset, the intensity early on and the growth of the COVID-19 epidemic in a given municipality. Research on the genoom of the virus in Belgium had discovered that not one but multiple strands of the virus have been introduced in Belgium, meaning there is not one person ‘zero’ who introduced the virus in Belgium but there are multiple ones that introduced it independently of each other. Virologists distinguish three types A, B and C, each following a different trajectory accross the globe. There are three well-known cases at the very start of the epidemic in Belgium: (i) the first person with contamination demonstrated in the laboratory was a businessman returning from Wuhan in China, (ii) the second was a business women returning from a business trip in France and other countries, and (iii) was a group of 13 friends and family who returned from a ski holiday from the same hotel in northern Italy.

All three contracted the virus while travelling abroad, with business trips and ski holidays linked to the richer part of the population. At the same time, once the virus was present in the municipality and the country went into lockdown, richer municipalities may have been better able to limit the spread of the virus, for reasons mentioned above. This allows us to derive the following hypotheses that will be put to the test with the data:

(1a) “The virus was earlier introduced in richer municipalities”

(1b) “Richer municipalities were bette able to limit the spread of the virus”

From an inspection of Map 1 it seems that there are three clusters of contaminations close to the French border, notably around the cities of Kortrijk, Mons and Arlon. Ginsburgh et al (2020) show that the northern part of France, bordering Belgium, has the highest death toll in France outside of Paris. Municipalities with a lot of cross-border labour and consumer migration may bring the virus home to their own municipalities. This may also have been the case with municipalities on the border with The Netherlands, Germany and Luxemburg. The city of Tilburg in the south of the Netherlands for example was the site of a major, early, outbreak.

On a global scale, the virus had several clusters of outbreaks before it arrived in Belgium via travellers, notably Wuhan in China and northern Italy. In the first weeks of March it was standard belgian policy not to test travellers who returned from business or holiday abroad when upon return they did not have any symptoms. The group of 13 friends and family for example needed to insist being tested because their holiday hotel was considered outside of the contaminated zone. Once the virus is introduced in the municipality the role of travellers, be them belgians, migrants or foreign residents returning to Belgium from visiting family or friends, may become less important for the subsequent evolution of the epidemic as the virus does not distinguish between nationalities. This leads to the second set hypotheses:

(2a) “Municipalities closer to a border had an earlier onset”

(2b) “Municipalities with more foreign nationalities had an early onset”

(2c) “The presence of foreign nationals does not play a role after the initial start”

There are also two variables whose likely effect is “obvious”, that is population size and population density. Municipalities inhabiting more people have defacto a larger probability to start the epidemic early, and given that contamination occurs trough close encounters with fellow citizens, population density is likely to play a role once the virus has been introduced in the municipality. This allows us to derive the next hypotheses:

(3a) “More populated municipalities had an early onset of the epidemic”

(3b) “More densily populated municipalities had a stronger growth of the epidemic“

The age-structure of the population and the size of households may also play an important role in the spread of epidemic. Elderly citizens are much more at risk, in particular when one considers the mortality risk, which we do not consider in this paper (Barnett and Grabowskoi, 2020). It is also well-known that the virus hit care homes very hard. Once the virus is introduced in a care home, via visitors or via a staff member, a large percentage of residents and staff may get infected. In addition, many people contract the virus via an infected household member. This leads us to the following three hypotheses:

(4a) “Municipalities with a larger share of elderly citizens had a faster growth”

(4b) “Municipalities with a larger share of their elderly population residing in care homes have a higher contamination rate”

(4c) “Municipalities with a larger share of singleton households had slower growth”

Belgium being a federal country with three entities that have their own socio-demographic and economic characteristics governed by regional public authorities next to the federal level, which may have its own effect on the epidemic, leading to the following hypothesis:

(5) “The location of a municipality in one of the three regions of the country, the Brussels Capital Region, The Flemish Region and the Walloon Region may effect the start day, the intensity early on and the growth of the epidemic. It is however likely that the effect of these dummy variables is captures, at least partly, by the socio-economic and demographic variables above.”

We will now take these hypothesis to the data by regressing the start day of the COVID-19 epidemic for each municipality, the intensity on 31/3 as well as the growth of the epidemic on a number of socio-economic and demographic characteristics. In particular, we will test the following equations:

$Start\ Day_i=\alpha_0+\sum\limits_{j=1}^n \beta_jX_{ij} +\varepsilon_i$ $Cases\_31\_March_i=\alpha_0+\sum\limits_{j=1}^n \beta_jX_{ij} +\varepsilon_i$

Whereby the dependent variable is the start day of the epidemic (equation 1) or the number of contaminated cases on March 31 (equation 2). Alfa is a constant, X a vector of socio-economic an demographic characteristics, beta the coefficient for each of these characteristics and e the error term.

Our third independent variable, the growth of the epidemic, defined as the cumulative number of contaminations on May 4 minus the cumulative number of contaminations on March 31 will be regressed as follows:

$\tag{3a} Growth\_Cases\_May\_4\_March\_31_i=\alpha_0+\delta_i Cases\_31\_March+\sum\limits_{j=1}^n \beta_jX_{ij} +\varepsilon_i$ $\tag{3b} (ln)Growth\_Cases\_May\_4\_March\_31_i=\alpha_0+\delta_i lagged\_cases+\sum\limits_{j=1}^n \beta_jX_{ij} +\varepsilon_i$

Whereby the dependent variable (Growth_Cases_May_4_March_31) is the change in contaminations between May 4 and March 31. As this is a change (or growth variable), we need to control for the initial value of the dependent variable, to wit the number of contaminations per 1000 inhabitants on March 31 (Cases_31_March). The other independent variables are the same as in equations 1 and 2, with X a vector of socio-economic and demographic characteristics. We also run a regression where we use the weekly observation of the contamination rate. To capture the exponential growth of the contamoinations we also run the model in log-linear form (equation 3b). To enable the transformation to logs, we add +1 to the number of cases (avoiding log 0 resulting in missing values).

As practised in the economic growth literature to deal with potential endogeneity, we will instrument the baseline value of the dependent variable, based on findings from estimating equation 2. Hence, we first estimate equation 2 to determine the initial level of the dependent (Cases_31_March) and then instrument it with one or more variables that affect(s) the growth of the epidemic only through their effect on the intial level. This two-stage estimation is then as follows:

$\tag{4a} Cases\_31\_March_i=\alpha_0+\sum\limits_{j=1}^n \beta_jX_{ij} +\varepsilon_i$ $\tag{4b} (ln)Growth\_Cases\_May\_4\_March\_31_i=\alpha_0+\delta_i \hat{Cases\_31\_March}+\sum\limits_{j=1}^n \beta_jX_{ij} +\varepsilon_i$

Results

Back to Table of Contents

In Table 1 we present the results of our regression analysis for the start day using ordinary least squares. We introduce our variables of interest gradually in order to find out if their impact is or is not affected when we introduce subsequent variables.

Starting with population size and population density it is clear that both have a statistically significant effect on the start date of the epidemic: the higher they are, the earlier the start date. I remind of the definition of ‘start date’ in this paper: the first date at which the municipality has at least 5 registered contaminations on a given day. This was used because of the limitations of the data published by Sciensano. It does not represent the date at which the very first contamination appeared in the municipality. Rather it is a indicator that combines the entry of the virus in a municipality with a first indication of the spread at a very early stage. Our dependent variable is here the number of days after February 29 and its computation was set out earlier in this paper. I use the results of method 1 here. The count for this variable has a minimum of 4 (March 4) and a maximum of 64 (May 3) and its distribution is shown on Graph 1.

The second set of regressors introduced are population composition variables, more in particular the share of population above age 65 and the share of singleton households. As outlined in the hypothesis section the first variable should expedite the date that the 5 cases threshold was reached, whereas the second variable is hypothesized as increasing (or better postponing) the date. Both variables do what they are supposed to do in regression 2 in table 1. Adding average income per capita net of taxes based on fiscal data, I find in regression 3 that richer municipalities were confronted with the virus in their municipality earlier on. The reasons for that have been set-out earlier, in particular business travel and ski-holidays.

In regresssion 4 we add the share of elderly persons that reside in a care home, computed as the number of officially registered and certified beds in care homes devided by the size of the population above age 65. The higher the share, the earlier the onset. While elderly persons are not responsible for the introduction of the virus in the municipality (as compared to business travellers and skiers), they may have spread it without knowing or before they knew they were contaminated. I refer again to the definition of ‘start date’ which does not identify the first day or patient zero, but rather the first five positive cases in a given day. I also refer to the restrictive test policy.

Table 1: the correlates of the onset of the epidemic at the municipality level, OLS

Note: ^*stat.sign at the 10% level, ^**at the 5% level and ^***at the 1% level. Robust standard errors between brackets.
Dep. Var.	Start date of the COVID-19 Epidemic, in #days after March 1
	R1	R2	R3	R4	R5	R6
Indepen. Var.
Pop.size (‘1000)	-.10^*** (.03)	-.12^*** (.04)	-.12^*** (.036)	-.11^*** (.036)	-.11^*** (.032)	-.10^*** (.03)
Population density (‘1000)	-.45^*** (.17)	-1.17^*** (.23)	-1.14^*** (.22)	-1.02^*** (.22)	-.54^** (.23)	-.001^*** (.0002)
% of population +65 of age		-.64^*** (.13)	-.45^*** (.15)	-.37^** (.16)	-.48^*** (.16)	-.48^*** (.16)
% singleton households		.46^*** (.10)	.31^*** (.12)	.27^** (.12)	.30^*** (.11)	.39^*** (.13)
Income per cap. (‘1000)			-.55^*** (0.18)	-.56^*** (.17)	-.40^** (.17)	-.52^*** (.19)
% of pop +65 in care homes				-.25^* (.13)	-.26^* (.13)	-.32^** (.13)
Municipality at the border					4.44^*** (1.36)	1.44 (1.45)
%Foreign Nationalities					-.20^** (0.08)
% French						.27^** (.12)
% German						.10 (.17)
% Dutch						.005 (.11)
% Italian						-1.28^*** (.19)
% Chinese						-14.3^*** (4.26)
Constant	29.91^*** (.64)	28.79^*** (3.21)	40.42^*** (4.80)	41.41^*** (4.79)	39.89^*** (4.59)	40.36^*** (5.21)
N	581	581	581	581	581	581
R2	0.14	0.21	0.23	0.23	0.26	0.30

In regressions 5 and 6, I add variables that capture exposure to foreign contacts. In regression 5 I introduce a binary variable that equals one if the municipality borders a neighbouring country (France, Luxemburg, Germany or the Netherlands) and equals zero otherwise. I expected that the proximity of the border would make these municipalities easy targets to import the virus from across the border. However, the opposite seems to be the case, these municipalities have on average a later start date, more than 4 days later to be precise, an effect that is statistically significant at the 1% level. The second variable I introduce here is the percent of in habitants who do not carry Belgian nationality. A belgian municipality has on average 7.8% inhabitants with non-belgian nationality, varying from 1% in remote rural areas to 48% in a municipality in Brussels. The variable is introduced here for the same reason as the border variable, with the twist that the exposure to the virus does not come form geographical proximity to another country, but rather from personal proximity to family, friends, migration, trade, holiday and travel to and from the country of nationality. I find that higher exposure (measured by a higher share of non-belgians) expedites the spread of the virus.

When we then investigate more carefully for which nationalities in particular exposure matters more than others, we do find a marginally significant effect of the share of persons with French nationality. I remark that the binary border variable now turns statistically insignificant, as its effect is at least partially captured by the newly introduced French variable. Two other variables however have a statistically significant effect at the 1% level, to wit the share of inhabitants with Italian and with Chinese nationality. The share of population in a municipality with Italian nationality varies from 0 to 10% and with chinese nationality from 0 to 1.2%. An increase by 1% in the share of italian inhabitants expedites the date with 1.3 days and we arrive at approximately the same result when increasing the share of chinese in the municipality with 0.1%.

COVID-19 spread from the city of Wuhan in China to the rest of the world and in Europe, the virus set hold in northern Italy before travelling to other places. It could be that Italians or Chinese residing in Belgium or travelling to Belgium to visit family and friends have imported the virus after visiting or residing in Italy or China, but it could also be that Belgians who entertain business, trade, friendship or marital relations with the Italian or Chinese inhabitants of their municipality imported the virus through travelling to either country. And it could be both, as there were multiple introductions of the virus in Belgium, as we discussed above.

Importantly, the effect of the other variables, in terms of the magnitude of the coefficient and its statistical significance does not change very much after the subsequent introduction of additional variables, indicating that each of them has an effect on the dependent variable after controlling for the other variables. The R-squared gradually improves after the introduction of new variables. Overall, the R-squared remains relatively low, meaning that our data are widely dispersed alongside the regression line. High-variability data nevertheless can have a significant trend. The trend indicates that the independent variables still provide information about the start date of the epidemic even though data points fall further from the regression line.

The start date was on average earlier for municipalities located in the Region of Brussels Capital (March 18), then Flanders (March 24) and than Wallonia (March 31). It however does not make much sense to introduce dummy variables for the region in the regression here as they account for part of the variation that we want to capture with the other variables. Rather, upon introduction (not shown) the effect of the population composition variables (density, young and elderly population) disappears, leaving the effect of the other variables intact.

Moving on to our second dependent variable, the contamination rate per 1000 inhabitants on March 31, we find several of the effects presented in Table 1, but with the opposite sign (see Table 2). In the first 5 regressions of Table 2 the effect of the regressors is as expected, with the only difference to Table 1 that the income variable is not statistically significant. Meaning that rigth after the initial introduction of the virus (as presented in table 1) income does not seem to affect the contamination rate on March 31.

In columns 6 and 7, I introduce the start date as an additional regressor. Most likely municipalities where the virus has been introduced earlier have a higher contamination rate on March 31, which is indeed born out by the OLS results in column 6. One may argue that the start date is endogenous to the characteristics of the municipality, which I have indeed demonstrated in Table 1. To that extent I instrument the start date with the share of Italians in the municipality, whereby I assume that the share only effects the contamination rate through its effect on the start date, which indeed seems to be the case here, with the test-statistic for underidentification as well as the F-test pointing in the right direction. Results of the IV are very similar tot he OLS. There is no difference in the magnitude of the coefficient in the OLS and IV regression, nor in its statistical significance.

Next, I discuss the results on the growth of the epidemic in the month of April. These are presented in Table 3. I remind here what is meant by growth: the change in the contamination rate between May 4 and March 31. In our analysis, the inclusion of a lagged dependent variable is obvious because the contamination rate at time t+1 depends on the contamination rate at time t. The relation between the lagged dependent and the dependent variable is modeled here in a linear way. In epidemiological models , their relation may be modeled in a different and more complex way. To capture the exponential growth of contaminations, I also use the log-linear model. Results do not differ much from the linear model.

