Cities with Cell Phone Records Appendix

A A.1 Appendix Estimation of Constrained Gravity Equation The structural gravity equation (2) implies a constraint between the destination fixed effects j , origin fixed effects µi , and the distance coefficient β in the empirical gravity equation (3). In our main specification, we do not impose these constraints, but we show that this does not affect the estimation results. Here, we describe the constrained estimation procedure in more detail. First, we estimate a version of gravity equation (3) without origin fixed effects by OLS: log(πij ) = j − β log(Dij ) + εij , (5) Next, using the estimated destination fixed effects ψ̂s1 and the distance coefficient β̂ 1 from  P  1 1 . the OLS regression, we compute the model-predicted origin terms µ̃2i = log s exp ψ̂s − β̂ dis Next, we estimate (5) using log (πij ) − µ̃2i as outcome variable. This leads to the step 2 estimates ψ̂s2 and β̂ 2 .  P  k k − β̂ d and ψ̂ In general, after step k, we construct model terms µ̃k+1 = log exp is s i s k+1 run (5) using log (πij ) − µ̃i as outcome variable. We iterate this procedure until the vectors     ψ̂sk and ψ̂sk+1 converge in L2 norm, with a tolerance of 10−6 . s s The procedure is identical to SILS (structurally iterated least squares) proposed in trade gravity literature (Head and Mayer, 2014), except that our model constraints are only on the origin fixed effects, but not on the destination fixed effects. Fally (2015) discusses the potential bias of the estimated gravity fixed effects without imposing these model constraints. A.2 Model Extension: Worker Heterogeneity in Effective Labor Supply In Section 3, we assumed that workers are ex-ante identical. However, in panel B in Table 2, we measure the model’s predictive power after netting out individual demographic characteristics from survey income. Here, we show how this validation regression arises directly in a specific model with worker heterogeneity. Assume that worker ω supplies ξω effective units of labor. ω 0 s income from working in j is ξω Wj instead of simply Wj . Otherwise, workers have the same disutility of commuting, and face the same profile of wages. This implies that workers living at the same location i face the same workplace location choice, regardless of ξω . Hence, in aggregate, the gravity equation (2) continues to hold unchanged. However, the average ξω of commuters working in j affects average income at that location. Hence, the correct validation regression should control for average ξω at location j from individual income. To the extent that ξω depends on observable characteristics (gender, age, education level, occupation, job sector), this is exactly what the specification in panel B, Table 2 achieves. A.3 Structural Estimation: How Much do Individual Shocks and Travel Time Affect Income In the main analysis, we assume that an agent earns income directly proportional to her wage. Formally, the Fréchet shocks Zijω and travel time Dij affect utility but not income. Here, we relax this assumption and allow Zijω and Dij to partly affect income; for example, they may affect productivity or labor supply. We derive a transparent method that allows survey income data to speak as to the role of shocks and travel time for income. αz ,αd −τ αd αz Model. Assume that income is given by Yijω = Wj Zijω Dij , where αz , αd ∈ [0, 1] respectively control the extent to which the shocks Zijω and travel time Dij affect income. For example, when αz = 1 and αd = 0, shocks affect utility and income equally, while travel time only affects utility. We derive formulas for expected income in the following four extreme extreme cases: 0,0 Eyijω = wj (6) 0,1 Eyijω = wj − τ dij ! 