Reducing societal impacts of SARS-CoV-2 interventions through subnational implementation

  1. Mark M Dekker
  2. Luc E Coffeng
  3. Frank P Pijpers
  4. Debabrata Panja
  5. Sake J de Vlas

  • Curated by eLife

    eLife logo

    eLife assessment

    This is a valuable contribution to the SARS-CoV-2 modelling literature that will be of interest to infectious disease modellers studying the impact of spatially heterogeneous interventions for transmission control. The calibration and analysis of the proposed model is sound and and the results provide convincing evidence that supports the claim that localised interventions could potentially reduce societal impact while maintaining outbreak control. However, the paper provides little insight into what drives the regional diffusion in the Netherlands and how that diffusion could be affected by local lockdowns and a more thorough exploration of the model is warranted. There is also an opportunity to consider behavioural consequences, feasibility, and potential ethical implications of the proposed approach in greater depth.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

To curb the initial spread of SARS-CoV-2, many countries relied on nation-wide implementation of non-pharmaceutical intervention measures, resulting in substantial socio-economic impacts. Potentially, subnational implementations might have had less of a societal impact, but comparable epidemiological impact. Here, using the first COVID-19 wave in the Netherlands as a case in point, we address this issue by developing a high-resolution analysis framework that uses a demographically stratified population and a spatially explicit, dynamic, individual contact-pattern based epidemiology, calibrated to hospital admissions data and mobility trends extracted from mobile phone signals and Google. We demonstrate how a subnational approach could achieve similar level of epidemiological control in terms of hospital admissions, while some parts of the country could stay open for a longer period. Our framework is exportable to other countries and settings, and may be used to develop policies on subnational approach as a better strategic choice for controlling future epidemics.

Article activity feed

  1. Version published to 10.7554/elife.80819 on eLife
  2. eLife

    Author Response

    Reviewer #2 (Public Review):

    This manuscript presents a rather technical modelling analysis of the impact of local lockdowns on Covid-19 hospitalisations in the Netherlands. The major strength of the study is that the authors attempt to calibrate their model to a novel data source, a commercial database of mobility patterns between municipalities. The major weakness is that the model seems overly complicated, many parameters seem to have been 'guessed' without a formal uncertainty analysis, e.g. within a Bayesian framework, so that it is impossible to judge how robust the results and therefore the conclusions are.

    Major points:

    1. In some aspects the structure of the model presented seems overly complicated: It is not clear why the authors chose the 1:100 population scale and why they didn't go directly for modelling the full population. Artificially reducing the population size has important stochastic effects at the early phase of the epidemic. Also it is not clear what it means when 1:100 of one municipality mixes with 1:100 of another municipality? The authors should at least attempt to see what impact this has on output, i.e. conduct a sensitivity analysis.

    The reason for choosing a 1:100 population scale instead of the full population is computational speed. Indeed, this (and its consequences) is not mentioned explicitly and will be added. Moreover, to identify the sensitivity of the results to population scale, we add runs on different population scales.

    • Added reasoning and consequences associated with the 1:100 population scale in SI C.1.

    • The sensitivity of the results to population scale is now discussed in SI C.1 using runs with other population resolutions.

    1. On the other hand the model goes into (too) much detail regarding mixing behaviour and attempts to model processes during each hour of the day. This does not seem to be informed by actual data, but the data seems to be made up e.g. as in A.6. As an ex-student and a father of a teenager I can tell you that the susceptibility profile guessed in Table 3 does not seem to be very realistic. As it is stated in the appendix, the Mezuro data set only provides daily averages of travelling between communitities, so it is not clear why the hourly resolution is actually needed in the model.

    Indeed, several aspects in the model are informed by “secondary statistics” which unfortunately add uncertainty. An example would be the normalization of the mobility matrices by means of data on how people spend their time (see SI A.3). Note that the example of the susceptibility profile that the reviewer mentions, however, does not involve such secondary statistics and happens to be exactly reported by the Dutch health agency (cited in SI A.5).

