Covid-19 across European regions: The role of border controls
The outbreak of the Covid-19-pandemic led to a massive return of the nation state. National governments around the world have taken far-reaching measures to control the spread of the disease, ranging from decisions to close shops, restaurants and schools to a full-blown lockdown of public life. In Europe, the crisis is a fundamental challenge to the principles of the EU – notably, solidarity, policy coordination, and free movement across national borders (Biancotti et al. 2020).
In particular, the temporal reintroduction of national border controls within the Schengen area may jeopardise the European project. According to Nicolas Schmit, Jobs and Social Rights Commissioner, the closure of borders such as that between Germany and Luxembourg was “just a reflex, which doesn’t add anything to health security” (New York Times, 17 April 2020). In a recent paper (Eckardt et al. 2020), we test for the treatment effects of border controls during the first wave of Covid-19.
This is related to many studies exploring which type of non-pharmaceutical interventions are most eﬀective to limit the spread of Covid-19. For example, Bonardi et al. (2020), Askitas et al. (2020) and Weber (2020) argue that the closure of borders or travel restrictions had little eﬀect. In contrast, studies on international air travel (Chinazzi et al. 2020, Keita 2020) find sizeable eﬀects, particularly if measures were implemented early. Also, Rothert et al. (2020) suggest that lax policies in the most lenient US states might translate into millions of additional infections in other parts of the country in the long run.
Treatment effects of border controls
We collected daily regional data on confirmed new Covid-19 cases from the respective statistical agencies of 18 Western European countries1 from calendar week 10 (starting 2 March 2020) to calendar week 17 (ending 26 April 2020). The data start roughly one week before the introduction of border controls, so we can test for treatment effects using within-country variation. Figure 1 shows the spread of Covid-19 across European regions. During the first two weeks of our sample (panel 1a), incidence was concentrated in northern Italy and parts of Spain. Calendar weeks 12 and 13 (panel 1b) saw a quick spread, in many cases across national borders (with interesting exceptions; see France–Spain). During this period, border controls were enacted. Weeks 14 and 15 (panel 1c) saw the apex of new cases, with incidence all over the map. Calendar weeks 16 and 17 (panel 1d) saw a reduction in new cases as most countries surpassed the height of their incidence curve during the first wave of 2020.
Figure 1 New confirmed Covid-19 cases per 1,000 inhabitants in specified calendar weeks
Note: see text for details.
Intuitively, the eﬀect of controls should vary with the extent of cross-border relations. Border regions should be more aﬀected than others, and border regions with intense cross-border relations before the controls should be aﬀected most. To test for the role of national borders for the spread of the disease, we estimate a series of diﬀerence-in-diﬀerences regressions with day and region fixed effects.2
In a first set of results, we distinguish regions located at controlled borders from all others. In a second specification, we restrict the treatment group to regions with an above-mean share of their workforce (> 0.9%) commuting to a workplace across a national border in 2019. For example, 30% of the workforce of Belgian Luxembourg and 11.3% of the workforce of bordering French Lorraine were cross-border commuters in 2019, hence both regions belong to our intensity-based treatment definition. In contrast, Spanish Aragon and bordering French Midi-Pyrénées both had no significant cross-border commuting in 2019, hence they are excluded from the treatment group. We also add a specification where we assume that controls have an eﬀect only with a time lag of at least one week, following Lauer et al. (2020).
Figure 2 shows our first set of results, using a PPML estimator. The first three models allow for immediate eﬀects of border controls. The last three models ‘shift’ the onset of border controls by seven days. 95% confidence intervals are based on heteroscedasticity-robust standard errors clustered on the region level.
Figure 2 Estimated treatment effects across specifications
Note: the reported coefficients are the percentage changes in cases relative to the control group due to border controls, together with 95% confidence intervals. Estimates 1 to 6 are based on a PPML estimator. Estimate 7 is based on our INLA approach.
All models suggest that border controls led to a reduction in the number of reported Covid-19 cases. In Figure 2 we transform the coefficients to report the percentage change in cases relative to the control group. We see that the size of the eﬀect is much larger once we use the narrow, intensity-based treatment definition (compare models 1, 2 and 4, 5). Intuitively, the introduction of border controls mattered much more for regions with a substantial number of cross-border commuters beforehand, compared to border regions with little or no commuting. The eﬀects become statistically significant once we use an intensity-based definition of the treatment together with country-specific time eﬀects.3 Figure 3 shows groupwise means and excess risks over time, based on models 2 and 3, respectively. Compared, for example, to the results in Weber (2020), this suggests that the use of regional instead of state-level data and accordingly the definition of the control group is quite important.
Figure 3 Groupwise means and excess risks over time
(a) Groupwise conditional means
(b) Periodwise treatment effects
Note: Panel 3a plots average daily new cases in the treatment and control groups, conditional on day and region fixed effects. Panel 3b shows (exponentiated) coefficients of the treatment group dummy for each day, conditional on country-day and region fixed effects. In both panels, grey areas show the 10 % confidence interval for robust standard errors clustered at the region level. Note that the “France spike” seen in panel 3a does not show up in panel 3b, because it is absorbed by the country-day fixed effects.
