Unequal response to mobility restrictions: evidence from COVID-19 lockdown in the city of Bogotá

ABSTRACT We study the effectiveness of the mobility restrictions imposed by governments to curb urban mobility. We use mobile phone-tracked movements to determine whether users left their homes and explore the role of socio-economic differences across neighbourhoods in explaining their unequal response to lockdown measures. We rely on novel data showing changes in movements in highly disaggregated spatial units in Bogotá, Colombia, before and during the first wave of the COVID-19 pandemic, matched with data on socio-economic characteristics and data on non-pharmaceutical interventions implemented in the period of analysis. We find that the general lockdown imposed in the city significantly reduced mobility (by about 41 percentage points). When looking at the unequal response across locations, we find that low-income areas, with higher population density, informality and overcrowding, reacted less to mobility restrictions.


INTRODUCTION
Mobility reduction was the main objective of the lockdowns implemented during the COVID-19 pandemic.They have also been documented as one of the most effective ways to reduce the spread of cases (Glaeser et al., 2020).Consequently, the ability to comply with lockdowns affects who remains shielded from contagion.We analyse lockdowns implemented in Bogotá, Colombia, from 20 March to 30 August 2020, during the first wave of the COVID-19 pandemic, and estimate the extent to which different areas in the city reacted to these policies.We then explore the role of spatial differences in socio-economic factors in explaining this unequal response.Understanding heterogeneous compliance is essential to guide policy responses both to address future public health crises and other types of urban problems with a spatial dimension.
Bogotá is a well-suited place to study heterogeneous impacts of lockdowns.First, it implemented a city-wide lockdown with uniform enforcement throughout the city for more than seven weeks, starting on 20 March.This general lockdown was eventually gradually lifted, with some economic sectors being allowed to start operations first.Subsequently, the city implemented district-specific lockdowns.We estimate and compare the impact of the city-wide coordinated lockdown with that of localised measures.The city also presents significant segregation of income over space (Castells-Quintana, 2019), which increases the expected unequal response to lockdown across households and locations in the city.
To analyse the heterogeneous impact of non-pharmaceutical interventions (NPIs), we build a unique dataset combining information on city dwellers exact residence location, mobility patterns and socio-economic characteristics at a disaggregated spatial level.We combine this with data on NPIs, including lockdown and cash subsidies, as well as the evolution of COVID-19 cases across the city.
Our findings suggest a significant impact of the general lockdown on mobility, with an average reduction of more than 40 percentage points.In contrast, the impact of district-specific lockdowns is less than 1/10th of the generalised lockdown impact.We find considerable compliance heterogeneity of the city-wide lockdown across neighbourhoods despite homogeneous enforcement.Key neighbourhood characteristics explain this heterogeneity.Neighbourhoods with lower income, higher population density and high informality rates tend to comply less.Overcrowding, measured as households per unit and persons per room, is also associated with lower compliance.
Internet access also proved key in improving neighbourhood compliance as well, most likely due to its connection to telework.
The remainder of the paper is organised as follows.Section 2 summarises the relevant literature and frames the contribution of this paper.Section 3 describes the data.Section 4 describes the COVID-19 pandemic evolution in Bogotá and the NPIs implemented.Section 5 first estimates the effect of lockdowns on mobility.It then estimates neighbourhood-specific compliance to lockdowns.Finally, we explain variation in compliance using neighbourhood's socio-economic characteristics.Section 6 concludes and discusses policy implications.