Lagged dependent variables can account for measurement error, noise and other effects. The question here is should we control only for the baseline value of the contamination rate (meaning on March 31) or should we include weekly lags of the contamination rate? In the first case I am only interested in estimating the growth of the contamination rate over the whole period of the strict lockdown, whereas in the other case I am regressing the contamination rate on a weekly basis. There are two main arguments to guide that discussion.

Table 2: the correlates of the contamination rate on March 31, OLS and IV regression

Note: ^*stat.sign at the 10% level, ^**at the 5% level and ^***at the 1% level. Robust standard errors between brackets.
Dep. Var.	Contamination rate per 1000 inhabitants on March 31
	R1	R2	R3	R4	R5	R6	R7-IV
Ind. Var.
Start day of epidemic						-.05^*** (.004)	-.05^*** (.012)
Pop density (‘1000)	.06^*** (.01)	.06^*** (.01)	.05^*** (.01)	.03^** (.017)	.04^*** (.01)	.01 (.01)	.007 (.02)
% pop +65	.041^*** (.011)	.042^*** (.007)	.034^*** (.012)	.039^*** (.012)	.041^*** (.013)	.028^** (.011)	.026^*** (.010)
% singleton households	-.017^*** (.006)	-.015^** (.007)	-.013^* (.007)	-.014^** (.007)	-.023^*** (.009)	-.024^*** (.008)	-.023^*** (.006)
Income per capita		.006 (.013)	.006 (.013)	-.002 (.014)	-.001 (.015)	-.02 (.015)	-.02 (.014)
%pop+65 in care homes			.017^* (.010)	.017^* (.01)	.018^* (.009)	-.002 (.01)	-.0003 (.009)
Border				-.22^** (.09)	-.06 (.10)	-.0002 (.09)
%Foreign Nationalities				.006 (.005)
% French					-.015 (.014)	-.0008 (.01)
% German					-.014^** (.006)	-.01^* (.05)
% Dutch					-.006 (.008)	-.013^** (.007)
% Italian					.05^** (.02)	.-004 (.02)
% Chinese					.88^*** (.25)	-.09 (.45)
Constant	.67^*** (.25)	.54 (.36)	.49 (.35)	.60 (.36)	.73^* (.40)	3.12^*** (.45)	3.03^*** (.74)
N	581	581	581	581	581	581	581
R2	0.04	0.04	0.04	0.05	0.08	0.36	0.35
Underid. test							23.63^***
First stage							Start day
% italian							-1.24^*** (.23)
F-test							29.03^***

According to Angrist and Pischke (2009), the lagged dependent model should be preferred when the assumption “that the most important omitted variables are time-invariant doesn’t seem plausible”. For this paper it means the extent to which we believe the factors that change during the lockdown, may have contributed to the change in the contamination rate. This reasoning applies to both the inclusion of the 6-week lag and the 1-week lag. If we believe such time-variant factors accors municipalities may have played an important role, than we should prefer the lagged-dependent model with a lag of one week. Since I do not have time-variant regressors to include in the estimation, not at the weekly level nor during the entire period of the lockdown, I do not know how relevant they are. However, we do know that the lockdown was a national (meaning in Belgium the federal level) policy and plausibly applies to all municipalities in the same way. The (federal) Minister of the Interior, Pieter De Crem, who is in charge of te police force issued a decree at the start of the lockdown with clear, uniform instructions for the enforcement of the lockdown regulations.

Secondly, McKenzie (2012) argues that, in cases of low autocorrelation of dependent variables, controlling for the lagged dependent variable is more powerful than either employing the difference-in-difference estimator or the single difference estimator. His intuition is that, in cases where baseline data have little predictive power for future outcomes, it is inefficient to fully correct for baseline imbalances. In our data however, the correlation between the contamination rates at the municipality level on March 31 and the subsequent weeks of the epidemic is very high: The Pearson correlation coefficient is $.80^{}$ after one week and still $.60^{}$ on May 4th (see Graph 5). Hence, the inclusion of a weekly lagged dependent variable explains a large part of the variation in the dependent variable.

Since both the inclusion of a 6-week lag and a weekly lag may shed ligth on the correlates of the growth of contaminations, I present both models: in Table 3, we control for the baseline value of the dependent variable (the contamination rate on March 31) only in our estimation of the contamination rate on May 4th and in Table 4 we use weekly lags. We include these regressions to find out if the results differ. Results in both tables are very similar.

Graph 4

Graph 5

In Table 3, income seems to be more important for the Flemish Region as compared tot he Walloon Region, whereas in the latter the share of the elderly population in care homes as well as the share of inhabitants from neighbouring countries exercise a statistially significant effect.

The most used method in case of lagged dependent variables combined with problems of endogeneity is General Methods of Moments (GMM), the preferred workhorse for dynamic panel data estimation. The Arellano–Bond estimator (1991) sets up a generalized method of moments (GMM) problem in which the model is specified as a system of equations, one per time period, where the instruments applicable to each equation differ (for instance, in later time periods, additional lagged values of the instruments are available). The instruments include suitable lags of the levels of the endogenous variables (which enter the equation in differenced form) as well as the strictly exogenous regressors and any others that may be specified. With the data at hand however, as mentioned earlier, we do not have time-variant covariates, hence we can only use the values of the time-invariant covariates as instruments in a GMM level equation. Past values of those instruments are however the same as current values, making GMM not feasible here, and leading to overidentification problems when executed.

Results of the analysis, presented in Table 3 show that three variables are correlated with the change in contamination rate (apart from the lagged dependent variable of course) across specifications and models: the share of the population above age 65 in a municipality, the average income per capita in the municipality and the share of the population above age 65 residing in a care home. Richer muncipalities manage to slow down the epidemic wheres a higher share of elderly in the municipality and a higher share of elderly residing in care homes accelerate it.

The effects of singleton households and population density disappears once we introduce regional dummies for the Brussels Capital Region, the Flemish and the Walloon Regions, demonstrating that the latter are correlated with these regional effects. In seperate regression per region, these variables no longer turn up in a statistically significant way.

Table 3: the growth of the contamination rate in April, OLS and IV regressions

Note: ^*stat.sign at the 10% level, ^**at the 5% level and ^***at the 1% level. Robust standard errors between brackets.
Dep.Var	Growth of the Conta. Rate between May 4 and March 31
	Lag is 6 weeks
	Log-linear	Linear model
	R1	R2	R3	R4	R5 - IV	R5 - IV
Indep. Var.	Belgium	Belgium	Flemish Region	Walloon Region	Belgium	Belgium
Lagged Conta.Rate	.035^*** (.028)	2.07^*** (.16)	2.46^*** (.21)	1.63^*** (.18)	2.38^*** (.68)	2.14^*** (.62)
Pop density (‘1000)	-.004 (.01)	-.01 (.04)	.23 (.29)	.28 (.35)	-.003 (.037)	-.0003 (.036)
% pop +65	.015^* (. 009)	.06^* (. 03)	.08 (.05)	.09 (.06)	.07^** (.034)	07^** (.03)
% singleton households	-.005 (.006)	.017 (.025)	-.05 (.041)	.06 (.04)	-.003 (.019)	-.003 (.019)
Income per capita	-.017^* (.01)	-.082^** (.040)	-.18^*** (.06)	.04 (.075)	-.11^*** (.036)	-.11^*** (.036)
%p.+65 in care home	.023^*** (.007)	.086^*** (.031)	.047 (.03)	.11^** (.046)	.07^** (.03)	.07^** (.036)
Foreign nat.
% French	-.013^** (.005)	-.066^** (.027)	.09 (.16)	-.08^*** (.027)
% German	-.001 (.004)	-.02^* (.011)	-.22 (.33)	-.06^*** (.018)
% Dutch	.009^* (.005)	.044^* (.025)	.022 (.026)	.72^*** (.22)
% Italian	.028^** (.011)	.028 (.053)	-.07 (.13)	.07 (.075)
% Chinese	.047 (.17)	-.87 (.76)	1.06 (1.09)	-1.75^** (.70)
Regional D. Brus.=base
Flemish	.003 (.10)	.21 (.47)	-	-	.51 (.47)	.49 (.45)
Walloon	.18 (.11)	.98^* (.51)	-	-	1.2^** (.56)	1.1^** (.54)
Constant	1.33^** (.31)	1.12 (1.39)	4.47^*** (1.45)	-2.12 (2.44)	1.57 (1.39)	1.78 (1.35)
R2	0.32	0.41	0.50	0.39
N	581	581	300	262	581	581
Underid. test					13.2^***	15.0^***
Over ident. Sargan stat.						0.83 (0.36)
First stage					Cont.r. 31/3	Cont.r. 31/3
% Italian					-.08^*** (.02)	-.075^*** (.02)
% Chinese						.56^* (.33)
F-test					12.9^***	8.20^***

We have seen earlier on that the share of italian and chinese nationals had a statistically significant effect early on in the epidemic: a higher presence is correlated with an earlier start and a higher contamination rate on March 31. Beyond this initial effect, the share of foreign nationals may also affect the subsequent growth of the epidemic, eg. via the frequency of its interactions within the municipality, as the virus does not discriminate between nationalities. In R2 the italian variable is no longer statistically significant, but it is in the long-linear model in R1. This is one of the few occasions were the models yield a different result.

In order to deal with the endogeneity of the lagged dependent variable we introduce the shares hence as instrumental variables in the linear version of the model in R5 and R6, thereby assuming that they only affect the growth of the epidemic through their impact on the contamination rate of March 31. This makes sense as the country was in lockdown, so one could not anymore cross national borders and import the virus from abroad. The test-statistics (F-test, underidentification test and overidentification test) are all favourable. In R5 and R6 the three variables mentioned above: income per capita, share of population above 65 and share of elderly in care homes are all statistically significant, with similar magnitude as in column 2.

In Table 4 we account for weekly lags of the dependent variable and model the growth of the epidemic in a linear as well as a log-linear way. Results are similar to Table 3 in the sensse that the same variables turn up in a statistically significant manner: share of the elderly population, share of that population residing in care homes and income per capita. The percentage of the population with foreign nationalities proofs statistically significant in these regressions, in particular when analysing the Walloon Region separately.

Table 4: Random effects panel data estimation with 1-week lags

Note: ^*stat.sign at the 10% level, ^**at the 5% level and ^***at the 1% level. Robust standard errors between brackets.
Dep.Var	Growth of the Contamination Rate between 4/5 and 31/3 with one-week lag
	Linear model			Log-linear model
	R1	R2	R3	R4	R5	R6
Indep. Var.	Belgium	Flemish Region	Walloon Region	Belgium	Flemish Region	Walloon Region
Lagged Conta.Rate	1.^*** (.009)	1.02^*** (.01)	.99^*** (.01)	.68^*** (.013)	.67^*** (.017)	.68^*** (.02)
Pop density (‘1000)	-.001 (.01)	-.03 (.05)	.07 (.07)	.002 (.004)	.02 (0.26)	.06^** (.03)
% pop +65	.015^** (.006)	.02^** (.009)	.02 (.01)	.007^** (.003)	.012^** (.005)	.005 (.007)
% singleton households	.001 (.005)	-.014^* (.008)	.011 (.009)	-.0002 (.003)	-.005 (.004)	-.0007 (.005)
Income per capita	-.016^** (.007)	-.03^*** (.01)	.01 (.014)	-.005 (.004)	-.017^** (.007)	.006 (.008)
%p.+65 in care home	.019^*** (.004)	.01 (.007)	.025^*** (.009)	.008^*** (.003)	.009^** (.004)	.009^** (.004)
Foreign nat.
% French	-.014^*** (.005)	.004 (.02)	-.017^*** (.005)	-.007^*** (.002)	-.0003 (.001)	-.007^** (.003)
% German	-.006^** (.002)	-.017 (.05)	-.014^*** (.004)	-.002 (.003)	.003 (.03)	-.005^* (.003)
% Dutch	.005 (.005)	.002 (.005)	.15^*** (.04)	.002 (.002)	-.0006 (.002)	.064^*** (.017)
% Italian	.02^** (.009)	.02 (.026)	.02 (.014)	.015^*** (.004)	.026^* (.015)	.012^* (.006)
% Chinese	-.047 (.12)	.27 (.19)	-.23 (.16)	.053 (.064)	.068 (.097)	.09 (.085)
Regional D. Brus.=base
Flemish	.02 (.088)	-	-	-.01 (.04)	-	-
Walloon	.11 (.096)	-	-	.02 (.04)	-	-
Constant	.42 (.28)	1.14^*** (.30)	-0.41 (.49)	.48^*** (.13)	.67^*** (.18)	-.026 (.23)
N	2905	1500	1310	2905	1500	1310

Conclusion

Back to Table of Contents

I performed a number of analyses in this paper, which focussed on socio-economic and demographic correlates of the contamination rate at the municipality level. I employed linear as well as log-linear models. The latter capture the exponential growth of the contaminations better, but results are very similar to the linear model. The correlation between the dependent and lagged dependent is very high, understandably in an epidemic. In Belgian municipalities during COVID-19, it is as high as 0.80 between March 31 and April 7 and still remains 0.60 between March 31 and May 4.

The inclusion of the lagged dependent variable thus explains a lot of the variation of the dependent in my regression models. Hence it is important to find the correlates of the contamination rate early on in the epidemic. I do that be estimating the correlates of the start date of the epidemic as well as the correlates of the contamination rate on March 31. I also turn my attention to the evolution of the epidemic by analysing the contamination rate at the end of the strict lockdown (May 4) and I find the same pattern. Income per capita, the share of elderly in the population, the share of elderly in home care and the exposure of the municipality to foreign travel, business and migration show up statistically significant in the analysis. Income in particular in the Flemish Region and foreign nationalities in particular in the Walloon Region.

The paper benefited from the data collected and released by Sciensano, but also faced the limitations of this provision as the exact number of cases on a given day was not published when this figure was below 5, for privacy reasons.

References

Back to Table of Contents

[1] Arellano, Manuel; Bond, Stephen (1991). “Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations”. Review of Economic Studies. 58 (2): 277.

[2] Angrist, Joshua D., and Jörn-Steffen Pischke. 2009. Mostly Harmless Econometrics: An Empiricists Companion. Princeton: Princeton University Press.