Our preferred specification is shown in Figure 2, model 6, where we control for region eﬀects and country-specific time eﬀects, use a narrow definition of the treatment group and test for lagged eﬀects. A value of -25.08 means that daily cases are reduced by about 25% due to border controls.
Spatio-temporal dynamics: An INLA approach
Spatial interactions in the data (suggested by Figure 1) and local variation of containment policies challenge this type of treatment analysis. Notably, missing local measures might bias our estimated treatment eﬀects upwards.
To control for both the spatio-temporal dynamics of the data and for potentially unobserved spatio-temporal heterogeneity (e.g. due to time-varying local policies), we specify a Bayesian spatial-temporal count data model, using the INLA formalism for Bayesian inference in latent Gaussian models (Blangiardo et al. 2013, Bakka et al. 2018). Further, we include a spatial random eﬀect assuming an iid Gaussian distribution. This allows for both structured and unstructured spatial eﬀects such that the model also absorbs unobserved spatial heterogeneity (Fahrmeir et al. 2004). Again, treating the number of new confirmed Covid-19 cases as outcome, the spatio-temporal count model includes time eﬀects, the distance from a continental border, the share of commuters in the workforce, a time-varying dummy for border controls and an oﬀset.4 The outer right line in Figure 2 shows the resulting (statistically significant) point estimate.
Clearly, spatio-temporal heterogeneity matters a lot. Our PPML approach with nationwide time eﬀects must have missed the impact of local containment policies, but also spatial spillovers. However, even if we allow for a very flexible form of unobserved spatio-temporal eﬀects, we find that border controls reduced the number of confirmed Covid-19 cases significantly. According to the INLA approach, the introduction of border controls reduced the number of daily new cases by roughly 6% (compared to 25% suggested by the PPML estimator).
The temporal reintroduction of border controls within the Schengen area is clearly costly, and their benefits have been disputed. We show that border controls helped to contain Covid-19, but only for regions with a substantial number of cross-border commuters prior to the crisis. As a robustness check, we use a Bayesian INLA approach to take unobserved spatio-temporal heterogeneity into account. With this we find smaller, but still significant, eﬀects in the area of 6%. We conclude that the temporal introduction of border controls made a measurable contribution to contain the spread of Covid-19. At the same time, it is likely that better policy coordination at the European level could have generated these benefits at lower economic (and political) cost, for example if based on a closer monitoring of cross-border commuting flows (Caldera and Koirala 2020).
Askitas, N, K Tatsiramos, and B Verheyden (2020), “Lockdown strategies, mobility patterns and Covid-19”, Covid Economics 23: 263–302.
Bakka, H et al. (2018). “Spatial modeling with R-INLA: A review”, WIREs Computational Statistics 10:e1443.
Biancotti, C, A Borin, F Cingano, P Tommasino, and G Veronese (2020), “The case for a coordinated COVID-19 response: No country is an island”, VoxEU.org, 18 March.
Blangiardo, M et al. (2013). “Spatial and spatio-temporal models with R-INLA”, Spatial and Spatio-temporal Epidemiology 4: 33–49.
Bonardi, Jean–Philippe et al. (2020), “Fast and local: How lockdown policies aﬀect the spread and severity of covid-19”, Covid Economics 23: 325–351.
Caldera, A and S Koirala (2020), “Eight priorities to strengthen international cooperation against Covid-19”, VoxEU.org, 30 June.
Fahrmeir, L, T Kneib, and S Lang (2004), “Penalized Structured Additive Regression for Space-Time Data: A Bayesian Perspective”, Statistica Sinica 14(3): 731–761.
Keita, S (2020),“Air passenger mobility, travel restrictions, and the transmission of the covid-19 pandemic between countries”, Covid Economics 9: 77–96.
Lauer, S A et al. (2020), “The Incubation Period of Coronavirus Disease 2019 (Covid-19) From Publicly Reported Confirmed Cases: Estimation and Application”, Annals of Internal Medicine 172(9): 577–582.
Rothert, J, R Brady and M Insler (2020), “The fragmented US: Local COVID-19 policies impact the rest of the country”, VoxEU.org, 22 September.
Weber, E (2020), “Which measures flattened the curve in Germany”, Covid Economics 24: 205–217.
1 Those are Andorra, Austria, Belgium, Denmark, France, Finland, Germany, Ireland, Italy, Liechtenstein, Luxembourg, the Netherlands, Norway, Portugal, Spain, Switzerland, Sweden and UK – in other words, all of Western Europe except for the isolated island of Iceland. For France, we approximate daily new cases by the number of hospitalized per day and region and rescale them to the number of confirmed cases with national data. We aggregate the data to the level of 213 roughly equally sized sub-national regions that closely follow the definition of European NUTS2 regions. Finally, we rescale all regional daily case counts to match the national totals reported by Johns Hopkins University’s Coronavirus Resource Center.
2 We note that due to the staggered treatment timing, the estimated coeﬃcients present weighted averages of the underlying group-time average treatment eﬀects that are likely to underestimate the actual average treatment eﬀect (Callaway and Sant’Anna 2019)
3 The number of observations decreases because three countries composed of a single region (Andorra, Liechtenstein, Luxembourg) drop out of the sample.
4 We model both the temporal dependence and the duration of the border controls by a second-order random walk specification of the eﬀect.