Cities with Cell Phone Records

Measuring Commuting and Economic Activity inside Cities with Cell Phone Records Abstract We show that commuting flows constructed from cell phone transaction data predict the spatial distribution of wages and income in cities. In a simple workplace choice model, commuting flows follow a gravity equation whose destination fixed effects correspond to wages. We use cell phone data from Dhaka and Colombo, covering hundreds of millions of commuter-day observations, to invert this relationship. Model-predicted income at the workplace level predicts self-reported survey workplace income, and model-predicted residential income predicts nighttime lights. In an application, we estimate that predicted commuter income is 4-5% lower on days with hartals (transportation strikes) in Dhaka. 1 1 Introduction Measures of urban economic activity at fine temporal and spatial scales are important yet scarce. Such data is necessary to understand how cities respond to localized shocks such as changes in transportation infrastructure or floods, and to help governments target scarce public resources. These issues are especially salient in large cities in developing countries, which are growing fast yet are least covered by conventional data sources. For example, less than 10% of the urban population in sub-saharan African countries is covered by a census of firms with wage data.1 At the same time, comprehensive new data sources on urban behavior, especially individual mobility, are becoming available across the world. In this paper, we provide a theory-based method to predict the spatial distribution of urban economic activity from commuting choices.2 The revealed-preference logic of our approach is simple. A core function of cities is to connect workers and jobs. While many factors enter into workplace choice decisions, areas with high wages should disproportionately attract workers, keeping distance and home locations fixed. We propose inverting this reasoning to infer the relative average wage at a location based on how “attractive” it is as a commuting destination. We use tools from urban and trade models to formalize this intuition. In the model, work location decisions aggregate up to a gravity equation on commuting flows, and destination fixed effects are proportional to log wages. This property holds for a general class of urban models developed to evaluate urban policies and transport infrastructure (e.g., Ahlfeldt et al. 2015; Heblich et al. 2018; Tsivanidis 2018; Severen 2019). We implement our approach using call detail record (CDR) data from two large metropolises: Colombo, Sri Lanka and Dhaka, Bangladesh. CDR data is a prototypical example of “big data” available in developing countries, and it contains phone user location for every transaction (phone call or text message). We construct individual commuting trips by observing a user’s location at different times during a single day, which leads to almost half a billion days with commuting information. We show that commuting flows constructed this way correlate strongly with commuting flows from a transportation survey from Dhaka, while additionally offering very fine geographic resolution and daily time variation. We pool the data over time and estimate the gravity equation implied by the model. We return to the high-frequency temporal aspect of the data in an application below. Estimated wages by location are derived solely from observed commuting decisions and 1 Authors’ calculation (Appendix A.7). Moreover, government statistics may sometimes be unreliable. For example, Nigeria and Japan recently made consequential revisions to jobs and output data (Financial Times, 2014; Japan Times, 2019). 2 Several papers use big data sources to empirically predict economic indicators (Blumenstock et al. 2015; Jean et al. 2016; Glaeser et al. 2017). Our approach is to use economic theory to guide how we interpret the data. 2 data on travel times, without any model training with actual wage data. We next assess how well this simple measure captures real differences in wages, using two income proxy data sources. First, we compare average model-predicted workplace income with commuter income data from a large transportation survey covering 1% of the population of Dhaka. Second, we compare in each city model-predicted average residential income with high resolution satellite nighttime lights, an intuitive proxy for residential income. Model workplace income is significantly positively correlated with survey workplace income, including after controlling for location employment density and distance to the central business district (CBD). These results hold robustly when repeating the exercise with residual log income, after projecting out demographic and occupation covariates from the survey. Hence, commuting choices encode valuable information on which urban areas are productive. At the same time, model-predicted income only explains around a quarter of the spatial variation in survey income. These results are related to a contemporaneous conversation on the ability of structural urban models to fit urban data. Interestingly, Severen (2019) finds that model wages estimated using census tract commuting flows barely predict tract-level wages in Los Angeles; this result may be related to the small size of the spatial unit, as well as non-wage factors determining commuting choices. Tsivanidis (2018) calibrates and estimates a general equilibrium model in Bogotá, and finds that model-predicted wages across 19 urban areas predict survey wages. By comparison, our focus in this paper is using a parsimonious model together with detailed commuting data to predict relative income. In the second validation exercise, in both cities, model-predicted residential income is a robust predictor of satellite nighttime lights. This relationship remains stable within sub-districts and after controlling for residential density and distance to the CBD. The explanatory power is significantly higher (R2 over 0.8), in part due to the inclusion of peri-urban areas in the analysis. A key advantage of the model is that we can compute how income “moves” around the city. In both validation exercises, we perform a horse-race between residential- and workplace-income. While the two measures are highly correlated, we find suggestive evidence that model workplace income better correlates with workplace survey income data, and model residential income better correlates with nighttime lights. The ideal application of our income-prediction method and of the high-frequency commuting data is to trace out the spatial and temporal impact of urban events and policies. To illustrate this potential, we estimate the economic cost of hartals, a type of strike intended to disrupt transportation and economic activity in Bangladesh. The daily commuting data exhibits visible differences between hartal and workdays, and model-predicted income falls by around 4-5% on hartal days. This effect is driven mostly by the extensive margin (fewer trips), and, 3 to a lower extent, by larger effects for high-income commuters. While precisely estimated, these changes are relatively small; these results are in line with previous studies of hartal. In the conclusion, we revisit how our project fits into a larger approach of using economic theory together with the rich choice information made available by “big data.” 2 Cell-Phone Data and Commuting Flows 2.1 Data Sources Cell phone transaction data. We use call detail record (CDR) data from large operators in Sri Lanka and Bangladesh to compute detailed commuting matrices. CDR data includes an observation for each transaction, such as outgoing or incoming voice call and text messages, or GPRS internet connections. Each observation has a timestamp, the anonymized participant user identifiers, and their cell tower locations. Towers are unevenly distributed in space; they are denser in urban and developed areas. We focus on the greater metropolitan areas around the capital cities of Colombo and Dhaka. The data covers a little over a year in Sri Lanka and four months in Bangladesh in the early 2010’s.3 We construct commuting trips by observing a phone-user connect to towers at different times of the day. On a given day, we define a user’s origin as the location of the first transaction between 5am to 10am, and the user’s destination as the location of the last transaction between 10am and 3pm.4 By definition, a user has at most one commuting trip per day. If the origin and destination correspond to the same cell tower, we say that the user was stationary, otherwise the user made a proper commuting trip. If transaction data is missing in either time interval, commuting behavior is not observed for that user-day.5 We then aggregate over users and non-holiday weekdays to obtain an origin-destination (OD) matrix of commuting flows between every pair of cell towers. One potential concern is that cell phone data is not representative of urban commuters. Cell phone ownership is high in both countries, but the ownership may vary systematically by demographic characteristics. A separate concern is that calling behavior may be correlated with mobility. To address these issues, in section 2.2 we compare the commuting flows constructed here with a representative transportation survey. 3 In Bangladesh, the data only covers outgoing voice calls. Our sample covers the Western Province in Sri Lanka, and the Dhaka, Narayanganj, and Gazipur Districts in Bangladesh. 4 We focus on the morning commute, as other types of trips (e.g., shopping) are more likely in the evening (Frank and Murtha, 2010). 5 Commuting data (including stationary trips) is available for 16% (in Dhaka) and 29% (in Colombo) of the theoretical maximum number of user-days, that is, if we observed each user on every day in the sample (Table C.1). 4 For robustness, we also construct commuting flows using individual-level “home” and “work” locations, defined as the modal locations during weekday nighttime and daytime periods, respectively, using all data for an individual. The flows using the two methods are strongly correlated and all our validation results are robust to this choice (Appendix Figure B.3, panel B). Google Maps travel time. As a proxy for travel costs, we obtain estimated typical driving travel times between pairs of cell towers using the Google Maps API. Because of the large number of bilateral pairs, in each city we obtain Google data for 90,000 randomly selected pairs of towers, and interpolate to pairs with nearby origin and nearby destination.6 Household transportation survey. We use individual survey data from the 2009 Dhaka Urban Transport Network Development Study or DHUTS (JICA 2010). The survey covers 16,394 randomly selected households in the Dhaka City Corporation (DCC), Dhaka’s urban core, as well as a sample of 1,716 households outside the DCC. Home and work locations are recorded at the level of 108 “survey areas.” Our main analysis sample consists of 12,510 commuters who live and work within the 90 survey areas inside the DCC and who report positive income from work, excluding students, homemakers, and the unemployed. In the main analysis, we exclude households outside of DCC, because the 18 corresponding survey areas are significantly coarser and detailed information on sampling is not available. Our results are robust to including all commuters who live and work inside the DCC (Appendix Table C.5). 2.2 Cell Phone Data Captures Aggregate Commuting Behavior Accurately In Dhaka, commuting flows derived from cell phone data are strongly related to those from the DHUTS commuting survey. To show this, we aggregate the cell phone commuting data up to the level of survey areas. The sample consists of 7,915 pairs of distinct survey areas with positive cell phone commuting flows, with a total of 12,510 trips in the DHUTS survey (this includes government workers) and over 18 million trips between 1.5 million tower pairs in the cell phone data. Commuting flows from the two data sources are strongly related, including when we control for log travel time, origin and destination survey area fixed effects (Appendix Table C.2), consistent with previous research validating cell-phone-based commuting flows (Calabrese et al., 2011; Wang et al., 2012; Iqbal et al., 2014). In addition, the decay of commuting flows with travel time is virtually identical between the two data 6 Appendix A.6 describes the interpolation procedure in detail. We collected data for Sri Lanka in 2016 and for Bangladesh in 2017. We extracted travel time without traffic congestion (Google did not provide travel time with traffic congestion in Bangladesh in 2017). 5 sources (Appendix Figure B.3, Panel A). Commuting flows constructed using “home” and “work” locations also decay at the same rate with log travel time (Panel B). Residential population density from from cell phone data is also strongly correlated with population density from the census (Appendix Table C.3). The adjusted R-squared is 0.61 in Dhaka and 0.49 in Colombo. The slope is 1.16 for both cities, hence cell phone data slightly over-represents population in denser areas. 3 Model: Commuting Flows, Gravity, and Wages Is it possible to infer the spatial distribution of wages from commuting flows? The interaction between wages and commuting costs to determine urban structure is fundamental in classical urban economics models (Alonso, 1960; Mills, 1967; Muth, 1968). Here, we explore this insight using a new generation of models inspired from the trade literature, designed to better match spatially disaggregated urban data (Ahlfeldt et al., 2015). In the model, commuters decide their work location taking into account wages at different potential work locations, commuting costs, and destination-specific idiosyncratic utility shock. Together with a parametric assumption on utility shocks, this implies that log bilateral commuting flows follow a linear gravity equation, with destination fixed effects capturing log wages. Furthermore, this relationship holds in equilibrium regardless of how wages are determined. 3.1 Workplace Choice Model Space is partitioned into a finite set of locations L, which may serve as both residential locations and work locations. In our application, these correspond to Voronoi cells around cell phone towers (depicted in Appendix Figures B.1, B.2 for the two cities). There is a unit mass of workers, and each worker ω sequentially decides where to live, and then where to work. We do not impose any restrictions on the home location choice. (Assuming joint home and work location choice leads to the same gravity equation (Ahlfeldt et al., 2015).) Given her residential location (or origin) i, the worker chooses her work location (or destination) j. The utility of worker ω residing in location i if she chooses destination j is: Wj Zijω Uijω = (1) τ Dij Wj is the wage per effective unit of labor supply at location j (all firms at location j offer the same wage), Dij is the travel time between i and j, and Zijω is an idiosyncratic utility shock that is i.i.d. following the Fréchet distribution, with scale parameter T and shape parameter 6 . In our baseline model, we assume that each worker supplies one unit of labor, and hence earns income Wj if she works in location j. In particular, we abstract from heterogeneity due to skill or other worker attributes.7 Each worker observes the shocks Zijω and chooses the work location j where Uijω is maximized. The probability that a worker commutes to j conditional on residing in i is given τ  P τ  by πij = (Wj /Dij ) / s (Ws /Dis ) . Taking logs, and denoting log quantities by lowercase letters: ! log(πij ) = wj − τ dij − log X exp (ws − τ dis ) (2) s 3.2 Estimating the Gravity Equation We estimate equation (2) through the following empirical gravity model: log(πij ) = ψj − β log(Dij ) + µi + εij (3) where µi and ψj are origin and destination fixed effects, and εij accounts for measurement error or other unmodeled factors. Most importantly, ψj is proportional to the (relative) log wage at j with a factor of , the Fréchet dispersion parameter. (As usual in discrete choice models, we can only identify wages up to scale.) Our main goal is to recover the ψj ’s from observed commuting choices. Note that, for this purpose, it is not necessary to model explicitly how wages are determined in equilibrium. In other words, this mapping between commuting choices and wages holds in any general equilibrium model that micro-founds the gravity equation for commuting flows with a discrete commuting choice model. We implement two approaches to estimate ψj . First, we simply estimate (3) by OLS. Second, we estimate the equation imposing the formula for µi given by the structural gravity equation (2).8 The two approaches yield very similar results (Section 3.4). Lacking detailed bilateral commuting flow data, some authors estimate log wages with an exactly identified procedure using residential and employment populations, and calibrated or separately estimated parameters (Ahlfeldt et al., 2015; Tsivanidis, 2018). Our approach 7 We model and investigate empirically two extensions where labor supply varies across individuals. First, in Appendix A.2, labor supply (and hence income) depends on observable demographics. Second, in Appendix A.3, Zijω and Dij partly affect labor supply, rather than only affecting utility, as in the main analysis. We develop a method to estimate how much Zijω and Dij affect income using survey income data. The results are consistent with Dij being a pure utility shock, and Zijω partly affecting income (Appendix Table C.11). 8 Specifically, we estimate (3) following an iterative procedure (Appendix A.1). In each iteration, we estimate (3) without origin fixed effects, after subtracting the model origin terms from the previous iteration from the left-hand side. We iterate until the vector of destination fixed effects converges. The procedure is identical to SILS (structurally iterated least squares) proposed in the trade gravity literature (Head and Mayer, 2014), except without destination fixed effect constraints. See Fally (2015) for potential bias without imposing model constraints. 7 using commuting flows is more robust against noise in the gravity equation (3), and allows us to test the stability of our results (Appendix A.4). 3.3 Mapping Model Locations to Geographic Areas A key advantage of the model is that model locations can be mapped directly to two-dimensional urban data. However, since in the model productivity shocks are independent across locations, the choice of location units matters. Empirically, larger Voronoi cells may mechanically yield larger destination fixed effects. We show that when the true model has multiple independent shocks at sub-locations within a given location, commuting flows defined at the larger location level still follow gravity equation (2) using an “effective” wage at that location. In particular, assume that location j is divided into Nj smaller areas where workers draw independent shocks, and all areas have the same “true” wage WjR . By standard Fréchet properties, the commuting probability to j 1/ is equivalent to a model with a single shock at j and “effective” wage Wj = Nj WjR .9 From equation (3) we estimate ψj =  log Wj and we recover the true wage as the area-adjusted destination fixed effect ψ̂jR = ψ̂j − log (Nj ) . (4) In robustness exercises, using un-adjusted destination fixed effects does not affect results, except when including distant peri-urban areas where cell phone towers are very sparse. 3.4 Estimation Results: Gravity and Wages We estimate gravity equation (3) using cell phone commuting flows and Google Maps travel times. Our goal is to recover the destination fixed effects, which in the model are proportional to workplace log wages. The estimation sample is non-holiday weekday commuting trips between pairs of towers excluding nearby and very distant towers.10 Table 1 reports the results, based on almost 20 million commuting flows between ∼ 1, 900 locations in Dhaka (columns 1-2) and 130 million flows between ∼ 1, 200 locations in Colombo (columns 3-4). The gravity equation is estimated with unconstrained OLS (columns 1 and 3) and by imposing model constraints (columns 2 and 4). See footnote 8 and Appendix A.1 9 In Appendix A.5 we prove a more general approximate invariance-to-aggregation result for destination fixed effects. Redding and Weinstein (2019) prove a related result for gravity models in trade. 10 In Dhaka, we further exclude 31 days with transportation strikes (hartals). Tower pairs closer than 3 minutes are excluded as they may capture calls randomly connecting to different towers (“tower-bouncing”) rather than real commuting. Destination fixed effects estimated including nearby tower pairs are virtually identical (Appendix Table C.4). Towers over the 99th percentile of the travel time distribution are also excluded (137 and 96 minutes in Dhaka and Colombo, respectively). 8 for estimation details for the latter. Commuting probability decreases strongly with travel time. Interestingly, although the average commuting trip is 25% longer on average in Sri Lanka, once we adjust for residential locations (i.e., gravity equation with origin fixed effects), the coefficients for Dhaka and Colombo are very similar, -1.65 and -1.76. This is a substantive finding, as the two cities differ in terms of economic development, population, and urban structure (mono- vs poly-centric). Figure 1 displays smoothed estimated wages in Dhaka and Colombo using choropleth maps. Estimated wages are higher near city centers and alongside some (but not all) major road corridors. Moreover, secondary centers are visible, especially in Dhaka. The next sections will compare these results with independent income proxies. Destination fixed effects using different estimation methods are highly correlated, including when we add 1 to tower pairs with zero commuting flows (57% of all possible tower pairs in Bangladesh and 15% in Sri Lanka) (Appendix Table C.4).11 For the rest of the paper, we use the destination fixed effects from the first estimation method of the gravity equation (columns 1 and 3) as model-predicted log wages (before area adjustment), and we report robustness results in the appendix. 4 Validation using Survey Income and Nighttime Lights The method above infers wages based on observed commuting choices. Does this approach predict real-world within-city income patterns? We now cross-validate with two alternate income data sources: self-reported income from a transportation survey in Dhaka, and high-resolution satellite nighttime lights data in Dhaka and Colombo. 4.1 Model-Predicted and Survey Workplace Income in Dhaka Our first validation exercise compares income from the model and survey income from the DHUTS survey (Section 2.1).We compute average income at the workplace level in each survey areas in the DCC, the finest geographic location available in the DHUTS survey. The model-predicted income measure is the area adjusted destination fixed effects ψ̂jR . In the model, this equals log labor income divided by , the Fréchet shape parameter of worker’s unobserved preferences. Hence, we expect a regression coefficient of around 1/. Since survey areas are coarser than cell phone towers, we average model income within each of the 88 survey areas with non-government workers, weighting each tower by its workplace population from the cell phone data. 11 Implementing the Poisson pseudo-maximum-likelihood estimator from Silva and Tenreyro (2006) to deal with zero flows is difficult due to the large number of locations. 9 In our main exercise, we correlate survey income and model-predicted income, at the workplace survey area location level. This is a transparent way to check how our measure – which is obtained solely from commuting choices and data on travel times – lines up with real data. We benchmark results in two ways. First, we consider two alternative variables to check that our results are not driven by simpler measures: employment density (computed from cell phone data), an established empirical predictor of income, and distance to the Central Business District (CBD), the main dimension of variation in monocentric urban models. Second, we repeat the entire exercise replacing log income with the residual after partialing out demographic and job variables from the survey (age, gender, years of education, occupation and job sector). We are interested whether predicted income from the model with ex-ante homogenous workers is confounded by observable worker characteristics.12 Table 2 presents the main results. Given that government jobs are typically paid less and are centrally located yet not market allocated, our estimation sample excludes government workers. Including them significantly weakens the correlation between model and survey income, unless we include occupation dummies. Model-predicted income explains 26 percent of the variation in average income at the survey area level, and the coefficient implies a Fréchet shape parameter of ˆ = 7.1, similar to estimates in the urban economics literature (6.83 in Ahlfeldt et al., 2015). In columns 2-3, employment density and distance have slightly lower and slightly higher predictive power, respectively.13 The coefficient on model-predicted income is almost unchanged when controlling for these variables (column 4), showing that the model contains information not available in these other measures. In column 5, we include model-predicted residential income. While the two model measures are highly correlated, the positive correlation with survey workplace income is loaded onto model workplace income. The coefficient on residential income is negative and significant, yet less precisely estimated. Results are broadly similar with residual income as the dependent variable (panel B). All coefficients are roughly half the size of those in Panel A, which suggests that part of the variation in average income captured by any of the three measures is due to sorting. However, model workplace income remains significant in all specifications, and the explanatory power is similar to Panel A. Hence, our revealed-preference approach has similar predictive power after controlling for workplace sorting along observable worker characteristics. Results are robust to several alternate specifications of the gravity equation estimation 12 This specification is exact in a model where income heterogeneity due to demographic factors enters multiplicatively in equation (1) (Appendix A.2). 13 Note that averaging within relatively coarse geographic areas favors the distance to CBD measure. Indeed, when averaging, while the range of distance to CBD remains roughly unchanged, the variance of average model-predicted income goes down, which tends to decrease R2 . 10 (Appendix Table C.5). Appendix Table C.6 uses individual survey data and shows that our main result is robust to controlling for origin survey area fixed effects, geographic area of destination location, and travel time. In terms of predictive power, our measures capture about one-quarter of the spatial variation of surveyed income. This suggests that factors outside the model are also important. In particular, distance to CBD has a significant, negative coefficient even after controlling for model income (Table 2, column 4). This may happen if we underestimate travel costs towards the CBD in our gravity equation, for example if traffic jams are valued beyond the time lost, or if locations near the CBD exhibit workplace disamenities, so that higher wages are required to compensate. By comparison, Tsivanidis (2018) uses a similar model estimated with population counts and finds R2 ≈ 0.3 between model-predicted wages and independent survey wage data in 19 urban zones in Bogotá, Colombia. Severen (2019) performs a similar exercise in Los Angeles on a model estimated with commuting flows at the census tract level and finds no correlation with real wage data, citing unobserved location-pair-specific factors as a potential explanation. 4.2 Model-Predicted Residential Income and Nighttime Lights In our second validation, we use high-resolution nighttime light satellite data (Visible Infrared Imaging Radiometer Suite; VIIRS) to assess the model’s performance in Dhaka and Colombo.14 Nighttime lights are an established income proxy in data-scarce environments (Henderson et al., 2010; Chen and Nordhaus, 2011), and the high resolution of VIIRS data (approximately 500 meters) makes it ideal for studying urban areas. We use the model to predict residential (take-home) income at the cell tower level. We test the assumption that nightlights measure residential income by also including model workplace income in our regressions. We also include two alternate variables: residential population density, and distance to the CBD. Model residential income is strongly related to nightlights at the cell tower level (Table 3). The R2 is high at 0.71 and 0.86 in Dhaka and Colombo. Residential density and distance to CBD are also highly correlated with nightlights, and they explain smaller shares of the variance. The model performs well at fine spatial resolution. The coefficient on model-predicted income remains large when including sub-district fixed effects (55 units in Dhaka and 42 units in Colombo), and when controlling for residential density, distance to CBD, and 14 We use a monthly cloud-free composite (January 2014) based on nights with zero moonlight and areas without clouds, curated by the Earth Observation Group (EOG). 11 model-predicted workplace income (column 4 in Table 3). In particular, the model-predicted residential income is robustly correlated, while workplace income is not statistically significant. Together with results in Table 2, this suggests that the model picks up real patterns of how income “moves” around the city. Our results are robust to using the gravity equation estimated with model origin terms, using home-work commuting flows, constructing residential income excluding nearby destination towers, and using log mean residential income (instead of mean log income) (Appendix Table C.7). In Dhaka not area-adjusting destination fixed effects reverses the sign of the correlation with nightlights, because of very large cells far away from the city center. Our results are broadly robust to only using data from the urban core of Dhaka (the area covered by the transport survey used in Section 4.1), except that the slope is shallower (Appendix Table C.8). We also repeat the residential income validation exercise using survey income. We do not find any correlation between model income and survey income (Appendix Table C.9, Panel A). This may be due to lower underlying differences in average income at the residential level (compared to the workplace level), and hence a more noisy measure when using survey data. Consistent with this idea, density and distance to CBD also explain very little variation in average residential survey income. 5 Application: The Economic Costs of Hartal We now illustrate how high-frequency commuting data and our detailed model-predicted income measure can be used in an application. Hartals are a form of political strike that involves a partial shutdown of urban transportation and businesses. They are common in South Asia, and especially in Bangladesh (UNDP 2005). On hartal days, typically announced a few days in advance, groups of people (some paid) enforce the transportation shutdown, especially on major roads and in certain locations. We use our data and model estimates to quantify the short-term impact of hartal on forgone income. We analyze how predicted income and travel patterns differ on hartal days compared to other days. We focus on the overall effect of hartals on commuting behavior, inclusive of potential changes in traffic congestion and commuting routes. Note, our empirical strategy cannot quantify direct impacts of hartals on worker productivity, nor long-term adaptation costs. We use daily individual commuting data from cell phone records. The sample covers commuters with distinct long-term home and work locations (towers) identified with the 12 procedure in Section 2.1, accounting for 27% of all users in the data.15 We only observe travel behavior if a user makes calls on a given day, and call behavior itself may differ on hartal days. We include commuter fixed effects to ensure that our results are not driven by selection across different types of commuters. Moreover, restricting to frequent callers yields very similar results (Appendix Table C.10). Given that users may travel to different destinations on different days, we use two definitions of predicted income. First, we assign the (model-predicted) wage of the worker’s destination j on that particular day, as long as she makes a proper trip (a trip with distinct origin and destination). We assign zero income to stationary (non-proper) trips. This is a way to capture that workers may earn income from different destinations on different days. Second, we use the wage from her long-term destination if the worker makes a proper trip to that location. We assign zero income to stationary trips and trips to any other destination.   In both cases, predicted income is exp ψ̂jR / where log wages ψ̂jR are estimated using only non-hartal weekdays, and  = 6.4.16 Since we are interested in changes in income during hartal days, the scale of ψ̂jR , which is not identified, does not matter for this exercise. Our main specification is: yct = β H Hartalt + β F F ridayt + β S Saturdayt + β Ho Holidayt + µc + γM onth(t) + εct where c denotes a commuter, t denotes a calendar date, and the outcome yct is predicted income or a dummy for making a proper trip. The unit of observation is a commuter times calendar date, for dates when we observe a trip (either proper and stationary). The main coefficient of interest is β H , which measures the difference in outcome on hartal days relative to workdays (non-holiday weekdays). We use β F , the effect on Fridays, the main free day in Bangladesh, as a benchmark. We use the data set of hartal dates in Dhaka from Ahsan and Iqbal (2015). They identify 33 hartal days over the 4 months in our sample. The study period preceded parliamentary elections and was marked by general instability, and hence hartals were more frequent than in previous years. Hence, our results may not directly generalize to periods with lower hartal intensity. Cell phone data picks up stark differences in behavior on hartal, weekends and holidays. Appendix Figure B.6 plots the change in predicted income relative to workdays, by calendar date. Predicted income is systematically lower during hartals compared to weekdays, yet 15 We are interested in canceled trips due to hartals, which are difficult to observe for users with identical home and work towers. Results with all users are qualitatively similar and smaller in magnitude. 16 This is our point estimate of the Fréchet parameter using a structural estimation method (Appendix A.3). The regression coefficient from Table 2 of log survey income (measuring ωj ) on the destination fixed effect (ωj ) implies a very similar number (ˆ  = 7.1 = 0.14−1 ). 13 not as much as on Fridays and on some important holidays, such as the end of Ramadan. Predicted income is higher than usual on the two Fridays that immediately follow long hartal spells (December 6th and 20th), consistent with temporal substitution. The results are also consistent with longer hartals having lower impacts. Table 4 shows the main results. To facilitate interpretation, all coefficients indicate proportional changes relative to the outcome mean on workdays. Odd columns assign predicted income to any proper trip, while in even columns outcomes are non-zero only when the commuter travels to her long-term workplace. There are four key insights. First, commuters in Dhaka earn on average 4.4 to 4.8% less on hartal days compared to workdays. These effects are significantly smaller compared to Fridays, when predicted income is 20 to 45% lower on average (panel A). Second, hartal days affect “all trips” and “work trips” roughly equally, while on Fridays, work trips are disproportionately affected. This suggests a limited “destination selection” effect of hartals; on average, commuters do not switch to lower-income destinations. Third, the reduction in predicted income is driven primarily by the extensive margin, namely fewer trips (panel B). Fourth, commuters working in high-income destinations reduce trips relatively more. In columns 3-4, we fully interact the model with an indicator for commuters whose long-term workplace location is below median in the predicted wage distribution. The interaction results show that the proportional reduction in trips is concentrated among high-income commuters, both on hartal and Fridays. Columns 5-6 document that this heterogeneity is not due to commute distance (as long-distance commuters are also more affected). These results show that commuters broadly succeed to maintain their workday travel routines on hartal days, which limits the short-term impact of hartal on economic activity. These results are consistent with previous studies on hartals in more specific settings (Ashraf et al., 2015; Ahsan and Iqbal, 2015). 6 Conclusion This paper provides a theory-based toolkit for using cell phone data to understand the spatial distribution of economic activity in cities. This framework is especially suited to measuring and interpreting the short-term impact of urban shocks such as floods, or of transportation incidents or improvements, on commuting and economic activity. Together with officials statistics, they can be used to investigate spatial discrepancies between formal and informal economic activity. Big data, such as cell phone or smartphone mobility records, credit card transactions, or user-generated reviews, are rapidly gaining popularity due to their ability to predict behavior, 14 individual characteristics and economic conditions (Blumenstock et al., 2015; Jean et al., 2016; Glaeser et al., 2017; Björkegren and Grissen, 2018). However, big data also contain a wealth of information regarding individual choices. This allows economists to apply revealed preference techniques to infer attributes of choice options, such as workplace wages in our paper or spatial aspects of consumption behavior (Athey et al., 2018; Davis et al., 2018; Agarwal et al., 2018). We believe that this type of applications is a promising path for using “big data” in economics. 15 References Agarwal, S., F. Monte, and B. Jensen (2018): “The Geography of Consumption,” NBER Working Paper No. 23616. Ahlfeldt, G. M., S. J. Redding, D. M. Sturm, and N. Wolf (2015): “The Economics of Density: Evidence from the Berlin Wall,” Econometrica, 83, 2127–2189. Ahsan, R. and K. Iqbal (2015): “Political Strikes and its Impact on Trade: Evidence from Bangladeshi Transaction-level Export Data,” IGC Working Paper. Alonso, W. (1960): “A Theory of the Urban Land Market,” Papers and Proceedings Regional Science Association, 6, 149–157. Ashraf, A., R. Macchiavello, A. Rabbani, and C. Woodruff (2015): “The Effect of Political and Labour Unrest on Productivity: Evidence from Bangladeshi Garments,” IGC Working Paper. Athey, S., D. Blei, R. Donnelly, F. Ruiz, and T. Schmidt (2018): “Estimating Heterogeneous Consumer Preferences for Restaurants and Travel Time Using Mobile Location Data,” AEA Papers and Proceedings, 108, 64–67. Björkegren, D. and D. Grissen (2018): “The Potential of Digital Credit to Bank the Poor,” AEA Papers and Proceedings, 108, 68–71. Blumenstock, J., G. Cadamuro, and R. On (2015): “Predicting Poverty and Wealth from Mobile Phone Metadata,” Science, 350. Calabrese, F., G. Di Lorenzo, L. Liu, and C. Ratti (2011): “Estimating Origin-Destination Flows Using Mobile Phone Location Data,” IEEE Pervasive Computing, 10, 36–44. Chen, X. and W. D. Nordhaus (2011): “Using luminosity data as a proxy for economic statistics.” Proceedings of the National Academy of Sciences of the United States of America, 108, 8589–8594. Davis, D., J. Dingel, J. Monras, and E. Morales (2018): “How Segregated is Urban Consumption?” Accepted, Journal of Political Economy. Duncan, C. (2005): Beyond Hartals: Towards Democratic Dialogue in Bangladesh, United Nationas Development Programme. 16 Fally, T. (2015): “Structural gravity and fixed effects,” Journal of International Economics, 97, 76–85. Financial Times (2014): “Nigeria almost doubles GDP in recalculation,” https://www. ft.com/content/70b594fe-bd94-11e3-a5ba-00144feabdc0. Frank, P. and T. Murtha (2010): “Trips Underway by Time of Day by Travel Mode and Trip Purpose for Metropolitan Chicago,” Tech. rep., CMAP Congestion Management Process, Chicago Metropolitan Agency for Planning. Glaeser, E. L., H. Kim, and M. Luca (2017): “Nowcasting the Local Economy: Using Yelp Data to Measure Economic Activity,” Harvard Business School Working Paper, No. 18-022. Head, K. and T. Mayer (2014): “Gravity equations: Workhorse, toolkit, and cookbook,” in Handbook of international economics, Elsevier, vol. 4, 131–195. Heblich, S., S. Redding, and D. Sturm (2018): “The Making of the Modern Metropolis: Evidence from London,” Working Paper. Henderson, J. V., A. Storeygard, and D. N. Weil (2010): “Measuring Economic Growth from Outer Space,” American Economic Review, 102, 994–1028. Iqbal, M. S., C. F. Choudhury, P. Wang, and M. C. González (2014): “Development of Origin-destination Matrices Using Mobile Phone Call Data,” Transportation Research Part C: Emerging Technologies, 40, 63–74. Japan International Cooperation Agency (2010): “Preparatory Survey Report on Dhaka Urban Transport Network Development Study (DHUTS) in Bangladesh : Final Report.” Tech. rep., Japan International Cooperation Agency, http://open_jicareport. jica.go.jp/pdf/11996774_03.pdf. Japan Times (2019): “Government to review statistics after revelation that faulty Japan jobs data spanned 14 years,” https://www.japantimes.co.jp/news/2019/01/11/national/ government-review-statistics-revelation-faulty-japan-jobs-data-spanned-14-years/. Jean, N., M. Burke, M. Xie, W. M. Davis, D. B. Lobell, and S. Ermon (2016): “Combining satellite imagery and machine learning to predict poverty,” Science, 353, 790–794. 17 Mills, E. S. (1967): “An Aggregative Model of Resource Allocation in a Metropolitan Area,” The American economic review Papers and Proceedings of the Seventy -ninth Annual Meeting of the American Economic Association, 57, 197–210. Muth, R. (1968): Cities and Housing, Chicago: University of Chicago Press. Redding, S. and D. Weinstein (2019): “Aggregation and the Gravity Equation,” NBER Working Paper 25464. Severen, C. (2019): “Commuting, Labor, and Housing Market Effects of Mass Transportation: Welfare and Identification,” Working Paper. Silva, J. S. and S. Tenreyro (2006): “The log of gravity,” The Review of Economics and statistics, 88, 641–658. Tsivanidis, N. (2018): “The Aggregate And Distributional Effects Of Urban Transit Infrastructure: Evidence From Bogota’s TransMilenio,” Working Paper. Wang, P., T. Hunter, A. M. Bayen, K. Schechtner, and M. C. González (2012): “Understanding Road Usage Patterns in Urban Areas,” Scientific Reports, 2, 1001. 18 7 Figures and Tables Table 1: Gravity Equation and Destination Fixed Effects log Commuting Flow (1) (2) (3) (4) log Travel Time −1.65∗∗∗ (0.01) −1.93∗∗∗ (0.01) −1.76∗∗∗ (0.03) −2.30∗∗∗ (0.02) City Model Constraints Number of Destination FE Number of Trips Observations Adjusted R2 Dhaka No 1.9 · 103 19.3 · 106 1.5 · 106 0.61 Dhaka Yes 1.9 · 103 19.3 · 106 1.5 · 106 0.65 Colombo No 1.2 · 103 129 · 106 1.1 · 106 0.75 Colombo Yes 1.2 · 103 129 · 106 1.1 · 106 0.69 Notes. This table reports estimates of the gravity equation (3). The outcome variable is log total commuting flows log (Vij ) between a pair of cell phone towers, computed from cell phone data and aggregated over weekdays. (This is equivalent to using log commuting probabilities P log (πij ) = log (Vij / s Vis ) as in equation (3), as the origin fixed effects capture the denominator.) In Bangladesh, we exclude hartal days. For each trip, the origin is the first location (tower) between 5 am and 10 am, and the destination is the last location between 10 am and 3 pm. Travel time between towers from the Google Maps API. The sample is all tower pairs with travel time between 180 seconds and the 99th percentile. Columns (1) and (3) report the coefficient −β̂ from an OLS regression with origin and destination fixed effects. Columns (2) and (4) report the estimates where at step k + 1 the model-predicted origin term P from  an iterative  procedure, P k k k µi = log + log ( s Vis ) is subtracted from log flow before running an OLS s exp ψ̂s − β̂ dis regression without origin fixed effects (Appendix A.1). Two-way clustered standard errors at the origin and destination level are reported in parentheses. ∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01 19 Figure 1: Estimated log Wages in Dhaka and Colombo 20 (A) Dhaka 1.7 2.8 1.2 2.1 0.9 1.7 0.6 1.0 0.4 0.5 0.1 -0.1 -0.1 -0.6 -0.4 -1.1 -0.9 -1.7 -1.6 -2.3 -2.4 -2.8 -5.4 -4.0 (B) Colombo Notes. These figures plot area-adjusted destination fixed effects (the model measure proportional to log wages) at the level of cell phone tower Voronoi cells in Dhaka and Colombo. Log wages are kernel smoothed with an adaptive kernel bandwidth (proportional to the radius of the equivalent-area circle of the Voronoi cell, see Appendix A.6). Table 2: Average Workplace Income: Model Predictions and Survey Data in Dhaka (1) (2) (3) (4) (5) 0.12∗∗ (0.05) 0.27∗∗∗ (0.06) −0.05 (0.07) −0.01 (0.07) −0.14∗∗∗ (0.03) −0.18∗∗∗ (0.03) Panel A. Outcome: log Survey Income (workplace) log Model Income (workplace) 0.14∗∗∗ (0.03) 0.15∗∗∗ (0.04) log Employment Density −0.18∗∗∗ (0.03) log Dist. to CBD −0.32∗∗ (0.13) log Model Income (residential) Observations Adjusted R2 88 0.26 88 0.16 88 0.33 88 0.42 88 0.46 0.08∗∗ (0.03) 0.15∗∗∗ (0.03) −0.03 (0.03) −0.01 (0.04) −0.05∗∗∗ (0.02) −0.07∗∗∗ (0.02) Panel B. Outcome: Residual log Survey Income (workplace) log Model Income (workplace) 0.07∗∗∗ (0.02) 0.08∗∗∗ (0.02) log Employment Density −0.08∗∗∗ (0.02) log Dist. to CBD −0.16∗ (0.09) log Model Income (residential) Observations Adjusted R2 88 0.23 88 0.13 88 0.17 88 0.28 88 0.3 Notes. This table compares survey and model predictions of average workplace income. The unit of analysis is a survey area from the DHUTS survey. The survey sample is 11,006 commuters who live and work inside the Dhaka City Corporation, who report positive income, excluding students, homemakers, the unemployed, and government workers. The outcome in panel A is the average income of survey respondents who work in a survey area, using log income truncated at the 99th percentile. In panel (B), it is the residual of log income on gender, age, years of education, occupation and job sector dummies. Model-predicted workplace income in survey area P b is j∈b yj VjW /VbW where j is a cell phone tower, yj = ψ̂jR is the area adjusted destination P P fixed effect at j, VjW = i Vij and VbW = j∈b VjW denote workplace population in tower j and survey area b, respectively. Regressions are weighted by survey area employment population. The Central Business District (CBD) is Shapla Chatter in Motijheel. Appendix Figure B.4 shows the corresponding scatter plots. Robustness in Appendix Tables C.5 and C.6. The analogous exercise at the residential level in Appendix Table C.9. Conley standard errors with 5 km distance cutoff shown in parentheses. ∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01. 21 Table 3: Average Residential Income: Model Prediction and Nighttime Lights (1) log VIIRS Nighttime Lights (2) (3) (4) Panel A. Dhaka, Bangladesh log Model Income (residential) 0.57∗∗∗ (0.04) 0.53∗∗∗ (0.07) 0.26∗∗∗ (0.06) log Residential Density 0.01 (0.02) −0.52∗∗∗ (0.10) log Dist. to CBD −0.11 (0.09) log Model Income (workplace) −0.04 (0.05) Sub-district FE (count) X (55) Observations Adjusted R2 1,868 0.71 1,868 0.31 1,868 0.34 1,868 0.84 Panel B. Colombo, Sri Lanka log Model Income (residential) 0.79∗∗∗ (0.02) 0.49∗∗∗ (0.05) 0.63∗∗∗ (0.03) log Residential Density −0.07∗ (0.03) −1.08∗∗∗ (0.12) log Dist. to CBD −0.30∗∗∗ (0.10) log Model Income (workplace) 0.01 (0.04) Sub-district FE (count) X (42) Observations Adjusted R2 1,199 0.87 1,199 0.66 1,199 0.76 1,197 0.93 Notes. This table compares high resolution nighttime lights with model predicted average residential income. The unit of analysis is a cell phone tower in the greater metropolitan area of each city. The outcome variable is log mean light intensity in the VIIRS nighttime light data inside each Voronoi cell (weighting each VIIRS cell by its overlap with the Voronoi cell). Average model residential P (take-home) income at tower i is j yj Vij /ViH where j indexes workplace towers, yj = ψ̂jR is the area adjusted destination fixed effect at j, and ViH is total residential population at i. Regressions are weighted by tower residential population. Column 4 controls for 55 sub-district (thana) fixed effects for Dhaka (panel A), and 42 sub-districts (Divisional Secretariat) fixed effects for Colombo (panel B). Appendix Figure B.5 shows corresponding scatter plots. Robustness to definitions of model income in Appendix Table C.7. Conley standard errors with 5 km distance cutoff shown in parentheses.∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01 22 Table 4: Impact of Hartal on Predicted Income, Travel Behavior, and Workplace Attendance (1) (2) (3) (4) (5) (6) All Coefficients: % Change From Workday Mean All Trips Work Trips All Trips Work Trips All Trips Work Trips Hartal −0.048∗∗∗ (0.009) −0.044∗∗∗ (0.015) Friday (free day) −0.208∗∗∗ (0.007) −0.452∗∗∗ (0.014) Observations 22.5 · 106 22.5 · 106 Hartal −0.038∗∗∗ (0.007) −0.037∗∗∗ (0.014) −0.045∗∗∗ (0.008) −0.068∗∗∗ (0.017) −0.050∗∗∗ (0.008) −0.079∗∗∗ (0.018) Friday (free day) −0.159∗∗∗ (0.006) −0.430∗∗∗ (0.013) −0.190∗∗∗ (0.007) −0.543∗∗∗ (0.018) −0.209∗∗∗ (0.008) −0.602∗∗∗ (0.019) Hartal x Low Income 0.015∗∗∗ (0.003) 0.067∗∗∗ (0.010) 0.013∗∗∗ (0.003) 0.063∗∗∗ (0.010) Friday x Low Income 0.066∗∗∗ (0.004) 0.239∗∗∗ (0.013) 0.058∗∗∗ (0.004) 0.215∗∗∗ (0.012) Hartal x Short Commute 0.014∗∗∗ (0.002) 0.031∗∗∗ (0.004) Friday x Short Commute 0.053∗∗∗ (0.003) 0.165∗∗∗ (0.005) 22.5 · 106 0.79 22.5 · 106 0.37 Panel A. Predicted Income Panel B. Make a Trip Observations Workday Mean 22.5 · 106 0.79 22.5 · 106 0.37 22.5 · 106 0.79 22.5 · 106 0.37 Notes. This table shows differences in predicted income and travel probability on hartal days and Fridays relative to workdays. All coefficients show proportional changes relative to workdays. The sample is all days with commuting data (including stationary trips) for commuters with distinct long term home and workplace towers (27% of all users). For commuter c on calendar date t, denote their trip origin by ict , destination by jct , and c’s long-term workplace by jcW. In panel  A, the outcome is predicted income. In R column (1), commuters earn the destination wage exp ψ̂jct / for any proper trip and zero otherwise. In column (2), commuters earn positive income only when ict 6= jct = jcW . In both cases, the gravity equation is estimated on non-hartal weekdays, and we use  = 6.4. In panel B, the outcome is a dummy for proper trip (jct 6= ict ) in odd columns, and a dummy for proper workplace trip (jct = jcW 6= ict ) in even columns. All regressions include commuter and month fixed effects, and dummies for Saturday and holidays. In columns (3)-(6), we fully interact the model with dummies for low-wage commuters (c’s long-term workplace wage ψ̂jRW is below-median) and short-commute commuters (c’s travel time between long-term home and work is c below-median). Reported coefficients are proportional changes relative to non-hartal, non-holiday weekday mean. Standard errors clustered at the calendar date level in parentheses. ∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01 23