[3] Barnett, M. L. and D. C. Grabowski (2020), “Nursing Homes Are Ground Zero for COVID-19 Pandemic,” Health Forum, Journal of the Amercan Medical Association, vol. 1, no. 3

[4] Bartscher, Alina Kristin, Sebastian Seitz, Sebastian Siegloch, Michaela Slotwinski and Nils Wehrhöfer (2020), Social Capital and the Spread of COVID-19: Insights from European Countries, IZA Discussion Paper No. 13310, IZA Institute of Labor Economics, May.

[5] Dellicour S, Durkin K, Hong SL, Vanmechelen B, Martí-Carreras J, Gill MS, Meex C, Bontems S, André E, Gilbert M, Walker C, De Maio N, Hadfield J, Hayette MP, Bours V, Wawina-Bokalanga T, Artesi M, Baele G, Maes P (submitted). A phylodynamic workflow to rapidly gain insights into the dispersal history and dynamics of SARS-CoV-2 lineages. biorXiv 2020.05.05.078758; doi: https://doi.org/10.1101/2020.05.05.078758

[6] Desmet, K. and R.Wacziarg, 2020. Understanding Spatial Variation in COVID-19 across the United States, NBER Working Paper No. 27329 June

[7] Ginsburgh, Victor, & Glenn Magerman & Ilaria Natali, 2020. “COVID-19 and the Role of Economic Conditions in French Regional Departments,” Working Papers ECARES 2020-17, ULB – Universite Libre de Bruxelles.

[8] Laboratory for spatial epidemiology at ULB, https://spell.ulb.be/news/covid19_analyses

[9] McKenzie, David. 2012. Beyond Baseline and Follow-up: The Case for more T in Experiments. Journal of Development Economics, 99(2): 210-21.

Dellicour S, Durkin K, Hong SL, Vanmechelen B, Martí-Carreras J, Gill MS, Meex C, Bontems S, André E, Gilbert M, Walker C, De Maio N, Hadfield J, Hayette MP, Bours V, Wawina-Bokalanga T, Artesi M, Baele G, Maes P (submitted). A phylodynamic workflow to rapidly gain insights into the dispersal history and dynamics of SARS-CoV-2 lineages. biorXiv 2020.05.05.078758; doi: https://doi.org/10.1101/2020.05.05.078758 ↩
Using weighted average whereby the start date in the 205 municipalities is weighted with the number of registered contaminations on that day does not change the results, it moves the start date at the district level with 0.5 or 1 day compared to method 1. ↩
This extreme case occurs 6 times in a total of 43 districts. On average a district has 5.2 municipalities ( standard deviation 4.2) with start date before March 31. ↩

COVID-19 Pandemic and Pollution: Evidence from Belgium and France

2020-06-29T17:22:10+00:00

Table of Contents:

Introduction
Data sources and methodology
- Pollution Data
- COVID-19 Data
  - Belgium
  - France
Results
- Graphical Analysis
- Econometric Analysis
Conclusions
References
Additional Material

Introduction

Back to Table of Contents

The outbreak of COVID-19, which is the most serious public health crisis in decades, is bringing complex challenges worldwide. Fighting this pandemic requires a better understanding of the so-called ‘risk factors’ associated with the spread of the virus. Identifying the driving forces favoring the diffusion of the virus is a topic of primary importance, not only to contain the rapid spread of the current pandemic, but also to prevent the occurrence of future pandemics.

Up to date, there is solid evidence that the elderly are the most vulnerable age group. Patients with some underlying health conditions might be affected more severely. Recently, concerns arose as poorer areas might be more affected by the pandemic than more affluent areas.¹ There are also cases of people with no risk factors who become critically ill. Given the novelty of the virus, however, more research about ‘risk factors’ related to the COVID-19 virus is needed.

Some researchers have postulated the existence of a relationship between the level of air pollution and the pace at which the virus spreads. In a recent position paper, a group of researchers at SIMA (Italian Society for Environmental Medicine) has shown, mainly through graphical analysis, the existence of a positive association between the number of COVID-19 cases and $PM_{10}$ (Particulate Matter 10) daily limit exceedances in Italian provinces (SIMA, 2020). The hypothesis, here, is that the particulate matter may act as a ‘vector’, facilitating transmission and allowing the virus to spread quickly. Soon after, the presence of the virus was uncovered on 34 samples of $PM_{10}$ in the province of Bergamo (Setti et al., 2020).

Even though some first evidence on the relationship between pollution and COVID-19 is emerging, research is still clearly scant and mostly limited to graphical analysis. In this paper, we contribute to the literature by taking a closer look at the relationship between air quality (as measured by the presence of Particulate Matter, PM10 and PM2.5, in the air) and COVID-19-related harm. We focus on two hardly-hit European countries: Belgium and France. We consider measures of $PM_{10}$ and $PM_{2.5}$ concentrations at the province level for Belgium, and at the regional level for France. Our outcome variables include the number of COVID-19 cases, the number of hospitalizations, the number of individuals in Intensive Care Units (ICU in what follows) and the number of deaths (at the province or regional level). We exploit variation both across regions (or provinces) and across time to study the association between the two pollutants and the virus. Through graphical analysis, we provide suggestive evidence of a positive association between the two. We further corroborate this evidence through simple regression analysis.

Understanding the association between pollution and the COVID-19 pandemic is important from a public health and a regulatory point of view. Indeed, this enables to provide both the scientific community and decision-makers with valuable knowledge and tools to react more promptly in case other epidemics arise in the future and on which regulatory interventions would be useful for their prevention and containment. Moreover, this type of analysis provides regulators with information on which factors should be taken into account when examining possible lockdown easing policies and exit strategies.

The rest of the paper is organized as follows. In Section 2, we describe the data sources and the methodology used to clean and aggregate the available information. Section 3 shows and discusses the results of our analysis for Belgium and France. Section 4 concludes.

Data sources and methodology

Pollution Data

Back to Table of Contents

Pollution data was collected from the European Environment Agency (EEA) website. This database provides information on concentrations of selected pollutants for each sampling point in each member state. EEA receives up-to-date data from most of its member states on an hourly or daily basis, depending on the country. We select two pollutants, $PM_{10}$ and $PM_{2.5}$, which are measured in $\mu g/m_{3}$.

We consider the average daily concentrations across sampling points in the same province or region, in order to obtain an average province/regional measure of the quality of the air. We have data for each province in Belgium (for a total of 11 provinces) and each (old) region in France (excluding overseas).

We focus on $PM_{10}$ and $PM_{2.5}$ for two reasons.² First, the particulate matter contribute predominantly to the development of respiratory and cardiovascular diseases. Its severe effects on human health and mortality are largely due to the small size of these particles, which allows them to penetrate deep into the lungs (WHO, 2013). Second, prior research has found a positive association between these two pollutants and the spread of the COVID-19 pandemic. As mentioned, one hypothesis is that the particulate matter acts as a ‘vector’ that transports the virus and, hence, facilitates its circulation. In this paper, we aimed at verifying the validity of this hypothesis.

COVID-19 Data

Back to Table of Contents

We use different data sources for COVID-19 in Belgium and France, as described below. Each country has been affected by the virus in different moments of time. As a consequence, each country started to count cases, hospitalizations and deaths at different dates. Therefore, we perform our analysis for each country separately.

Moreover, our (simple) econometric analysis exploits the panel structure of our data. Hence, the unit of analysis is province-day for Belgium, and region-day for France. We always consider outcome variables standardized by the province or regional population and lagged by 20 days. That is, each COVID outcome variable is regressed on the level of pollution 20 days before. The rationale is that, according to the currently available medical knowledge, the virus takes between 14 and 20 days before showing the first symptoms. As a consequence, each panel includes pollution data starting from 20 days before the COVID-related harm counts (see below).

Finally, the dependent variables are always expressed in logarithmic terms, while the independent variables (pollution concentrations) are kept in levels.

Belgium

Data concerning coronavirus-related harm are collected from Sciensano.be. This database provides information on the number of COVID-19 cases, as of March 1, and on hospitalizations, as of March 15, both at the province level and on a daily basis. Data on deaths are, unfortunately, only available at the regional level. Hence, for Belgium, we consider the following outcome variables: total number of individuals currently hospitalized per 100,000 inhabitants; total number of individuals currently in ICU per 100,000 inhabitants; number of cases per 100,000 inhabitants. Population data are collected from Statbel and refer to population on January 1, 2020.

The panel containing information on the number of cases, includes data from February 10 (20 days before the count, starting on March 1) to April 22. The panel containing information on the number of hospitalizations, instead, includes data from February 24 to April 23.

France

Data concerning coronavirus-related harm are collected from ’Sante Publique France’ website. This database gives information on the total number of individuals currently hospitalized, on the number of individuals currently in ICU and on the cumulative number of deaths since March 1, 2020. These data are available starting from March 18 for each department and on a daily basis. In order to obtain regional data, we simply aggregated information on hospitalizations and deaths for departments belonging to the same region. Hence, for France, we consider the following outcome variables: total number of individuals currently hospitalized per 100,000 inhabitants; total number of individuals currently in ICU per 100,000 inhabitants; cumulative number of deaths per 100,000 inhabitants. Population data are collected from the INSEE (Institut National de la Statistique et des études économiques) website and refer to population on January 1, 2020. The French panel includes data from February 28 to April 21.

Results

Graphical Analysis

Back to Table of Contents

Results for our graphical analysis are included in Figures 1 to 4. Each figure shows the relationship between our outcome variables (one for each panel of each figure) and the two pollutants, $PM_{10}$ and $PM_{2.5}$, for each country under investigation. For these plots, we consider one date and exploit cross-province or cross-regional variation in pollution and COVID-related outcomes. For each country, we choose the day before the beginning of lockdown for two main reasons: first, confinement measures forced people to stay at home, thus inevitably reducing exposure to pollution. Second, the decision to implement a lockdown policy can be viewed as a signal that the pandemic has reached the country and the number of cases, hospitalizations and deaths started to be considerably high, thus raising concerns for public health.³

Each point on each plot represents one province or region and refers to the real observations. The red line, instead, represent the fitted values, that is, the predictions for the outcome variable from a linear regression of the latter on the independent variable (pollution, in our case).

This set of graphs uncovers a positive relationship between COVID outcomes and pollution. This seems more evident in the two figures concerning France. For example, in panel (b) and (c) of Figure 4, the dots are (almost) all located very close to the red line of fitted values.

Even though this graphical analysis provides suggestive evidence of a positive association between pollution and the virus adverse outcomes, this is limited to only one day and cannot be considered as rigorous as econometric analysis. For this reason, we subsequently perform regression analysis.

Econometric Analysis

Back to Table of Contents

For the regression analysis of this section, we exploit the panel structure of the data described in Section 2. We run separate regressions for each country and each pollutant. The econometric specification is as follows:

$log Y_{r, t} =\beta_{0}+\beta_{1} X_{r, t-20}+\alpha_{r}+u_{r, t},$

where $Y_{r, t}$ is the COVID-related outcome of interest in province or region $r$ on day $t$; $X_{r, t-20}$ is the pollutant concentration in province or region $r$ at time $t-20$, while $\alpha_{r}$ are the province or region fixed effects. $\beta_{1}$ is the parameter of interest to be estimated. Finally, $u_{r, t}$ is an idiosyncratic error term.

Results are shown in Tables 1 to 4. Each column of each table reports estimates of $\beta_{1}$for each outcome variable. In our baseline specification (first row of each table), we simply regress $Y_{r, t}$ on the pollutant concentration. This allows us to exploit both cross-regional and time variation. In the second row of each table, we add the province or region fixed effects. This allows us to control for province or region-specific characteristics that do not vary over time. Given our log-level specification, estimates should be interpreted as the percentage variation in the outcome variable associated with a one unit increase in the independent variable. When fixed effects are included, estimated coefficients should be interpreted as the percentage variation in the outcome variable associated with a one unit increase in the independent variable, within a province or region and across time. The estimated coefficients for Belgium are always positive and statistically significant at 1% level. In France, COVID-related outcomes do not seem to be significantly associated with $PM_{2.5}$ concentration. However, there exists a strong positive correlation between the spread of the virus and $PM_{10}$ concentration. Also, note that estimated coefficients are bigger in magnitude for Belgium than for France. In Belgium, an increase of one $\mu g/m^3$ in $PM_{2.5}$ concentration is associated with a 6.8% increase in the number of cases, a 4.1% increase in the number hospitalizations and a 4.2% increase in the number of individuals in ICU. This results are consistent with prior research, even though estimates are smaller than those found by previous studies in the US (Xiao Wu et al., 2020).

Table 1: COVID-19 and Particulate Matter 2.5 in Belgium

Note: Robust standard errors in parenthesis.
^*p < 0.1, ^**p < 0.05, ^***p < 0.01
	(1) Cases (100,000 inh.)	(2) Hospitalizations (100,000 inh.)	(3) ICU (100,000 inh.)
PM 2.5	0.0633^*** (0.0052)	0.0407^*** (0.0048)	0.0416^*** (0.0049)
PM 2.5 (Province fixed effects)	0.0682^*** (0.0036)	0.0410^*** (0.0036)	0.0427^*** (0.0045)

N	803	660	660

Table 2: COVID-19 and Particulate Matter 10 in Belgium

Note: Robust standard errors in parenthesis.
^*p < 0.1, ^**p < 0.05, ^***p < 0.01
	(1) Cases (100,000 inh.)	(2) Hospitalizations (100,000 inh.)	(3) ICU (100,000 inh.)
PM 10	0.0337^*** (0.0033)	0.0268^*** (0.0032)	0.0265^*** (0.0033)
PM 10 (Province fixed effects)	0.0413^*** (0.0020)	0.0272^*** (0.0022)	0.0280^*** (0.0029)

N	803	660	660

Table 3: COVID-19 and Particulate Matter 2.5 in France

Note: Robust standard errors in parenthesis.
^*p < 0.1, ^**p < 0.05, ^***p < 0.01
	(1) Deaths (100,000 inh.)	(2) Hospitalizations (100,000 inh.)	(3) ICU (100,000 inh.)
PM 2.5	0.0241 (0.0204)	0.0111 (0.0099)	0.0081 (0.0088)
PM 2.5 (Region fixed effects)	0.0255 (0.0218)	0.0107 (0.0106)	0.0083 (0.0095)

N	1,188	1,188	1,188

Table 4: COVID-19 and Particulate Matter 10 in France

Note: Robust standard errors in parenthesis.
^*p < 0.1, ^**p < 0.05, ^***p < 0.01
	(1) Deaths (100,000 inh.)	(2) Hospitalizations (100,000 inh.)	(3) ICU (100,000 inh.)
PM 10	0.0353^*** (0.0071)	0.0139^*** (0.0049)	0.0100^** (0.0046)
PM 10 (Region fixed effects)	0.0403^*** (0.0062)	0.0155^*** (0.0040)	0.0117^*** (0.0039)

N	1,188	1,188	1,188

Conclusion

Back to Table of Contents

This paper corroborates existing evidence linking pollution levels to COVID-19 pandemic propagation. Together with prior research, our results suggest that pollution may be an important factor contributing to the diffusion of the virus that policy-makers should take into account.