1,1 Eyijω X K 1 exp (wj − τ dij ) − for some absolute constant K = log   s 1,0 1,1 Eyijω = Eyijω + τ dij When neither shocks nor travel time affect income, income is simply the destination wage. In the second case, travel time fully affects labor earnings. When the shocks Zijω affect income, as in the third and fourth cases, log income for a worker commuting between i and j depends on the distribution of the shock conditional on destination j being chosen. By virtue of the Fréchet distribution, the conditional distribution yijω |j ∈ arg maxs Uisω is  P P  −τ  also Fréchet with the same shape parameter  and scale Ti = s Tis = s Ws Dis . In particular, this distribution only depends on the origin i and thus expected log income is the same for all destinations j. In the general case, log income is a convex combination of the following four extreme cases: αz ,αd 1,1 1,0 0,1 0,0 yijω = αz αd · yijω + αz (1 − αd ) yijω + (1 − αz ) αd · yijω + (1 − αz ) (1 − αd ) yijω . (7) Using (6) and dropping the constant K, this simplifies to " αz ,αd Eyijω ! # X αz αd 1 − αz = log [wj ] + [−τ dij ] exp (wj − τ dij ) + τ dij +    s (8) The intuition of this expression is as follows. For the third term, if travel time affects income, we expect that people who commute further away have lower income. The difference between the first two terms is more subtle. If Fréchet shocks affect income, then the first term is the best explanatory variable for income.17 If shocks do not affect income, the wage at the destination should be the best predictor of income. Estimating Parameters αz , αd , . We are now in a position to estimate the parameters αz , αd and . Specifically, we estimate by OLS the equation: S yijω = ρ1 X̂ij1 + ρ2 X̂j2 + ρ3 X̂ij3 + εSijω , (9) S where yijω is survey-based income of commuter ω who lives at i and works at j, and X̂ij1 =  P  + β̂dij , X̂j2 = ψ̂j and X̂ij3 = −β̂dij are estimators of the three terms log s exp ψ̂s − β̂dij in square brackets in (8), computed using the gravity equation estimates. (Recall that ψ̂j is a consistent estimator for wj , and β̂ is a consistent estimator for τ .) Asymptotically, we have ρ̂3 1 ρ̂1 , α̂d = , and ˆ = . (10) α̂z = ρ̂1 + ρ̂2 ρ̂1 + ρ̂2 ρ̂1 + ρ̂2 Table C.11 reports the estimates of αz , αd , and  based on estimating equation (9) with OLS, and using transformation (10). We report two types of standard errors: based on the Delta method (in round parentheses) and based on bootstrapping at the origin survey area level (in square parentheses).18 In columns 1-2, we estimate the full equation (9), and we find that α̂d is close to zero with a small and insignificant negative value, and the other 17 The first term is analogous to the market access term in gravity trade literature, except that it includes the compensation income from commuting cost in utility. 18  For the Delta method, following the equations in (10), define (, αz , αd ) = h (ρ1 , ρ2 , ρ3 ) = ρ1 ρ3 1 , α̂z , α̂d ) is given ρ1 +ρ2 , ρ1 +ρ2 , ρ1 +ρ2 . The Delta method states that the asymptotic covariance matrix of (ˆ T by J (h) ΣJ (h) where J (h) is the Jacobian of h and Σ is the asymptotic covariance matrix of (ρ̂1 , ρ̂2 , ρ̂3 ). The Jacobian is obtained by differentiating h with respect to the ρ’s:   −1 −1 0 −2   ρ2 −ρ1 0 J (h) = (ρ1 + ρ2 ) −ρ3 −ρ3 ρ1 + ρ2 coefficient αd is constrained to zero ρ3 is also zero, we have (, αz ) = h (ρ1 , ρ2 ) =   When the distance ρ1 1 and the Jacobian is ρ1 +ρ2 , ρ1 +ρ2   −1 −1 −2 J (h) = (ρ1 + ρ2 ) ρ2 −ρ1 parameters are imprecisely estimated when using bootstrapped standard errors. Given that the model restricts ρ3 ≥ 0 (from αd ∈ [0, 1]), in columns 3-4 we restrict the coefficient on travel time to be equal to zero (ρ3 = 0) and estimate the other two parameters. This does not affect the point estimates for α̂z and ˆ while improving precision. These results show that idiosyncratic shocks partly affect income, while travel time is most consistent with a pure utility cost. Robustness exercises using “home” and “work” commuting flows and iterated gravity equation give qualitatively similar results (not reported). A.4 Discussion: Estimating Wages without Commuting Flows In the absence of detailed commuting flows data, Ahlfeldt et al. (2015) use an exactly identified procedure to infer wages from total residential and total workplace population counts at each location, as well as knowledge of the distance coefficient β (which in turn is estimated from a gravity equation at a coarse level). While the two procedures are equivalent if equation (3) does not have noise, estimating wages using bilateral commuting flows is more robust against measurement error and idiosyncratic departures from the model.19 If