AdobeStock_347892553_Editorial_Use_Only.jpeg
VoxEU Column COVID-19 Health Economics

MIDIS: A measure for latent social distancing

As governments begin to ease lockdown measures over the coming months, understanding how effectively countries have applied social distancing practices will be essential. This column introduces a Model-Inferred measure of latent social DIStancing (MIDIS) and calculates the measure for 44 countries using daily data and an epidemiology model. Mobility data from Apple and Google indicate that the measure can accurately measure distancing, and the measure also reflects governmental and behavioural responses while maintaining a robust relationship with daily output losses.

Measuring social distancing is critical in the context of the COVID-19 pandemic. Until a vaccine or an effective antiviral treatment is developed, policymakers as well as individuals will be caught between the necessity and tolerability of social distancing (Baldwin 2020). On the one hand, there is a need to sufficiently contain the spread of the disease through social distancing to prevent an overrun on healthcare facilities. On the other hand, it is impossible to continue social distancing—either through governmental or individual measures—indefinitely, since people’s livelihoods depend on being able to work.

To inform and facilitate social distancing-related decision making processes against a background of necessity/tolerability trade-offs, it would be desirable for policymakers and individuals to have access to a relatively reliable and robust distancing measure with minimum requirements of high-frequency data. In a recent paper, we attempt to identify a social distancing term for each country and each day in order to construct such a measure (Attar and Tekin-Koru 2020).

Identification strategy

We derive a Model-Inferred DIStancing (MIDIS) measure using an extended version of the workhorse Susceptible-Exposed-Infected-Recovered-Deceased (SEIRD) framework, which was originally proposed as a SIR model by Kermack and McKendrick (1927). In a typical SEIRD model, a nonlinear dynamical system explains the spread and eventual containment of an infection over time.

We extend the simple SEIRD model with a time-varying and country-dependent social distancing term. Our core idea is to identify this distancing term, MIDIS, for each country and each day by exploiting the fact that the pure probability of transmission and the average incubation period are constant and common across countries.

For all countries, the model horizon is the first 30 days after the 500th COVID-19 case is confirmed. Hence, time periods are not synchronised across countries with respect to calendar time. The need to restrict our analysis to the period after the 500th case for each country originates from the fact that the official COVID-19 statistics for China start with 548 cases on 22 January 2020. We choose to restrict the analysis to the first 30 days after the 500th case to disregard the effects of a partial removal of social distancing measures.

The resulting solution expresses MIDIS as a function of observable epidemiological data and thus provides a model-inferred measure of a latent variable that can be tracked over time. An important advantage of our identification strategy lies in the ease with which it can be put into practice by other researchers because it employs a relatively simple epidemiological model and readily available data.

Another advantage of our identification strategy is that MIDIS captures a wide range of social distancing components. These include not only policy interventions (school/work closures, bans on traveling and mass gatherings, or stay-home orders) but also behavioural responses such as fear, trust, or reciprocity, which cannot be measured in a straightforward way.


A detailed account of the identification strategy as well as all data and codes we use are available on the MIDIS website.


MIDIS in selected countries

We apply MIDIS to the data compiled by Johns Hopkins University (JHU 2020) and compute it for 44 countries with a total number of confirmed COVID-19 cases that exceeds 10,000 as of 11 May 2020.

For the immediate 30-days in the aftermath of the 500th case, our results show that countries exhibit considerable variation in average social distancing levels. Figure 1 illustrates the 30-day average MIDIS values (varying between zero and 100) for the 44 countries we investigate.

The variation apparent in Figure 1 has its roots in the initial values and the evolution of social distancing in these 30 days. To be precise, with the exceptions of the US and Spain, countries start with an initial social distancing level that is larger than the Chinese benchmark. Furthermore, in a large number of countries, there is a minor decline of MIDIS within the first week followed by a slow but persistent increase later on.

South Korea is the country that sustains the highest average level of distancing, and the US is the least effective country in this respect. The considerable cross-country variation in social distancing levels and the South Korean success relative to European countries are consistent with the SIR-based empirical evidence presented by Chudik et al. (2020).