LITERATURE REVIEW
The literature on NPIs and their impacts on mobility and the evolution of COVID-19 has increased significantly.However, the large majority of studies have focused on developed countries (Dave et al., 2020) and cross-country or cross-city comparisons (see Brodeur et al., 2021, for a survey).Studies focused on developing countries, especially showing analysis for within-city locations, is relatively scarcer.Barnett-Howell and Mobarak (2020) discuss the differences in trade-offs between the benefits and costs of social distancing experienced by developing and developed countries, highlighting the need to study countries of different income levels to fully understand dynamics and policy consequences.This work suggests that NPI compliance can be significantly different in developing countries as mobility restrictions impose a more substantial burden on poorer households.
Even studies that focus on developed countries have found that income affects the impact that NPIs have across subnational regions.For the United States, Wright et al. (2020) find that counties with above-median income comply more with lockdowns, reducing mobility by 60% more than poorer counties.Furthermore, richer locations even reduce their mobility before mandatory measures are put in place.Goolsbee and Syverson (2021) analysed the number of consumer visits to businesses for US counties and did not find evidence of large temporal or spatial shifting in response to lockdown policies once they control for the cumulative number of COVID deaths, suggesting that people were more concerned about exposure to the disease than to mobility restrictions.Gupta et al. (2021) found that the average hours spent at home grew by about 53% between the first week of March and the second week of April, where state-level lockdowns can account for about 55% of the change, with the remaining 45% attributable to private responses of individuals to the pandemic.
For developing countries, as noted, there are fewer studies.Bargain and Aminjonov (2020) show that the overall mobility reduction was smaller in Africa (between 30% and 40%) compared with Latin America (more than 40%).Using a difference-in-difference approach, the authors find that an increase in 1 SD (standard deviation) in regional poverty rate (23.6 points) increases in 7.8 points the work-related mobility during the lockdowns.Carlitz and Makhura (2021) explore the response of different population segments to lockdown orders at the subnational level in South Africa.They find that compliance with confinement orders presented a greater challenge for rural population and poorer regions.Lee et al. (2021) find a reduction of 50% on trips in Pakistan during the period of highest COVID-19 cases surge, measured through a web-based questionnaire.
As mentioned previously, most papers studying the impact of NPIs on mobility reduction during the pandemic have either focused on cross-country or cross-region analyses.Some do analyse mobility reductions at the city level.Ruiz-Euler et al. ( 2020) look at differences in mobility reduction across US cities, finding a mobility gap between higher and lower income groups.Song et al. (2022) use cell-phone data to study mobility fall in the Texan city of El Paso, showing heterogeneous declines depending on the type of destination.Aloi et al. (2020) focuses on the city of Santander, Spain, documenting how the availability of different means of transportation was associated with different magnitudes of mobility changes.
A few papers explicitly analyse heterogeneity in mobility behaviour across different locations within a city.Ramani and Bloom (2021) quantify the effect of COVID-19 on migration patterns and real estate markets within and across US cities, and document a relocation of economic activity away from the city centres and towards lower density areas.Trasberg and Cheshire (2021) analyse mobility in the Greater London of the UK area using mobile applications and explore how differences in socio-economic characteristics may explain the heterogeneity in mobility patterns across neighbourhoods.During the national lockdown, activities declined more in wealthy neighbourhoods (from −20.11% to −11.52%) compared with the London average, while the decrease was significantly less in suburban areas classified as intermediate lifestyle (11.53%) and multi-ethnic suburbs (9.72%).In a developing country context, Pinchoff et al. (2021) conducted a household survey across five urban slums in Nairobi, Kenya, and analysed their mobility patterns, showing that most mobility was limited to locations within the slums during lockdowns.
We contribute to this literature on NPIs and mobility by (1) focusing on a large city of a developing country, (2) exploring heterogenous compliance across neighbourhoods and (3) analysing the potential socio-economic reasons behind this heterogeneous response.To do so, we rely on a novel and unique dataset that includes information on mobility at a highly disaggregated spatial level for the city of Bogotá, before and during the first wave of the COVID-19 pandemic.
We extend the set of socio-economic characteristics that have been studied elsewhere as potential determinants of lockdown compliance and add multidimensional poverty measures, informality, and sectoral composition, as well as average information on a household's home infrastructure.Trasberg and Cheshire (2021) presents the closest approach to ours.We complement their analysis by providing methodological improvements, such as an alternative approach to geographically weighted regression, and the analysis of individual socio-economic determinants of compliance, instead of only indices.We also use mobility measures that identify residential location.In this way, we measure mobility changes in a neighbourhood corresponding to inhabitants that live in that neighbourhood, as opposed to changes general observed traffic that combines local traffic and traffic from other neighbourhoods.This allows us to better connect the local mobility decline to the demographic characteristics of those that see their mobility changed.

