The Covid-19 pandemic has prompted governments around the globe to establish an arsenal of nonpharmaceutical interventions (NPIs). NPIs are community mitigation strategies citizens must follow to keep the spread of coronavirus at bay, apart from getting vaccinated. Whether NPIs came about as pure recommendations (e.g., wash your hands frequently) or mandatory measures (e.g., evening curfews), we can probably all agree that these interventions have radically changed how we have been going about our daily lives for almost two years now. In my previous post, I evidenced the impact of governmental intervention on the example of a single policy in Germany and Sweden, two countries taking pandemic responses in very different directions. But how do such strategies play out in aggregate? Existing research alludes to considerable variation by when and in what cadence national governments roll out NPIs (Hale et al., 2020). Intrigued by these findings, I tackle the following question next:
How did Germany and Sweden differ in enacting stringent social and physical distancing measures in response to daily COVID case dynamics early into the pandemic?
Why should we care? In the wake of the COVID-19 pandemic, all we had was intuition. Almost two years into the pandemic, what we have is data, lots of data. Associating policy deployment schedules with case dynamics in the aftermath holds invaluable learnings for governments to put together more evidence-based response strategies down the line.
Setup
I collect data from late February until late November 2020 to analyze policy mobilization in response to COVID-19 case dynamics throughout the early phases of the pandemic.
Collection of epidemiological data. I tap into the publicly accessible OxCOVID19 database (Mahdi et al., 2020) to query the number of confirmed COVID-19 cases accruing in both countries for the timespan of interest. (Read more about this repository in my previous article). Since incident data is returned as cumulative counts, I disaggregate all values into daily figures to start with. This exercise reveals a suspicious portion of 88 consecutive instances of 0 daily cases for Sweden. By conducting spot-checks against case reports from worldometer, I expose these data points as missing values. In addition, 20 instances of negative daily case counts raise red flags. I resort to incident data from the European Centre for Disease Prevention & Control (ECDC, 2021) to address missing and erroneous values. This is essential to prevent such records from distorting any data manipulations down the line.
The take-away from this detour is twofold: First, even data provisioned through the most reputable sources does not release us from the task of rigorous validation before jumping headfirst into the analysis. Second, if a quick look around might reward us with "organic" substitution values, we should resist the temptation of adopting more sophisticated imputation strategies. Why? Data-driven imputation alleviates the pain but does not provide a cure considering that arithmetically synthesized values are pure approximations of an undisclosed reality.
After removing all mentioned inconsistencies, another issue comes to mind. The day a given COVID-19 case is confirmed is not necessarily the day this incident transitions into the database (bureaucracy is our worst enemy, and data engineers also have weekends…). To smooth out the effects of reporting latency, I transform case counts into 7-day moving averages. It is worth mentioning that resampling to higher levels of granularity (e.g., from days to weeks or months) provides us with an alternative means to tame noise. Yet, I deliberately decided against this measure because my objective is to explore case-policy dynamics on the lowest granularity level possible. Lastly, I normalize observations to daily case counts per 100k citizens to factor out stark discrepancies in overall population sizes between Sweden and Germany (Germany has approximately 8 times as many residents as Sweden).
Collection of policy data. The World Health Organization’s PHSM API was queried to access data on Public Health and Social Measures undertaken by both countries over time (WHO, 2021). Whereas OxCOVID19 pre-bins policy data into 8 coarse-grained categories, PHSM sets forth a more granular classification scheme. Using categorical attributes _whocategory, _whosubcategory, _whomeasure and enforcement, I subset those policies qualifying as stringent social and physical distancing measures. In addition, attributes _prev_measurenumber and _following_measurenumber allow me to trace back parent-child dependencies between policies. This is incredibly valuable to characterize policy deployment schemes adopted by either country in greater depth. In that sense, any policy without an entry in _prev_measurenumber qualifies as a newly introduced intervention. Conversely, any policy associated with a _prev_measurenumber refers to an offspring of an existing policy.


I created a 2-by-2 small multiples plot to frame my analysis of policy deployment dynamics relative to daily case counts in Germany and Sweden. The bottom-two Gantt Chart-like plots illustrate the commencement and duration of policy interventions in either country over time. Colour encodings distinguish between parent policies without predecessor and child policies representing a preceding intervention’s modification, extension, or re-introduction. For visual clarity, the number of parent and child policies active at any given point in time are projected as stacked bars into the upper two quadrants depicting 7-day moving average case counts for both countries at available dates.
