Using mobile phone data for epidemiological simulations of lockdowns: government interventions, behavioral changes, and resulting changes of reinfections

2.1 Disease progression model

The disease progression model is taken from the literature [17, 18, 19, 20, 21, 22] (also see [23]). The model has states exposed, infectious, showing symptoms, seriously sick(= should be in hospital), critical (= needs intensive care), and recovered. The durations from one state to the next follow log-normal distributions; see Fig. 2 for details. We use the same age-dependent transition probabilities as [16], shown in Tab. 1.

• Download figure
• Open in new tab

Figure 2

State transitions [17, 18, 19, 20, 21, 22],

Infecting another person is possible during infectious, and while showing symptoms, but no longer than 4 days after becoming infectious. This models that persons are mostly infectious relatively early through the disease [18], while in later stages the infection may move to the lung [19], which makes it worse for the infected person, but seems to make it less infectious to other persons.

2.2 Contact model

Encounters in “containers”, i.e. in facilities and vehicles, are directly taken from the data. Synthetic persons in the same container at the same time have a certain probability to interact; if they interact, the probability of an infection is given by Eq. (1) if one person is contagious (maximally until 4 days after becoming infectious; see above) and the other is susceptible. In particular, the duration of the interaction to feed into Eq. (1) is also provided by the data.

The algorithm looks at agents when they leave a facility. At this point, it randomly selects 3 other agents which are at the facility at the same time, and with each selected agent computes a possible infection if either the leaving agent is susceptible and the other agent is contagious, or the other way round. The infection model is Eq. (1), the time r is the time (duration) that both agents were simultaneously at the same facility. We bound the infection dynamics at 3 other agents, since we assume that persons do not interact with everybody in the facility. If there are fewer than 3 other agents at the facility, then interaction simply happens with everybody. Clearly, we do not want the model to interact with everybody at the facility. Using “10” instead of “3” in trial runs did not make a noticeable difference (except that the 0 parameter needed recalibration).

The same algorithm is used for interaction in (public transport) vehicles. In terms of implementation, the model uses a generalized dynamics for containers, and treats both facilities and vehicles as such containers.

Note that the number “3” is used when the agent leaves. There is, however, also interaction when other agents leave. Thus, this rather models the interaction with “available spaces” than with persons. E.g. assume the following time line (from left to right) and consider in particular the agent X:

This can, for the ease of interpretation, be interpreted as three seats in a row in a public transport vehicle. We have the following events with respect to person X:

• At time 11, X enters the vehicle, and takes a seat between A and B.

• At time t2, B leaves the vehicle. In consequence, a possible infection of X from B is computed with Eq. (1), with t2 – t1 as the duration of the interaction. Also, C enters the vehicle, and takes the seat of B.

• At time t4, C leaves the vehicle. A possible infection of X from C is computed, with t4 –t2 as the interaction time. At the same time, D enters the vehicle, and takes the seat of C.

• At time t5, X leaves the vehicle. A possible infection of X from A is computed, with t5 –t1 as interaction time, and from D, with t5 –t4 as interaction time.

In the algorithm, however, we do not explicitly model seating positions, but just assume that every synthetic person that leaves interacts with up to 3 randomly selected other agents.

The variation over time of the density of persons inside the container is currently not taken into account. This could, however, be done in a further modelling step, given data of facility and vehicle sizes. What is, however, taken into account is the thinning out of persons in the container when they are reducing their activity participation (see Sec. 2.5) no longer there: The algorithm still computes interaction with up to 3 randomly selected persons, but if these persons do not participate at that activity on that day, then there is no infection dynamics in either direction.

2.3 Multi-day modelling

Optimally, one would have multi-day trajectories. In our case, the data that we have ends at the end of the day. Our simulations thus run the same person trajectories again and again. This presumably underestimates mixing. However, there is still strong mixing because the synthetic persons interact with other persons at the same facility every day. For example, a train may have 1000 passengers. Out of these, any of our agents would only see a small number other agents. This leads to different encounters in every synthetic day, even when repeating the same trajectory over and over.