It is important to note that our study has some limitations. First, our econometric analysis does not identify causal effects. We believe that reverse causality does not represent an important issue here.⁴ However, an omitted-variable problem may persist: even though our specification with province or region fixed effects allows us to control for province or region-specific characteristics that do not vary over time, there may still exist some time-varying local characteristics that are spuriously correlated with pollution and, hence, would bias our estimates.

Second, the use of up-to-date pollution data has pros and cons. On the one hand, these data allow us to exploit recent daily information. On the other hand, although the European Environment Agency (EEA) is a reliable source, this information is not revised. Indeed, the data received by the EEA from each sampling point in each member state are revised and made publicly available with one-year delay. In more detail, in September of each year, the EEA publishes revised data concerning the previous year: hence, pollution data for 2020, for example, will be available in September 2021.

Fighting the current pandemic requires a better understanding of the disease and corresponding propagation mechanisms. More research on this topic is welcome.

References

Back to Table of Contents

[1] Ginsburgh, V., Magerman, G. and Natali, I. (2020), COVID-19 and the Role of Economic Conditions in French Regional Departments. Working Papers ECARES 2020-17, ULB (Université Libre de Bruxelles.

[2] Office for National Statistics (2020), https://www.ons.gov.uk/peoplepopulationandcommunity/birthsdeathsandmarriages/deaths/bulletins/deathsinvolvingcovid19bylocalareasanddeprivation/deathsoccurringbetween1marchand17april. Last access: June, 2020.

[3] Setti, L., Passarini, F., De Gennaro, G., Barbieri, P., Perrone, M. G., Borelli, M., Palmisani, J., Di Gilio, A., Torboli, V., Fontana, F., Clemente, L., Pallavicini, A., Ruscio, M., Piscitelli, P., Miani, A. (2020), SARS-Cov-2RNA found on particulate matter of Bergamo in Northern Italy: First evidence. Environmental Research.

[4] SIMA, Società Italiana di Medicina Ambientale (2020), https://www.simaonlus.it/?page_id=694. Last access: June, 2020.

[5] WHO, World Health Organization (2013), Health effects of particulate matter. Policy implications for countries in eastern Europe, Caucasus and central Asia. https://www.euro.who.int/__data/assets/pdf_file/0006/189051/Health-effects-of-particulate-matter-final-Eng.pdf. Last access: June, 2020.

[6] Xiao Wu, M.S., Nethery, R. C., Sabath, M. B., Braun, D. and Dominici, F. (2020), Exposure to air pollution and COVID-19 mortality in the United States.

Additional Material

Back to Table of Contents

In this appendix, we provide an alternative version of Figure 1 in the main text. We still consider the number of cases, hospitalizations and individuals in ICU and the pollutant PM 2.5, but we now apply a time lag of 15 days between the observation of the pollutant concentrations and the measures of COVID-related harm.

For England and Wales, see Office for National Statistics. For France, see Ginsburgh, Magerman and Natali (2020). ↩
Note that the indices refer to the diameter of the particles. $PM_{10}$ stands for inhalable particles with a diameter of 10 microns or smaller, while $PM_{2.5}$ indicates particulate matter of diameter of 2.5 microns or smaller. ↩
Note that observations on the pollutants’ concentrations refer to the day before the lockdown and, hence, the corresponding observations for COVID-related outcomes refer to 20 days after lockdown. In appendix A, we also provide a version of Figure 1 for a time lag of 15 days. ↩
Reverse causality issue should be minimized by the use of lagged values of pollution. ↩

After the great lockdown: five uncomfortable truths to work out

2020-06-03T14:00:00+00:00

After the great lockdown: five uncomfortable truths to work out

May 24. Since its recognition as a human disease by mid January 2020, the Covid-19 has infected five million people, and killed more than 300,000 worldwide according to the official statistics.

Along the way, major lockdown has been put in place in more than 40% of countries, accounting for nearly 80% of the world population (Lipton and Prado, 2020). While this has brought a major drop in economic activity, which may lead to a global GDP shrinking by 3% this year, according to last estimates by the IMF,¹ the pandemic is now getting in some form of control five months later, with health systems able to breathe. The narrative around ther Covid-19 has evolved drastically across time. It was first “a bug that has no evidence to spread to humans” and “it is like another flu”, to, “the Covid-19 pandemic is real”; “social shutdown is the only way to flatten the curve of diffusion of the disease”, and now: “we have managed it, it is time to exit and to look forward”.

Figure 1: Country examples of Covid-19 peak time

While most of us have possibly followed the same cycle of narrative, from « this is nothing, to this is very scary, and we are now relieved », the evidence is indeed that wave 1 outbreak is getting under control, witth many but not all, countries having active cases peak (as early as Feb 17 for China, but just before May for Japan in Asia; in the second end of April for Europe, with Switzerland and Germany just ahead by early April, and May 21 for the US, see Figure 1).

But going back to the “new” normal is laudable, but must also keep a close management eye on both the disease evolution, as well as on the socio-economic burden that has come along the Covid-19 crisis. Here are five uncomfotable truths to cope.

1. We must get ready for a second wave

The odds of having a second wave are high. This is what we may learn from the past influenza-like pandemics, and what we might infer from the “official statistics” regarding the Covid-19 infection rate to date, - as the total infected are far away from the threshold of herd immunity that allows to control the disease.

But what else can we derive from the past experience of pandemics, and from own simulation post pandemic?

The past has shown that:

The pandemic often come back rather soon as wave 2, in the four to 8 months of the first. This means we may be hit still this year by the second wave.
The second wave can even be more lethal - for example, in the second wave of the 2009 H1N1, some countries like Canada, got hit… 5 times more - and the second wave in 1918 of the Spanish influenza seems to have killed at a multiplier the same size as the one found in Canada during the 2009 health crisis, see Figure 2.

Figure 2- Wave 2 can be dangerous

Simulation is also useful to see what might happen if social distancing is no longer imposed but made directly by citizens at 50% of the success of the lockdown and if the Covid-19 behaves under some extremes, such as the common cold (HCov-0C43, mild disease but seasonal and fast decay of immunity, disappearing in 4 to 6 months), or the SarS-Cov-1(more lethal, but with longer immunity, in the range of 1.5 to 3 years), and if we believe or not that cross-immunity within the family of coronaviruses might work.

Using parameters in the literature, such as seasonality that implies a peak - through of 20% gap for the reproduction rate, or cross-immunity between viruses of about 30% (see Kissler et al., 2020), the messages are clear that:

Wave 2 is always capable of happening, whatever assumptions taken
Short immunity (like the common cold) and limited cross-immunity leads to faster as well as larger occurrence than vice-versa—and possibly, or a size higher than wave
Social distancing, if too stringent, may however lead to a strong wave 2, as no built immunity is deployed, but we better need at least 50% reduction of contacts of effective lockdown to control the disease

The past as well as scenarios fine tuned to the socio-epidemiology of the Covid-19 and its coronavirus family leads to the uncomfortable truth that « one battle was won, not the war ». And the next battle may come soon and be big(ger than what we just won).

2. We may not be taken off-guard by being fooled by the wrong figures

It has been said multiple times that we are « flying blind, regarding key figures on the pandemic ». The current official figures suggest a rather low infection rate for a virus with such a difficult profile of latency and symptoms, and a high fatality burden.

We do not suggest that reported cases are intentionally under-reported. Those figures may be adequate if they only cover the people with severe symptoms, requiring hospitalization, and where systematic testing was made. The danger of not having a comprehensive view on figures is however many. The first is that we may forget other channels - the typical example has been not to look at home care, which happened to be a major source of deaths, - see my other article

The second risk is that the fatality rate at time of hospitalization is moving from virtually zero to 20%, at time of hospitalization, and possibly at 50% when someone is at ICU. Those stats are not good, as it may look like the odds of dying become as large as flipping a dice. This means we must look before hospitalization, when people critiically upgrade from death free to death liable. But without statistics, it is very difficult to prevent.

The third risk is that we may be put off-guard as to the timing of the wave 2. We have collected many statistics and run maximum likelihood models to find most probable estimates of infection, see among others my previous post.

Many new studies also reveal that the recorded cases may be off by a factor of up to 5-10. For example, one recent study done in France suggests that hospitalization is only 3.6% of total infected, or a ratio of 28, while the 85/15 rule (85% mild cases, 15% severe cases for covid-19), rather implies a ratio of 5, leading to an understatement of actual infection by 5-fold (Salje et al., 2020). The implication of such a gap is that fatality rates are clearly overstated, but also that people can be more often faced with getting the disease than said. At current transmission rate, and given that, despite a 5 to 10 fold adjustment, most countries are far off the herd immunity portion of infection, a higher contaminated stock means that the infection flow is larger per day, as the infection figure builds up like a power law. We might thus take high risk for wave 2, as bed capacity are fixed, or at best can be increased, but usually only linearly.

3. Health risk may only start for the recovered from the Covid-19

At current, a large focus has been on serious cases and fatalities in the Covid-19 crisis. The typical assumption is that recovered people are all « fit and proper ». But remember, that, for 1 death at hospital, 5 others survive—and possibly, for one person requiring hopsitalisation, possibly 5 times more officially, and possibly 25 - 50 times more did recover outside the hospital channels.

What if this massive number of people, - between 5 to 40 million people worldwide, recovery is not complete?

Those risks are not nil, as acknolwedged by Kelly Servick, a writer for Science, in an April piece entitled «For survivors of severe COVID-19, beating the virus is just the beginning». In fact, the history provides some clear evidence of short and long-term damages.

Regarding short-term damages, more than one third of people who got hospitalized for the 2003 SARS outbreak felt anxiety and depression disorders, still one year after the infection (Lee et al, 2007). Likewise, if pneumony is a marker for Covid-19, there is four times more probability to suffer a cardiovascular disease, for those getting hospitalized for acute pneumonia than not.

But those are only short-term effect; long term effects may be present too. In fact, multiple studies looking at in utero reaction of to be born kids from parents caught into the 1918 pandemics suggests large morbidity effects still 25 to 40 years after, affecting lung, kidney, and many other organs, with impact on productive and social life (Almond, 2005 and 2006). Compiling a series of in utero studies, the effects may be important, impacting the next generation, in a ratio of 1 to 9% of weight of new borns, see Figure 3.

Figure 3 - how adverse shocks may affect the long-term - in utero effects

4. We must be bold enough to relaunch inclusive economies

The burden is not only about health, it is socio-economic. Short-term costs linked to lockdown are not small, with eg in the US, average income and wealth lost estimated to be more than 5,000 US Dollars and 33,000 dollars respectively (Coibion et al, 2020).

If total burden may come to 5-10% of welfare lost, it is already clear that a « V » recovery may be rather optimistic. A « U » shape may be a better representation, as seen from the early data of China, where economic recovery has been slow pace.

A « L » shape is possibly not to be neglected, both because of the risk of Wave 2 still this year, and because crises of that size may lead to major distorsions, affecting investments and ultimately growth path in the future. Plans have been announced by a large set of countries to stimulate growth, and prevent the worst case of a « L- like » recovery. The key question remaining is: do they spend (fast) enough and inclusively?

To date, most countries have put a fiscal stimulus in the range of 2% of GDP, on top of facilities of re-payment. The later is de facto crucial as the OECD, leveraging Orbis data, finds that 1/3 of firms may run out of liquidity after after three months of lockdown. This liquidity crunch is thus massive and must be sorted out clearly (OECD, 2020). Regarding the former, the size of the stimulus might appear not bold enough.

Consider that the spent figure is higher than during the crisis 2008 (eg., G-20 spent roughly 1.4% of GDP in stimulus package), and at that time too US spent significantly more (like China) to reboost the economy. Today, we see the same pattern, with US, but this time Germany, spending close to 10% of their GDP (data courtesy of Bruegel, see Figure 4) for their stimulus package. Yet, there remain some clear surprises:

Why is it that countries with more covid damages seem to commit less amount of stimulus than others?
Why is the level of fiscal spent not bigger than the 2008 crisis, while the GDP shortfall of the pandemic may oscillate between -2 to -8% percent for 2020 alone in developed countries, that is, an effect that is larger than witnessed during the 2008 crisis? In fact, if consumption and private investment is 70% of GDP, the (undiscounted) multiplier effect from year 1 to 5 after the spent of 2%, would be 3.6%, with recovery of spent by 2022, and for a total of 5.6%, of just the mid range of risk of output contraction for Covid-19.

The EU Recovery plan proposed by the Franco-German team of a €500 billion of spending that has been recently backed by the EU budget a few days ago is a path to boldness and would represent an European-wide fiscal policy, that could be jointly spent on key forward looking infrastructure in sustainability theme, new investment in frontier technologies, among others.

Figure 4 - Immediate fiscal impulse commitment for Covid-19 risk

Finally, the question is not that we spend enough to restore - the question is whether we remain inclusive. There is clear evidence that the ethnic and socio-economic distribution of health impact of the Covid-19 is not favorable to lower socio-economic group and minorities. Eg lower socio-economic groups suffer more often from comorbidities and are thus more at risk of fatality from the disease; likewise those groups are more at risk of unemployment.

Or, if employed, they are more at risk of exposure with lower probabilities of remote working among others. Increase in inequality from the Covid-19 must be part of next agenda, as it ultimately would weight on prospect of recovery and growth.

5. We must secure democracy

Last but not least, it will be critical to fix the impact of the Covid-19, on the role of the the State, the dynamics of election, and public economy at large.

Many states have voted for exceptional power, in the first place to take measures linked to confinement. But it may remain crucial to consider those powers are temporary, as part of this exception — not as a new rule. Some governements are already tempted to take advantage of extra power given by the Covid-19.

More subtle is the issue of election. France is going for a second round of vote for municipalities, at “Covid-19 time, that may undermine the representation of election, for example. A recent study conducted in the US identified that counties that voted after Super Tuesday and which were then exposed to Covid-19 outbreak, were less likely to support Sanders , leading to 4 percentage points less support compared to Sanders 2016 vote (Bisbee and Honing, 2020). This effect is material and may mean election may be greatly influenced by the hazard of a pandemic wave like Covid-19. We may extrapolate this argument for the US presidency election, a fortiori if a new wave (is likely to) break around that voting time.