The US stands out among these 44 countries such that the initial decrease in MIDIS in the first five days is the fastest in the sample. Hence, the US is not only the country that records the lowest initial MIDIS value; on the sixth day, the MIDIS value of the US converges to the minimum of the entire sample. 

Figure 1 30-day average of MIDIS values

Source: Authors’ own illustration.

We also investigate whether the daily mobility data from Apple and Google validate our distancing measure because mobility has now been widely used in the burgeoning COVID-19 literature, sometimes as a proxy for social distancing (Alfaro et al. 2020, Coven and Gupta 2020, Durante et al. 2020, Doganoglu and Ozdenoren 2020).

Figure 2 compactly illustrates the information on the coefficient estimates of individual equations that regress MIDIS on different components of the mobility data. The results show that mobility indicators are strongly correlated with MIDIS and support the validity of our identification strategy.

As expected, increased mobility in the public sphere has a strong, inverse relationship with MIDIS. Among the indicators from the Apple data, we estimate the largest effect for transit stations. The magnitude of the estimated slope parameter is close to those we obtain for mobility indicators for transit stations and workplace indicators using Google data.

Figure 2 MIDIS and mobility data

Note: The circles are point estimates and the lines are 95% confidence intervals. The data sources are Apple (2020) and Google (2020).

Cross-country heterogeneity in MIDIS

What causes this variation in social distancing across countries? The natural candidates are governmental as well as individual responses to the pandemic, coupled with country-specific factors.

We argue that the behavioural response is at least as important as the governmental response in explaining the variation in MIDIS across countries and time. Figures 3 and 4 show the impact of governmental and behavioural responses on social distancing measured by MIDIS within the framework of a very basic panel data estimation.

The data for governmental response is the Stringency Index (varying between zero and 100) from the Oxford COVID-19 Government Response Tracker (Hale et al. 2020). We use JHU (2020) epidemiological data components as proxies for behavioural responses. It would not be hard for anyone who consciously experienced the COVID-19 pandemic to recall that they had to drop everything to get news of the numbers of infected, deceased, and recovered for the COVID-19 cases in their cities, countries, and the world every day to prepare for the next day. In other words, people use daily epidemiological data – particularly the numbers of infected or deceased individuals that headline all types of news outlets – to inform their behaviour on the following day.

Figure 3 MIDIS and governmental response

Note: Marginal effect of lagged values of Stringency Index on MIDIS with a 95% confidence interval, for 44 countries in the 30-day period, in a country fixed effects specification as discussed in Attar and Tekin-Koru (2020).

Figure 4 MIDIS and behavioural response

Note: Marginal effect of lagged values of the number of deceased people (in thousands) on MIDIS with a 95% confidence interval, for 44 countries in the 30-day period, in a country fixed effects specification as discussed in Attar and Tekin-Koru (2020).

Our results show that MIDIS varies positively with containment measures taken by governments and people’s reaction to the pandemic in a robust manner. Indeed, the impact of behavioural responses measured by the numbers of deceased on the previous day is stronger than the impact of containment measures as shown in Figures 3 and 4. While a one-point increase in the Stringency Index in the previous day increases MIDIS by 0.15 points, every 1,000 people deceased a day before increases MIDIS more than three points.

A basic application using MIDIS

Finally, we use MIDIS to study the economic costs of social distancing. While doing so, we stay oblivious to supply or demand side dynamics of these economic costs and focus on their outcome in terms of output loss only.

Unlike epidemiological data, it is impossible to come by daily data for output, which makes it necessary to use a proxy instead. We use the Bruegel Electricity Tracker of COVID-19 Lockdown Effects compiled and calculated by McWilliams and Zachmann (2020) to approximate output losses experienced during the pandemic based on the premise that significant economic activity relies heavily on the use of electricity.

Figure 5 Output loss and MIDIS (selected countries)

Note: Authors’ calculations.