random shocks affect bilateral commuting flows proportionally, as in equation (3), then in the procedure using total counts, commuting pairs with high commuting flow (e.g. nearby pairs) have a large, noisy influence on estimated wages. Indeed, we find that the gravity equation explains only between 0.6 and 0.75 of the variation in bilateral log commuting flows. Severen (2019) documents that time-invariant link-specific effects explain a large share of the variance in commuting flows in Los Angeles. Related, bilateral commuting data allows us to test the stability of our results, i.e., the estimated wages should be similar when using a sub-sample of location pairs for estimation. Indeed, we show that results are similar when estimating (3) with or without nearby pairs of towers (for which commuting flows are more likely to be mismeasured). Finally, with the gravity equation we can in principle estimate the effect of distance non-parametrically and simultaneously with destination fixed effects. Appendix Figure B.3 shows that a linear fit as a function of log travel time is, in fact, appropriate. A.5 Model: Approximate Invariance to Aggregation Level The model has a general (approximate) invariance property with respect to the level of geographic aggregation, both at the origin and at the destination level. 19 The constrained gravity equation (2) yields asymptotically identical results as the procedure in Ahlfeldt et al. (2015) Pgiven the knowledge of β = τ if we estimate the model using the employment distribution as moments ( i Hi πij ), where Hi is the number of residents in location i, instead of the bilateral commuting probability log(πij ) as moments, as in our procedure. At the origin level, the model is approximately invariant with respect to the origin aggregation level, because the basic discrete choice problem is individual specific. At the destination level, the aggregation level affects the interpretation of wages Wj in a straight-forward way. Assume that location j is in fact composed of several sub-locations k1 , k2 , ..., kNj , and we estimate the model at the higher level (j) and ignore the sub-locations. N P 1/ j  The wage we obtain, Wj = , represents a C.E.S. aggregate with elasticity of the `=1 Wk` true underlying wages at all sub-locations within j. (This is easy to prove using the standard properties of the Fréchet distribution.) In particular, this implies a simple adjustment for the destination fixed effect j = wj estimated using the gravity model. Assume that the “real” underlying wage is constant and denoted by WjR within each location j, then the 1/ C.E.S. relationship becomes Wj = Nj WjR , or in logs the underlying wage is given by 1/ wjR = wj − log(Nj ). In terms of estimated quantities, this becomes ψ̂jR = ψ̂j − log (Nj ). The underlying destination fixed effect ψ̂jR is obtained from the fixed effect ψ̂j , estimated ignoring sub-locations, minus an adjustment factor equal to the log of the number of true underlying locations where shocks are realized, Nj . This relationship is exact if the distances between each sub-location in location j and all other locations do not depend on the sub-location. Redding and Weinstein (2019) derive an exact relationship by using all the distance profiles in the context of gravity equations of trade models. A.6 Data: Smoothing procedure for Google Map Travel Time Due to the large number of bilateral tower pairs (on the order of ∼ 106 ), we obtain travel time estimates from the Google Maps API for 90,000 randomly selected pairs in each country, and interpolate the travel time for remaining tower pairs. The data was collected in June 2016 for Colombo and in July 2017 for Dhaka. Specifically, for each pair we query the Google Maps Distance Matrix API for the typical driving time on a weekday with departure time at 8 am.20 To interpolate the travel time for tower pairs without actual data, we use the following procedure. For any such tower pair i, j, we compute the kernel smoothed speed (travel time divided by straight line distance) of trips starting at origins a near i and ending at destinations b near j, for pairs a, b for which we have Google Maps data. Formally, the smoothed speed