DATA
We track mobility using mobile phone pings emitted outside of the users home.Lockdowns were aimed directly at reducing urban mobility and, thus, reductions in mobility are the best measure of their efficacy. 1Our phone mobility data come from GRANDATA, a data laboratory. 2 The United Nations Development Programme (UNDP) in Latin America and the Caribbean and GRANDATA produced these data.
We get access to a measure of changes in mobility for every census tract between any date and a baseline date of 2 March 2020.Since lockdowns began in Bogotá on 20 March, these data allows us to measure the actual mobility impact of mobility restrictions.Our analysis ends on 31 August.
Mobile phones generate pings or signals associated with the user's location at different points in time.The location of this ping is related to a hash of the MADID (Mobile Advertising ID).The average number of geolocation pings per user per day is 130.In Bogotá, residency is determined as the location where the user was present more often in an initial baseline period during the night-time.The residence location is assigned to a hexagon with a diameter (of the hexagon's circumscribed circle) of 40 m.The location hexagons follow the Geohash location system. 3 Those users who have had fewer than 10 daily pings, adding in and out of home pings, for example, because the mobile phone remained off for a long time, or for whom the data did not capture an entire day because all their pings have happened in less than eight hours, were filtered out.Mobility events are defined as pings detected outside of the home residence hexagon; this allows us to identify if people left their homes and is our measure of mobility.No location individual device information is reported to protect the privacy of individual users living within each one of the hexagons.We access the data at the census tract level.
To match the geography of our socio-economic data, we average the traffic growth for all the census tracts in each zoning planning units (UPZs for their acronym in Spanish).UPZs are the smallest unit of analysis for urban planning and zoning in the city.There are 112 UPZs in the 19 districts (i.e., localidades) of Bogotá.UPZs are urban areas smaller than districts but larger than a neighbourhood.Despite their urban policy role, UPZs are very heterogeneous.For instance, UPZs areas range from 0.8 to 9.2 km 2 and their population ranges from 63,000 to 262,000 inhabitants.We then average the growth rates observed every week.This last averaging controls implicitly for day of the week heterogeneity.In the end, we have a weekly panel of traffic at the UPZ level.
An initial concern with our data is how representative phone mobility is of people's total mobility. 4This boils down to smartphone penetration.Recent data show this is very high in Colombia, with 84.5% of households having at least one smartphone at home in Bogotá (according to DANE).These households use these devices for around 4.3 hours daily, while 85% of the population state they never leave their phones at home.This supports the use of smartphone technologies to track population mobility.Another significant caveat to the accuracy of the statistics we present here is related to systematic sampling errors.This is equivalent to asking whether phone tracking and device ownership rates are homogeneous across our spatial units of analysis (i.e., the UPZs).The Multipropósito Survey indicates high smartphone presence across UPZs despite significant differences in socio-economic indicators.The average UPZ has 64% of households with smartphones.Penetration ranges between 37% and 85% across the UPZs distribution.Furthermore, all our mobility measures are normalised using baseline movement.If movement changes, conditional on residing in a UPZ, are similar between those who use a mobile device and those who do not, our measure should capture general mobility declines accurately across UPZs with different device usage rates.
GRANDATA has a random sampling of mobile phones in the smaller available Geohash hexagons (which have a circumscribed circle with a diameter of 40 m).These units are significantly smaller than the UPZs (the smallest (average) UPZ has an area equivalent to an associated circle with a diameter of 0.58 (2.05) km).This ensures comparable coverage across all UPZs.
Figure 1 shows the evolution of mobility in Bogotá.Mobility change is calculated for each UPZ relative to their own mobility benchmark in week 0. It shows for every week since March 2020, the average 5th, 25th, 75th and 95th percentiles of mobility growth.These statistics are calculated across UPZs mobility each week.Lockdown started in week 2. Positive values indicate increases in mobility with respect to the reference week and negative values show a decline.
Figure 2 maps the change in mobility by UPZ.Together, both Figures 1 and 2 show that mobility fell in all UPZs after lockdown, although with noticeable heterogeneity across locations.They indicate that some areas began mobility reduction even before the lockdown.The city's wealthiest areas, located in the east, had higher mobility reductions than other areas.People living in these places were able to work from home and were more adapted to comply with the national government's measures.Reductions in mobility last until week 15.In the meantime, in the south of the city, where poorer UPZs are located, we can find the lowest mobility reductions.In these areas, most people do not have formal jobs and mandatory lockdown means no income whatsoever for these households.High mobility at the south of the city persisted in weeks 3-6 and worsened by week 15, compared with baseline.
To explore the role of socio-economic characteristics, we match our mobility data with data from the metropolitan 2017 household-level survey, called the Multipropósito Survey, and carried out by DANE.This survey includes data on the labour market, housing conditions, poverty and demographic characteristics.Information about households and individuals is representative at the UPZ level for 73 out of the 112 UPZs. 5  Although our focus is on mobility, we also look at the evolution of COVID-19 cases.However, using data on cases has multiple challenges, as COVID testing is not random nor uniform, affecting disease prevalence measurement (Badr et al., 2020;Niehus et al., 2020).Besides, in Colombia, the distortions, especially at the beginning of the pandemic, were large because there were long waiting lists for tests processing due to a shortage of locations able to process them.With these caveats, we explore the connection between mobility reduction and the evolution of cases.We use COVID-19 daily cases from Bogotá's Secretariat of Health.Around 70% of the cases reported the patient's residence address.We geocoded this address and aggregated it at the UPZ level.The database contains the date of the laboratory result's diagnosis.
Table 1 presents some descriptive statistics at the UPZ level.Variables gather information about COVID cases, change in mobility, subsidies, demographics, labour market, household infrastructure and sector-of-employment-specific data.Starting with COVID-19 cases, we can see large heterogeneity across locations, from 151 in San Isidro Patios-Chapinero, in the east of the city, to 7040 in El Rincon-Suba, in the north-west.This reflects how the pandemic had different influence among locations within the same city.The possibility to stay at home is likely to be one of the main factors for this heterogeneous impact.Regarding mobility, as it can be seen, there were important differences across locations: while the average reduction in mobility during the entire period for poorer zones was around 7%, in more affluent zones it was 55%. 6 And, as it can be seen, there are significant differences in terms of income.This is also reflected looking at incidence of monetary poverty.The Usme and Ciudad Bolivar districts, with a combined population of more than 1 million people (around 13% of the population), and located in the south, are among the poorest areas of the city.Labour market outcomes are also quite heterogeneous.Informality in poorer UPZs is higher than 47%, while in more affluent UPZs is below 20%.We find a similar pattern in unemployment: in poorer areas it can be as high as 14.4%, while in the more affluent zones as low as 2%.Regarding economic sectors of employment, richest UPZs concentrate workers who can do their jobs at home and health workers essential to contain the virus.For instance, the share of jobs in the education and healthcare sectors is positively correlated with per capita income, with a correlations of 0.7.By contrast, poorer UPZs concentrated more workers with less possibility of teleworking.There is also a larger proportion of construction workers in the poorest UPZs, around 11%, compared with less than 3% in the richest ones.Regarding demographics, it is also important to note that the population is much younger in poorest UPZs.In the wealthiest five UPZs, for instance, 10% of the population is above 65 years, while this is only 5% in the poorest UPZs.Poorer UPZs are also denser and more overcrowded, have fewer years of education, and have worse housing infrastructure.Finally subsidies were distributed by the government to poor households as a measure to help them stay at home.As expected, city authorities gave more subsidies to poorer UPZs.Virtually all UPZs in Bogotá had households that received subsidies.There are areas with a higher concentration of subsidies, mainly those in the south and south-west, and some of lower income neighbourhoods in the north-west (see Figure A5 in Appendix A in the supplemental data online). 7

