Right and yet wrong: a politico-economic perspective on Germany’s early COVID-19 policy

ABSTRACT

Applying a spatio-temporal endemic–epidemic forecasting model, we evaluate different perspectives on the adequacy of COVID-19 containment policies. Using Germany’s early containment policy as an example, we show that containment policies judged as rational based on the real-time perspective of policymakers may be deemed unnecessary or ineffective in ex-post evaluations. We also demonstrate that one-size-fits-all policies implemented in Germany early in the pandemic are likely suboptimal.

KEYWORDS:

• containment measures
• real-time analysis
• COVID-19
• SIR model
• spatio-temporal modelling

JEL:

• C23
• C53
• I12
• I18

ACKNOWLEDGEMENTS

We thank the three anonymous reviewers and editor Ben Derudder for the excellent comments we received during the review process. Any remaining errors are our sole responsibility. This paper is based on an earlier working paper version (Berlemann & Haustein, 2020).

DISCLOSURE STATEMENT

No potential conflict of interest was reported by the authors.

Notes

1. The index shows the mean score of the nine metrics. When policies vary at the subnational level, the index is shown as the response level of the strictest subregion. To control for the vaccination status, the stringency index is calculated for vaccinated and non-vaccinated individuals. We show a weighted national average based on the share of vaccinated individuals.

2. We focus our analysis on the early epidemic phase because testing activity was stable throughout the first wave, such that the employed infection data had minimal testing biases. Until mid-June 2020, the number of weekly tests fluctuated around 350,000 (see https://www.rki.de/DE/Content/InfAZ/N/Neuartiges_Coronavirus/Testzahl.html).

3. We explain the RKI data in detail in section 4.2.

4. For a discussion of real-time data use in macroeconomic forecasting, see, for example, Croushore (2011).

5. Therefore, we do not have to take into account possible publication lags of infection numbers.

6. The SIR model assumes the population can be subdivided into at least three compartments: susceptible (S), infectious (I) and recovered individuals (R). When central parameters such as the likelihood of susceptible individuals to become infectious, the time infectious individuals remain infectious and the time until recovery are known, the SIR model can be formulated as a system of differential equations. After calibration, it can be used for forecasting or simulation.

7. The SEIR model extends the SIR model for a latent state of being exposed (E) before becoming infectious.

8. Numerous studies on the effectiveness of alternative containment measures have been conducted for countries such as, for example, China (Kraemer et al., 2020), Spain (Orea & Álvarez, 2020), Italy (Celani & Giudici, 2022) or the United States (Abouk & Heydari, 2021; Chernozhukov et al., 2021; Courtemanche et al., 2020). A multi-country study based on regional data was conducted by Hsiang et al. (2020).

9. The SIQRD model is an extension of the SIR model that additionally considers quarantines (Q) and deaths (D).

10. We explain the employed data in detail in section 4.2.

11. This approach has been used to model the spatio-temporal spread of COVID-19 in Italian provinces (Giuliani et al., 2020; Celani & Giudici, 2022).

12. The time trend also corrects for potential changes in the testing intensity.

13. In principle, e_r can vary over time. However, as local population counts are not available at high frequency at the county level, we employ the newest population counts from the end of 2019, which are available in the RKI data.

14. Held et al. (2005) provide a motivation for the modelling choices of the autoregressive component.

15. In contiguity matrices of the type ‘queen’, only directly neighbouring counties can influence the infection numbers. The selection of the contiguity matrix is based on computational considerations. Applying a power-law distance decay specification such as in Meyer and Held (2014) did not produce systematic convergence of the estimates in most models.

16. For this package, see https://github.com/jbracher/hhh4addon. See also Bracher and Held (2022) for an application to dengue fever in Puerto Rico and to viral gastroenteritis in Berlin.

17. See https://www.rki.de/DE/Content/InfAZ/N/Neuartiges_Coronavirus/Fallzahlen.html/.

18. COVID-19 often occurs without any or with only mild symptoms (e.g., Streeck et al., 2020). Thus, the factual number of infections is likely larger than the one reported in the RKI data. However, as we explain in section 3, underreported data are not an issue in our empirical approach.

19. Over the sample period, the share of observations with reference dates decreases to an extent. This is most likely due to shrinking capacities of the local public health departments to re-investigate missing data in the case reports. Overall, 70.25% of all observations in our sample have a reference date.

20. Not all counties were immediately affected by COVID-19. Hence, it is not always possible to calculate predictions of the reference date for these observations based on a fixed-effects model. Therefore, we use a simpler regression model including only the age group and gender for our real-time analysis mimicking the situation of the policymaker.

21. We skipped parts of the upper part of the prediction interval due to visibility of the mean prediction.

22. This does not hold true for the introduction of the obligation to wear facemasks, which differed substantially between regions. As a result, we do not study this measure here. For an analysis of the effects of face masks on infection dynamics in Germany, see, for example, Mitze et al. (2020).

23. The components displayed at the country level can be obtained by summing over all counties for a giving point in time t. For example, to obtain the endemic autoregressive component, we would calculate ${\sum }_{r=1}^R{\hat{\lambda }}_r\cdot {\sum }_{d=1}^D{u}_d\cdot Y_{r.t-d}$ for a given point in time t. ${\hat{\lambda }}_r$ corresponds to the estimate of ${\lambda }_r$. Thus, they represent the contribution to the number of predicted infections.

24. We refrain from further extending the forecast horizon as the next group of containment measures was announced as early as 16 March.

25. The forecast intervals were constructed via 10,000 Monte Carlo simulations.

26. This is because the third group of containment measures was already announced a week later, on the evening of 22 March.

27. For the United States, Goolsbee and Syverson (2021) use cellular phone records data to show a change in mobility before a lockdown. Similarly, Chetty et al. (2020) observe changing spending behaviour of high-income people by avoiding high-incidence areas. For the UK, Wood (2022) argues to observe falling incidence numbers before a full lockdown was imposed. For Japan, Watanabe and Yabu (2021) show that providing appropriate information can induce behavioural changes and is effective in containing the spread of the virus.

28. For example, to obtain the endemic autoregressive share of county r at a point in time t, we calculate: ${\displaystyle \frac{{\hat{\lambda }}_r\cdot {\sum }_{d=1}^D{u}_d\cdot Y_{r,t-d}}{{\hat{\mu }}_{r,t}^Y}.}$

29. Saxony was one of the least affected states by the first infection wave. In addition, a relatively large share of individuals supporting COVID-19 scepticism and the popularity of the far-right Alternative für Deutschland (AFD) party might be relevant for the evaluation (e.g., Reuband, 2022).