ŝij between i and j is given by ŝij = 20     d di,a 1 K hj,bb sa,b GMa,b a6=i,b6=j ha hb K ha     , P dj,b di,a 1 K GM a,b a6=i,b6=j ha hb K ha hb P Specifically, we queried Google Maps for travel times on Friday, August 26, 2016, in Colombo, and Wednesday, September 13, 2017, in Dhaka. The queries were sent about one month before these dates. where GMa,b is a dummy indicator for having Google Maps data between a and b, and dx,y is the distance in decimal degrees between towers x and y. We use the two-dimensional Epanechnikov kernel K (d) = π3 |1 − d2 |+ , together with an “adaptive” kernel bandwidth √ hx ≡ h ĀAx where Ax is the area (in square kilometers) of Voronoi cell of cell phone tower x, Ā √ is the average of Ax over all towers x, and h = 0.03 and h = 0.1 are the (manually-chosen) bandwidth for average-radius towers in Dhaka and Colombo, respectively. A tower with average square root area will have a bandwidth of about 3.5 kilometers in Dhaka, and 11 kilometers in Colombo. We evaluate the predictive power of this procedure using a leave-one-out test using only the tower pairs with actual Google Maps data. The R2 of ŝa,b, on sa,b is 0.98 and 0.96 and Colombo and Dhaka, respectively. Note, the prediction ŝa,b is computed without using sa,b . A.7 Availability of Conventional Data Sources Fine-grained spatially disaggregated data on wages at the firm location is rare and difficult to access in developing countries. For example, the Bangladesh economic census does not include labor costs data, and we were not able to acces Sri Lanka economic census microdata. As a case study, here we document the availability of firm census data in Sub-Saharan Africa, a region undergoing rapid urban growth and urban transformation. We collected data on the 27 largest countries that account for over 95% of the population in the region. Of these, 16 ever had an economic census, 11 covered informal firms. However, at most 4 included wage data, which accounts for between 5.6 and 8.6% of the urban population of all countries in the sample. (The 2014 Ghana and 2015 Zimbabwe censuses included wage data, while for the ongoing censuses in Mali and Togo we do not know if wage data was collected.) Table A.1: Sub-Saharan African Countries with Economic Cenuses Country Nigeria South Africa D.R.C. Ethiopia Angola Tanzania Ghana Cameroon Kenya Ivory Coast Uganda Mozambique Madagascar Mali Senegal Zambia Burkina Faso Zimbabwe Benin Guinea Niger Chad Togo Malawi South Sudan Rwanda Burundi Urban Population 98,525,244 38,112,552 37,382,220 22,367,255 20,157,104 19,972,890 16,529,104 13,918,524 13,756,737 12,652,168 10,536,394 10,990,322 9,769,765 8,101,667 7,690,895 7,659,992 5,806,985 5,446,070 5,432,724 4,711,991 3,659,066 3,546,586 3,332,216 3,238,839 2,532,134 2,150,199 1,458,139 Year of Last Census 2004 2002 2014 2014 2016 2017 2010 2014 2019 2016 2011 2009 2015 2008 2018 2017 - Covers Informal Wage Data Yes No Yes Yes No No No No Yes Yes Yes No No ? Yes ? Yes Yes Yes ? No No No Yes ? ? Yes No Notes: For each country, we checked the national statistics agency website as well as the Google Search results for the terms “economic census,” “firm census,” “establishment census,” “enterprise census,” and “business registry,” in English, French or Portuguese. We could not find official census reports for Ethiopia and Zambia, while the Mali and Togo censuses are still ongoing. Detailed results available upon request. Data on urban population from https://en.wikipedia.org/wiki/Urbanization_by_country and https://en.wikipedia.org/wiki/List_of_sovereign_states_and_dependent_territories_ in_Africa. B Additional Figures Figure B.1: Administrative Units and Cell Phone Voroni Cells in Dhaka Notes. This figure shows the map of cell phone tower Voronoi cells in Dhaka, Bangladesh. The yellow shaded area is the Dhaka City Corporation (DCC), the urban core of Dhaka, the main sample in the DHUTS transportation survey. The overall study area covers three districts in Bangladesh: Dhaka, Gazipur, and Narayanganj. The Voronoi cell of a tower is the locus of all points closer to that tower than to any other tower. Figure B.2: Administrative Units and Cell Phone Voroni Cells in Colombo