The lower two charts (Figure 1, Subplots 3 & 4) depict a contrasting narrative. After the first few cases were recorded, the German government released a cascade of 68 mandatory distancing policies, more than 4 times the amount introduced by Sweden. In fact, 54% of all compulsory distancing policies enforced by German authorities are still active to the end of the observation horizon. Remarkably short intervals of less than 4 days on average between the introduction of two interventions and extensive average policy run times of 101 days emerge as leading themes in Germany’s policy mobilization apparatus. In sum, visual cues and summary stats allude to a very systematic and ample policy response. Federal government structures are often considered to pose coordinative hurdles to swift crisis containment. In light of this analysis, I dare to argue that Germany’s decentralized executive structures might have eventually favored a fast and comprehensive rollout of these interventions.
These findings stand in sharp contrast with inferences drawn from the corresponding chart for Sweden. All of Sweden’s 16 mandatory distancing policies were phased out by the end of September. Policies were not only introduced at much greater intervals of 11.53 days on average but also lasted only for a mean duration of 42 days. Supporting these observations, the associated line chart (Subplot 2) reveals that Sweden barely changes gears when daily case counts change. By contrast, Germany’s policy portfolio steadily expands after an initial ramp-up almost in perfect sync with increasing case counts up to the point where incidents first peak in mid-March (Subplot 1). Even when newly confirmed cases remained low, Germany upheld what appears to be a 2:1 blend between numerous parent and child policies.
Statistical Analysis
Using the Pearson correlation coefficient, I correlate the number of active policies with daily confirmed case and death counts to look at identified dynamics with a statistical lens. In plain English, a correlation coefficient quantifies the strength of a linear relationship between two different variables. A value close to 0 hints at a weak relationship, while a value close to absolute 1 suggests that two variables move in near-perfect sync. A positive coefficient means that both variables move in the same direction, whereas a negative coefficient indicates movements in opposite directions. It is worth mentioning that the assumption of a linear association between variables remains a naive simplification of reality, but that’s okay. It is our awareness of these underlying assumptions that matters when interpreting results. Most importantly, correlation coefficients are easy-to-compute and conceptually accessible association metrics for us to draw preliminary conclusions about potentially more complex relationships.
Correlation analysis results showcased in Table 1 corroborate inferred differences in how Swedish and German policy deployment schemes attune to daily case dynamics. Associating the number of policies active with normalized daily case and death counts for Germany yields correlation coefficients of 0.33 and 0.36, respectively. Coefficients pair with low p-values attesting statistical significance. We might mourn the magnitude of these coefficients. Yet, findings deserve contextualization in the realm of national policymaking. When considering that bureaucratic hurdles undoubtedly curb the flexibility of re-adjusting a democratic, federally organized government’s policy agenda from one day to another, identified correlations are deemed substantial.
The same analysis yields a comparatively small, non-significant correlation coefficient of -0.1 for Sweden. Once more, these findings resonate with literature emphasizing the strikingly minimalistic policy response of the Swedish government, which opts for voluntary self-policing rather than mitigation through explicit legislative action (Capano et al., 2020). Correlating number of policies active with normalized daily death counts however yields a statistically significant correlation coefficient of 0.3. This somewhat surprising finding might root in Sweden’s tendency to pivot from lighter to more stringent policy interventions as cumulative death tolls rise (Cheibub et al., 2020).
What does it all mean?
Uncovered differences in pandemic policy response resonate with Cheibub et al.’s (2020) findings that national democracies are by no means similar in the way they respond to the pandemic. The driving forces behind these heterogeneities remain subject to debate. I reason apparent differences between Germany and Sweden with recourse to Hofstede’s cultural dimension theory. This theory qualifies the notion of Uncertainty Avoidance as "(…)the extent to which the members of a culture feel threatened by ambiguous or unknown situations and have created beliefs and institutions that try to avoid these" (Hofsede-insights.com, 2020).
Scoring 29 out of 100, Sweden is portrayed as an uncertainty-tolerant nation that prefers to steer its society on a minimum necessary set of rules. On the other hand, Germany, scoring 67 out of 100, relies on planning and systematic anticipation of uncertainty. The two Gantt charts (Figure 1, Subplots 3 & 4) provide compelling support for this reasoning. Germany issued a set of closely clocked parent policies early in the pandemic, which has not only been largely sustained but extended into a vast body of social distancing policies as both, new parent and child policies were added over time. While this can be understood as an act of radical uncertainty mitigation to counter a pandemic of unknown proportions, the fact that Germany maintains an average Parent-to-child policy ratio of 2.23 to 1 (Figure 1, Subplot 1) portrays a continuous yet controlled expansion of policy intervention. In contrast, a ratio of 0.53 to 1 for Sweden rather epitomizes a reactionist approach (Figure 1, Subplot 2). Eventually, Sweden’s policy apparatus, which falls short in all thematized dimensions – policy duration, deployment frequency & total interventions to date – hints at a policy footprint unique to a high-trust society defying excess governmental control.
Limitations
My analysis employed two external data feeds to address shortcomings within the OxCOVID19 database and procure additional variables required to "datafy" the research question at hand. However, integrating data from multiple sources implies integrating different information taxonomies, data collection, and pre-processing methodologies, which are only tentatively disclosed on the data controllers’ websites. With my analysis being at the mercy of several unknowns regarding data provenance and methodology, findings must be taken – as always – with a healthy pinch of salt.
While this analysis ascertained significant correlations between the number of policies active and daily case counts, I caution you not to mistake correlation for causation. This analysis assembles a largely descriptive entry point for readers to witness different social distancing policy strategies at work. Yet, this convenience is established at the cost of omitting numerous factors, which can easily nudge us into erroneous attributions of cause and effect. For instance, other factors such as weather not controlled for by the analysis might have facilitated social distancing over the summer months. Such circumstances remind us not to label the number of social distancing policies as a stand-alone driving factor behind lowering daily case counts. I briefly talk about the concept of omitted variable bias in my previous article in case this little reflection sparked your interest.
Conclusions
We set out by asking: How do Germany and Sweden differ in enacting stringent social and physical distancing measures in response to daily COVID case dynamics early into the pandemic? Capano et al.( 2020) call the pandemic as a "natural experiment" during which governments show remarkably different policy responses to the same problem. My analyses reveals stark differences between how the governments of Germany and Sweden mobilize against the pandemic. At the same time, I arrive at the conclusion that these responses are the complex product of national leadership, government organization, and political capacity to conceive and deploy policies on a national scale.
The morale: It is not the hammer or the dance but the carefully crafted blend between press and release that needs to be understood against the backdrop of national and cultural particularities. But if we can say one thing for certain, it is probably this: An answer to how much hammering and dancing we need in which situation should not forgo the opportunity to put the data this pandemic has provided us with to the test.
References
Capano, G., Howlett, M., Jarvis, D. S. L., Ramesh, M., & Goyal, N. (2020). Mobilizing Policy (In)Capacity to Fight COVID-19: Understanding Variations in State Responses. Policy and Society, 39(3), 1–24. https://doi.org/10.1080/14494035.2020.1787628
Cheibub, J. A., Hong, J. Y. J., & Przeworski, A. (2020). Rights and Deaths: Government Reactions to the Pandemic. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.3645410
Hale, T., Angrist, N., Petherick, A., Kira, B., Phillips, T., & Webster, S. (2020). Variation in government responses to COVID-19. Blavatnik School of Government Working Paper Series, 32. www.bsg.ox.ac.uk/covidtracker%0Ahttps://www.bsg.ox.ac.uk/research/publications/variation-government-responses-covid-19
Data
ECDC (2021).Daily Update of new reported cases of COVID-19 by country worldwide.[Data set].https://www.ecdc.europa.eu/en/publications-data/download-todays-data-geographic-distribution-covid-19-cases-worldwide
Mahdi, A., Blaszczyk, P., Dlotko, P., Salvi, D., Chan, T. S. T., Harvey, J., & Zarebski, A. E. (2020). OxCOVID19 Database: a multimodal data repository for better understanding the global impact of COVID-19. medRxiv.
WHO (2021). PHSM: Tracking Public Health and Social Measures. [Data set & Code Book]. https://www.who.int/emergencies/diseases/novel-coronavirus-2019/phsm