The same holds for facilities, where the data is actually constructed such that there are up to 400 persons per facility. That is, any work, shopping, or leisure facility can have up to 400 visitors at the same time, and every person that has that facility in its trajectory can interact with all of these persons. This number of interacting agents on each single day is, however, limited (see infections at facilities and in vehicles). This results in an issue for home locations: Our data model does not differentiate between “persons living in the same block” and “persons living in the same household”. For this reason we split persons living in the same block into realistic household sizes with a maximum number of six people per household [24].

As stated before, activities are of a certain type. These types are used in this research to determine the effects that certain strategies have on the epidemic spreading. If an activity is removed from a schedule, agents will still be put in the respective containers, but do not take part in the infections dynamics. Such agents essentially represent “holes”, since they may still be picked as one of the three agents to interact with, but there is no infection.

2.4 Base calibration

Most parameters of the model are taken from the literature, as explained above and in the additional material. The remaining free parameters are, from Eq. (1), Θ · q · ci · in. Since none of these numbers seem to be known for COVID-19, we have first set the base values of q = ci = in = 1. We keep these parameters separate, since they later allow the modelling of different contact intensities and of masks. The remaining Θ parameter was then calibrated so that in the base case, the number of cases doubles every three days. This is a plausible fit to the initial days which were totally without distancing measures, which is what our base case represents. Calibrating Θ in our model corresponds to calibrating β, the infection rate, in a compartmental model.

Another free parameter is when the first infection occurs. This corresponds to the initial number of infected persons, I_o, in the compartmental models. This date was set such that our number of persons showing symptoms, which corresponded to the sampling strategy in Germany during the early phase, is 50% above the reported case numbers. Evidently, this has to do with the fraction of non-reported cases, which is considered in Section 2.7. We also feed the simulation one additional infected case per day, since Berlin is embedded into a system where disease import is probable. The result of this base calibration can be seen in Fig. 5 (blue).

2.5 Reductions in activity participation

During the unfolding of the epidemics, people decided to no longer participate in certain activities. We model this by removing an activity from a person’s schedule, plus the travel to and from the activity. In consequence, that person no longer interacts with people at that activity location, and in consequence neither can infect other persons nor can become infected during that activity. Overall, this reduces contact options, and thus reduces epidemic spread.

A very important consequence of our modelling approach is that we can take that reduction in activity participation from data. Unfortunately, the activity type detection algorithm is not very good for these unusual activity patterns, as one can see in Fig. 3 when knowing that all educational institutions were closed in Berlin after Mar/15. What is reliable, though, is the differentiation between at-home and out-of-home time, as displayed in Fig. 4. One clearly notices that out-of-home activities are somewhat reduced after Mar/8, and dramatically reduced soon after. After some experimentation, it was decided to take weekly averages of the activity non-participation, and use that uniformly across all activity types in our model, except for educational activities, which were taken as ordered by the government.

• Download figure
• Open in new tab

Figure 3

Reduced activity participation over the course of the epidemics in Germany.

• Download figure
• Open in new tab

Figure 4

Change in activity participation compared to normal for normal workdays. All out-of-home activities are combined into one number. (*) denotes the first day of closures of schools, clubs, and bars; and (#) the first day of the so-called contact ban which came together with closures of all restaurants and non-essential stores. The data clearly shows that people had reduced their out-of-home activities before the government-ordered closures.

2.6 Calibration against case numbers in Berlin

The simulation is calibrated against the Berlin case numbers (Fig. 5, violet dots) [25]. Each reported case comes with two dates: reporting date, and so-called reference date, which refers to the onset of symptoms. Other than Dehning et al. [26], we use the reference date, which is closer to the real dynamics.

• Download figure
• Open in new tab

Figure 5

RKI case numbers [25] together with simulated result for different values of a, and final calibration result. The lines show the mean of 300 simulations; the error bands denote 5% and 95% quantiles. The calibration assumes that 1/3 of new infections remain undetected.