Figure 5 depicts output loss against MIDIS along with a basic linear fit between the two. The results indicate a significant negative output response to social distancing during the COVID-19 pandemic. In other words, in countries with higher levels of MIDIS, there is a higher level of output loss in the 30-days following the 500th case. Indeed, a 10% increase in social distancing causes up to a 3.8% increase in output loss as reported in Attar and Tekin-Koru (2020).

Clearly, the effects of MIDIS on output loss documented in Figure 5 do not identify the structural mechanisms, but they serve as a reduced-form device that allows us to see the quantifiable output impact of distancing. We should also note that the strong relationship between distancing and output loss is expected to hold in the very short run, e.g. a month, but it may weaken and be reversed in the longer run. This is because both the centralised/optimal distancing policies and decentralised/individual distancing practices are time-dependent, changing daily as demonstrated in the related literature (Bethune and Korinek 2020, Farboodi et al. 2020).

Final words

We consider our way of identifying latent social distancing through MIDIS as an initial attempt to improve the SEIRD model and a contribution to the intense debate on the effects of the COVID-19 pandemic. We expect our qualitative results to be informative and useful. However, due to the highly stylised nature of the underlying epidemiological model we use to construct MIDIS, we urge our readers to interpret our quantitative results with care.

Since governments are taking steps to ease lockdown restrictions at the time of writing this column, we can safely say that there is much that will change in the near future in regards to social distancing and its diverse effects on individual lives. Our hope is to continue this line of work to incorporate what was missed in the current version of Attar and Tekin-Koru (2020) and new developments in the pandemic as they arise.

References

Alfaro, L, E Faia, N Lamersdorf and F Saidi (2020), “Social Interactions in Pandemics: Fear, Altruism, and Reciprocity”, NBER Working Paper 27134.

Apple (2020), Mobility Trends Reports.

Attar, M A and A Tekin-Koru (2020), “Latent Social Distancing: Identification, Causes and Consequences”, Covid Economics: Vetted and Real-Time Papers, 1(26).

Baldwin, R (2020), “COVID, remobilisation and the ‘stringency possibility corridor’: Creating wealth while protecting health”, VoxEU.org, 10 April. 

Bethune, Z and A Korinek (2020), “COVID-19 infection externalities: Pursuing herd immunity or containment?Covid Economics: Vetted and Real-Time Papers, 1(11).

Chudik, A, M H Pesaran and A Rebucci (2020), “Voluntary and mandatory social distancing: Evidence on Covid-19 exposure rates from Chinese provinces and selected countries”, Covid Economics: Vetted and Real-Time Papers, 1(15).

Coven, J and A Gupta (2020), “Disparities in Mobility Responses to COVID-19”, NYU Stern Working Paper.

Doganoglu, T and E Ozdenoren (2020), “Should I stay or should I go (out): The role of trust and norms in disease prevention during pandemics”, Covid Economics: Vetted and Real-Time Papers, 1(16).

Durante, R, L Guiso and G Gulino (2020), "Civic Capital and Social Distancing: Evidence from Italians' Response to COVID-19", VoxEU.org, 16 April.

Farboodi, M, G Jarosch and R Shimer (2020), “Internal and external effects of social distancing in a pandemics”, Covid Economics: Vetted and Real-Time Papers, 1(9).

Google (2020), COVID-19 Community Mobility Results.

Hale, T, N Angrist, B Kira, A Petherick, T Phillips and S Webster (2020), “Variation in government responses to COVID-19”, BSG Working Paper Series 2020/032.

JHU (2020), COVID-19 Data Repository.

Kermack, W O and A G McKendrick (1927), “A Contribution to the Mathematical Theory of Epidemics", Proceedings of the Royal Society of London. Series A, Containing Papers of a Mathematical and Physical Character, 115(772): 700-721

McWilliams, B and G Zachmann (2020), “Bruegel electricity tracker of COVID-19 lockdown effects”, Bruegel Datasets.

630 Reads