THE COVID-19 PANDEMIC IN BOGOTÁ
In this section, we provide a quick overview of the context of the first wave of the pandemic in the city of Bogotá, as well as the different NPIs implemented (at the city and district level).
The first reported case of COVID-19 in the country happened in Bogotá on 6 March.However, according to the city's administration, more than 210,000 people came into the country from Europe or the United States, where the virus was already circulating through air travel between January and February. 8The lack of international travel restrictions is blamed for the rapid spread of the virus.As of 30 November 2020, there were 374,074 COVID-19 cases, which led to 8505 deaths. 9Among the positive cases, 51.29% were women, and the average age was 39 years old.The spatial distribution of cases describes the uneven nature of COVID-19 and slightly follows poverty distribution (see Figure A1 in Appendix A in the supplemental data online).Poor areas saw more cases and are also the ones with fewer mobility declines.They also had higher population densities and worse household infrastructure.

Non-pharmaceutical interventions (NPIs)
The Bogotá government was the first to announce a lockdown drill for 20-23 March.This announcement was followed by a national lockdown presidential declaration, which began on 24 March at midnight and was planned to end on 12 April at midnight.As cases surged, the lockdown was extended to 27 April.During this lockdown, only sectors considered as fundamental were able to work, including transportation, food provision, healthcare and deliveries.Some banks and notaries were partially open also.
After 27 April, the national government allowed the reopening of activities for the construction and manufacturing sectors; companies were allowed to resume operations under local governments' surveillance and authorisation.Bogotá stayed closed for two more weeks while they approved the companies that were to start operations after visiting them to verify safety protocols.
Lockdown was extended for the general public until 11 May.
After the first city-level lockdown was lifted, cases surged.The city started implementing localised restrictions by district, according to the prevalence of cases.On 30 May, the first was implemented for Kennedy district, in the south-west of the city, for two weeks from 1 to 14 June.Afterward, restrictions for Ciudad Bolivar, Engativá and Bosa districts, in the south of the city, followed.These districts were closed until 30 June.On 13 July, Ciudad Bolivar, San Cristobal, Rafael Uribe, Chapinero, Santa fe, Usme, Martires and Tunjuelito, in the south and south-east of the city, started lockdown until 26 July.These district-level restrictions continued until 30 August.Figures A2 and A3 in Appendix A in the supplemental data online show the district-level restrictions timeline.
As a first exploratory analysis of lockdown in mobility, we show the relationship between the general lockdown and mobility in a binned scatterplot in Figure A4 online.Measurements in this graph are made weekly.The value of the horizontal axis denotes the share of days under lockdown in the considered week.Negative numbers in the vertical axis refer to mobility drops with respect to the baseline date of 2 March 2020.The graph shows that the fall in mobility was significantly larger during weeks with a higher proportion of days under lockdown, which indicates that the measures may have had an association with the decline in mobility.