References

Almond, D. (2006). Is the 1918 influenza pandemic over? Long-term effects of in utero influenza exposure in the post-1940 US population. Journal of political Economy, 114(4), 672-712.
Almond, D., & Mazumder, B. (2005). The 1918 influenza pandemic and subsequent health outcomes: an analysis of SIPP data. American Economic Review, 95(2), 258-
Bisbee, J and D. Honig, (2020), Flight to safety: 2020 Democratic primary election results and Covid, Covid Economics, Vetted and Real time papers, 19
Coibion, O. , Y. Gorodnichenko and M. Weber ( 2020), The costs of covid 19 crisis- lockdowns, macro-expectations and consumer spending, Covid Economics, vetted and real time papers.
Helferty, M., Vachon, J., Tarasuk, J., Rodin, R., Spika, J., & Pelletier, L. (2010). Incidence of hospital admissions and severe outcomes during the first and second waves of pandemic (H1N1) 2009. Cmaj, 182(18), 1981-1987.
Kissler, S. M., Tedijanto, C., Goldstein, E., Grad, Y. H., & Lipsitch, M. (2020). Projecting the transmission dynamics of SARS-CoV-2 through the postpandemic period. Science.
Lee AM, Wong JG, McAlonan GM, et al. Stress and psychological distress among SARS survivors 1 year after the outbreak. Can J Psychiatry. 2007;52(4):233‐240.
Lipton, Alex, and Marcos Lopez de Prado. “Exit Strategies for COVID-19: An Application of the K-SEIR Model (Presentation Slides).” Available at SSRN 3579712 (2020).
Mummert, A., Weiss, H., Long, L.P., Amigó, J.M. and Wan, X.F., 2013. A perspective on multiple waves of influenza pandemics. PloS one, 8(4).
OECD (2020a), “Corporate sector vulnerabilities during the Covid-19 outbreak: assessment and policy responses”, Tackling Coronavirus Series
Taubenberger, Jeffery K., and David M. Morens. “1918 Influenza: the mother of all pandemics.” Revista Biomedica 17, no. 1 (2006): 69-79.

See The great lockdown worst economic downturn since the great depression ↩

The case for more digitization at Covid time

2020-06-03T09:00:00+00:00

The case for more digitization at covid time

May 8. The Covid-19 pandemic has officially taken the life of hundred thousands people worldwide. It has put the economies in standstill, with about a 5% decline in GDP for this year, as a result of the lockdown imposed to crush the diffusion of the disease.

While we are at progressive exit now, the challenge remains large to prevent a “L-scenario” of economic recovery. We make the case here for including a digital toolbox to government stimulus plans, so as to maximize a more promising recovery emerging out of the Covid-19 outbreak.

A more pessimistic scenario

Covid-19 impact on economies is estimated to be gloomy. Yet, most estimates are however based on the assumption that the society will go soon “back to normal”, with a gradual and smooth reopening of economies, even if possibly some localized and less prolonged shutdowns are to be implemented to avoid a second wave of outbreak. Those estimates all are leading to a recovery in about one year, and a pandemic that is mastered without any large second or third wave.

Those two assumption may be optimistic. On the health side, there is still a possibility of a second and worst outbreak than the first, like the world had witnessed in major infleuneza pandemics, e.g. during the 1918-19 Spanish flu. On the economic side, there is also a risk of what is called a « L » economic recovery, where the disruptions may lead to a permanent output gap. This is not an “extra doom scenario” – after all, ten years after the sub-prime 2008 crisis and the Lehman Brothers bankruptcy, about 60% of countries in the world, still have an output trajectory below precrisis according to robust research by the IMF.¹

A major cause of this gap is a large cut in intangible capital by firms or reduced education due to school close that may impact long-term productivity.² Other factors are citizens behavior, among which, fear that may not easily go away. For example, more than 40% of US workers do not feel at ease to go back to work,³ while 65% of Chinese citizens might resist to consume as much as before the disease.⁴

A cost of another outbreak, versus the gradual case assumed, is likely to weight another 2-3 percent drop in GDP, should the risk materialize. The long-term impact, may be of the same magnitude, cumulatively after 10 years. That is, if we use the comparison with the 2008 crisis from the IMF that led to a mode of sustained drop of yearly output of minus 15 basis points a year, for countries affected by the sub-prime crisis.

Digitization as a way to limit covid-19 risk and boost our future

Given those major risks, Covid-19 is also making transparent how digital may be a formidable enabler to fight the disease in the short-term, while building a more productive, and safer society than currently in the long-term. Consider a few examples.

Tracing and testing for the disease. A challenge is that the Covid-19 is very contagious, and one needs to spot infected very quickly and timely, while also act upon the social chains of contacts. Digital technology tools, leveraging location services, big data and analytics, are critical for making this happen, and there is evidence that countries with such tools at hand, mostly in Asia, have been able to better curb the disease.⁵ This is not exclusive to Covid-19, as nowcasting was successfully used during the 2015 Zika virus, or even for the flu.⁶
Supply chain boost. A major issue during, and of, the pandemics has been supply chain disruption. Dun & Bradstreet had once reported that million of companies around the world have a first and second tier supplier in the Hubei region, the center of outbreak of the covid 19.⁷ Digital technologies here may help better anticipate those effects and react accordingly for much better resilience.⁸ Examples include digital twins as more pervasive simulation tools or effects, and better synchronize responses.⁹ A large set of companies are also using 3D print of health supplies (eg swabs) to circumvent shortage.
Effective R&D. As early as in February of this year, digital machine learning tools have identified multiple rheumatoid arthritis treatments as being powerfully repurposed for treating the virus. Such type of drugs have been recently confirmed as effective in random health trials by end of April, or two months later.¹⁰
Working with, rather, than against the machine. Many companies are now adopting tele-working, and digital automation interfacing tools for work, replacing physically-exposed contacts. Without appropriate technology, the alternative would be not to work for 40% of non essential jobs time. For essential jobs, - those that allow the economy still to work, eg retail logistics, etc - the risk is to expose individuals to the risk of infection. Digitization of the full chain of retail, including e-commerce, has proven to be a very effective solution to the Covid-19 challenge.

In need of faster digital diffusion

The digital technologies above are part of frontier of digitization - as it includes big data, AI, 3D, IOT, digital twins, or still AR/VR. Those technologies, and related applications are however still far from being used extensively both by consumers and enterprises.

For examples, testing and tracking tools based on digital technology have demonstrated very good specificity, eg they may spot 70-80% of infected cases, despite a large set of non-symptomatic cases. The key challenge for the case of digital tracking is that it must have sufficient reach - eg more than 50% of citizens must it to spot infected and warn their social ties. While this means a penetration à la Facebook in the US, this rate is in practice not easy to achieve. Even in countries promoting digitla tracking, the adoption rate is not at this level - it is reported to be at 40% in Iceland, and about 20% in Singapore and Israel.¹¹

Another example is the use of teleworking technology to put to work the workforce, with non essential interaction tasks, while automation may be the solution for virus-prone interactions. About 50% of people never worked from home before Covid-19.¹² Yet, again, here, teleworking is usually used by less than 30% of employees in countries such as US or Japan, to date, and half of them only do it for one day a work week.¹³ Similarly advanced automation is also relatively low, eg about 10-15% in entreprises are implementing, as it is badly perceived as a way of machines against employment.¹⁴

Regarding digital supply-chain tools, the effects may be in the range of boosting productivity by 10-30%.¹⁵ There as well, the challenge is adoption — only about 40% of companies worlwide have been digitizing their supply chain.

We conclude that digitization holds the promises of both fighting Covid-19 in the short-term as well as offering a solution for faster and stronger recovery of our economies, against the worst case of « L-scenario ». The current economic stimulus plans advocated by many governements must include a digitization tool box - a message to relay in the face of limited awareness of the digital lever in the current discussion of those plans.

References

Chen W., Mrkaic M., and M. Nabar; “The global economic recovery 10 years after the 2008 financial crisis”, IMF, April 2019. ↩
Baker S., Bloom N., Davis S. J., Kost K., Sammon M., and Viratyosin T. (2020). “The Unprecedented Stock Market Reaction to COVID-19”,Covid Economics: Vetted and Real-Time Papers1, 3 April. ↩
How employees feel about coronavirus now: A pandemicex survey update ↩
Covid-19 economic impact ↩
See The western world should urgently play the asian smart route to control the covid-19 ↩
Akhtar M., Kramer M., and Gardner L. (2019). “A dynamic neural network model for predicting risk of zika in real time”, BMC Medecine, 17. ↩
Ivanov D. (2020). “Predicting the impacts of epidemic outbreaks on global supply chains: A simulation-based analysis on the coronavirus outbreak (COVID-19/SARS-CoV-2) case”. Transportation Research Part E: Logistics and Transportation Review, 136, 101922. ↩
Ivanov D., Dolgui A., Sokolov B.; “The impact of digital technology and Industry 4.0 on the ripple effect and supply chain risk analytics”. Int. J. Prod. Res. 2019;57(3):829–846 ↩
Ivanov D., Dolgui A., Sokolov B.; “The impact of digital technology and Industry 4.0 on the ripple effect and supply chain risk analytics”. Int. J. Prod. Res. 2019;57(3):829–846 ↩
Beck B., Shun B., Choi Y., Park S., and Kang K. (2020). ”Predicting commercially available antiviral drugs that may act on the novel coronavirus”, BioRxiv, 2, Feb. ↩
Thorneloe R., Epton T., Fynn W., Daly M., Stanulewicz N., Kassianos A., Shorter G. W. et al. “Scoping review of mobile phone app uptake and engagement to inform digital contact tracing tools for Covid-19”, (2020). ↩
Covid-19 is a before and after moment in the digital transformation ↩
Higa K., Wijayanayake J. I. “Adoption of Telework by Japanese Organizations: A Survey Study” ↩
EIBIS 2019 Report on Digitalisation ↩
Bughin J., and van Zeebroeck N. (2019). “The right response to digital disruption”, Sloan Management Review, The right response to digital disruption ↩

We badly need a dynamic dashboard of behavioral changes to best navigate our way out of Covid-19

2020-06-02T14:00:00+00:00

May 8th. A large part of the discussion related to how the Covid-19 outbreak will evolve is based on a set of crucial epidemiologic models. Among those, the most classical one is the S-I-R model (« S » for susceptible, « I » for infected and « R » for recovered), which tells that the dynamics of the disease is closely related to the basic reproduction rate, $R_0 = b/y$ where $b$ is the rate of effective contacts between the infected and the susceptible, and $y$ is the rate at which infected individuals recover.

In particular, for the Covid-19, the rate of infection seems at least twice faster than recovery ($R_0>2$). Furthermore, complex cases requiring hospitalisation in intensive care are large enough (between 15% to 20% of infected cases), to both create a pandemic explosion and a major risk of excess demand for healthcare. In consequence, the SIR model has implied that a more or less stringent lockdown is required to slow the curse of the disease. If we note by « a » the parameter of lockdown, this is equivalent to build a bridge between the basic reproduction rate $R_0$, and the effective rate, so that $R = R_0*(1-a)$, with $0< a <1$ is a measure of intensity of social distancing success.

Different analyses suggest that for countries which had opted for lockdown, this has been very effective, with $a=0.64$ for Belgium (see my previous post); $a=0.8$ in France (see Angot, 2020), and $a>0.9$ in the case of Wuhan in China (Lin et al., 2020). But, at the time when we go to a relaxation of the rules of the lockdown, the dynamics of « a » added to the S-I-R model may be crucial to re-assess. What if some susceptible people feel that the outbreak of Covid-19 is controlled, and they start now to over-compensate with more social contacts? What if people feel that economic risks are too high versus the cost of being infected, and decide to go back to work without many protections (not necessarily that they wil not try to self-protect, but the availability of masks may remain limited, etc)?

Thus, individuals will be the ones « in fine » to make $0 < « a » <1$, high enough for a sufficiently long time to crush the pandemic, with little room for Covid-19 to come back as a revenge, in a new wave. In general, we may think that « a » will naturally decrease for the susceptible with time, as the higher the number of infected, the higher expected cost of the disease if infected. This behavior is possible, but the pace of increase in « a » is likely not enough to avoid the heath care crunch (see also Bethune, and Korinek, 2020). For a significant and structurally long term decrease in « a », we need more, and possibly some incentives.

Regarding incentives, the economic theory of moral hazard is a good avenue to look at designing optimal incentives. One example of (negative) incentives is to tax inadequate risky behavior. In the case of Taiwan, it looks like the penalty for infringement has been put at the expected value of a lost life from Covid-19, as result of misplaced social behavior, - and whether you are contagious or not. In the spirit of optimal theory, this multi-thousands money fine aims to have people fully internalizing the risk by making people pay for the ex post materialisation of risk.

Today, many actions launched by governments, during or after lockdown can be tested through the theory of optimal (dis)incentives. But usually, this is not done. Rather, the idea has been to hope for the best, that is, governements keep recalling people of their « moral obligations » to comply with distancing rules. While necessary, it is by far not sufficient, and is clearly important that we know more about how individuals perceive the disease and its evolution, and how they react to consequences of pandemics and policiy actions, be it for health and for socio-economic reasons, to design such incentives.

While we defer a post to this later on, we are here more concerned as to what citizens do and think, so that we have a clear basis as to how to thing about the right focus of incentices. Surprisingly, the data remain sparse.

We have collected a few data, courtesy of Neurohm, from a survey being rolled out across a large set of countries worldwide on the Covid-19 situation. The interesting element of the data is that AI is used in order to ensure a base line of probing survey answer, so that we only use survey answers that look like sufficiently solid for the findings below (see here). We here use France as an example, via a sample of 1,300 individuals. The data below may have some bias, so we caution that we should not be taken the findings as fully representative. For example, we find that 0.8% of the sample population claims to have been positively tested. This figure is higher than currently reported in France (roughly 0.3% of population). But the idea is also to show how crucial it is to understand behaviors and behavioral changes, - and rather quickly, to secure the crush of the disease.

Behavior at time of Covid-19

Here are five facts of importance we collected:

Finding 1: Own and collective infection perception match

We know that the disease is difficult to be diagnosted, with formal test. We find that the perception of being infected by the disease, is large (about 9% outside testing), but this also is quite close to how people report on their knowledge of third parties being (said to be) infected in one citizen’s community. Otherwise, stated, expectations are likely built as a convergence between personal and social perceptions, in line thus with the SIR model that the dynamic of contagion is linked to the stock of infected.