Since our simulation explicitly models the state of “showing symptoms”, we can directly compare this to the case numbers (the sampling strategy in Germany at that time was to not test persons without symptoms). Using the model as explained so far leads to the orange curve in Fig. 5. This is clearly too high. Multiplying the reductions in activity participation by a factor (green curve) does not help: This now gets the downward slope right, but the curve implies that the fraction of nonreported cases would be smaller before mid-march than afterwards. Since this is implausible, we set a back to one and introduce a behavioral change point earlier in the simulation, see the red curve. This is consistent with Dehning et al. [26], who report a first change point around Mar/7 which “matches the timing of the first governmental intervention which included cancellations of large events, as well as increased awareness”.

We model this by introducing a change of contact intensity ci, early in March, for all out-of-home activities, public transport, and quarantined-at-home. We calibrate the day and magnitude of the ci change by minimizing the Root Mean Squared Logarithmic Error (RMSLE) between detected cases and persons showing symptoms in the simulation adjusted by an estimate of 50% unrecorded cases; the two free parameters are the new value of ci, and the first day with new value (constrained to be between Mar/3 and Mar/9). As the objective function is expensive to evaluate, we chose the Tree-structured Parzen Estimator [27]. The red curve in Fig. 5 shows the final result of the calibration process with the ci change point on Mar/7 and ci = 0.32. Other than Dehning et al. [26], we do not need a second and third change point; the activity non-participation from our data is sufficient. For a comparison with hospital numbers see Sec. 2.7.

2.7 Comparison with hospital numbers in Berlin and resulting statement about undetected cases

Fig. 6 compares our simulation, with α = 1 and the additional change point on Mar/7, with reported hospital cases in Berlin [28]. As a part of the calibration process, the hospitalization probabilities of Tab. 1 are multiplied by 1.6. One observes that our model is somewhat “early” when compared to the Berlin hospital admissions. In reaction, one could shift the starting date of the infection process to later. This would improve the timing of the simulated hospitalizations, but the maximum would be lower, which could then be compensated for by a larger multiplier for the hospitalization probabilities. However, then the simulated cases showing symptoms would be lower than the reported case numbers, which seems implausible. Alternatively, one could change the progression model, and assume more time between showing symptoms and hospital admission. This would mean a deviation from the values reported in the literature, which we did not want to do.

• Download figure
• Open in new tab

Figure 6

Reported hospital cases [28] together with simulated results. The lines show the mean of 300 simulations; the error bands denote 5% and 95% quantiles.

Given current knowledge, this points to a rather little underreporting in Berlin - since more underreporting would mean an earlier date for the initial infection, which in turn would mean even earlier dates for the hospitalization numbers, contrary to what is observed. This is consistent with other studies [29]. Even with the assumption of little underreporting, however, the disease progression model from the literature is too “fast” when compared to the data in Berlin.

2.8 Reductions of the reinfection rate R

In the following, we investigate the contributions of certain non-pharmaceutical interventions towards reducing the reinfection rate R. This is obtained by running two different simulations, one without the measure and one with the measure, taking note of the corresponding reinfection numbers, dividing them by each other, and averaging. The following interventions are considered: (1) Certain types of activities are removed from the infection dynamics by removing these activity types from each synthetic person’s trajectory, as described above. Note that a 100% reduction of activity participation seems quite unrealistic except for schools; our mobility data (Fig. 4) implies that despite the perceived strength of the measures in Berlin, the drop was never more than 50%. (2) If activities are removed only with a certain probability, a random draw is made every time a synthetic person has that activity type in its plan. This means that the model assumes that, say for a 50% work reduction, it will be another 50% subset of persons at work every day. This intervention, in consequence, does not sever infection networks, but just slows down the dynamics. (3) Masks are added into the infection model of Eq.1 by reducing the shedding and/or the intake accordingly: cloth masks reduce shedding and intake to 0.6 and 0.5 of their original values, surgical masks to 0.3 and 0.3, and N95 (FFP) masks to 0.15 and 0.025 [30, 31]. Evidently, it is now possible in our model to assign masks to activity types, to persons, or both. (4) Contact tracing is implemented in the following way: If a synthetic person shows symptoms, then after a configurable delay d and with a configurable probability γ, contact persons of the last 14-days are put into quarantine at home. The parameter d parametrizes the delays caused by testing and by tracking down contact persons, the parameter γ the probability that a person cannot be contacted or that it does not follow the request to stay at home. Persons in the same household as the symptomatic person are always put into quarantine; contacts during shopping, while using public transport, or with contacts shorter than 15-minutes are ignored.