Mobility and COVID-19 cases
As mentioned above, the relationship between mobility and COVID-19 cases has been widely established.In Bogotá, as simple exploration also suggests that COVID-19 cases are associated with mobility.We run a regression of one week lagged mobility on COVID-19 cases, controlling for week and UPZ fixed effects, and find a significant elasticity of mobility to cases of 1.23.This suggests that a drop in mobility of 1% is associated with a decline in the growth of cases of 1.23%.Results are provided in Figure A6 in Appendix A in the supplemental data online. 10 But working with COVID-19 cases is difficult, as discussed. 11Besides these difficulties, our focus on mobility stems from an interest in place based policies.One of the contributions of the paper is to analyse the compliance of a city-wide policy that has significant heterogeneity in adoption.This, together with the established importance of mobility as a way to reduce the disease incidence, and the measurement issues, motivates our choice of mobility as the primary dependent variable.

EMPIRICAL ANALYSIS
In a first stage, we estimate the impact of lockdown restrictions on mobility by estimating different specifications of the following equation: (1) using a weekly panel of mobility data at the UPZ level.M it denotes the mobility change with respect to the baseline in the week t for UPZ i. LockDown t is an indicator variable for the weeks in which the city-wide lockdown was active.g i and t t are UPZ and week fixed effects that control for time invariant UPZ characteristics and common time trends, respectively.h captures the average effect of the general lockdown on mobility.b i measures the mobility change observed in UPZ i, above or below the average change h, that is, a UPZ compliance premium. 12 For robustness, in some specifications we further include UPZ-specific time trends g i × t t .
We estimate equation (1) using two main samples.One uses all weeks starting from our baseline week (the week starting on 2 March 2020) includes the eight-week general lockdown period, and extends to 12 weeks after the baseline, right before district lockdowns began.The other sample includes all weeks in our data, from baseline week 0 to week 30.The first more restricted sample allows for the cleaner analysis, as the lockdown measure applied to all UPZs is homogeneous.The extended sample is larger, which improves power, but requires that we control for the district lockdowns.However, our main focus remains on the effect of the general lockdown. 13 The general lockdown, the treatment of interest, was applied at the same time to all our UPZs.Thus, identification of the policy impact relies on the differential mobility evolution across UPZs.A critical assumption is that of no differential anticipatory effects across UPZs.This assumption is supported by the simultaneous implementation and equal levels of enforcement for all UPZs with only a couple of days' notice. 14As in the first waves of the pandemic the situation was evolving rapidly, this is a plausible assumption. 15 Then, in a second stage, we explain the variation in the estimated coefficients b i , or UPZ compliance, using UPZ's socio-economic characteristics by estimating the following equation: where P i , L i , D i and S i are vectors of variables measuring UPZ's income, aggregate poverty, labour market, demographics, infrastructure and other characteristics, as presented in Table 1.
The u parameters explain the role of the initial socio-economic characteristics in explaining the heterogeneity in the mobility changes across UPZs as a reaction to the general lockdown. 16 Note that the location compliance premia, b i , is itself an estimated variable and therefore has some estimation error.In the second stage, we are not taking that estimation error into account.This can be overcome by estimating both equations in a single stage, where equation ( 2) substituted into equation ( 1).This solves the problem of the estimation error of the b i used as dependent variable in the second stage.We provide these results as a robustness and obtain almost identical coefficients.A single stage, however, does not deliver the valuable estimates of heterogeneous compliance across UPZs that the two-stage approach does.Thus, we present the twostage version as the benchmark.