Finding 2: Voluntary reduction in social contagion expected to be > 35%

Based on epidemilogical data, and supposing that there « a > 0 » is only linked to imposed lockdown, simulations suggest that lockdown effective at 50%, will stabilize the outbreak if imposed for a period in the range of 50 to 80 days. The total span of the disease may be then be close to 6 months, at which time the disease dies out (see Alvarez et al., 2020). As discussed above, however, the good news is that individuals may be increasing protection with the disease, so that benefits of lockdown is not fully lost and the pandemic may be finishing faster and with lower infections.

At the time of survey in France, the individuals were already about 2 months in the pandemic. But looking at their expectations on the likely extra duration for the outbreak to disappear, the average is above 1.5 months extra, for a doubling of people being infected, versus the current perception of 11%. Calibrating the risk of getting the Covid-19 to the expected cost of life, so that susceptible people reduce their exposure intentionally without lockdown measures, and using the S-I-R model, we find that the above is compatible with « $a=0.26+0.7* portion\ of\ Infected$ » going forward, leading to an effective reproduction rate decline to $R_t=0.7$ at the time the disease fully dies out, and $a=36%$ (Figure 1)

Figure 1: what “a” to anticipate from survey expectations?

Finding 3: Typical behaviors barriers well, but not perfectly understood/accepted

Reducing the rate of effective contacts can be done through a large set of behaviors, among which washing hands, etc. In general, we find that a population adhering to a few protective behavior may indeed make a growing to 40-50%, pending obviously on the availability of supplies (eg availability of masks, etc). The data survey clearly show that the bulk of, but not all individuals have understood the importance of those behaviors. In fact, while 90% clearly see the danger linked to the disease, 80% actually clearly state that washing their hands is key, and fewer than 60% understand the home insulation, the importance of mask wearing, or restrictions in public transport.

Assuming that people acknowledgement leads to behavior, a would just indeed just go up to 30%. However, this may be far stretched - if behavior is done in 2/3 of cases (as often noticed in behavioral studies), the self protective mechansims will not be enough to match the expectations of the population, and would be also relatively « tied » to ensure that the disease does not have a second wave.

Finding 4: Average is not necessarily representative

The above has looked at the average, but the distribution may be skewed in terms of perception. Regarding the expected duration of the disease, 40% said that the pandemic has been crushed already, for a median that is less than 3 weeks, or more than two times faster than the average. To get to this figure, and number of expected extra infection at the median, « a » will need to double from 0.25 to 0.50, which is not likely to happen based on finding 3, as the protective behavior importance is also reduced at the median by about 25% versus the average.

The risk here is that some part of population do have too optimistic behavior—which may lead to failure of expectations and a disease that will last longer.

Finding 5: the importance of social circles

If our curse is related to our social behavior, we may also see how social consideration may play to counter-act the wrong behavior. Here is a good news: close circles health is as important, if not more, than own health on average. In particular, people are especially worried about the health of old members of their families and about their children health. Also important, people worry about their own citizens for 60% of cases Figure 2.

Figure 2 - self or third party caring?

Action plan

The above data are usually not available - but they bring a key message. When we get out of the lockdown, we will need to make sure we track (and guide if any) people behaviors to ensure to crush the disease. Self- interest, of possibly, third party interest may be already a good incentive to reduce social contacst, but it is likely to be not enough to navigate through the disease. We need to create a dynamic dashboard on how people may think and act along the way - without this, drivers to ex post measures of recovery, infections and the like, may not be understood, or understood too late versus the fight of the pandemic we still need to have.

The french data also make clear that a few assumptions made in the literature are too simplistic. Do not think average, as the distribution of beliefs and behaviors do not obey a typical distribution; do look at how people loop back social circles to influence their behaviors and beliefs; do not hope for the a sustained reduction as close to the lockdown. If we make the point that beliefs are not necessarily biased, one can really infer how people may adapt behavior in the medium term. Their behaviors often do not match the core assumptions of policy makers. In the case of the French data, the fit between compliance to self protection and beliefs of the pandemics suggests that the effort to be driven by the Franch citizens is less than half of what has been triggered by the imposed lockdown.

References

Alvarez, F. E., Argente, D., & Lippi, F. (2020). A simple planning problem for covid-19 lockdown (No. w26981). National Bureau of Economic Research.

Anastassopoulou, C., Russo, L., Tsakris, A., & Siettos, C. (2020). Data-based analysis, modelling and forecasting of the COVID-19 outbreak. PloS one, 15(3), e0230405.

Angot, Philippe. “EARLY ESTIMATIONS OF THE IMPACT OF GENERAL LOCKDOWN TO CONTROL THE COVID-19 EPIDEMIC IN FRANCE.” (2020).

Bethune, Z. A., & Korinek, A. (2020). Covid-19 infection externalities: Trading off lives vs. livelihoods (No. w27009). National Bureau of Economic Research.

Lin, Qianying, Shi Zhao, Daozhou Gao, Yijun Lou, Shu Yang, Salihu S. Musa, Maggie H. Wang et al. “A conceptual model for the outbreak of Coronavirus disease 2019 (COVID-19) in Wuhan, China with individual reaction and governmental action.” International Journal of Infectious Diseases (2020).

Leung, K., Wu, J. T., Liu, D., & Leung, G. M. (2020). First-wave COVID-19 transmissibility and severity in China outside Hubei after control measures, and second-wave scenario planning: a modelling impact assessment. The Lancet.

Why our social future is changing and how Covid-19 will accelerate the pace of change

2020-06-02T09:00:00+00:00

May 7th. Our lives have always been closely linked to a social structure, enforced by law, or informally by norms and habits that evolve through times. This blog is the follow up blog to the previous one written for Ubiverse, -a community around key matters for our society, which I encourage you to look at and join (Why our social future is changing and how covid-19 will accelerate the pace of change), and it tries to make the case that Covid-19 might be the catalyst to accelerate towards our new social future.

In the previous blog (Why our social future is changing and how Covid-19 will accelerate the pace of change), we have shown that the social norms of the past which led to a long period of enlightenment and happiness growth, have become challenged, due to the combination of three forces: digitization, globalisation, and sustainability. While we are at the point of crisis leading to large social divide, we believe that the Covid-19 crisis may well be the case in point for an acceleration of changes, and for a new solution space.Here is why.

Covid-19 interaction with forces at work

Not only the emergence, but the consequence of the Covid-19 pandemic is linked to our social model, and the forces at work that are reshaping it.

Regarding the emergence of Covid-19, more than 75% of the new communicable diseases have been zoonoses in the recent past. If we recognize that zoonoses are complex and dynamic diseases, part of their outbreak are clearly linked to human-related factors, such as ecotourism, or culinary habits (see how the Covid-19 patient zero originated most likely from the Wuhan food and animal market). But it may well be that the virus is the long term result of human intrusion into life ecosystems as well our new social way of leaving, that is, increased urbanization.¹

Urbanization on the other hand, is not only a cause of epidemic occurrence. Covid-19 becomes more of a problem in cities, as close contacts increase with inhabitants density and because urban areas employ more workers in customer facing occupations, that will be affected by more or less social distancing measures put forward by many governments to flatten the curve of the coronavirus pandemics.² Covid-19 also brings some advantage to cities. Because of the business model, the sharing economy is more developed in high density of population, helping to find a solution to the logistics of retail trade and food services, two sectors most affected by the containment measures of social distancing. Finally, cities concentrate more on the new skill set in demand, which are now using digital platforms as new models of social interactions.

We all know that globalisation directly affected the spread of the Covid-19 to become a pandemic, affecting more than 200 hundred countries by now. We also know how pollution and sustainability also interact with Covid-19. It has been recently claimed that pollution increases the morbidity risk of Covid-19. In Italy, for example, it has been reported that cities that exceed twice more often the limit set for PM10 than others have also registered twice more cases of Covid-19 infections.³

We have seen how digital technologies are a blessing to trace contaminants timely and put those infected in appropriate quarantines, like done in most Asian countries.⁴ We have witnessed how digital technologies can be used to resolve supply chain issues (eg 3D print of swabs for Covid-19 testing), and how AI has quickly helped identify the genome of the Covid-19 (with the hope) to accelerate the effective race for antivirals and vaccines.⁵

Finally, the social norms leading to more individualism have been tested heavily as being moral enough in the case of deciding whether to comply with quarantine rules. One case study illustrates the point. When Italy decided for a full containment measure in the hope of curbing the Covid-19 disease, a few individuals decided to move from a high-risk zone such as Milan to southern regions of Italy. 40 of those individuals were enough to cause the sudden spread of the virus observed afterwards in the South.⁶

Of course “cheating” has been part of life, and can be tolerated. But exceptions may have to prevail, especially in situations such as Covid-19, creating major externality risks. This is especially so, as the fatality rate associated to contagion of the Covid-19 disproportionately affects the old population (about 12% more than average, after controlling for other factors such as comorbidities), while the largest contributor to close contacts is the young generation, also the one less inclined to obey to the rules of social distancing.⁷

This being said, the episode of the Covid-19 crisis is likely to be more than just another force at work. It may be the catalyst to a new social model of interactions for multiple reasons:

Covid-19 is only at the start of its pandemic status, with a first outbreak wave slowly ending in a few countries, but with more and more evidence that likely less than 10% of the population got infected currently by the virus. This level is proof of the effectiveness of the lockdowns, but in the meantime, is also a level, far away from the implied immunity threshold of between 35–55% that will ensure a control on the disease spread, and the insurance of no new outbreak like we have just witnessed. In passing, this level of immunity is based on initial reproduction rates of $R_0=2.2$, and accounting for behavioral changes, as well as asymmetric distribution of contagion⁸; infections however do not stop at this stage, but the flow is slowly declining to stop at possibly 65% to 85% of the population.
Given the above, the risk of multiple waves for Covid-19 is non nil, so that, if we phase down lockdown, we are not going back to the same social interaction model as before, at least in the foreseeable future. The new normal would likely entail a model with lower physical interactions at work, for education, and for economic exchange. The new normal may also make lurking and cheating less acceptable to many, supporting the idea of more coercitive measures (eg Taiwan put fines on those who were noncompliant of containment, at level in the thousands of USD). Likewise, while there has been a large movement towards keeping our strict privacy, tracing to prevent the large outbreak of Covid-19 might be accepted under certain conditions , eg more US citizens accept it than refusing it, for example.⁹
Another consequence of Covid-19 is the boost towards further enterprise digitization, which has been largely lagging the one on the consumer side. Many more companies are likely to adopt remote working, and digital interfacing for work, replacing physical contacts but recognizing that progress in enterprise platforms has made the digital remote experience as good—if not better—than physical encounters.
Large pandemics in general affect social capital for a long time. For example, the Spanish flu of 1918 led to a significant burden of deaths that affected a decrease in social trust and in turn led to lower growth and prosperity. Of importance, the descendants of people who face the pandemic did reveal this social behavioral change.¹⁰ There is thus a risk that after Covid-19, our social capital will be in part depleted, leading to lower economic welfare. As social trust has been damaged recently with multiple crises (bombing, migration, financial crisis of 2018), Covid-19 will only add to the fire, and must require active actions to rebuild the role of our institutions as trusted partners. In particular, a major recovery plan will have to address all the leftovers, with a notion of prioritization towards the most fragile segments of the population. The home care crisis, as well as the lack of equipment support to hospital workers will have to be turned around soon, and vocational careers built up, as well as stock equipment buffers made up, to face the ongoing waves to come of Covid-19.
Last but not least, Covid-19 is likely a catalyst for how our model of interactions will evolve towards a more global model, one which will embrace—rather than be afraid of—machines. Covid-19 has proven that automation may easily complement our work (eg, video conferencing), and health (eg, AI-based genomics).
The new model of interactions will also be more respectful of our earth resources, as Covid-19 has shown how pollution can be dramatically reduced under economic shutdown and how pollution may be an important driver of Covid-19 infection. The fact that the EU commission is likely to rebuild the economy after the Covid-19 shutdown, through a major green infrastructure plan, is proof that people can learn to turn a crisis into a necessary green opportunity to come by 2030.

Cascio, A., Bosilkovski, M., Rodriguez-Morales, A. J., & Pappas, G. (2011). The socio-ecology of zoonotic infections. Clinical microbiology and infection, 17(3), 336-342. ↩
Koren, M, and R. Pito, 2020, Business disruptions from social distancing, Covid Economics, Issue 2. ↩
Coccia, M. (2020). Two mechanisms for accelerated diffusion of COVID-19 outbreaks in regions with high intensity of population and polluting industrialization: the air pollution-to-human and human-to-human transmission dynamics. medRxiv ↩
The western world should urgently play the asian smart route to control the Covid-19 ↩
Koyama, T., Platt, D., & Parida, L. (2020). Variant analysis of COVID-19 genomes. Bull World Health Organ. http://dx.doi.org/10.2471/BLT, 20. ↩
Donnarumma, F. and G. Pezzulo, (2020), Morla decisions in the age of covid 19 : your choices really matter, Vox. ↩
Olsen, Asmus Leth, and Frederik Hjorth (2020), “Willingness to Distance in the COVID-19 Pandemic.” And Ivchenko, A., J. Jachimowicz, G. King, G. Kraft-Todd, A. Ledda, M. MacLennan, L. Mutoi, C. Pagani, E. Reutskaja, C. Roth, et al. (2020). Evaluating covid-19 public health messaging in italy: Self-reported compliance and growing mental health concern. ↩
Three key Covid-19 indicators to curb a potential of 20 million human fatality ↩
Yougov Survey: Would you support or oppose the US government using a GPS tracking app to ensure Americans diagnosed with coronavirus remain quarantined? ↩
Aassve A., Alfani, G., F. Gadolfi, M. Lemoglie (2020), pandemics and social capital,- from the Spanish flu to covid, Vox. ↩

Covid-19 and the Role of Economic Conditions in French Regional Departments

2020-05-29T09:00:00+00:00

Table of Contents:

Introduction
The econometric model and data
Results
Analysis of Covariance
Conclusions
References
Tables
Figures

Introduction

Back to Table of Contents

The outbreak of the Covid-19 pandemic has found both the scientific community and the general public unprepared. Its rapid spread and the skyrocketing number of individuals dying for the virus have caused deep concerns, profound uncertainty and anxiety around the globe. The pandemic not only affects individual health and health care systems, but also economic and sociologic ecosystems. Flattening the curve affects our behavior, but our behavior also affects the curve. In this paper, we aim at understanding how existing socio-economic disparities might contribute to differences in the spread of the virus.

