Regional policy coordination of pandemic responses using an iterative mobility-driven algorithm

ABSTRACT The COVID-19 experience has shown that horizontal cooperation of non-pharmaceutical interventions across jurisdictions is crucial to combat pandemics. However, the question of how to construct policy coordination regions has not yet received enough attention. In this study, we develop an iterative mobility-driven community detection algorithm based on the modularity function to identify multilevel public policy coordination regions. As a case study, we use movement of people in the United States to identify regions of interconnected locations at different levels. We argue that pre-emptively designed community structures based on mobility can be more appropriate to meet critical preparedness goals than existing jurisdictions.


INTRODUCTION
The COVID-19 pandemic has caused widespread loss of lives and economic activity across the world since early 2020 (Cutler & Summers, 2020;International Monetary Fund (IMF), 2021; World Health Organization (WHO), 2021).Before the adoption of effective therapeutics and/ or mass vaccination, the best measure to reduce and control the spread of the disease has been non-pharmaceutical interventions (NPIs) that limit social contact through closures, mask mandates, etc. (Chu et al., 2020;Courtemanche et al., 2020;Flaxman et al., 2020).NPIs are expected to play an important role even when high vaccination rates are achieved due to potential virus mutation, as evidenced by the spread of the Delta variant across the world in 2021 and the subsequent variants such as Omicron.The lack of vaccines in many countries and vaccine hesitancy, coupled with inevitable delays in understanding new variants' impact on transmission and vaccine efficacy, further highlight the importance of using NPIs along with any available pharmaceutical interventions.Given that there is little appetite for broad-based, nationwide lockdowns in most countries, it is more effective to have a targeted regional NPI policy based on transmission patterns.
In some cases, NPIs have been successful, as evidenced by the significant drop in incidence in countries such as Australia (Zachreson et al., 2021), China (Kraemer et al., 2020) and New Zealand (Jefferies et al., 2020), where social distancing was extensively imposed and strictly enforced for significant periods of time.In each of these settings, NPIs have been tailored to target human mobility for mitigation, suppression and even elimination of COVID-19.Such successes in combating the pandemic are not universal, however.The United States, for instance, has experienced high rates of infection since March 2020 with multiple surges, despite NPIs imposed by states, counties and even municipalities.Aside from political factors, one of the plausible explanations for this general lack of success is the absence of public health policy coordination.Studies have shown that variations in the levels or timing of NPIs among neighbouring jurisdictions can lead to the spillover of spatial movements and spread of infections across boundaries (Brulliard & Weiner, 2020;Elenev et al., 2021;Holtz et al., 2020).This phenomenon is not limited to the United States.Italy became a hotspot early in the pandemic where the worst-hit areas were not only those densely populated, but also the most globally connected (Bourdin et al., 2021).Ruktanonchai et al. (2020) examine movement data across European countries and report that community transmission of COVID-19 can accelerate if countries with high interconnections fail to coordinate their NPIs.While it has become clear that collaboration in terms of NPIs is critical, the question of how to define policy coordination regions has not yet received enough attention.In the UK, coordination across colour-coded local tiers or regions constructed based largely on transmission intensity alone may not have helped much in controlling the outbreak (Laydon et al., 2021).Identifying policy coordination regions based on ex-post epidemiological criteria, no matter how objectively determined they may be, forces public health response to play catch-up with waves of infections across areas to break the chain of transmission.A more robust coordinated policy response to pandemics would be grounded in contagious disease preparedness.
In a recent editorial, Bailey et al. (2020) suggest that researchers take a 'safe distance' from the pandemic itself by examining what was measured before COVID-19.
Here we argue that ex-ante measures such as the pre-pandemic movement of people may be more appropriate to inform policymakers in identifying policy coordination regions well before an outbreak occurs.Recent COVID-19 outbreaks have highlighted the fact that pandemics can strike across borders in a way that leaves local governments unprepared as to which other jurisdictions it should have a coordinated policy response with (Dzigbede et al., 2020).A combination of national governance along with decentralised regional and local governance may be more effective in combating COVID-19 (McCann et al., 2022).Therefore, the critical first step is to identify appropriate local and regional jurisdictions that need to coordinate.Since viruses know no boundaries, people's movement, and thus the ex-ante rate of contact, drives the spread.As such, the governments of those localities whose residents tend to have high incidents of contact as established at a pre-pandemic baseline should be prepared to impose coordinated NPIs as soon as an infectious disease becomes a threat.
In this study, we develop an iterative mobility-driven community detection algorithm to identify multilevel coordination regions.The algorithm establishes a network of locations based on mobility data, and the modularity maximisation model is used to identify multilevel interjurisdictional regions.We use pre-pandemic movement of people in the United States, as captured by smartphone data, to identify regions of interconnected locations.Since people move relatively more frequently within than they do across these regions, such a spatial structure can guide the local governments within the same regions to coordinate their NPIs.This study extends the research stream of policy coordination and integration by providing a replicable and straightforward methodology for policymakers and practitioners to identify those local jurisdictions to coordinate with in addressing transboundary public health crises.
The remainder of the paper is structured as follows.In the next section, we provide the background, highlighting the need for valid methodologies to identify partner local regions at different levels to collaborate on policies in the context of a pandemic.We then discuss community detection specifically for this purpose.This is followed by a description of the dataset and the methodology.Our results show that the proposed multilevel regions have more consistent community structure with regard to movements than existing ad-hoc administrative borders.Further, we find that our proposed regional structure exhibits lower levels of variation as well as higher levels of spatial and temporal synchrony in COVID-19 incidence and, therefore, provides a more appropriate basis for pandemic NPI coordination, compared with the existing administrative state borders.Finally, we discuss our contribution to the literature and the practical implications for policymakers.

BACKGROUND
While the debate about optimal pandemic management continues, it has become clear that an effective public health response requires policy coordination across agencies and levels of government (Holtz et al., 2020).For instance, in the United States, public health departments are often organised within city, county and state jurisdictional borders.While this system works well in the case of a localised disease outbreak or natural disaster, coordination challenges in large-scale epidemics may occur as 'many countries have experienced coordination challenges between national and subnational governments' (Organisation for Economic Co-operation and Development (OECD), 2020).The idea of ad hoc jurisdictions is difficult to reconcile with infectious disease epidemiology: a contagious virus can easily travel across such arbitrary borders, as people move around for work commuting, shopping or recreational purposes.In fact, many studies have already established a strong link between human mobility and the spread of the SARS-CoV-2 virus (e.g., Zhang et al., 2022), and daily movement data have been used to explain changes in COVID-19 cases and deaths in the United States (Brinkman & Mangum, 2022;Tokey, 2021;Yilmazkuday, 2020).
Recently, Laroze et al. ( 2021) investigated spatial contagion across local authority districts in England between March and June 2020 using data on historical work commuter flows.They emphasise the difference between 'spatial clustering' and 'spatial dependence', which is akin to the distinction between correlation and causation.While spatial clustering refers to some geographical areas displaying similarities in disease outcomes, spatial dependence means that the outcome in one geographical area depends on the outcome or the determinants of the outcome in other areas.Laroze et al. present empirical evidence on how regional units spatially depended on each other during the first wave of COVID-19 in England.Following up on this idea, we argue that identification of communities based on the fundamental dynamics of human-to-human transmission is critical for contagious disease preparedness, surveillance and response efforts.Ansell et al. (2010) identify two main coordination challenges in a transboundary health crisis: vertical or intersectoral coordination; and horizontal or interjurisdictional coordination.Much of the existing literature explores the former, which points to the need for collaboration across a myriad of government agencies at different levels, along with engagement of private and non-profit organisations (Dzigbede et al., 2020).In this article, we instead focus on horizontal or interjurisdictional coordination and address the need for valid methodologies to identify which jurisdictions should be prepared to collaborate well before a transboundary health crisis hits.
Examining ex-ante mobility patterns can help outline maps of potential viral transmission across jurisdictions and geographical areas and allow better healthcare and economic resource alignment.For example, in March 2020, New York City in particular and New York State in general became an early COVID-19 epicentre, prompting the governor to impose a strict lockdown of non-essential businesses to slow the contagion.Recently, however, Birge et al. (2022) show that the zip codes with the most infections did not necessarily overlap with those areas of highest commercial activities.They conclude that the economic cost of the uniform lockdown in the New York metro area could have been significantly reduced by implementing spatially differentiated lockdowns and strategically tailoring mobility restrictions for specific neighbourhoods.Similarly, Fajgelbaum et al. (2021) show that economic activity and potential for virus spread is not uniform in space, and spatially differentiated lockdowns achieve substantially smaller income losses compared with blanket lockdown mandates.Tailoring NPIs in a spatial structure can be achieved by examining the baseline mobility patterns of the population that exist before the pandemic, as captured by smartphone data (Holtz et al., 2020).In so doing, the contagion can be curbed without imposing stringent measures on a large scale, thus allowing policymakers to contain the outbreak without wrecking the economy of entire regions.
Studies have shown that existing jurisdictional boundaries can even become a hinderance to intergovernmental cooperation and information sharing (e.g., Ki et al., 2020).The idea of redesigning pre-existing jurisdictions into more effective and efficient regional configurations is not new either; Kong et al. (2010) call this an 'optimal region design problem' and show that liver transplant efficiency in the United States can be significantly improved by re-clustering geographical areas under the existing ad hoc organ allocation regions into an optimal set of regions.Here, we claim that the first step for successful coordination in the context of combating pandemics is to identify the appropriate local jurisdictions to cooperate with based on ex-ante metrics.We posit that multilevel regions of horizontal cooperation can be pre-emptively designed according to how a contagious disease propagates through human interaction and movements, which by their nature command and demand certain patterns of reciprocity.To this end, we use ex-ante mobility data to identify coordination regions within which a highly contagious disease such as COVID-19 can spread easily.
The use of mobility data to identify coordination regions has been promoted in recent studies.For example, Monte (2020) considers an algorithm that connects counties under a given coordination region when mobility between them is above an arbitrary threshold.Ruktanonchai et al. (2020, suppl. material) apply a random walk-based algorithm, known as Walktrap, to detect communities based on the movement among subdivisions of countries.Baghersad et al. (2023) propose a heuristic and a column-generation algorithm to improve the computational efficiency in detecting communities.Although these studies highlight the idea of using mobility to identify coordination regions, they do not provide iterative heuristic algorithms that guarantee finding highly interconnected communities at multiple levels.We extend the existing literature by designing an iterative mobility-driven community detection algorithm that aims to identify optimal communities based on the modularity maximisation function.Our algorithm ensures that the constructed coordination regions are highly interconnected on the basis of spatial movement patterns.Furthermore, the proposed iterative algorithm, as described in the following section, identifies coordination regions at different geographical levels, which subsequently may be used to implement multilevel targeted NPIs.

METHODS AND DATA
Two broad categories of methods are commonly used to classify objects (such as events or locations) into homogeneous and distinct groups.The first, known as cluster analysis, group observations based on similarities and/or differences of attributes.For example, k-means, a nonhierarchical clustering method, aims to group objects into k distinct groups while minimising the squared Euclidean distances between objects and their clusters' centroids.Recently, Abdullah et al. (2022) use k-means clustering to group provinces in Indonesia based on the COVID-19 confirmed, recovered and death cases and identify three groups of provinces with different COVID-19 risks.Another approach is to group adjacent neighbours by investigating the existence of spatial autocorrelation using measures such as Moran's I.Such spatial clustering methods have been used to find COVID-19 hotspots (Fonseca-Rodríguez et al., 2021;Kang et al., 2020;Kim & Castro, 2020).In the context of pandemics, these cluster analysis methods can identify the spread of disease based on ex-post measures of morbidity and mortality.
While cluster analysis groups objects based on their similarity, the second categorycommunity detection methodsconsiders the interactions between objects.These algorithms aim to identify communities based on the connectivity or interactions of nodes in a network of objects.The community detection problem is commonly encountered in many distinct fields, from biology for identifying protein complexes and segregated subnetwork in the human brain to community health partnerships involving cross-sector collaboration among diverse organisations (e.g., Nepusz et al., 2012;Wig, 2017).Since the contact rate is a key factor in communicable disease transmission (Buckee et al., 2020), such community detection algorithms can identify coordination regions based on ex-ante movement of individuals before an outbreak.
In the past few decades, different techniques have been developed to identify potential subnetworks of a network, such as Walktrap (Pons & Latapy, 2006) and Louvain (Blondel et al., 2008), to find the optimum set of communities according to predefined quality functions.Modularity, introduced by Newman and Girvan ( 2004), is the most commonly used quality function for finding the communities with high interconnections and has been successfully applied in different disciplines, such as social science, computer science, marketing and economics (e.g., Fortunato & Hric, 2016;Good et al., 2010).Intuitively, modularity measures the strength of a community structure in a network: high modularity implies that nodes are highly connected within modules and sparsely connected between modules.For a weighted network, the modularity function calculates the difference between observed connectivity among constructed communities in the actual and a random network, defined by: where A ij is the weight of edge between nodes i and j, k i = j A ij , m = 1/2 i,j A ij ; and X i,j is a binary variable equal to 1 if nodes i and j belong to the same community, and 0 otherwise.Modularity can take values between −0.5 and 1, with higher values indicating better clustering or a more defined community structure.Modularity value is 0 when the number of within-community edges is no better than random, while in practice values in the 0.3-0.7 range indicate a strong community structure (Newman & Girvan, 2004).

Iterative community detection algorithm
NPIs are targeted strategies to slow the spread of viruses in affected regions.These policies can be applied at different geographical levels, such as county, state or nation, depending on the nature of each specific policy.This study uses an iterative mobility-driven community detection algorithm based on the modularity function to identify coordination regions at different geographical levels.
Figure 1 summarises the developed iterative algorithm, and the steps are described below.

Step 1: Establish a global mobility network
The iterative algorithm starts with establishing a network of locations using mobility data.Locations in the network are connected based on the real movement of people.In this study, following the digital epidemiology approach (Salathé et al., 2012), we use smartphone data to track movement of individuals in a completely anonymous fashion.In addition to being de-identified, the mobility data were accessed through remote secure access and big data query under a data usage agreement.Camber Systems, affiliated with the COVID-19 Mobility Data Network formed by a group of infectious disease epidemiologists at universities around the world, 1 makes these data publicly available for research. 2In this dataset, the movement of multiple mobile devices is aggregated across space and time to reflect an approximation of population-level mobility, instead of using individual travel or behavioural patterns (Maas et al., 2019).Limitations of using mobile phone data to understand infectious disease transmission and dynamics are discussed in the last section.
This study focuses on between-county movements in the United States.To capture the baseline movement of people before the impact of ex-post pandemic restrictions, we use mobility data in the eight-week period (20 January-15 March 2020) before a national emergency was declared and movements were affected by varying NPIs.The number of unique movements from one county to another is captured in 4-hour blocks (six blocks per day).This mobility dataset includes activity date, start of the 4-hour block, Federal Information Processing Series (FIPS) codes of from and to counties, and number of movements (transitions), yielding 43,340 records and more than 500 million between-county movements.An example is presented in Table 1.
Next, we calculated the total number of movements between any two counties during the eight-week period to arrive at a dataset containing 3107 unique counties (including Washington, DC and excluding US territories) and 24,639 unique, undirected between-county links.The distribution of the number of movements is significantly skewed to the right (mean ¼ 22,560, median ¼ 233, SD ¼ 121,195).Well-defined subnetworks in the form of strong connections between some counties allow the identification of regions, despite most relationships being weak.Table 2 displays the overall top-20 between-county and interstate movements and their share in overall movements.

Step 2: Establish the highest-level coordination regions
After establishing the mobility network, the first level of coordination regions is created by solving the modularity maximisation problem (equation 1).Modularity optimisation is a hard-computational problem, and an approximation method is often required to find optimal or nearoptimal results in a reasonable time (Blondel et al., 2008).We apply an efficient heuristic method developed by Blondel et al. (2008), known as the Louvain method, to identify communities of highly interconnected counties.This method has two phases (see Appendix A in the supplemental data online for its pseudo-code).In phase 1, each node is assigned to its own community.Then, for each node i, gain in modularity is calculated for joining i to the neighbouring node j into a community C using 396 Milad Baghersad et al.

REGIONAL STUDIES
the formula: where in is the total weights of edges in C; tot is the total weights of edges from other nodes to nodes in C; and k i,in is the total weight of the edges from node i to nodes in C.Then, node i is placed into the community with the maximum gain of modularity, if at least one gain of modularity is positive.This process is repeated for all nodes until no improvement in modularity can be achieved.
The second phase starts with creating a new network using communities found in the first phase as new nodes.The weights between new nodes are equal to the sum of the weight of the edges between (original) nodes in two communities.The process in phase 1 is then Regional policy coordination of pandemic responses using an iterative mobility-driven algorithm 397 REGIONAL STUDIES reapplied to the new network created in the second phase.
The phases are iterated until no further improvement is attained.The output of step 2 is the upper-level coordination regions (ULCRs).
Once the structure of coordination regions is established, if smaller regions are needed for more targeted interventions, then the algorithm moves to step 4. Otherwise, the algorithm continues to step 5.

Step 4: Establish the next-level coordination regions
To find smaller communities, the modularity maximisation problem is applied within each existing coordination region, where the Louvain method (as described in step 2) is applied to identify communities of highly interconnected counties.The output of step 4 is the lower-level coordination regions (LLCRs) within each established region, and the algorithm returns to step 3.

3.1.5.
Step 5: Report the final set of coordination regions Once the desired level of coordination is established, the regional structure is reported, and the algorithm terminates.
Depending on the needs of decision-making, this iterative algorithm can produce multiple levels of coordination zones.Here, we apply this algorithm to identify regions in the United States at two different levels.The results are described in the next section.

Upper-level coordination regions (ULCRs)
Applying the Louvain method to the mobility network of 3107 nodes (counties) and 24,639 edges (between-county movements), we find that US counties are grouped in a multilevel structure, with 44 ULCRs, as shown in Figure 2. The modularity of this ULCR structure is 0.938, indicating a well-defined community structure.The median number of counties in a region is 66 (SD ¼ 37), and the largest region has 156 counties.Using the 'spatial association between zones' tool in ArcGIS Pro, we compared the ULCR structure with states boundaries and received a global association of spatial association score of 0.77, indicating that the ULCR structure is not well-aligned with state boundaries.Indeed, 39 out of 44 regions include counties from more than one state, and the average number of states a region covers is 3.25.On average, states belong to 2.75 regions, with a median of 3.00 (SD ¼ 1.43).This multi-state nature of the region structure highlights the importance of interjurisdictional policy coordination.
Although it is natural to assume that counties within a state to be well-connected (i.e., high between-movements), we expect that the ULCR structure based on the Louvain method to display even higher levels of connectivity.To verify, we compare the modularity of ULCRs to that of states as the alternative community structure.We find that the ULCR structure with a modularity score of 0.938 is indeed a better community structure with regard to movements than states with a modularity score of 0.867.Figure 3 also shows that, in general, the within-to-between mobility ratios for ULCRs are higher than those for the states.Since movement is a critical factor for pandemic management, we argue that the ULCR structure is a superior foundation for public health policy coordination than arbitrary state borders.

Lower-level coordination regions (LLCRs)
Experience of COVID-19 demonstrated that cross-border coordination should be carefully developed and supported at different levels of governments (OECD, 2020).With a multilevel structure, more locally targeted policies may be applied in subcommunities within each of the 44 ULCRs, especially in those regions that encompass a large number of counties or even states.We apply the same community detection algorithm to identify LLCRs within each ULCR.
Examples of two large regions are used to demonstrate this process.First, we consider the large ULCR located in Southwestern United States that include counties from Arizona, California, Nevada, New Mexico and Utah.Applying the Louvain algorithm to this ULCR as a network, five LLCRs are detected, as shown in Figure 4.These LLCRs tend to be anchored by major metropolitan areas such as Los Angeles, San Diego, Las Vegas and Phoenix.Although all these five subcommunities are part of a large mobility region at the national level, this LLCR structure further refines the movement pattern and thus can allow for more localised policy coordination.398 Milad Baghersad et al.

REGIONAL STUDIES
As a second example, we consider the large ULCR containing much of Florida.This region includes several large cities, with diverse populations that face possibly differing local issues.After treating this region as one network and applying Louvain algorithm to it, six LLCRs emerged, as shown in Figure 5.These six also tend to be anchored by metropolitan areas (such as Miami, Orlando, Tampa and Jacksonville) and their neighbouring counties.
Interestingly, some of these LLCRs overlap with existing combined statistical areas (CSAs) as defined by the US Census (e.g., the Miami-Fort Lauderdale-Port St. Lucie CSA in Florida falls within a single region in Figure 5), but others do not (e.g., the Lakeland-Orlando-Deltona CSA in Florida is covered partly by two separate regions in Figure 5).As Fortunato and Hric (2016) point out, both the internal cohesion among community members Regional policy coordination of pandemic responses using an iterative mobility-driven algorithm 399

REGIONAL STUDIES
and their separation from the rest of the network play a role for communities to be properly defined.

Application of regions to the COVID-19 pandemic
The identified multilevel interjurisdictional regions have high level of movements within them.As a result, one would expect that counties within the same region exhibit a similar impact by a contagious disease transmitted through contact.To test this, we retrieved US countylevel COVID-19 infections data from the 'Daily Reports' compiled by the Center for Systems Science and Engineering at Johns Hopkins University (JHU), publicly available from the repository. 3We obtained the daily number of confirmed cases by county, as reported by state and local health departments in each of the 50 states (Dong et al., 2020), restricting the frame of analysis to the 17week period from 1 June to 27 September 2020 when all states started to reopen after the initial and most restrictive lockdowns (Alexander et al., 2021).This time frame also corresponds to the period when testing availability had greatly improved across the country after an initial period of deficiency in testing capacity (Food and Drug Administration (FDA), 2021).US territories and 65 counties with missing mobility data or no cases reported in the JHU database, along with a few cases that could not be positively assigned to a particular county (e.g., cruise ships and federal correctional facilities), are excluded.The resulting dataset covers 3069 counties, out of 3141 counties and county equivalents in the United States.
We first compare the variations in COVID-19's impact on ULCR and states to see if the former structure is more homogenous than the latter.Table 3 shows the number of confirmed COVID-19 cases (daily and seven-day moving average) in ULCRs and states.In addition to the overall standard deviations, we also include the between-and within-region variations.The within-region variation gauges how much variation exists among all counties within the same region over time, ignoring all the between-region variations.The within-region variation is far lower than between-region variation, revealing more uniform COVID-19 impacts within ULCRs than between them.Furthermore, when counties are grouped into their corresponding states, the ratio of within-tobetween region variation is 12% higher than the same ratio for ULCRs.
We further test the variations using the Caliński and Harabasz (1974) index, also known as the variance ratio criterion.A high Caliński-Harabasz index implies a well-defined clustering structure with high within-regions homogeneity and between-regions heterogeneity (see Appendix B in the supplemental data online for a detailed discussion).Using total number of confirmed COVID-19 cases by county over 17 weeks, the Caliński-Harabasz index for ULCRs and states turn out to be 4.30 and 2.37, respectively, suggesting that the ULCR structure is almost twice as homogeneous within each region (and heterogeneous between them) in COVID-19 impact compared with the state borders.This finding further validates that the ULCRs are more appropriate than states for coordinating COVID-19 policies.
Lastly, we compare the spatial correlation of COVID-19 cases in counties within the same ULCRs versus state  Milad Baghersad et al.

REGIONAL STUDIES
borders using Moran's I, a commonly accepted measure of spatial correlation which has been used for studying the spread of COVID-19 (Bourdin et al., 2021;Kang et al., 2020).To examine not only spatial correlation but also temporal synchrony, we follow Bi et al. ( 2022) and adopt the global spatio-temporal Moran's I index, calculated as: where k denotes the number of regions; x i,t represents COVID-19 incidence at time t in region i; l is the time lag; and x t shows the average COVID-19 incidence at time t.v ij is the normalised spatial weight: v ij = 1 if both nodes (i and j) belong to the same region, and 0 otherwise.The global spatio-temporal Moran's I index is between −1 and 1, indicating perfect negative and positive relationship, and 0 shows no autocorrelation.To validate our results, we compare the Moran's I index based on ULCRs and LLCRs versus the existing administrative jurisdictions.
Figure 6 depicts the time-series diagram of spatiotemporal Moran's I index for both ULCRs and states.The values of global spatio-temporal Moran's I based on the ULCR structure are higher than those based on state boundaries.This suggests that ULCRs exhibit a higher level of spatial and temporal synchrony in COVID-19 incidence, thus offering a more appropriate basis for pandemic NPI coordination compared with states.
To study LLCRs, we use the state of Florida as an example because there is an existing jurisdictional structure as a direct comparison.Florida's local health councils 4 engage in regional health planning and implementation activities, in particular for communicable diseases.Figure 7 demonstrates the global spatio-temporal Moran's I for LLCRs in Florida is higher than that for the 11 local health councils during the entire 17-week period.This finding is consistent with those from the comparison between ULCRs and states.

DISCUSSION AND CONCLUSIONS
Horizontal cooperation across local or regional jurisdictions has been called for when dealing with the COVID-19 pandemic (Chandrasekhar et al., 2021;OECD, 2020).Interjurisdictional coordination is particularly crucial in decentralised and federally structured geographies such as the United States and European Union (Chandrasekhar et al., 2021;Holtz et al., 2020;McCann et al., 2022).Cheng et al. (2021), for example, show that COVID-19 spreads at higher speeds in decentralised countries, where it takes longer to implement NPIs.On the other hand, the experience of COVID-19 pandemic has shown that a broad, unified approach applied to a whole country or a state is not necessarily ideal (Dhillon & Karan, 2020), and painting with a broad brush can have more severe economic consequences (Birge et al., 2022).Hence, a coordinated approach taken up by local jurisdictions may be most appropriate.
Policy coordination within and across traditional boundaries has garnered considerable attention in recent years, and a call for more general theories, empirical cases, causality analysis and outcomes has been issued (Trein et al., 2021).Before engaging in any integration efforts, however, jurisdictions need to identify partners to coordinate and collaborate with.However, there has been a dearth of inquiry, empirical or theoretical, on how such groups may be defined.Our study aims to fill this gap of identifying which local jurisdictions should be working together to combat transboundary health crises.Recent studies have shown that mobility is a critical component in transmission of a contagious diseases such as COVID-19 (Chang et al., 2021;Kraemer et al., 2020).Although there are other existing functional region definitions based on movements, such as US labour market areas and travel-to-work areas in the UK, they typically focus on commuting flows between work and residential areas based on census data.We recognise that organising policy across geographies that are connected through mobility is essential.In this study, we use real-time big data on mobility at a finer level for all movements including educational and recreational activities.We apply modularity maximisation to mobility data to discover highly distinctive regions that are at times significantly different within or across existing jurisdictional state lines.Our proposed methodology is grounded on the nature of pandemics along with a theoretically established approach of community detection.
We applied the algorithm to group the US counties based on mobility and verified the resulting multilevel regional structure using actual COVID-19 morbidity data.Based on the nature of NPIs, coordination may be required at different geographical levels, and we show that ULCRs can be divided into smaller LLCRs.The ULCR structure can be used to apply policies that need  Milad Baghersad et al.

REGIONAL STUDIES
broader coordination such as travel restrictions and quarantine requirements, while LLCRs can be used for more targeted policies such as lockdowns that are too expensive or difficult to apply at a broader scale.As such, this study extends the application of community detection using Louvain algorithm to effectively identify multilevel structure in a nested network.
Our results suggest that regions created based on mobility are often not aligned with pre-existing jurisdictions.Indeed, in many cases, ULCRs (and even LLCRs) contain counties from multiple states, and often a single state belongs to several different regions.The practical implication is that a state-wide public health policy, commonly implemented independently from other states, is not necessarily an effective strategy.For instance, a policy response such as bar and restaurant closing in one state may not be successful if counties in a neighbouring state do not follow suit, as movement of people between counties in the same region will continue to spread the virus.Our findings provide a guideline for identifying partners of interjurisdictional coordination to more effectively combat a pandemic.
Similar practical implications can be drawn for COVID-19 vaccination policies, as the goal of maximising the vaccination rate can be achieved by targeting regions, rather than individual counties or states.Although the starting point for this study is to identify regions of local jurisdictions for coordination efforts to address a contagion, the resulting community structure can inform vaccination efforts as well.Identifying highly connected communities irrespective of existing jurisdictions will help better coordinate their efforts to improve vaccination rates and address any access inequities (Bubar et al., 2021).With the focus gradually shifting to post-pandemic policies, the public health challenges will turn towards how to deliver vaccinations, monitor the acquired immunity, and conduct surveillance of new virus variants and their spread (Kissler et al., 2020).As communities reopen and population mobility returns to pre-pandemic levels, inter-jurisdictional policy coordination will be critical to avoid repeating past errors.
This study makes a significant contribution to the literature by providing several important theoretical, managerial, and policy implications.To the best of our knowledge, our study is the first to develop public health coordination regions using community detection methodology for managing NPIs in a pandemic.The proposed iterative algorithm is applied to identify regions within the United States at two different levels.Using COVID-19 data, we confirm that counties within these regions show significantly more homogeneous disease incidence than counties within existing administrative borders.Given that the threat of an infectious disease outbreak is a major challenge for public health, investigating population movements to identify optimal policy coordination regions can help meet critical preparedness goals.Our findings provide policymakers and practitioners with a replicable methodology to identify groups of jurisdictions for joint management of transboundary crises such as pandemics.
As with all research, this study has limitations.First, the movement data collected from smartphones has its intrinsic shortcomings.Since the data are anonymised and captured in 4-hour blocks, analysing individual  movements over time or on long trips is not possible.
A more comprehensive analysis of mobility for interjurisdictional policy coordination would also include long-range travel (e.g., by air).Additionally, by its very nature, access to smartphones is also subject to disparities due to age, race/ethnicity, or income (Kishore et al. 2020).While smartphone data may not provide a complete picture of all movements in some areas, it has been shown to be a reliable measure of mobility and social contact (Couture et al., 2022).Therefore, the regions created should be considered as a starting point for policy coordination, subject to adjustments based on local knowledge or more refined mobility data.Second, the COVID-19 morbidity data that we use to verify the regional structure is somewhat limited, subject to varying practices of testing and data reporting across states and local governments.
In addition, the validation exercise is limited by the fact that it does not consider variability in factors such as population, policy and surveillance strategies.
Although our proposed methodology can develop multiple levels of coordination regions, the question of how the counties in the same regions should coordinate their efforts is beyond the scope of this study.As a critical research direction, future studies should analyse the cost and benefit of coordinating different policies at multiple levels.Another important research advance would be to incorporate social determinants of health into the analysis.It has been shown that both disease transmission (e.g., Amdaoud et al., 2021) and the impact of NPIs (e.g., Guaitoli & Pancrazi, 2021) vary along the socioeconomic dimensions.Accounting for socioeconomic conditions would allow for a more comprehensive and deeper understanding of disease and mortality dynamics and in turn lead to a more effective public health policy response.
While our analysis is based on mobility across counties, the same framework can be applied at finer geographies (such as at the zip code or census tract level), depending on data availability, to effectively customise targeted NPIs.Similarly, our analysis can be extended to show how public health policies should adapt to any changes in population movement over time, as recently highlighted by Shi et al. (2020).By examining the differences between the spatial contagion of the severe acute respiratory syndrome (SARS) in 2003 and that of COVID-19, Shi et al. (2020) suggest that the unique patterns observed in each case should be attributed not only to contrasts in disease dynamics but also to changes in population movements such as migrations and the rapid growth of recreational and business trips over time.They argue that disease prevention and control strategies should be adaptive as the changes in population movement are expected to continue.Finally, while our methodology is specifically designed for public health policy in response to pandemics, it can be extended to other transboundary health crises, such as the spread of illicit substances and related addiction.Each case, of course, may require a unique dataset for detecting communities, but the idea of identifying regions based on grounded theories and empirical evidence is universally applicable as the critical first step of effective health policy coordination.

Figure 3 .
Figure 3. Ratios of within-to-between mobility, states (a) versus upper-level coordination regions (ULCRs) (b).Note: Whites patches are counties with missing data.

Figure 6 .
Figure 6.Time-series diagram of the spatio-temporal Moran's I index.

Table 1 .
Transitions between two counties on 20 January 2020.
b FIPS, Federal Information Processing Series.

Table 3 .
Between versus within variation in confirmed COVID-19 cases.Note: Authors' calculations using US county-level COVID-19 infections data from the 'Daily Reports' compiled by the Center for Systems Science and Engineering at Johns Hopkins University (JHU) for the 17-week period between 1 June and 27 September 2020.The number of observations reflects 17 weeks × 7 days for states or upper-level coordination regions (ULCRs) without missing data.