    We agree that the model includes much detail, which potentially has weaknesses as the reviewer rightfully mentions. However, one of the main points of this paper is that in order to address the questions of local interventions, geographical spread and associated hospital admissions, we simply need this level of detail, or even higher. In other words, assessments of such mechanics would be even more uncertain if this level of detail is not included.

    We agree that the motivation for hourly resolution is not well described in the manuscript – this will be added. The reasoning is that mixing of the population is highly heterogeneous throughout the day: clearly, seen in Fig. S5 (SI A.7), mixing at work is fully different from mixing at school or at home.

    Moreover, people meet at work in different municipalities and then return to home to potentially spread the disease further. It is exactly such mechanics that we are after in our analysis.

    • Added a more in-depth discussion of the mobility data in SI C.2.

    • Added the motivation for hourly resolution in SI A.1-A.3.

    1. It is not clear why the authors rely on only one short period of the Mezuro data set in March 2019 and not investigate the same data source during the actual lockdown in 2020, or even for the full year, as travelling is likely to be very season dependent. This would provide much better estimates of the effects of lockdown on travel patterns. The analysis presented and categorisation into frequent, regular and incidental also need further explanation. It is not clear how international travel is accounted for in the mobility data.

    The reviewer is correct that using a longer mobility dataset or one that is exactly addressing the period of the actual lockdown would be beneficial. The reason we are not doing so is simply that this data is not available.

    The model accounts for international travel by means of its initialization, but not further. In practise, international travel got severely reduced throughout this period. Hence, we deem the uncertainties associated with not accounting for international travel limited.

    • Added a discussion on the effect of using this mobility dataset in SI C.2. • Added a further explanation of categorizing the movements (in SI C .2).

    • Added a discussion on international travel in SI C.2.

    1. Beyond the technical points on the modelling, the main hypothesis of whether local lockdowns may work has also not been sufficiently discussed outside of the Dutch context. The authors fail to mention that this was the approach chosen in Northern Italy at the start of the epidemic (https://en.wikipedia.org/wiki/COVID-19_lockdowns_in_Italy) where it didn't work, as we all know. On the other hand, more recent local lockdowns in China appeared to be successful, albeit at a great societal cost in terms of restrictions to freedom (https://en.wikipedia.org/wiki/COVID19_lockdown_in_China).

    The reviewer is correct that we only show this in the Dutch context. We can reason about other situations, but clearly these situations differ vastly from country to country.

    Reviewer #3 (Public Review):

    This work uses an agent-based model of SARS-CoV-2 transmission (calibrated to the first wave in the Netherlands) to examine how the societal impact of interventions could have been reduced - while maintaining epidemiological impact - if they were implemented at a subnational (eg, municipality) rather than a national level. After more than two years of lockdowns and mobility restrictions, the societal cost of such measures is becoming better understood, and it is important to evaluate the effectiveness of such measures and reflect upon how they can be deployed in a minimally disruptive fashion. Mathematical and computational models are a natural choice for such investigations as they enable researchers to explore counter-factual scenarios ("what might have happened had we acted differently?")

    The authors conclude that subnational interventions, triggered via prevalence in a particular municipality, could have controlled the first wave of SARS-CoV-2 in the Netherlands with minimal health cost but less societal disruption than national interventions. This claim is supported by reference to Figure 4 showing the impact on (a) hospital admissions and (b) municipalities without interventions through different phases of the outbreak. For more remote/rural municipalities, the use of interventions is delayed by ~1 week, although some (6%) of municipalities avoid interventions altogether.

    Strengths:

    As noted above, the general objective of this study is important and of potentially broad interest. The agent-based model is complex, but not unreasonably so, and makes good use of rich demographic, mobility, epidemiological/clinical, etc. data for calibration. The simulations conducted using the model support the specific conclusions of the manuscript.

    Weaknesses:

    While the motivation and approach are strong points of this work, the analysis and interpretation would benefit from further development. The robustness of model behaviour to the threshold used to trigger subnational interventions is explored; however, there are other aspects of the model that are not subjected to sensitivity analysis, including:

    1. The impact of imperfect surveillance (eg, due to asymptomatic transmission, reporting delays, etc);
    1. The impact of non-compliance, which could potentially differ for subnational versus national interventions;
    1. The impact of pathogens/variants with transmission/severity characteristics different from the original SARS-CoV-2 strain.

    In the absence of such analyses, it is difficult to generalise the findings beyond "this is how subnational interventions could have been used to control the first wave of SARS-CoV-2 in the Netherlands" to "this is how subnational interventions could be used effectively in the event of future outbreaks" (of a SARS-CoV-2 variant or other pathogen).

    The discussion focuses on limitations associated with the model but does not consider other potential implications of subnational interventions. For example:

    1. Subnational interventions may produce unintended consequences if populations respond by relocating from regions with interventions (high prevalence) to regions without interventions (low prevalence).
    1. Subnational interventions would require extremely effective public health messaging to avoid confusing populations. Particularly in densely populated regions where municipalities may be small and tightly connected, the feasibility of communicating (and enforcing compliance with) interventions may be challenging.
    1. A proposal to implement subnational interventions - following the results of this work - may raise ethical questions about cost-benefit trade-offs (eg, whether 355 additional hospital admissions is an acceptable price to pay for 36 million person-days without interventions; ie, two days per citizen, on average). The fact that such decisions would (in the even of a future outbreak) need to be made rapidly, in the face of potential uncertainty about pathogen characteristics, heightens the need for clear understanding of how situational factors may affect the likely effectiveness of interventions (at any scale).

    Impact and broader utility:

    As noted, the question addressed - how we can reduce the disruption caused by interventions for transmission control - is important. Thus, the work presented in this manuscript has the potential for broad utility. Currently, this is limited by the focus on specific outbreak instance.

    In general terms, we agree with the reviewer. That said, the “possibility space” of policymaking is infinite dimensional, in the sense that the intervention measures can take an infinitely many forms, starting times and durations. The framework that we have built upon combining data sources such as demography, mobility, interactions and disease parameters now makes it possible to explore these possibilities. These will be explored in future work.

  3. eLife

    eLife assessment

    This is a valuable contribution to the SARS-CoV-2 modelling literature that will be of interest to infectious disease modellers studying the impact of spatially heterogeneous interventions for transmission control. The calibration and analysis of the proposed model is sound and and the results provide convincing evidence that supports the claim that localised interventions could potentially reduce societal impact while maintaining outbreak control. However, the paper provides little insight into what drives the regional diffusion in the Netherlands and how that diffusion could be affected by local lockdowns and a more thorough exploration of the model is warranted. There is also an opportunity to consider behavioural consequences, feasibility, and potential ethical implications of the proposed approach in greater depth.

  4. eLife

    Reviewer #1 (Public Review):

    This study will be of interest to those who want to understand how non-pharmaceutical interventions in response to epidemics in human populations (here using the example of SARS-CoV-2 in the Netherlands) might be designed to reduce negative societal impacts through subnational implementation. In most countries, including the Netherlands, non-pharmaceutical interventions such as movement restrictions and school closures were implemented simultaneously throughout entire jurisdictions. The authors of this paper investigated whether subnational heterogeneities in the prevalence of infection could be exploited to develop local level control measures that varied according to local changes in prevalence of infection, an innovation that could potentially reduce negative societal impacts of interventions while maintaining similar levels of epidemiological control. Using simulations from a carefully parameterised agent based model of SARS-CoV-2 transmission in the first wave in Netherlands , the authors generated convincing evidence to suggest that this would be, at least in theory, possible, though practical difficulties of implementing such a control policy were not explored.

  5. eLife

    Reviewer #2 (Public Review):

    This manuscript presents a rather technical modelling analysis of the impact of local lockdowns on Covid-19 hospitalisations in the Netherlands. The major strength of the study is that the authors attempt to calibrate their model to a novel data source, a commercial database of mobility patterns between municipalities. The major weakness is that the model seems overly complicated, many parameters seem to have been 'guessed' without a formal uncertainty analysis, e.g. within a Bayesian framework, so that it is impossible to judge how robust the results and therefore the conclusions are.

    Major points:

    1. In some aspects the structure of the model presented seems overly complicated: It is not clear why the authors chose the 1:100 population scale and why they didn't go directly for modelling the full population. Artificially reducing the population size has important stochastic effects at the early phase of the epidemic. Also it is not clear what it means when 1:100 of one municipality mixes with 1:100 of another municipality? The authors should at least attempt to see what impact this has on output, i.e. conduct a sensitivity analysis.

    2. On the other hand the model goes into (too) much detail regarding mixing behaviour and attempts to model processes during each hour of the day. This does not seem to be informed by actual data, but the data seems to be made up e.g. as in A.6. As an ex-student and a father of a teenager I can tell you that the susceptibility profile guessed in Table 3 does not seem to be very realistic. As it is stated in the appendix, the Mezuro data set only provides daily averages of travelling between communitities, so it is not clear why the hourly resolution is actually needed in the model.

    3. It is not clear why the authors rely on only one short period of the Mezuro data set in March 2019 and not investigate the same data source during the actual lockdown in 2020, or even for the full year, as travelling is likely to be very season dependent. This would provide much better estimates of the effects of lockdown on travel patterns. The analysis presented and categorisation into frequent, regular and incidental also need further explanation. It is not clear how international travel is accounted for in the mobility data.

    4. Beyond the technical points on the modelling, the main hypothesis of whether local lockdowns may work has also not been sufficiently discussed outside of the Dutch context. The authors fail to mention that this was the approach chosen in Northern Italy at the start of the epidemic (https://en.wikipedia.org/wiki/COVID-19_lockdowns_in_Italy) where it didn't work, as we all know. On the other hand, more recent local lockdowns in China appeared to be successful, albeit at a great societal cost in terms of restrictions to freedom (https://en.wikipedia.org/wiki/COVID-19_lockdown_in_China).

  6. eLife

    Reviewer #3 (Public Review):

    This work uses an agent-based model of SARS-CoV-2 transmission (calibrated to the first wave in the Netherlands) to examine how the societal impact of interventions could have been reduced - while maintaining epidemiological impact - if they were implemented at a subnational (eg, municipality) rather than a national level. After more than two years of lockdowns and mobility restrictions, the societal cost of such measures is becoming better understood, and it is important to evaluate the effectiveness of such measures and reflect upon how they can be deployed in a minimally disruptive fashion. Mathematical and computational models are a natural choice for such investigations as they enable researchers to explore counter-factual scenarios ("what might have happened had we acted differently?")

    The authors conclude that subnational interventions, triggered via prevalence in a particular municipality, could have controlled the first wave of SARS-CoV-2 in the Netherlands with minimal health cost but less societal disruption than national interventions. This claim is supported by reference to Figure 4 showing the impact on (a) hospital admissions and (b) municipalities without interventions through different phases of the outbreak. For more remote/rural municipalities, the use of interventions is delayed by ~1 week, although some (6%) of municipalities avoid interventions altogether.

    Strengths:

    As noted above, the general objective of this study is important and of potentially broad interest. The agent-based model is complex, but not unreasonably so, and makes good use of rich demographic, mobility, epidemiological/clinical, etc. data for calibration. The simulations conducted using the model support the specific conclusions of the manuscript.

    Weaknesses:

    While the motivation and approach are strong points of this work, the analysis and interpretation would benefit from further development. The robustness of model behaviour to the threshold used to trigger subnational interventions is explored; however, there are other aspects of the model that are not subjected to sensitivity analysis, including:

    1. The impact of imperfect surveillance (eg, due to asymptomatic transmission, reporting delays, etc);

    2. The impact of non-compliance, which could potentially differ for subnational versus national interventions;

    3. The impact of pathogens/variants with transmission/severity characteristics different from the original SARS-CoV-2 strain.

    In the absence of such analyses, it is difficult to generalise the findings beyond "this is how subnational interventions could have been used to control the first wave of SARS-CoV-2 in the Netherlands" to "this is how subnational interventions could be used effectively in the event of future outbreaks" (of a SARS-CoV-2 variant or other pathogen).

    The discussion focuses on limitations associated with the model but does not consider other potential implications of subnational interventions. For example:

    1. Subnational interventions may produce unintended consequences if populations respond by relocating from regions with interventions (high prevalence) to regions without interventions (low prevalence).

    2. Subnational interventions would require extremely effective public health messaging to avoid confusing populations. Particularly in densely populated regions where municipalities may be small and tightly connected, the feasibility of communicating (and enforcing compliance with) interventions may be challenging.

    3. A proposal to implement subnational interventions - following the results of this work - may raise ethical questions about cost-benefit trade-offs (eg, whether 355 additional hospital admissions is an acceptable price to pay for 36 million person-days without interventions; ie, two days per citizen, on average). The fact that such decisions would (in the even of a future outbreak) need to be made rapidly, in the face of potential uncertainty about pathogen characteristics, heightens the need for clear understanding of how situational factors may affect the likely effectiveness of interventions (at any scale).

    Impact and broader utility:

    As noted, the question addressed - how we can reduce the disruption caused by interventions for transmission control - is important. Thus, the work presented in this manuscript has the potential for broad utility. Currently, this is limited by the focus on specific outbreak instance.

  7. Version published to 10.1101/2022.03.31.22273222v2 on medRxiv
  8. ScreenIT

    SciScore for 10.1101/2022.03.31.22273222: (What is this?)

    Please note, not all rigor criteria are appropriate for all manuscripts.

    Table 1: Rigor

    NIH rigor criteria are not applicable to paper type.

    Table 2: Resources

    Software and Algorithms
    SentencesResources
    First, we reduce inter-municipality mobility as reported by Google [34] in the four phases of the first wave in the Netherlands.
    Google
    suggested: (Google, RRID:SCR_017097)

    Results from OddPub: We did not detect open data. We also did not detect open code. Researchers are encouraged to share open data when possible (see Nature blog).

    Results from LimitationRecognizer: We detected the following sentences addressing limitations in the study:
    Another limitation is that mobility in our framework is quantified based on mobile phone signal data that only provide anonymized movements between pairs of locations. As such, the data do not provide identifiers to link multiple movements into one itinerary, which means that in our analysis, agent movements are somewhat shorter on average than in reality, but agents also visit more different locations than in reality. We further assume that agent movements vary randomly day-by-day, whereas in reality commuting means that an agent would repeatedly travel to the same location. However, the impact of this simplifying assumption is limited as, at the start of an epidemic, the distribution of movement over agents is of relatively low importance, especially in the case of a relatively small and highly connected country as the Netherlands. This is in contrast to situations towards the tail of an epidemic or in larger geographies (e.g., Brazil [24, 25] and India [26]), where the transmission potential of “high-mobility corridors” can eventually dry up as a result of rising immunity among high-mobility individuals. Finally, we adopted data on national patterns in mobility (Google mobility), meaning that it was not possible to account for changes in mobility by geography or demographic group. The geographical aspects could be addressed by using longitudinal mobile phone signal data or individual-level self-reported data via mobile phone apps [27–29]. This would require that such data ...

    Results from TrialIdentifier: No clinical trial numbers were referenced.

    Results from Barzooka: We did not find any issues relating to the usage of bar graphs.

    Results from JetFighter: We did not find any issues relating to colormaps.

    Results from rtransparent:
    • Thank you for including a conflict of interest statement. Authors are encouraged to include this statement when submitting to a journal.
    • Thank you for including a funding statement. Authors are encouraged to include this statement when submitting to a journal.
    • No protocol registration statement was detected.

    Results from scite Reference Check: We found no unreliable references.

    About SciScore

    SciScore is an automated tool that is designed to assist expert reviewers by finding and presenting formulaic information scattered throughout a paper in a standard, easy to digest format. SciScore checks for the presence and correctness of RRIDs (research resource identifiers), and for rigor criteria such as sex and investigator blinding. For details on the theoretical underpinning of rigor criteria and the tools shown here, including references cited, please follow this link.

  9. Version published to 10.1101/2022.03.31.22273222v1 on medRxiv