Our concern is to study whether regional variations in socio-economic conditions (poverty level, education level, density of doctors), as well as geographic circumstances (the East and North-east borders with Germany, Luxembourg and Belgium) have an influence on the pattern of the pandemic in continental France.¹ Figures 1 and 2 show the 94 French continental departments, and the intensity of the outbreak in these regions. Figure 1 depicts the total number of Covid deaths by department on April 20, 2020. Figure 2 shows the total number of discharged Covid patients on the same day. Clearly, the North (Belgian and Luxembourg borderline) and North-East regions have been hit harder, relative to the South-West areas, as can be explained by the initial hotbeds of Alsace and Bas-Rhin (Low Rhine) regions. France recently decided to pursue confinement in the Eastern regions as well as in Paris and its surrounding departments, but unlock in the Western and central parts of the country. On May 5, 2020, the President of the Bas-Rhin department in Alsace claimed that it would be ‘pure madness’ to unlock his department. The French Prime-Minister unveils, at the same time, that France ‘is cut into two pieces’ (Huffpost, 2020). Hervé le Bras (2020), researcher at the Institut national d’études démographiques, analyzed the dynamics of the pandemic, comparing how the virus developed in two regions: the Haut-Rhin department, where the number of cases of Covid is large and somewhat out of control, and Bouches-du-Rhône in the South, where the epidemic took off much later.

In this paper, we use information on socio-economic variables and Covid data to understand how socio-economic variation might contribute to these differences. We implement a simple estimation setup, with lagged socio-economic indicators (year 2017), and current Covid data (April 2020). We do not look at the dynamics, nor at demographic variation at the regional level. We run regressions of both deaths and discharged (D & D in what follows) separately, as well as together (using an analysis of covariance framework) on a certain number of departmental geographic and socio-economic characteristics, such as total population, Gini coefficients to measure inequality, level of education, and doctors’ density.

Unfortunately, it was not possible to know with certainty the exact moment at which the epidemic started in each department. Though some figures on cases are available, this information is rather scant before March 18, 2020. In addition, it may have been difficult for public authorities themselves to identify the so-called ‘patient 0.’ This poses, of course, reliability issues for data proceeding March 18. However, we downloaded data on the cumulative number of deaths and patients discharged from hospitals on April 20 when both authorities and the general public were well aware of the pandemic and the collection of data had become more rigorous and well organized. The problem here is that the number of days between ‘patient 0’ and April 20, is not the same in all departments. It is even suggested that a few cases of Covid appeared in France in December 2019, or even earlier on November 16, 2019, long before blazing, in the department of Haut-Rhin (Peillon, 2020).²

The paper is organized as follows. In Section 2, we describe the econometric model and the variables that are used. Section 3 is devoted to econometric results, and Section 4 concludes.

The econometric model and data

Back to Table of Contents

We believe that, in this specific context, reverse causality is not an issue, though our results do not identify causal effects since an omitted-variable problem may still exist. The econometric model we use is simple. We regress the number of Covid deaths and discharged from hospitals on departmental variables:

$y_{ik}= R_i\alpha_k + X_i\beta_k + \epsilon_{ik}, i =1,..., 94; k= 1,2,$

where $i$ is one of the 94 departments, $y_{ik}$ is either the number of deaths ($k=1$) or of discharged inhabitants ($k=2$) in department $i$, $R_i$ is a vector of two dummy variables which represent geographical characteristics (Northern and Eastern France, and Ile de France, that is Paris and its surroundings), $X_i$ is a vector of departmental socio-economic characteristics, $\alpha_k$ and $\beta_k$ are vectors of parameters and $\epsilon_{ik}$ is the error term. Variables $y_{ik}$ and population, which is one of the variables in vector $X_i$, are expressed in logarithms.

The dependent variables were downloaded on April 20 from the website Santé Publique France. The choice we had was to take the more recent data at the time we started our analysis, though they were slightly declining afterwards. We chose the high point of the pandemic.

The vector of two dummy variables $R_i$ includes the following borders: (a) Northern and Eastern departments that have a border with Southern Belgium, Luxembourg and Germany,³ as well as (b) Ile de France, a group of departments, with Paris (75) as center.⁴

Data for the variables in vector $X_i$ are all downloaded from the INSEE (Institut National de Statistique et des Etudes Economiques) website and include: number of inhabitants in logs, Gini coefficient, basic education,⁵ number of doctors per 100,000 inhabitants.

We ran five regression for D & D, by introducing variables one after the other, in the order described above. The first two contain the dummies North-East and Ile de France; next comes population,⁶ inequality within regions measured by the Gini coefficient, education and density of doctors.

Results

Back to Table of Contents

Baseline Regressions

Results appear in Tables 1 (for deaths) and 2 (for discharged) and are very similar across the two tables. Clearly, the North-Eastern border and Ile de France (column 1) have the largest number of D & D people. This may be partly due to the fact that departments in the North-East and, especially, in Ile de France, which includes Paris and the surrounding departments, are among the most populated ones. This becomes obvious in column (2), where we add the variable population, which causes a drop in the magnitude of the coefficients for both dummies (and a larger drop for the dummy Ile de France). Despite this, coefficients picked up by the two dummies remain significantly different from 0 at the 0.01 probability level. Population only can thus not fully capture the extension of the virus in these two regions.

Next, we add the Gini coefficient. In both regressions the variable picks up positive effects that are all significantly different from 0, at the 0.01 or 0.05 probability level. This means that a larger level of inequality is associated with a larger number of deaths and severely ill individuals. This is an important result and seems to be in line with previous findings in the UK. According to an article published on The Guardian, poorer areas in England and Wales are significantly more affected by the pandemic, with twice a death toll as more affluent neighborhoods. This may be due to a few reasons: for instance, individuals in a disadvantaged economic status are more likely to have pre-existing conditions, they are more likely to live in worse quality housing, they are more likely to have jobs that cannot be done through smart-working (by staying at home). These are, evidently, factors that contribute to expose more the most vulnerable populations to the virus.

The effect of poor education, as described earlier, is probably overshadowed by the effect of inequality, that is, large differentials in incomes within each department. People without higher education are likely to remain poor. The coefficients picked up by the variable are not significantly different from 0.

Finally, we get to those who have been of great help in the corona pandemic. One expects that more physicians per inhabitant would help containing the outbreak, by allowing people to enter a hospital quickly enough and by providing them with the necessary treatment. Indeed, here, the density of doctors (both generalists and specialists) is negatively associated to both the number of deaths and the number of individuals severely affected by the virus, as proxied by the number of discharged. The coefficients are, however, not statistically different from zero, which may be due to the presence of some other (confounding) factors that we are not taking into account.

Analysis of Covariance

Back to Table of Contents

The estimated parameters displayed in Tables 1 and 2 do not seem to be very different, though Table 1 deals with death, while Table 2 deals with those who were discharged from hospitals. To check whether they are significantly different, we opted using an analysis of covariance, which implicitly assumes that the distribution of errors is the same in both subsamples (D & D). The model is now:

$y_{i0}= R_0 \alpha_0 + X_0\beta_0 + \delta R_0\alpha + \delta X_0\beta + \epsilon_{i0}, i =1,..., 186.$

In this formulation, $y_{i0}$ is a vector constructed by piling up each department’s deaths followed by each department’s discharged and is regressed on a matrix $R_0$ formed by piling up two matrices $R_i$. Matrix $X_0$ is constructed in the same way by repeating twice $X_i$. Finally, $\delta$ is a dummy variable equal to 1 for observations related to $y_{i1}$, that is, deaths, and 0 for discharged. The coefficients on the interaction terms $\delta R_{0}$ and $\delta X_{0}$ will tell us whether the effect of the covariates is different for deaths and discharged.⁷ The results that we now analyze can be found in Table 3.

As can be checked, the coefficients picked up by the variables North-East, Ile de France, Population, Gini, Education and Doctors’ density, as well as the value of the intercept, are exactly the same as those in Table 2. This is due to the fact that our dummy, $\delta$, is equal to zero for discharged and, hence, these coefficients pick the effect of the covariates on the number of discharged individuals. Those coefficients that were significantly different from 0 remain so, and those that were not, remain so as well. Standard errors are also approximately the same.

The estimates for $\alpha$ and $\beta$, instead, will tell us the difference in the effect of each independent variable across the two groups (deaths and discharged). To make this clearer, consider the following example: in Equation (5), results show that the effect of North-Eastern regions is equal to 0.955 for the group of discharged (as in Table 2), to which 0.242, picked up by the interaction $\delta R_0$, should be added for those who died. The sum is equal to 1.197 and it is identical to the coefficient associated to North-East in column 5 of Table 1. This shows that the coefficient associated to each interaction term will yield the difference in the effect of each covariate across the two groups: if this coefficient is not statistically different from zero, we conclude that this difference is not significant.

The estimates of $\alpha$ (associated to the interactions North-East*Dummy and Ile de France*Dummy) are positive and significantly different from 0 at the 0.01 or 0.05 level, which implies that they increase the role of the two regions for those who have died, in all regressions.

The effect of Population is common across the two groups (D & D) since the coefficient for Dummy*Population is not significantly different from 0. Analogous reasoning applies to the Gini since the Dummy*Gini coefficients are not different from zero. Finally, Education and Doctors’ density do not contribute to the fits. It is also interesting to note that the coefficient for Dummy*Intercept is negative (but small and not significantly different from 0 in Equations (2) and (3)) which indicates that the number of deaths is (fortunately) smaller that the number of discharged, on average.

It is also worth noting that all fits of equations (3) to (5) are good since the adjusted $R^2$ are larger than 0.65 and increase to 0.73 in the analysis of covariance in Table 3.

Conclusions

Back to Table of Contents

There is a clear pattern of heavy infections (deaths and discharged patients) along the French border with Belgium, Luxembourg and Germany, and in the departments that surround Paris. It is not clear whether the effect of the border is due to the countries that border France (and people passing from one country to the other), or to a cause that we did not find. This is different for Ile de France with over 12 million inhabitants, who, before confinement, were traveling, usually using metros or trains in and out of Paris, where they work (or vice-versa). Note that more recently, that is, after inputting the numbers of D & D, the virus moved to more central and western regions,⁸ though with less virulence than in the North-East and Ile de France. We will need, however, to wait a couple of weeks, to check whether this will remain milder.

The fact that population is related to D & D is obvious, but far from being the only factor, as we showed above.

Finally, it should be clear that more inequality means that the population is not homogeneous, and that richer people live in one part of a town or a village, are probably more careful, and may have gardens to be able to breath, while poor people have little choice, live in another part of the town and are more likely to walk on the street and in parks. This is what a British report also points out (Improvement Service, 2020, p. 3):

“People living in socio-economic disadvantage are more likely to be working in the low paying jobs which are keeping the country going in supermarkets, as cleaners, delivery drivers and home care workers, and a significant proportion of these low paid workers will be women. The four ‘C’s’ of cleaning, care, cashiering and catering, commonly seen as women’s work are now massively important, and those working in these areas are being exposed daily to the risk of contracting Covid-19.”

They are also more likely to have lost (at least for some time) their job, which is a very grim perspective.

As we said, the estimated coefficients picked up by education, which are not significantly different 0, are probably in the shadow of unequal incomes measured by the Gini coefficient. The fraction of poor people who usually have a low level of education are less likely to escape the pandemic.

Most crises are likely to increase inequality, given increases in the rate of unemployment and lower wages that follow, as well as difficulties to get loans from banks to pay their mortgage, even if they are only temporary. And Covid-19 will probably not be different. This means that the pandemic hits harder areas in socio-economic disadvantage today and will probably exacerbate disparities in the near future (Furceri et al., 2020). This highlights the importance of policy interventions aimed at helping individuals living in poor conditions, in order to (i) attenuate the impact of the pandemic today and (ii) attenuate the (potential) negative consequences of the pandemic in the near future.

References

Back to Table of Contents

Furceri, D., Loungani, P. L., Ostry, J. D. and Pizzuto, P. (2020). COVID-19 will raise inequality if past pandemics are a guide. VOX CEPR Policy Portal [web: last accessed May 20, 2020].

Le Bras, Hervé (2020). On entrevoit trois stades de l’épidémie de Covid-19 en France. Le Monde, April 30, 2020.

Huffpost (2020). Déconfiner les départements rouges? De la ‘pure folie’ pour le président du Bas-Rhin, Le HuffPost, May 8, 2020.

Improvement Service (2020), Poverty, inequality and Covid-19 [web: last accessed May 20, 2020].

Peillon, Luc (2020). L’origine de l’épidémie de Covid en France peut-elle remonter à l’automne 2019? Libération, 21 mai 2020.

Pidd, H., Barr, C. and Mohdin, A. (2020). Calls for health funding to be prioritised as poor bear brunt of COVID-19. The Guardian, May 1, 2020.

Tables

Back to Table of Contents

Table 1: Regression Results. Number of Deaths.

Note: Two of the 94 departments had no deaths on the 20th of April.
Robust standard errors in parenthesis. ^*p < 0:1, ^**p < 0:05, ^***p < 0:01
	(1)	(2)	(3)	(4)	(5)
North-East	1.78493^*** (.4041796)	1.362145^*** (.2496312)	1.320606^*** (.2371874)	1.199691^*** (.2474712)	1.196585^*** (.2473837)
Ile de France	2.658232^*** (.2022909)	1.434573^*** (.1942773)	1.212202^*** (.1892048)	1.402966^*** (.2181983)	1.301863^*** (.3032876)
Population		1.063447^*** (.1446876)	.9891452^*** (.1619487)	1.158865^*** (.225545)	1.150728^*** (.230567)
Gini			.0538155^** (.0268751)	.0981521^*** (.0295869)	.1110501^*** (.0363411)
Education				.0470603 (.0304477)	.0407087 (.0304477)
Doctors’ Density					-.000752 (.0011276)
Intercept	3.670147^*** (.1283489)	-10.23082^*** (1.931404)	-10.65228^*** (1.852999)	-16.09139^*** (4.318988)	-15.81489^*** (4.498292)

R²	.3922114	.6461379	.652798	.665427	.6662798
Adjusted R²	.3785533	.6340744	.6368347	.6459751	.6427231
N	92	92	92	92	92

Table 2: Regression Results. Number of Discharged.

Robust standard errors in parenthesis. ^*p < 0:1, ^**p < 0:05, ^***p < 0:01
	(1)	(2)	(3)	(4)	(5)
North-East	1.498502^*** (.3883368)	1.013639^*** (.2502798)	.9433316^*** (.2346669)	.9574594^*** (.2418449)	.954981^*** (.2415764)
Ile de France	2.423485^*** (.1784774)	1.100784^*** (.159507)	.7230497^*** (.1747128)	.6993664^*** (.2000135)	.6430911^** (.264656)
Population		1.112531^*** (.1123722)	.993284^*** (.1211013)	.9752686^*** (.1553269)	.9719336^*** (.1580135)
Gini			.0902798^*** (.0292446)	.0848994^** (.0330915)	.0923457^** (.0361311)
Education				-.0055659 (.0224333)	-.0089142 (.025325)
Doctors’ Density					-.0004245 (.001146)
Intercept	4.907909^*** (.1208542)	-9.592106^*** (1.489271)	-10.37125^*** (1.389197)	-9.751786^*** (3.047651)	-9.626842^*** (3.148192)

R²	.3571441	.7118813	.733274	.7334894	.7338071
Adjusted R²	.3430154	.7022774	.7212863	.7183468	.715449
N	94	94	94	94	94

Table 3: Results of the Analysis of Covariance.