The impact of lockdown on mobility
Using our weekly panel, we start by looking at the impact of general lockdown measures on mobility estimated through parameter h in equation ( 1).Table 2 shows the results.In columns 1-2 we include location fixed effects at the UPZ level, while in columns 3-4 we further include week fixed effects.Once we control for UPZ and week fixed effects (see column 3), results suggest a decrease of around 41 percentage points in mobility compared with baseline mobility (week 0).In columns 2 and 4 we include district-specific lockdowns in the set-up of equation ( 1).Both general and district-specific lockdowns cause a decline in mobility.However, after controlling for the general lockdown's impact, UPZ and week fixed effects the effect of district-specific restrictions' is minor.It stops being significant if we control for lockdown heterogeneous effects instead of the UPZ specific trend (see column 8).Table A1 in Appendix A in the supplemental data online further explores the behaviour of mobility a week before and after the lockdown.The coefficient in column 4 for the week before lockdown with UPZ and week fixed effects shows that even before the implementation of mobility restrictions, voluntary reductions of mobility are detected.Also we find that the average effect on mobility lowers but stays relatively high even a week after the lockdown (see column 6).Table A2 in Appendix A in the supplemental data online checks the robustness of our results to measuring lockdown continuously from 0 to 1 depending on the share of days within the week affected, with 1 being complete lockdown the whole week.We also test the robustness of our results to the inclusion of general time and UPZ-specific time trends.In all cases, we find a significant reduction of mobility due to lockdown.Finally, we perform a simple placebo test by generating random assignment of lockdowns across weeks and UPZs.As expected, we find no significant effect of the placebo.
Table A3 in Appendix A in the supplemental data online includes subsidies provided by the government as a control variable.As can be seen, the main effect in our preferred specification of general lockdown in reducing mobility is not affected by the inclusion of this policy.In column 4, we allow for potential non-linearities by introducing the square of subsidies per capita.The interaction between linear and square subsidies with lockdown are not statistically significant.

The role of socio-economic characteristics on the unequal response to lockdown
So far, our analysis has shown a very consistent and robust impact of the general lockdown in reducing mobility in Bogotá.In this section, we now explore the role of socio-economic characteristics in explaining the heterogeneous response to lockdown measures across within-city locations.We recover differential effects on the change in mobility for each location (i.e., b i coefficients from equation 1).The dependent variable in these regressions is the change in mobility in the UPZ with respect to the baseline week.The estimated b i s compare mobility between weeks that were part of the lockdown and weeks that were not, controlling for the average impact of lockdown, h, and taking into account the first 12 weeks of NPIs.These results correspond to column 7 of Table 2 where we obtain a fall in the average mobility of 56.5 percentage points after the general lockdown was established.b i therefore gives the additional percentage growth or decline in mobility with respect to the average effect of the lockdown.These differential effects are shown in Figure 3. 17  The left panel of Figure 3 shows the b i point estimates in the city map, and the right panel shows the distribution of these estimates.The strongest decline in mobility happened in high- income UPZs, most located in the north of the city.In some of these wealthy areas it was notorious for many households fleeing to their vacation homes, a pattern detected, for instance, in the wealthy neighbourhoods of New York City (Coven et al., 2020;Kim et al., 2021).In these locations, the point estimate for the change in mobility beyond the average is around −25 percentage points, which amounts to a total decline of about −80 percentage points with respect to the baseline date.By contrast, the weakest decline in mobility is observed in the south-west and south-east of the city, where there is a high concentration of low-income neighbourhoods with high informality levels.In these locations, the point estimate for the UPZ specific mobility change is around 35 percentage points, implying a total change in mobility of around −21 percentage points.Despite the heterogeneity in compliance, all UPZs reacted to the lockdown with an absolute decline in mobility.
We now estimate specifications of the general form of equation ( 2) and explore how different socio-economic characteristics explain these differential responses to lockdown.The socio-economic characteristics are standardised to ease comparison.Figure 4 shows the main results of regressions in Tables A4 and A5 in Appendix A in the supplemental data online.Each group of coefficients, identified by colour and marker, comes from a separate regression. 18 First, we find that higher income per capita is associated with a stronger decline in mobility following the lockdown: an increase of 1 SD in this variable is associated with an additional decline of 6 percentage points in mobility.Similarly, an increase of 1 SD in the multidimensional poverty index (MPI) (Human Development Initiative, 2018; UNDP, 2015) is associated with 5 percentage points higher mobility. 19Informality, as expected, is also associated with lower compliance.Together, these results show the decisive role of income in lockdown compliance, in line with Bargain and Aminjonov (2020) 2).Each group of coefficients, identified by colour and marker, comes from a separate regression.For regression statistics, see Tables A4 and A5 in Appendix A in the supplemental data online.
with different income levels.Results are also in line with the idea of a relative higher cost of staying at home for lower income households (Wright et al., 2020).In a developing country context, the relative cost of staying at home for lower-income individuals is expected to be even higher.
Second, we analyse the sectoral composition of workers across UPZs.Related work has highlighted that professions associated with teleworking capabilities are concentrated in higher income households (Papanikolaou & Schmidt, 2022), providing another possible mechanism for higher compliance beyond income.Using the sectoral information from the Multipropósito survey, we build a measure of the share of workers per sector in each UPZ.We remove any sector that presents less than 1% of workers.Only construction resulted to be positively related with mobility which is consistent with the fact that this was one of the first sectors to start operations.Otherwise, we find little significant role of sectoral composition in explaining heterogeneous compliance across UPZs.However, as hinted by Table 1, a possible explanation is the relative dispersion of individuals in different sectors across the areas of the city.
Third, and in terms of demographics, we find that education, measured as the average number of years of education for household members above 12 years of age, is significant and associated with higher compliance.An increase of 1 SD, corresponding to one additional year in UPZ average education, is associated with 4 percentage points lower mobility, in line with Haug et al. (2020) and Brzezinski et al. (2020).Other demographic characteristics show behaviour corresponding to different levels of risk: UPZs with a higher share of younger people comply less, while those with a higher share of older and married people comply more.
Fourth, we look at the role of home infrastructure.The city shows a significant prevalence of precarious home conditions.For example, the share of households without a fridge rises to 15% in the UPZ with the worst numbers, while the share of households without a cooking stove rises to 10%.But these factors do not seem to critically influence compliance.Access to home internet, however, is highly associated with compliance: an increase of 1 SD in the share of households with internet access is associated with a decline of mobility of an additional 7 percentage points, the strongest of all factors studied.The access to home internet variable still shows large variation across households and UPZs, ranging from 36% to 85%.
Finally, population size and density, as well as overcrowding, are all associated with lower compliance.An increase of 1 SD in overcrowding, for instance, is associated with 5 percentage points higher mobility in the UPZ.This result is in line with previous evidence showing that the sharing of a smaller space makes COVID-19 incidence higher (Furman Center, 2020).The result for overcrowding is a much stronger factor than mere residential density, as for instance in results for the United States.
One additional concern is related to using compliance measures, b i , that are themselves estimates.As discussed in section 5, a single-stage regression, where equation ( 2) is substituted into equation ( 1), overcomes this issue as it allows direct estimation of the effect of socio-economic characteristics on mobility.Single-stage regressions are shown in Figure A7 and Tables A6 and  A7 in Appendix A in the supplemental data online.Point estimates and confidence intervals are very nearly identical to the two-stage results, providing reassurance that the estimation error of b i was not a significant concern.A single-stage regression increases the number of observations and improves power.A single stage, however, does not deliver the valuable estimates of heterogeneous compliance across UPZs that the two-stage approach does.Thus, we keep the twostage version as the benchmark.

DISCUSSION AND CONCLUSIONS
In this paper, we have studied the effect of lockdowns on mobility in Bogotá, a large city in the developing world.We have analysed this measures' impact at a detailed spatial level, looking at differences in levels of compliance across locations within the city.To do so, we relied on a unique and novel dataset merging localised data mobility, policies implemented during the pandemic, socio-economic characteristics and the evolution of COVID-19 cases.According to our most preferred specification, we have found that the city-level lockdown reduced mobility, on average, by around 41%.Beyond this effect, district-specific lockdowns seem to have had small marginal effects on mobility.
Our results are consistent with those of Trasberg and Cheshire (2021), a study with the most comparable approach to ours.They find a larger drop in average mobility in Greater London of 80%.Like us, they also find heterogeneity in compliance across locations.In our case, we find an even sharper difference in compliance between Bogotá's most affluent and poorest neighbourhoods.Surprisingly, the reduction in mobility in wealthier areas was roughly the same in both studies, while there is a contrasting difference in the compliance of the poorest neighbourhoods in both cities.While in Greater London the compliance, or decline in mobility, was 9.7% less in lower income areas than in the city average, in Bogotá it was 35% lower.This difference highlights that poorer citizens are less able to comply with lockdown restrictions and maintained relatively higher mobility in less developed contexts, making them more vulnerable to the virus.In contrast, the context of a developing country did not make that much of a difference for the wealthiest families.
Our analysis suggests that differences in socio-economic characteristics across within-city locations partly explain this unequal response in mobility.The lockdown impact on mobility was smaller in areas with higher population density, informality and overcrowding.There is some degree of spatial dependence across districts; mobility reductions in one spatial unit seem associated with mobility reductions in adjacent units.However, when taking into account spatial correlation, our main results hold.
This study provides additional evidence with respect to similar results found by Kim et al. (2021) for New York City, a big city in a developed country.Regarding the determinants of mobility, we provide evidence that for Bogotá, income-related variables such as education have a positive effect on lockdown compliance, while poverty has a negative impact.These results are in line with Bargain and Aminjonov (2020) at a regional level or Wright et al. (2020) for US counties, and Ruiz-Euler et al. (2020) comparing across cities of different income levels.Also, results showing that overcrowded units with no remote work (construction or transport sectors) are more prone not to comply were highlighted by the Furman Center (2020) for New York City.
In large developing cities such as Bogotá, where inequalities are already high, the unequal impact of the COVID-19 pandemic is at the same time reflecting and aggravating the reality of socially fractured urban areas.Addressing this urgent challenge has become more evident than ever.Understanding differences in response to policies can help better target public spending and government interventions during critical moments, such as new general or localised lockdowns.
These results also entail new challenges for policymakers in order to develop effective solutions in situations of emergency.Intrinsic limitations of people given by their income, by the characteristics of their houses or by the economic sector in which they are working must be taken into account to design lockdown policies.Otherwise compliance with NPIs will be small representing a risk, not only to these populations but also to the entire society.
13 a is the average effect of district-specific restrictions.Recall that districts are areas that include multiple UPZs and that all district-specific lockdowns were implemented after the citywide lockdown ended, with no time overlap. 14Anecdotally, the short notice even caused waves of people to encounter problems when returning to the city from weekend trips. 15In contrast, district specific lockdowns were implemented according to local COVID cases surges, an endogenous variable to mobility.This makes anticipatory effects likely.We focus on the effect of the main lockdown and use the UPZ-specific lockdown indicators as a control and do not interpret them causally. 16The two-stage strategy is similar to that followed in other urban economics papers such asCombes et al. (2008) and Roca and Puga (2017), where city productivity is estimated in a first stage and then regressed on explanatory variables in a second stage. 17Hainmueller et al. (2019) highlight two problems of this type of multiplicative interaction models.First, these models assume a linear interaction effect that changes at a constant rate with the moderator.This is an issue when the treatment variable (lockdown in our analysis) is either binary or continuous and the moderator (UPZ indicator in our analysis) is continuous.When both are binary, as in our case, they suggest using a fully saturated model that dummies out the treatment and the moderator and includes all interaction terms, the approach we follow.Second, estimates of the conditional effects of the independent variable can be misleading if there is a lack of common support of the moderator.This is not an issue, again, because all levels in the interaction terms are parameterised as dummy variables. 18We estimate independent regressions taking into account that there is a high correlation between several of the socio-economic variables presented in Table 1. 19The MPI is a comprehensive poverty measure that complements traditional monetary poverty measures with deprivations in health, education, labour market outcomes and living standards.

Figure 1 .
Figure 1.Evolution of mobility in Bogotá.Note: Mobility change is calculated for each zoning planning units (UPZ) relative to their own mobility benchmark in week 0. Average and percentiles are calculated across UPZs mobility change each week.

Figure 2 .
Figure 2. Mobile phone mobility growth.Note: The map shows the average weekly percentage growth rate with respect to the baseline date (2 March 2020).Week number is determined relative to the baseline week.Week 1 is the week previous to the general lockdown.Week 3 is the first after lockdown.Week 9 is the last week before the lockdown lift.

Figure 3 .
Figure 3. UPZ relative reaction to the general lockdown.The values for each UPZ come from the coefficients that allow for a heterogeneous response to the general lockdown (the b i location mobility premia estimated in equation 1).
when comparing regions of developed and developing countries, Wright et al. (2020) for US counties, and Ruiz-Euler et al. (2020) analysing cities

Figure 4 .
Figure 4. Results from the second-stage regressions that explain the heterogeneous reaction to the general lockdown across UPZs.Note: The coefficients are equivalent to the us from equation (2).Each group of coefficients, identified by colour and marker, comes from a separate regression.For regression statistics, see TablesA4 and A5in Appendix A in the supplemental data online.

Table 1 .
Descriptive statistics for the zoning planning units (UPZ) in our estimation sample.

Table 1 .
Continued.: Mobility data range from 2 March to 30 August 2020 and are available for 112 UPZs.Case data range from 2 March to 4 October 2020.Socio-economic data come from the 2017 Metropolitan Household Survey (Encuesta Multipropósito) and are available for 73 UPZ. Note

Table 2 .
Impact of general lockdown on mobility.Robust standard errors are shown in parentheses.UPZ Time Trends interacts with UPZ and week effects: γi × τt.UPZ Lockdown Effect allows for the effect of the general lockdown to be heterogeneous by adding a term that interacts the lockdown indicator with UPZ indicators: LockDowntt × γi.