The world’s leading publication for data science, AI, and ML professionals.

How do people move when social distancing becomes the rule

Uncovering the impact of shelter-in-place requirements on mobility behavior among German and Swedish residents early into the Covid-19…

Hands-on Tutorials

Photo by Timon Studler on Unsplash
Photo by Timon Studler on Unsplash

Many countries have enforced shelter-in-place requirements to contain the spread of COVID-19 on national territories. These non-pharmaceutical interventions (NPIs) have impacted people’s mobility routines in ways that policy researchers hurry to decipher. The aim is not only to be better equipped for new pandemic episodes but also to uncover timely learnings for those countries which are not over the worst. Motivated by Bryant & Elofsson’s (2020) claim that people in different countries will alter their mobility patterns differently in response to similar NPIs, I decided to add to the debate with an exploratory analysis framed by the following question:

How do citizens in Germany and Sweden reduce personal mobility in response to shelter-in-place requirements enacted by their respective governments over time?

The comparison between Germany and Sweden promises interesting findings. Both countries adopted seemingly opposing strategies early into the pandemic to contain the virus. Sweden recurrently bannered the news with its "lockdown-light" strategy placing great trust in voluntary self-policing among its residents. On the other hand, Germany launched early and stringent governmental interventions which were not seldomly praised as role model response.

Setup

I turn to the publicly accessible OxCOVID19 database (Mahdi et al., 2020) to query mobility and policy data on both countries. The OxCOVID19 database represents a large, multi-model repository collating epidemiological, government response, mobility, and sociodemographic data across a vast spectrum of nations. All available data is curated as part of the OxCOVID19 Project which aims to advance our understanding of the pandemic and motivate the development of mitigation strategies based on statistical analysis. I factor in data from late February until late November 2020 to explore policy enforcement dynamics together with changes in public mobility throughout the early phases of the pandemic in both countries.

My analysis draws on two tables within the OxCOVID19 database. Shelter-in-place policy enactment data is sourced from the Government Response table collating country-level ordinal data on containment policies enacted. The stringency at which policies are enforced at any given point in time spreads across three levels. A value of 0 denotes that a given policy is currently not active, whereas a value of 1 signifies that policy compliance is recommenced. Lastly, a value of 2 implies that compliance is mandatory. I retrieve Google mobility trends data from the Mobility table, capturing % daily mobility deviations from pre-pandemic baseline values for different location categories (e.g. workplace, transit station, parks). I aggregate mobility data to the nation-level to enable a cross-country comparison and match the granularity of available policy data for subsequent linkage. I further resample mobility data into weekly bins to deflate the impact of weekends, bank holidays, and other contextual factors (e.g., weather) not explicitly controlled for by the analysis.

Descriptive Analysis

It’s never a bad idea to kick it off with some descriptive stats. Table 1 records the percentage distribution of shelter-in-place stringency levels for either country for the timespan in scope. A few things stick out.

Germany suspended shelter-in-place orders for 66% of time while Sweden issued shelter-in-place recommendations for 86% of the dates in scope. Yet, Sweden has never turned this recommendation into a requirement whereas Germany enacted the most stringent policy level 20% of the time, spanning almost two-thirds of the timeframe the policy was issued.

Next, I plot % mobility changes amongst German and Swedish citizens by location category over time. I further overlay each time series plot by vertical bars whose width and colour correspond to the duration and stringency with which the two governments enacted shelter-in-place policies.

Figure 1. Relationship between changes in personal mobility and shelter-in-place requirements enacted
Figure 1. Relationship between changes in personal mobility and shelter-in-place requirements enacted

Germany encased periods of greater freedom of mobility in shorter, contracted periods of stringent policy enactment relative to Sweden, which made a mere recommendation of sheltering-in-place the new normal ever since the introduction of the policy in mid-March. This visual narrative coincides with voices in academia and media highlighting the existence of two radically different virus containment roadmaps. Sweden adopts a unique policy blend of mild mandates paired with public appeals to self-policing without curtailing personal freedom in a society that dislikes explicit governmental remote control. In contrast, Germany advocates early, stringent intervention followed by a "(…) safe dance around the infection curve" through softer measures (Jung et al., p.363, 2020).

Notable % mobility decreases from pre-pandemic reference values across location categories for both countries emerge whenever shelter-in-place requirements are in place regardless of stringency level. The summary statistics next to each plot highlight notable drops in average mobility changes when shelter-in-place policies are active vs. lifted. Expressed in relative terms, double- and triple-digit dips in mobility surface for both countries across location categories whenever shelter-in-place policies are in motion.

Significance Testing

Okay, we can say so far that mobility patterns differ between Germany and Sweden for the timeframe in scope. But can we be confident that looming differences are not simply due to chance? To address this question, I proceed with the calculation of 95% confidence intervals. This is to assess whether the average change in personal mobility significantly differs between German and Swedish residents. In layman’s terms, a 95% confidence interval tells us that we are 95% confident that the true difference in average mobility between both nations situates somewhere between the lower and upper interval boundary that I will compute for each of the 4 location categories. Now suppose the bounds of a given confidence interval do not include 0. In that case, we can be confident that the relative reduction in mobility among German residents is effectively higher (i.e., if both interval boundaries are negative) or lower (i.e., if both interval boundaries are positive) compared to their Swedish peers. Table 2 reports the confidence intervals across the 4 location categories in scope.

Germany, which introduces a blend of shelter-at-home recommendations and requirements, experiences steeper relative mobility decreases on average than Sweden across all categories. Differences are particularly pronounced for "high footfall" location categories Retail & Recreation (-29.1%) and Transit Stations (-11.3%). Conversely, differences are less marked for categories Workplace (-3.1%) and Grocery & Pharmacy (-5.7%). These differences make sense since going to the office, buying food, and getting medicine remained fundamental rights irrespective of the strigency with which shelter-in-place policies were enacted, especially at the beginning of the pandemic where alternatives such as remote working have not yet come of age. Most importantly, I find that my 95% confidence intervals ascribe statistical significance to identified differences for all location categories except for category Workplace. All things equal, the time series plots in Figure 1 reveal that German citizens rediscover their joy for personal mobility strikingly quickly upon policy suspension at the beginning of May. These upswings even surpass the mobility levels of their Swedish peers, who remain subject to shelter-at-home recommendations during late summer. I recommend you to revist Figure 1 and compare how German and Swedish mobility time series evolve for categories Transit Stations (Subplots 1 & 5) and Workplace (Subplots 2 & 6) to witness these dynamics in action.

To get to the bottom of these observations, I compute another set of 95% confidence intervals to assess whether identified mobility differences prevail when Germany suspends shelter-in-place policies while Sweden does not.

The stunning result: German residents indeed show significantly less reduction in mobility for 3 out of 4 location categories. Changes in mobility are between 9.9% and 2.7% less removed from pre-pandemic baseline values for the categories Workplace, Transit Stations and Grocery & Pharmacy than for their Swedish peers.

What does it all mean?

The outcomes of my analysis affirm that reduced personal mobility coincides with policy deployment in both countries but that different deployment strategies regarding timing and stringency are likely to elicit different responses over time. Murphy et al.’s (2020) study on citizens’ containment policy adherence motivations claims that compliance is not here to last once regulations are lifted. Building upon this argument, normative commitments to government-induced health interventions might be the main motive behind temporary reductions of personal mobility among German and Swedish citizens. In that sense, German citizens might return to more opportunistic behavioral patterns as the normative pressure evaporates when restrictions are lifted. Conversely, ongoing shelter-in-place recommendations continue to moderate mobility patterns among Swedish residents.

At the same time, instances where mobility patterns among Swedish citizens show high variability indicate that dynamics are far less black and white than it seems at first glance. Let’s consider the change in mobility among Swedish residents for location category Retail & Recreation (e.g., Subplot 7 in Figure 1) for instance. Mobility levels at times even surpass pre-pandemic levels in the absence of any policy shift. To reason this finding, let’s recall that shelter-in-place requirements have always been exercised as a recommendation for Sweden. Consequently, people’s mobility behavior might gravitate to changes in concurrent, more stringently enacted policies not accounted for by my study. Such situations typify what statisticians commonly refer to as "omitted variable bias". It occurs when a model or an analysis leaves out one or more relevant variables which potentially confound the true association between the variables included. In our case, the relationship between shelter-in-place requirements and changes in personal mobility.

Limitations

Every study comes with its own limitations. It is time to tell you about mine. You might recall that I have deliberately engaged in data aggregation and resampling. While aggregation meant to level mobility and policy data in terms of granularity, resampling meant to smooth out the impact of day-of-week effects or special events not controlled for by the analysis. At the end of the day, there is no free lunch when it comes to data preprocessing. In this sense, resampling and aggregation measures are double-edged swords that elevate legibility and tame noise at the expense of obfuscating subtle signals, which only become evident at lower levels of granularity. Lastly, a few limitations relate to the data I use. Studying the Understand the data section in Google’s (2021) COVID-19 Community Mobility Report, I learned that mobility data is derived from smartphone users who have a Google account and gave access to their location history. In precluding smartphone users without Google account, smartphone users with Google account but different location sharing settings, and people without a smartphone (yes, they still exist!), sample representativeness is at risk. Similarly, Google states that data is withheld if places are not busy enough to prevent re-identification. This well-meant privacy threshold backfires on the actual purpose of reliably capturing footfall. Lastly, Google benchmarks mobility changes against a fixed five-week period between January 3rd and February 6th. Using two winter months as a reference frame to evaluate mobility changes for the remainder of the year raises serious data validity concerns.

Conclusion

This article began by asking: How do citizens in Germany and Sweden reduce personal mobility in response to shelter-in-place requirements enacted by their respective governments over time? By analyzing 8 months’ worth of mobility and government policy data on Germany and Sweden, I showed that the stringency with which shelter-in-place requirements are enforced seemingly impacts people’s mobility behavior. A more conservative enactment of the shelter-in-place policy adopted by Germany coincides with steeper reductions in personal mobility upon enforcement, followed by even steeper rebounds in mobility levels upon relaxation. Eventually, findings allude to a complex relationship between changes in personal mobility and policy enactment that my analysis has only began to uncover. In highlighting the limitations at play, I aspired to critically reflect on my analysis outcomes and, in a more general sense, provide a kind reminder that seemingly convincing findings must always be taken with a grain of salt. Above all, I hope this article inspires further explorations of this kind to improve governmental decision-making in the face of threatening infectious disease outbreaks like COVID-19.

References

Bryant, P., & Elofsson, A. (2020). Estimating the impact of mobility patterns on COVID-19 infection rates in 11 European countries. medRxiv.

Jung, F., Krieger, V., Hufert, F. T., & Küpper, J. H. (2020). How we should respond to the Coronavirus SARS-CoV-2 outbreak: A German perspective. Clinical hemorheology and microcirculation, 74(4), 363–372.

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.


Related Articles