Note: Two of the 94 departments had no deaths on the 20th of April. This results into 2*94-2 observations.
Robust standard errors in parenthesis. ^*p < 0:1, ^**p < 0:05, ^***p < 0:01
	(1)	(2)	(3)	(4)	(5)
North-East	1.498502^*** (.3894372)	1.013639^*** (.2510045)	.9433315^*** (.2353612)	.9574594^*** (.2425761)	.954981^*** (.2423228)
North-East*Dummy	.2864277^*** (.0957056)	.3485058^*** (.0861668)	.3772743^*** (.0952936)	.2422315^** (.0939364)	.2416038^** (.0957339)
Ile de France	2.423485^*** (.1789831)	1.100784^*** (.1599689)	.7230497^*** (.1752297)	.6993664^*** (.2006182)	.6430911^** (.2654737)
Ile de France*Dummy	.2347463^** (.0927611)	.3337892^*** (.1135917)	.4891525^*** (.1457478)	.7035998^*** (.1643193)	.658772^*** (.2096761)
Population		1.112531^*** (.1126976)	.993284^*** (.1214596)	.9752686^*** (.1557965)	.9719336^*** (.1585017)
Population*Dummy		-.0490841 (.1035451)	-.0041388 (.1141379)	.1835962 (.1461286)	.178794 (.1472504)
Gini			.0902798^*** (.0293312)	.0848994^** (.0331916)	.0923457^** (.0362427)
Gini*Dummy			-.0364643 (.0223767)	.0132527 (.0248453)	.0187044 (.0257238)
Education				-.0055659 (.0225011)	-.0089142 (.0254033)
Education*Dummy				.0526262^*** (.0189861)	.049623^** (.0210125)
Doctors’ Density					-.0004245 (.0011496)
Doctors’ Density*Dummy					-.0003275 (.0008917)
Intercept	4.907909^*** (.1211966)	-9.592106^*** (1.493583)	-10.37125^*** (1.393307)	-9.751787^*** (3.056866)	-9.626843^*** (3.157919)
Intercept*Dummy	-1.237761^*** (.0738208)	-.6387168 (1.390247)	-.2810338 (1.372104)	-6.339603^** (2.82479)	-6.188046^** (2.888944)

R²	.4825624	.7321625	.7433982	.749052	.7495514
Adjusted R²	.4681891	.7216296	.7302765	.7331875	.7306222
N	186	186	186	186	186

Figures

Back to Table of Contents

We do not include Corsica, la Réunion, islands in the Atlantic Ocean, and French Guyana. ↩
As will be seen, the pandemic was, and still is very serious in the Eastern France. ↩
The following French departments are part of this border: Nord (department number 59), Ardennes (68), Meuse (55), Meurthe-et-Moselle (54) Bas-Rhin (67), Haut-Rhin (68) and Moselle (57). We excluded a certain number or Eastern departments, that border Switzerland (essentially mountains, though Geneva is quite close to France) as well as the Italian and the Spanish borders, for the same reason (the Alps and the Pyrenees), though Italy and Spain were hardly hit by the virus. ↩
The other departments are Essone (91), Hauts-de-Seine (92), Val-de-Marne (94), Oise (60), Seine-Saint-Denis (93), Val d’Oise (95), and Yvelines (78). ↩
Education levels are census data available for 1999, 2010 and 2015. Data after 2015 are extrapolated. INSEE provides the number of individuals older than 16 who do no longer attend school (‘population non-scolarisée’) in each education group. Basic Education, here, is defined as the share of those with no diploma or with a Diplôme National du Brevet (DNB), which is granted after completion of the first cycle of education, or with a Brevet d’etude professionnelle (BEP) or Certificat d’apritude professionelle (CAP), which are obtained after completing the first two years of a professional high school. ↩
We also tried population density and GDP in place of population (all combinations), but population performs best. ↩
Note that the number of observations should be equal to $2\times 94$, since there are 94 departments, but two observations on the variable $y_{i1}$ are missing (see above). ↩
Auvergne, Côte-d’Armor, Franche-Comté, Loiret, Pays de Loire, Vendée, and other regions. See Direct Coronavirus en France : bilan, nouveaux cas et foyers https://www.topsante.com/medecine/maladies-infectieuses/zoonoses/coronavirus-en-direct-nouveaux-cas-foyers-en-france-634781 [last consulted on May 19, 2020]. ↩

Monitoring the macroeconomic impact of Covid-2019 in real-time: insights from unconventional data

2020-05-25T11:00:00+00:00

"~~Philosophy~~ Economics is written in this grand book - the universe - which stands continually open to our gaze."

Adapted citation from Galileo Galilei's, “Il Saggiatore”.

How big is the economic impact of Covid-19 on the economy? This question has been giving economists and policymakers a massive headache for some weeks now.

During the recent crisis, policymakers found themselves to be taking decisions about economic stimulus packages without having a precise idea of the dimension of the contraction in economic activity caused by the coronavirus. Take the case of the US, for instance. The CARES Act, a major piece of legislation aimed at stimulating economic activity, was introduced to Congress on January 24th and ultimately signed into law on March 27th. Nevertheless, up to April 17th, Federal Reserve estimates suggested the US economy was experiencing only a moderate contraction of -0.4%. A similar case could be made for each of the European Union member states.

It goes without saying that having a preliminary assessment of the economic impact of the virus in real time is crucial in order to accurately design effective stimulus packages. This necessity, nevertheless, clashes with the scarcity of proxies of economic activity available in real time. In this article, we explore the potential of some unconventional data to monitor the state of economic activity before official data are released.

The sky over Los Angeles before, during, and after the lockdown.

Conventionally, real time evaluation of economic activity is performed via a class of models which go under the name of nowcasting models. These models, which were developed inter alia thanks to the contribution of some former ULB scholars in the 2000s,¹ are based on the idea that there is a number of common shocks which affect the economy at several layers. Such models are able to exploit the data flow in real time to produce timely forecasts of economic activity in real-time.

As a concrete example, consider what these researchers at the Federal Reserve are doing here. What you find below is an indicator of weekly economic activity developed by Daniel Lewis, Karel Mertens, and Jim Stock at the Fed to track the status of the economy in real time. What these researchers at the Federal Reserve did is very simple. They used weekly data on steel production, fuel consumption expenditure, electricity consumption, consumer surveys, staffing indices, and unemployment claims in order to construct a weekly index of economic activity. On May 12th, the US weekly economic index says the US economic is shrinking at the rate of -11.2%.

NY Fed’s Weekly Economic Index, May 12th. Source: Federal Reserve of New York.

Unfortunately, in Europe, we do not have the same data at a weekly frequency. There are, however, some indicators which allow us to evaluate the state of economic activity in real-time.

In this article we are going to analyse three indicators which will give us a sense of the level of economic activity in real-time in the EU: (i) electricity data, (ii) flight departures, and (iii) mobility indicators.

1) Electricity Data

Since 2015, ENTSO-E, the European Network of Transmission System Operators, publishes the daily electricity consumption for all European countries. Preliminary evidence suggests the electricity consumption to be a strong predictor of the level of economic activity in EU countries. In particular, the year-on-year change in Industrial Production in March 2020 can explain up to the 35% of the year-on-year change in the temperature-adjusted Electricity Consumption at a cross-country level.

Electricity Consumption and Industrial Production in March 2020 (year-on-year). Sources: Eurostat (Industrial Production), ENTSO-E (Electricity Consumption)

Our preliminary estimates suggests that the year-on-year changes in industrial production in March 2020 covaried almost one-to-one with the year-on-year change in electricity consumption. During the month of March 2020, the temperature-adjusted electricity consumption in Belgium declined by 6.9% whereas industrial production declined by 5.5 % compared to February 2020, and by 3% compared to March 2019. Bruegel provided a detailed documentation of the electricity fallout in the bulk of European and non-European countries here and here.

Electricity demand is typically determined by a household and an industrial component. The household demand, once adjusted for the temperature fluctuations and long-run trends, can be considered acyclical (i.e. not subject to business cycle fluctuations). On the other hand, the industrial demand for electricity is highly dependent on the level of economic activity.

Industrial demand for electricity essentially depends on three factors. The first is the sectorial composition of the economy. For instance, an economy characterised by heavy industrial sectors will have a sizable demand for electricity, whereas a service-based economy will have a much lower demand. The second factor is the energy efficiency of the economy, which measures, ceteris paribus, how much energy is needed in order to produce a certain variety of goods. The last one is the level of economic activity. The industrial sector will employ an amount of electricity proportionate to the volume of the national production.

Therefore, if the first two factors are stable, the drop in electricity can contain precious information concerning the level of economic activity in the economy. Nevertheless, the asymmetric sectorial impact of the Covid shock might represent a confounding factor in the analysis.

2) Flight Departures

It looks like some sectors are not having an easy time. Take airlines for instance. In March 2020, departures from EU airports plummeted by over 40% compared to March 2019. The drop in the flight departures is correlated with the Industrial Production Index and can explain up to the 47% of its variance. Daily departures and arrivals data are made available at daily frequency by Eurocontrol for all EU countries.

Industrial Production and Flight Departures. Source: Eurocontrol

The aviation sector, despite being a relatively small sector - it contributes to the 2.1% of the European GDP - gives a key indication about the mobility of goods and people. This statistics includes commercial air transport services, and general aviation. This indicator, therefore, contains joint information on the flows of commercial goods and passengers for touristic and business reasons. Nevertheless, the correlation between these two variables, might be considerably influenced by the latest data points, and should be handled with care.

While Eurocontrol statistics come with some lag, FlightRadar24 publishes every day the number of planes flying in the world. According to their estimates, the global number of flights declined by 62% in March. Their impressive infographics are definitely worth a click.

The sky over Europe one year ago and today. Source: Eurocontrol

3) Mobility Indicators

Google, CityMapper, and Here have developed daily indicators to track the degree of mobility of the urban population. Such indices are useful to get a sense of the degree of dynamism of the population.

Industrial Production versus Mobility in Retail and Grocery. Sources: Google Mobility Report, European Commission

Industrial Production versus Mobility in Stations. Sources: Google Mobility Report, European Commission

Some preliminary evidence suggests that some of these mobility indices might be useful to keep track of the drop in economic activity. In the figure below, we plot on the y-axis, the year-on-year percentage change in Industrial Production in March 2020, and on the x-axis the average drop in Retail&Grocery mobility and the average drop in the Transit via Stations in March 2020 compared to the average mobility in the previous year. Such indices can account for respectively 36% and the 37% of the variance of the drop in Industrial Production in March 2020 at a cross-country level.

Such correlation might suggest that the drop in the time spent in retail and grocery shops might be a reasonable proxy for the drop in the purchase transactions for consumption goods, whereas the drop in the average transit via stations might contain some information regarding the drop in active employment. The Transit via Station indicator might be more relevant for urban contexts, in which citizens tend to move prevalently via public transports, whereas it might be less informative for extra-urban areas. Real-time urban mobility data from CityMapper suggest that countries like France and Germany might be ahead in the recovery compared to the other EU countries.

CityMapper Mobility index (February 5th to March 22nd). Source: CityMapper

Some good reads on the macroeconomic impact of Covid-19:

Vasco Carvalho et al. look at the drop in Spanish consumption through the lens of 1.4 billion transactions.

Sinem Hacioglu, Diego Kanzig, and Paolo Surico estimate that Covid implied a 30% contraction in the median income in the UK using granular transaction data.

Petr Sedlacek and Vincent Sterk argue that if the Covid crisis affects start-up creation (as it is likely), it will have a protracted effect on the economy’s growth rate.

Here the Financial Times collected data from restaurants, cinemas, shops, roads, flights and energy in order to illustrate the slowdown in global economic activity in real time in the UK.

Ravenpack publishes on a daily basis a series of indicators to track the sentiment of the media and the general public.

See: Giannone, Domenico & Reichlin, Lucrezia & Small, David, 2008. “Nowcasting: The real-time informational content of macroeconomic data,” Journal of Monetary Economics, Elsevier, vol. 55(4), pages 665-676, May. ↩

Learning from the curve

COVID-19 in Belgium - what to do now? A commentary from the GEES experts

Une évaluation de la contribution du premier tour des élections municipales à l’épidémie de COVID-19

Référence de l’étude

Les auteurs

References

The Spread of COVID-19 in Belgium: a Municipality-Level Analysis

Abstract

Introduction

Description

Onset of the epidemic

Intensity of the epidemic on March 31

Growth of contaminations in the month of April

Analysis

Results

Conclusion

References

COVID-19 Pandemic and Pollution: Evidence from Belgium and France

Introduction

Data sources and methodology

Pollution Data

COVID-19 Data

Belgium

France

Results

Graphical Analysis

Econometric Analysis

Conclusion

References

Additional Material

After the great lockdown: five uncomfortable truths to work out

After the great lockdown: five uncomfortable truths to work out

1. We must get ready for a second wave

2. We may not be taken off-guard by being fooled by the wrong figures

3. Health risk may only start for the recovered from the Covid-19

4. We must be bold enough to relaunch inclusive economies

5. We must secure democracy

References

The case for more digitization at Covid time

The case for more digitization at covid time

A more pessimistic scenario

Digitization as a way to limit covid-19 risk and boost our future

In need of faster digital diffusion

References

We badly need a dynamic dashboard of behavioral changes to best navigate our way out of Covid-19

Behavior at time of Covid-19

Action plan

References

Why our social future is changing and how Covid-19 will accelerate the pace of change

Why our social future is changing and how Covid-19 will accelerate the pace of change

Changing social structures

Covid-19 interaction with forces at work

Covid-19: a catalyst for a new social future

Covid-19 and the Role of Economic Conditions in French Regional Departments

Introduction

The econometric model and data

Results

Baseline Regressions

Analysis of Covariance

Conclusions

References

Tables

Figures

Monitoring the macroeconomic impact of Covid-2019 in real-time: insights from unconventional data

1) Electricity Data

2) Flight Departures

3) Mobility Indicators

Some good reads on the macroeconomic impact of Covid-19: