Do regionally targeted lockdowns alter movement to non-lockdown regions? Evidence from Ontario, Canada

Regionally targeted interventions are being used by governments to slow the spread of COVID-19. In areas where free movement is not being actively restricted, there is uncertainty about how effective such regionally targeted interventions are due to the free movement of people between regions. We use mobile-phone network mobility data to test two hypotheses: 1) do regions targeted by exhibit increased outflows into other regions and 2) do regions targeted by interventions increase outflows specifically into areas with lesser restrictions. Our analysis focuses on two well-defined regionally targeted interventions in Ontario, Canada the first intervention as the first wave subsided (July 17, 2020) and the second intervention as we entered into new restrictions during the onset of the second wave (November 23, 2020). We use a difference-in-difference model to investigate hypothesis 1 and an interrupted time series model to investigate hypothesis 2, controlling for spatial effects (using a spatial-error model) in both cases. Our findings suggest that there that the regionally targeted interventions had a neutral effect (or no effect) on inter-regional mobility, with no significant differences associated with the interventions. We also found that overall inter-regional mobility was associated with socio-economic factors and the distance to the boundary of the intervention region. These findings are important as they should guide how governments design regionally targeted interventions (from a geographical perspective) considering observed patterns of mobility.


Introduction
Non-pharmaceutical interventions, such as those that limit mobility and thus reduce contacts, can help curb the spread of COVID-19 (Ferguson et al., 2020;Flaxman et al., 2020). At a global scale, national level restrictions have proven successful at limiting the spread of COVID-19 across borders that have had strict policies to limit movement between countries (Wells et al., 2020). Within countries, regionally targeted lockdowns have been implemented worldwide in an attempt to curb the spread of COVID-19 via contagion effects (Bourdin et al., 2021;Desjardins et al., 2020;Paez et al., 2020). For example, China ordered a mandatory lockdown in Wuhan, associated with the origin of Covid-19 cases from January 23, 2020 (Ren, 2020). In Italy, eleven municipalities were initially quarantined on March 1, then, a lockdown order was imposed on March 9 for 26 northern provinces (Caselli et al., 2020). During the second wave of COVID-19, the Spanish government imposed local lockdowns in Barcelona, Madrid, and Zaragoza ("Coronavirus," 2020) and local lockdowns have also been widely imposed in the UK (Gathergood and Guttman-Kenney, 2021).
Empirical work has demonstrated that areas with lower interregional travel have lower COVID-19 case and death rates (Yilmazkuday, 2020). Further, modelling studies have demonstrated regionally targeted restrictions to be effective at reducing the overall spread of COVID-19 (Karatayev et al., 2020). However, the efficacy of regionally targeted intervention strategies is at least partially dependent upon people from targeted regions not moving to other areas. In some cases, so-called hard lockdowns have been implemented and enforced by local law enforcement which severely limit the free movement of people (Lau et al., 2020), whereas other jurisdictions have enacted softer lockdowns where people are still freely allowed to move between lockdown and non-lockdown regions (Mahase, 2020). In areas where movement between regions is not actively restricted, there has been widespread speculation about the efficacy of such regionally targetedor locallockdowns, due largely to the free movement of people. To date, there have been no studies that specifically measure how inter-regional mobility changes as a result of regionally-targeted lockdowns (where no active enforcement of such restrictions exist).
In this paper, we test two hypotheses concerning changes in interregional mobility in regions targeted by locally-specific government interventions to limit the spread of COVID-19: 1) do regions targeted by interventions exhibit increased outflows into other regions and 2) do regions targeted by interventions increase outflows specifically into areas with lesser restrictions. We test whether distance to the boundary of the lockdown region and other socio-economic indicators were associated with any observed changes in outflows. For each intervention period, we study the changes in outflows derived from network mobility data over the 7-day period before and after the intervention date.

The COVID-19 pandemic in Ontario, Canada
In Ontario, Canada, the first case of COVID-19 was reported on January 25, 2020 (see Fig. 1 for a timeline of the COVID-19 pandemic in Ontario). The premier of the province ordered a state-of-emergency beginning March 17, 2020 which represents the start of the first lockdown in Ontario. By June and July, case counts had reduced, and the province began its phasing out of lockdown restrictions, ending the first wave of the pandemic. By October 2020 cases had risen again substantially, and by late November the province was back into strict lockdowns in specific regions, with the rest of the province joining on December 26, 2020.
Regionally targeted restrictions were implemented by the provincial government both during the first wave of COVID-19 and in the second wave. Specifically, here we study two well-defined intervention dates corresponding to interventions that led to regional differences in lockdown conditions in Ontario, Canada at the end point of the first lockdown on July 17, 2020 (Intervention 1) and the beginning of the second lockdown on November 23, 2020 (Intervention 2; Fig. 1). Regionally targeted lockdowns were organized at the Public Health Region (PHR) level of which there are 35 PHRs in Ontario (Fig. 2). In the first intervention, 10 of 35 PHRs were held back under more strict lockdown conditions for a minimum of 7 days while the remaining 25 PHRs were allowed to open more fully. In the second intervention, 2 of 35 PHRs (Toronto and Peel) were forced to go back into a stricter lockdown, ahead of the rest of the province, as a result of climbing COVID-19 case counts. Therefore, our two interventions also represent two different types of interventions. The first intervention was associated with the lessening of lockdown restrictions at the termination of a pandemic wave, and the second was associated with the onset of restrictions during the initiation of a pandemic wave.
In both cases political and public support for regionally targeted lockdowns wavered due to an assumption that those under stricter lockdown conditions would readily move to other areas of the province to access services (Crawley, 2020). The premier of Ontario initially suggested that a regional approach to reopening after the first lockdown in Ontario would not be used, citing the potential for people to travel between regions to access certain businesses where they are open. Despite stating that this option was off the table in early May 2020, the province did end up implementing different reopening timelines for different health regions, with dates for reopening stages in different regions generally no more than a month apart. Following both lockdowns, when there have been varying levels of restrictions in different health regions, there have been concerns about people traveling between regions to access businesses that are only able to open in certain regions, and that this could lead to virus spread from higher transmission areas into lower transmission areas. Many rural communities in Ontario were requesting that people from COVID-19 hot zones not travel to their area to avoid the potential spread of the virus in their communities where transmission was lower. This especially became a concern in the summer when restrictions on out-of-country travel prompted people to look for places closer to home to go on vacation.

Measuring inter-regional mobility (outflows)
Inter-regional mobility (or external mobility) refers to trips that are made that start and end in different regional units, which are defined depending on the specific study area. Studies looking at inter-regional mobility in the United States often use counties as the geographic unit representing a region, where trips originating in one county and ending in a different county are considered to be inter-regional trips (Xiong et al., 2020). Similarly, coarser analysis may simply record if individual has been in two different counties within a specified time range (Yilmazkuday, 2020). The geographic units should generally be large enough that travel between them is not a common occurrence for day-to-day activities. If the geographic units being used are smaller, a moving regional boundary may be applied where trips are only considered inter-regional if they are between two geographic units that are greater than a specified distance apart (Pullano et al., 2020). Analysis of inter-regional mobility must therefore also account for the fact that individuals closer to a regional boundary are likely to have more inter-regional trips than those living further from the boundary. Some definitions of a regional trip focus on what are termed inflows, or the amount of people traveling into a region (Xiong et al., 2020). However, it is more common to focus on trips aggregated by origins, termed outflows to measure outward inter-regional mobility patterns (Pullano et al., 2020).
Data from mobile phones are now widely used to study inter-regional mobility (i.e., inflows and outflows) and these data are commonly used in epidemiological studies. Due to privacy concerns, mobile-phone data are typically aggregated to coarser geographical units used to develop geographical flow networks, which can be used to capture inter-regional mobility patterns (e.g., Sekulić et al., 2021). In the context of spatial epidemiology, inter-regional mobility patterns, derived from mobile-phone data, have been previously used to study, for example, patterns of spread of cholera in Haiti (Bengtsson et al., 2015), malaria in Kenya (Wesolowski et al., 2015), and asschistosomiasis in Senegal (Mari et al., 2017).
Following the COVID-19 outbreak, mobility flows derived from mobile phone data were identified as having a crucial role to play in understanding the dynamics of COVID-19 spread and the effectiveness of different mitigation strategies, guided by the recommendations in (Oliver et al., 2020). The sheer volume of research to date and variety of datasets used to study mobility patterns during the COVID-19 pandemic is extensive. Mobile phone data, and specifically inter-regional flows, were used to improve spread models (e.g., Arenas et al., 2020), and study the efficacy of non-pharmaceutical interventions (e.g., Aleta and Moreno, 2020; Gatto et al., 2020;Pepe et al., 2020). Further, many studies have examined the substantive socio-economic impacts of the COVID-19 pandemic by looking at changes in mobility behaviour using mobile phone data (e.g., Bonaccorsi et al., 2020;Lee et al., 2021;Long and Ren, 2021). However, to date there have been no studies that have specifically used large mobile-phone datasets to study the impact of regionally targeted lockdown policies on changes to inter-regional mobility (specifically outflows) in regions without strict enforcements limiting mobility.

Network mobility data
To measure outflows, we used de-identified mobile-phone network mobility data from TELUS communications Inc through their Insights platform which is a privacy preserving system for analyzing mass-mobility patterns within Canada. These data comprise a sample of ~3.5 million devices in Ontario Canada, representing a sample of ~25% of the adult population (Long and Ren, 2021). The devices included in our sample remain relatively consistent throughout the year with only a small percentage of turnover throughout (occurring through both gain and loss). The TELUS data comprise of connections to the cellular tower network over time. There are approximately 90000 unique receivers within the cellular tower network in Ontario. As a device is moved, it will switch its connection from receiver to receiver (termed handovers), generally taking the most proximal receiver. For each connection, the data contain the start and end time of the connection, along with the geographical coordinates associated with that receiver, thus representing a near-continuous record of location while a device is powered on. From the sequences of receiver connections, we can determine estimates of movement (i.e., which receivers each device were connected to). In urban areas the density of towers is greater and thus we capture more detailed pattern of movement, in rural areas the density of towers is very sparse and thus we capture a less detailed pattern of movement.
To estimate the home neighbourhood of each device, we used the network data to identify a proximal subset of network tower receivers associated with the largest cumulative dwell time of that device. This subset of receivers is termed the cluster of home neighbourhood receivers. In the urban areas, these clusters can consist of many receivers, while in the rural areas of only a few. For each device we then computed the weighted average of the geographical coordinates of the cluster of home neighbourhood receivers, where the weights were established proportional to the dwell time at each receiver. We then identified the Aggregated Dissemination Area (ADA) associated with the cluster of home neighbourhood receivers for each device using a spatial intersection. We repeated the above process of identifying a home neighbourhood receiver set and their home neighbourhood for each month under study to capture potential (month-by-month) changes. The ADA geographical units are used by Statistics Canada and represent suitable geographical units for studying inter-regional flows as they roughly capture neighbourhoods (populations ca. 5000-15000) in urban regions. There are 1534 ADA units in Ontario (Fig. 2). In our analysis we include only those ADA regions that contain a minimum of 100 devices.
Our analysis focuses on inter-regional flows which we define as a trip or visit from one public health region to another. Specifically, we count the total number of trips or visits to a different public health region (PHR) for every ADA in Ontario. We defined a visit as any instance of a device spending a minimum of 10 min at a cellular receiver located in a different public health region from its home neighbourhood (ADA). With this definition, we capture the volume of outflows from each ADA to other public health regions. To account for the fact that our sample size varies between ADAs we calculate the number of visits per 1000 devices as a population adjusted measure of inter-regional outflows. Because we are interested in mobility from regions under more restrictive lockdown conditions we focus only on outflows and ignore inward flows.

Statistical analysis
We examined two intervention periods; Intervention 1: July 17, 2020, associated with regionally targeted easing of lockdown restrictions, and Intervention 2: November 23, 2020, regionally targeted re-introduction of lockdown restrictions coinciding with the second wave. For each intervention, we performed two pieces of analysis. First, we performed a spatial difference-in-difference (DID) model (Dubé et al., 2014) to test whether regions that were placed under more restrictive lockdowns had higher rates of outflows to other PHRs relative to regions with lesser restrictions. Second, we performed an interrupted time-series (ITS) model to test whether regions under more restrictive lockdown restrictions specifically increased their outflows to regions under lesser lockdown restrictions in the 7-day period immediately following the intervention.

Difference in difference (DID) model
We used a difference in difference (DID) model to test if regions under more restrictive lockdown conditions increased (or decreased) their outflows (visits to other PHRs) relative to regions under lesser restrictions following the regionally targeted lockdown. The DID model takes the following form where F is the 7-day average outflow rate for each ADA region i (i = 1, …, n) at time t (t is the time subscript with 0 before the intervention and 1 after the intervention), P is a dummy variable indicating with 0 indicating the pre-and 1 the post-intervention period, R is a dummy variable with 0 indicating the region was not targeted by the intervention and 1 indicating the region was targeted by the intervention, X is a matrix of region specific covariates, ε is the error term, and the β are the estimated coefficient parameters. Interpretation of the intervention effect in a DID model is inferred via the β P*R parameter, which in our case if β P*R is significantly positive provides evidence supporting the hypothesis that outflows increased in areas targeted by regional lockdowns following the intervention relative to the control regions.
To account for the presence of spatial autocorrelation in our data we implemented a spatial-version of the DID model (Delgado and Florax, 2015). Spatial DID models have been implemented in a number of previous applications where interventions are studied in the context of regional differences for example in the study of regeneration of urban greenspaces (Heckert and Mennis, 2012), transportation developments (Comber and Arribas-Bel, 2017), and impacts of windfarms on property values (Sunak and Madlener, 2016). We implemented the spatial DID model by adjusting the form of the error term to account for spatially autocorrelated model residuals (termed a spatial error model -SEM; Anselin, 1988). Specifically, the SEM version of the DID model takes the same form as in (1) but includes a spatially autocorrelated error term: where λ is the spatial autocorrelation parameter, W is a n × n spatial weights matrix, τ is the number of time points (here τ = 2), u are the spatially structured error term and v the remaining, independent aspatial error term (Delgado and Florax, 2015). We used an asymmetric, k-nearest neighbour definition for the spatial weights matrix with k = 15, and based on preliminary testing of the models found the results were not sensitive to small changes in k.

Interrupted time series (ITS) model
We used an interrupted time series (ITS) model to test if those regions that were targeted in intervention 1 and intervention 2 increased their outflows specifically to other regions not targeted in those interventions. An ITS model takes the following form where G is the population adjusted outflows to PHRs under lesser restrictions of ADA region i at time t, T is an integer (1, …, 14) specifying each day in the 14-day period associated with each intervention (sevendays before and after the intervention), P is a dummy variable for the pre-and post-intervention time periods, X is a matrix of (optionally time varying) regionally specific covariates. With the ITS model, β T is interpreted as the overall time trend, β P is used to infer whether there is a noticeable change in the level (i.e., intercept) of the relationship associated with the intervention (termed a level-change), and β T*P is used to infer if there is a change in the slope of the relationship following the intervention (termed a slope-change). This form of the ITS, which includes both level-and slope-change terms, is called a level and slope change ITS model (Bernal et al., 2017). We implemented a spatial error version of the ITS model in a similar fashion to the DID model (spatial error model (SEM); Equation 2). In the spatial version of the ITS model we have 14 time periods and therefore τ = 14 in equation 2. This results in a block diagonal spatial weights matrix consisting of 14 identical spatial weights matrices arranged along the diagonal and zeros elsewhere (though they need not identical). Like with the spatial DID model we again used an asymmetric, k-nearest neighbour definition of the spatial weights choosing k = 15.
The preceding formulations of the DID and ITS models are very similar (Benmarhnia and Rudolph, 2019;Lopez Bernal et al., 2019). However, the different forms of the DID and ITS models allow us to address two slightly different questions in relation to how mobility behaviour (i.e., outflows) changed as a result of regionally targeted lockdown interventions. The DID model allows us to compare both those regions targeted by the intervention and those in a control group (i.e., not targeted). In this analysis, we can test if areas under stricter lockdown restrictions had different outflow behaviour relative to those areas under lesser restrictions. Whereas, in our case, the ITS model only includes those regions targeted by the intervention. In the ITS analysis, we specifically study changes in the outflows from regions with stricter lockdown restrictions to areas with lesser restrictions. ITS models can include control regions (termed controlled ITS models; Lopez Bernal et al., 2018), but the dependent variable in this case (i.e., counts of flows from within lockdown PHRs to non-lockdown PHRs) is by definition always 0 for non-lockdown regions. Therefore, we implement the ITS model without controls focusing only on regions targeted by the intervention.
In each of the DID and ITS models we account for spatial autocorrelation present by including a spatial effect (λ) in the model residuals (termed a spatial error model; SEM). In practice, both the spatial DID and spatial ITS models are a form of a spatial panel model (Elhorst, 2014), with τ = 2 time units in the DID case and τ = 14 in the ITS case.
For each model, we first compared the 7-day period before the intervention with the 7-day period after the intervention. To account for the fact that these interventions were typically communicated ahead of their implementation (thus potentially leading to a pre-intervention change in behaviour) we further ran the model comparing the 7-day period after the intervention with the 7-day period beginning 28-days before the intervention. The results for the period 28-days before the intervention are found in the Supplementary Material.
We ran the models both with and without further covariates associated with each ADA region (Table 1). We used three components from the Canada Index of Multiple Deprivation to describe socioeconomic characteristics of the origin ADA (Government of Canada, 2019; Matheson et al., 2012) which we believed may impact external mobility. A further explanation of these three components is included in Table 1. Further, we included the distance to the boundary of the public health region, which we would hypothesize to impact outflows due to distance decay effects (Fotheringham, 1981). Finally, we included population density which is sensitive to both the urbanness of the ADA and the overall size of ADA region.
When interpreting our statistical models, we use α = 0.05 to denote statistically significant associations. We then focus our discussion of the results on the direction and magnitude of the effect sizes. All analyses were conducted using the R statistical computing environment. We used the package spatialreg (Bivand and Piras, 2015) for fitting the models, the package spdep (Bivand et al., 2013) for the construction of spatial weights. Code is provided in the supplementary material demonstrating the process used to build and fit the models.

Intervention 1: July 17, 2020
Although only 10/35 PHRs were targeted with the intervention on July 17, 2020, they accounted for 901/1523 of the ADAs, as these ten PHRs comprise the more densely populated areas in Ontario. We found that intervention regions generally had higher change in outflows compared to the control (no intervention) regions on average (4% compared to 8%; Table 2). The intervention regions had a much lower standard deviation than the control group indicating less variability in outflows ( Table 2).
The spatial DID analysis of the first intervention revealed no significant difference in outflows associated with the regionally targeted lockdown effect both without and with controlling for area-level covariates (Table 3). We found that the area level-covariates were strong predictors of outflows. The direction of this relationship for distance to boundary was negative and for population density was positive. We found a positive relationship between outflows and the economic dependency indicator, suggesting areas with higher economic deprivation had more outflows. The negative relationship between outflows and the ethno-cultural index suggests areas with greater ethnic diversity had lesser outflow rates. Finally, the negative relationship between residential instability index and outflows indicates that areas with more unstable housing characteristics also had lower outflow rates. The AIC and λ values support the DID model that includes the spatial effects and the additional socio-economic covariates.
We found that regions held back under tighter lockdown restrictions increased their outflows to control (non-lockdown) PHRs by approximately 15% (Table 4), which was almost twice the overall increase in outflows to all PHRs (8%; Table 2). There was a large degree of variation in the rate of outflows across the different ADA regions targeted by the lockdown (Table 4).
From the interrupted time series model, we see that β T was negative during the summer intervention on July 17, 2020 (Table 5) suggesting that over this two-week period outflows were decreasing slightly. We see the positive coefficient associated with the intervention which suggests an immediate level-change associated with increased outflows following the intervention (Table 5). The magnitude of this step-change effect is substantial, and remains consistent after controlling for distance and socio-economic covariates (β P = 186). We also see that weekend days are associated with a similar magnitude increase to outflows during the July 17, 2020 intervention (β weekend = 151; Table 5). The slope change after the intervention was also negative but was not significant in either ITS model.
From the socio-economic indicators we see that economic dependency and the ethno-cultural index were negatively associated with outflows from lockdown to non-lockdown regions, but there was no association between residential instability and outflows from lockdown Table 1 Area level covariates used in the difference-in-difference (DiD) and interrupted time series (ITS) analysis during two regionally targeted interventions in Ontario Canada.

Variable
Explanation Source

Distance to Border
Distance from centroid of ADA to boundary of public health region or regionally targeted intervention  Table 2 Population adjusted outflows defined as flows from each aggregate dissemination area (ADA) in Ontario to a different public health region (PHR) from the origin. Flows were averaged across the during the 7-day period before (Pre) and after (Post) the intervention (July 17, 2020). Ten PHRs were targeted to remain under stricter lockdown conditions intervention which occurred at the termination of the first wave in Ontario.

Table 3
Spatial Difference in Difference (DID) analysis using a spatial regression (spatial error model) framework for 7-day average per-capita outflows to other public health regions during the 7 days before and after the intervention on July 17, 2020 in Ontario, Canada (targeting 10 Public Health Regions). Mobility outflows and socio-economic covariates calculated at the aggregate dissemination area (ADA) level (n = 1523). to non-lockdown regions. Again, like with the DID we find evidence of a strong distance-decay effect with areas further from the boundary of the intervention associated with lower outflows (Table 5). Population density was also again positively associated with outflows. Similar to the DID model in the first intervention the AIC and λ support the most complex model including both the spatial effects and the region-specific covariates.

Intervention 2: November 23, 2020
In the second intervention the Toronto and Peel PHRs were moved into lockdown on November 23, 2020 ahead of other PHRs, these two health regions account for 436/1516 of the ADA regions under study. Outflows to other PHRs remained stable in both the control and intervention regions between the week before and after the intervention. Outflow rates were similar (~700 outflows per 1000 population) in the control and intervention regions (Table 6), however the variation (standard deviation) in outflow rates was lesser in the intervention regions (Table 6).
From the spatial DID analysis (Table 7) we can see that there is no significant effect associated with the DID parameter without regionspecific socio-economic covariates (β P*R = 19.5, p = 0.627) or with covariates (β P*R = 19.0, p = 0.573). We see a significant region effect, where the intervention regions were associated with lower outflows overall. We again see a strong distance decay effect where regions closer to region boundaries had more outflows than regions further away. Finally, in the second intervention the higher values on the ethnocultural index and residential instability index were both significantly associated with lower outflows, whereas economic dependency and population density were not significantly associated with outflows. The observed AIC and λ values again support the model including both the spatial effects and the region-specific covariates.
When looking at the outflows specifically from the lockdown regions to non-lockdown regions before and after the intervention we see a low increase in outflow rates (~4%; Table 8). The standard deviation stayed relatively consistent between the two time periods.
We again applied interrupted time-series analysis to the second regionally targeted lockdown intervention where two PHRs (Toronto and Peel) were moved into stricter lockdown restrictions on November Table 4 Changes in outflows (7-day average before and after the intervention) from aggregated dissemination areas (n = 901) in public health regions targeted by first intervention (July 17, 2020)

Table 5
Interrupted time series analysis using a spatial regression (spatial error model) framework for daily outflows from the 10 public health regions held back under tighter restrictions to other public health regions during the 7 days before and after the intervention on July 17, 2020 in Ontario, Canada. Mobility outflows and socioeconomic covariates calculated at the aggregate dissemination area (ADA) level (n = 823).  Table 6 Population adjusted outflows defined as flows from each aggregate dissemination area (ADA) in Ontario to a different public health region (PHR) from the origin. Flows were averaged across the during the 7-day period before (Pre) and after (Post) the intervention (November 23, 2020). Two PHRs (Toronto and Peel) were targeted to move into stricter lockdown conditions in this intervention which occurred at the initiation of the peak of the second wave in Ontario.  23, 2020 ahead of the rest of the province (Table 9). The regionally targeted intervention strategy employed in Ontario resulted in a small but non-significant decrease (i.e., level-change) in outflows to PHRs in both the model without (β P = − 54.0, p = 0.220) and with the regionspecific socio-economic covariates (β P = − 53.8, p = 0.187). The time trend (β T ) and the time slope-change (β T*P ) was also not significant in both models (Table 9). We found that weekend days had significantly lesser outflows than weekdays, which is the opposite pattern in the first intervention, reflecting seasonal patterns of mobility in Ontario. We also found that the same strong distance-decay effect as in other models. Again, we find that the AIC and λ values support the inclusion of the spatial effects and the additional region-specific covariates.

Intervention 1: July 17, 2020
Based on the ITS analysis (models ITS1 and ITS2) we found evidence that outflows from lockdown regions to non-lockdown regions increased (i.e., a significant level-change effect in the ITS models) following the intervention which kept targeted regions under stricter conditions while lifting restrictions in other regions on July 17, 2020. However, the DID analysis revealed that these changes observed in lockdown regions were on par with the other regions in the province (i.e., no significant DID effect in the DID models 1 and 2). These results suggest that the regionally targeted relaxation of the lockdown at the end of the first wave did not significantly increase outflows to other regions. Instead, we found that patterns of movement in regionally targeted areas were consistent with other areas of the province. This result can be viewed as a neutral outcome in terms of the effectiveness of such regionally targeted measures in terms of their impact on flows of individuals between regions.

Intervention 2: November 23, 2020
The imposition of new lockdowns in two PHRs (Toronto and Peel) during the second wave of COVID-19 ahead of other regions (which eventually followed) did not result in any significant increase in outflows as shown in both the DID and ITS models. Interestingly, there was no significant period effect in the DID model, and there was also no significant time-trend, or slope-change in the ITS model. The results again provide strong support that regionally targeted lockdowns associated with the second intervention had a neutral effect on outflows from targeted areas. One interesting finding here was the significant region effect in the DID model which suggests that during the period studied, targeted PHRs were associated with lower levels of outflows compared to other regions, but this was an overall pattern and not associated with the specific intervention (a result further supported using the reference period 28 days prior to the intervention; see Supplementary Material).

General discussion
The differences between the results in the first and second intervention may be due to a number of factors, most notably the intervention on July 17, 2020 involved the relaxation of restrictions in a regionally targeted manner while the intervention on November 23, 2020 involved targeting regions to go back into more restrictive lockdown conditions. This may be evidence that regionally targeted interventions impact movement outflows differently depending on whether they occur at the outset, or end of lockdown periods. One explanation for increased outflows observed when regionally targeted interventions are implemented at the end of lockdowns vs the beginning is that there may be differential attitudes, compliances, and health behaviours going into a lockdown vs coming out of it because psychological factors relating to lockdown fatigue (Naumann et al., 2020). However, a further explanation of this effect may be the seasonal differences that exist in travel in Ontario between the summer and winter seasons (Abdelgawad et al., 2015).
The second intervention occurred during the lead up to the Christmas season, and at this time schools and many businesses were still open. Weekend travel was lower during this period relative to weekday travel. Local businesses, especially shopping centers, in areas outside of the restricted regions reported higher levels of activity from customers from within the two restricted areas (Hewitt, 2020;Hristova, 2020). This is in spite of widely documented shifting consumer patterns to online shopping (Bounie et al., 2020). While an increase in outflows was not observed during the second lockdown, there is alternative evidence that more targeted visits, for example to shopping centers, may have been occurring at the expense of other types of mobility.
We found that in both the first and second intervention, in the ITS model, day of the week (i.e., weekday vs weekend), was an important predictor of outflows. During the July 23, 2020 intervention, the restrictions were lifted on a Friday. Weekends in the summer period were associated with much higher mobility than weekdays. As mobility increased during the summer months people were visiting some of the less densely populated areas in Ontario, which included areas outside of the regions held back under more restrictive conditions for recreation and vacation opportunities. However, in the second intervention period we found that weekend days were associated with significantly lower Table 8 Changes in outflows (7-day average before and after the intervention) from aggregated dissemination areas (n = 437) in two public health regions targeted in the second intervention (November 23, 2020)   outflow rates. This is likely because during this period more businesses and workplaces were open, and therefore this difference between the two periods may be reflective of the different types of mobility observed during the fall relative to the summer period (Kubota et al., 2020;Long and Ren, 2021). We hypothesized that distance to the boundary of the PHR or targeted intervention would impact the rate of outflows, that is regions closer to the boundary between a lockdown and non-lockdown region would have higher rates of outflows (i.e. a distance-decay effect; Hipp and Boessen, 2017). We found strong evidence of such a distance decay effect in both DID and ITS models, for both the first and second intervention period, where distance to boundary was consistently negatively associated with outflow rates. Regionally targeted lockdowns were implemented based on PHR boundaries (see Fig. 2); however, these boundaries may poorly reflect the mobility patterns of Ontarians. As an alternative, regionally targeted lockdowns may be more effective at limiting outflows from targeted areas if they reflect observed mobility patterns, for example through the development functional travel regions based on mobility flows (Aguiar et al., 2020;Farmer and Fotheringham, 2011;Sekulić et al., 2021). We found that in the first intervention the economic dependency index was positively associated with outflows to other PHRs in the DID model, but negatively associated with outflows from targeted regions to non-targeted regions in the ITS model. This may reflect regional differences being captured in the DID model (which included all ADAs in province) vs the ITS model (which included only those ADAs targeted by the intervention). In the second intervention, we found that economic dependency showed no association with outflows to other PHRs in the DID model but was again negatively associated with outflows from targeted regions to non-targeted regions in the ITS model. These results suggest that within areas targeted by the interventions (in both the first and second interventions) areas with lower economic status (higher economic dependency) were associated with lower outflows to nonlockdown regions. Therefore, our findings on outflows and economic conditions are surprising because they conflict with previous analysis that found that areas with lower economic status were associated with higher relative daily mobility levels throughout the pandemic (Bonaccorsi et al., 2020;Dasgupta et al., 2020;Huang et al., 2021;Jay et al., 2020;Lee et al., 2021;Long and Ren, 2021;Weill et al., 2020). It could be that because these areas are found to be engaging in higher relative mobility overall, the regionally targeted intervention then does not lead to substantive changes in outflows associated with an intervention. Alternatively, higher relative daily mobility levels in more economically deprived areas may be simply occurring within the PHRs regionally targeted for interventions. Thus, this mobility pattern is not captures as outflows to other PHRs.
The ethno-cultural index was found to be negatively associated with outflows in both the DID and ITS models for both the first and second interventions. This consistent finding suggests that areas with more ethnic diversity have lower rates of outflows overall compared to areas with lower ethnic diversity. Previous research using the same index, found that these areas also have lower relative daily mobility (Long and Ren, 2021). Interestingly the effect size of the association between ethno-cultural index was much larger (for both the DID and ITS models) in the first wave compared to the second wave. One potential explanation for this difference in effect size could be the differences in the regions included in the targeted interventions in the July 17 intervention and the November 23 intervention. The first intervention contained both the ethnically diverse areas around the Greater Toronto Area, but also less ethnically diverse regions further in the suburbs, and in Southwestern Ontario. The second wave only included the Toronto and Peel PHRs, which may be a reason for this observed difference. A second explanation could be seasonal, where summer months were associated with different patterns of mobility relative to those in November.
The residential instability index was negatively associated with outflows to other PHRs in the DID models for both the first and second intervention periods. However, the residential instability index showed no association with changes in outflows from targeted areas to nontargeted areas as evidenced by the ITS models. The likely explanation for this is that the DID models comprise all regions of Ontario, whereas the ITS models only focus on the targeted regions. Therefore, residential mobility may be a stronger predictor of outflows in the larger context, but not when looking at specifically those areas targeted by the interventions.
We hypothesized that population density might be positively associated with outflows, both because it was believed that people from denser urban areas would look to utilize the parks and natural recreation areas found in less dense regions (Dewis, 2020;Geng et al., 2021). Similarly, population density is approximately inversely proportional to the size of the ADA region, and therefore, outflows may be more likely out of smaller regions due to the smaller travel costs (after controlling for distance to the boundary). We found that this hypothesis was true in the first intervention (in both the DID and ITS analysis), but no relationship was found between population density and outflows in the second intervention. These findings support the hypothesis here, but again the seasonal effect of the first intervention (during summer) vs the second intervention (late fall/early winter) likely play a major factor in how population density is associated with flows to other regions. These results make sense in the context that typical summer travel in Ontarioboth historically and in the summer of 2020 during the COVID-19 pandemic (Long and Ren, 2021) -is associated with longer distance trips compared to the winter season.

Conclusions
In this paper, we test two specific hypotheses about how regionally targeted interventions impact inter-regional mobility patterns (i.e., outflows) during two interventions in Ontario, Canada. The first, hypothesis was regions targeted by the interventions have higher outflows to other PHRs compared to control regions not-targeted by the intervention. The second hypothesis was that regions targeted by the intervention increase their outflows specifically to those areas not targeted by the intervention. We found little evidence to support these hypothesis in either the first or second intervention, although we did find that in the summer months at the end of the first wave, a regionally targeted lockdown-exit was associated with higher outflows to non-targeted areas relative to before the intervention. However, in the context of the pattern of change in the control regions this increase in outflows was likely in line with increases seen in other regions. Overall, our findings suggest that the regionally targeted interventions in Ontario, Canada had a neutral effect on mobility outflows from the targeted areas. We did find that socio-economic and geographical effects influenced these outflows more substantively. One of our more interesting findings was that outflows were negatively associated with economic dependency which contrasts much previous work about how socio-economic status and mobility were related in response to COVID-19. However, these effects were related to overall patterns of mobility outflows, and not specifically the interventions. Our findings suggest that regionally targeted interventions may thus have neutral effects if they are implemented without strict controls on individual mobility (as was the case within Ontario). These findings should be considered by local/regional governments as they look to implement regionally-targeted interventions in subsequent waves of COVID-19 and in future, similar epidemiological events.