We determine consequences of interventions by administering them to the unrestricted dynamics (i.e. the blue curve of Fig. 5), on Mar/7. We first collect for each infected person the number of reinfections. These numbers are then averaged over all persons becoming contagious on a certain day; this corresponds to the “case or cohort reproductive number” method of [32]. These numbers are averaged 100 Monte Carlo runs with different random seeds. The resulting R values can be plotted over time, see Fig. 7. Evidently, the interventions are visible in these numbers before the interventions are implemented, because the reinfections are registered backwards in time to when the person became contagious. However, only from the day of the intervention onward will they show the full effect. To further reduce Monte Carlo noise, we average ρ =1 − R_intervention/R_base over the first 14 days after the introduction of the intervention. For these values, we then compute the mean , the standard deviation , and finally the standard error is the number of different simulation runs (Monte Carlo seeds). Evidently, the standard deviation gives information about the variability of the Monte Carlo runs, while the standard error gives information about the probability that the true mean is different from zero.

• Download figure
• Open in new tab

Figure 7

Illustration of determination of different reinfection rates (top) and their ratios (bottom) for two different measures: reduction of out-of-home work activities by 50% (left), and reduction of all out-of-home activities by 50% (right). Each curve is averaged over 100-different Monte Carlo runs. The results given in Table 1 are additionally averaged over the first 14 days after the intervention (red line in bottom figures).

3.1 Behavioral interpretation of government interventions vs mobility data

One striking consequence of the activity participation data (Fig. 4) is that, after the initial government intervention from Mar/7 that cancelled large events and raised awareness, the population reaction in fact preceded the government interventions, rather than the other way around. Out-of-home activity participation was already reduced between Mar/7 and Mar/14, before the second government intervention that closed schools, clubs, and bars. Similarly, there was a further considerable drop between Mar/14 and Mar/21, again before the so-called contact ban (strongly reduced interaction between different households) and closing of all restaurants and non-essential stores in Germany. At least for Berlin and probably for Germany, it is thus not true that the government forced society and the economy to come to a halt; rather, society did this by itself, and the government presumably stabilized or reinforced behavior that was happening anyways.

3.2 Reductions of the reinfection rate R

Tab. 2 shows the reductions of the reinfection rate R for different policy interventions, as computed by our approach. As simulation output, all of our values are significantly different from zero at the 3a level. One notices the strong reduction of R from reducing all out-of-home activities by 50% - this roughly corresponds to the situation in Germany at the end of March (cf. Fig. 4). We also find that this is a combination of many much smaller contributions: only the reductions of work or of leisure activities reach two digit ranges for the reduction of R.

Similarly, masks help, but a two digit impact can only be reached when wearing masks at work (where our measure is maybe better translated as “90% of people wear N95 masks at work or use single person offices”).

Also contact tracing followed by quarantine-at-home has a strong impact. Evidently, this assumes that the contact tracing system is not overwhelmed by large infection numbers. We also find that even a system with infinite tracing capacity results in a finite reduction of R. In consequence, if contact tracing is not sufficient, it needs to be supported by other measures.

Assuming that the initial changes in behavior from Mar/7 are changes that will remain stable over the epidemics and the remaining R is around 2, then still a combination of relatively many measures is needed to bring R to below one. Note that this needs to be done multiplicatively, i.e. where (a), (b),... are the different measures. If contact tracing is part of this package, then if the contact tracing system becomes overwhelmed, situation where R was maybe slightly above one will suddenly jump to one where R is about 30% to 40% larger. The result will be a strong acceleration of the exponential growth dynamics, and fairly strong measures will have to be taken to get R back to below one. We think that successful contact tracing is a plausible explanation why the situation is currently (middle of July) under control in Germany despite the relatively open regime compared to other countries.