Notes. This figure shows the map of cell phone tower Voronoi cells in Colombo, Sri Lanka. The study area covers the entire Western Province in Sri Lanka. The Voronoi cell of a tower is the locus of all points closer to that tower than to any other tower. Figure B.3: Commuting Flows from Survey Data and Cell Phone Data Panel (A) Survey vs Cell Phone Data Dhaka Log commuting flow 5 10 15 90 survey wards −5 0 Cell data, log(mean()) Survey data, log(mean()) Bootstrapped CI 5 6 7 Log Travel Time 8 Panel (B) Commuting Flows vs Home-Work Flows 6 −2 −2 0 2 4 Log commuting flow 0 2 4 6 8 Colombo Commuting between cell phone towers 8 Dhaka 1,868 cell phone towers 5 6 7 8 Log Travel Time Commuting, log(mean()) 9 5 6 7 8 Log Travel Time 9 Home Work, log(mean()) Notes. This figure compares the decay of commuting flows with travel time in survey and cell phone data. The unit of analysis is 7,836 survey area pairs in Panel A, and 1.6 · 106 and 1.4 · 106 tower pairs in Dhaka and Colombo in Panel B, respectively. Panel A compares commuting flows from the DHUTS survey (red, dash) and from cell phone data (blue, solid) in Dhaka. Panel B compares daily commuting trips (blue, solid) and home-work commuting trips (black, dash). See Section 2.1 for the definition of home-work commuting trips. In each graph, commuting flows are first averaged within each of 100 equal bins of log travel time below the 99th percentile, and the plot shows the local linear regression of log mean commuting flow on log travel time. This procedure avoids the bias due to zero commuting flows, which is important for survey and home-work commuting data. The DHUTS sample (described in Table C.2) has 12,510 commuters. The cell phone data sample has 18 · 106 trips in Panel A, and 38 · 106 daily trip and 5.2 · 106 for home-work trips in Dhaka, and 237 · 106 daily trips and 2.6 · 106 home-work trips in Colombo, in Panel B. In Panel A, pointwise bootstrapped 95% confidence intervals clustered at the origin survey area shown in gray. Figure B.4: Average Workplace Income: Model Predictions and Survey Data (A) Model and Survey Workplace Income ● ● ● Log Survey Income ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●●●● ● ●●●● ●● 9.5 ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● 9.0 ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●● ● ●● ● ●● ● ● ●● ●● ● ● ● Log Survey Income (Residualized) ● 10.0 ● 0.3 ● ● ● ● ● ● ● ●●● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● 0.0 ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● −0.3 ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● −0.6 −1 0 1 Log Model Income (workplace, normalized) 2 −1 0 1 Log Model Income (workplace, normalized) 2 (B) Employment Density and Survey Workplace Income ●● Log Survey Income ● ● ●● ● ● ● ●● 9.5 ● ●● ●● ● ● ●● ● ● ●● ● ●● ● ● ●● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ●● ●●● ● ● ● ●● ● ●● ● ● ● ●● ● ● ● ● ● 9.0 ● ● Log Survey Income (Residualized) ● 10.0 ● 0.3 ● ● ● ● 0.0 ● ● ● ● ●●● ● ●● ● ●●● ● ●● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ●● ● ●●● ● ● ● ●●● −0.3 ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● −0.6 11 12 Employment Density 13 11 12 Employment Density 13 (C) Distance to City Center and Survey Workplace Income ● ● ● Log Survey Income ● ● ● ● ●● ● ●●● ● ● ●● ● ●● ●●● ● ● ●● ●●● ●● ●● ● ●● ● ● 9.5 ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 9.0 ●● ● ●● ● ● ● Log Survey Income (Residualized) 10.0 0.3 ● ● ● ● ●●● ● ● ● ● ●● ● ●● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● 0.0 ● ● ● ●● ● ●● ● ● −0.3 ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● −0.6 6 7 8 Log Distance to CBD 9 6 7 8 Log Distance to CBD 9 Notes. The scatter plots in the three columns correspond to columns 1, 2, and 3 in Table 2. The left-side graphs correspond to panel A (log survey income), and the right-side graphs to panel B (the residual of log survey income after regressing on demographics). Each circle represents a workplace survey area, with radius proportional to employment population obtained from the cell phone data. Figure B.5: Average Residential Income: Model Prediction and Nighttime Lights (A) Dhaka Log VIIRS Nighttime Lights 4 2 0 −2 −6 −4 −2 0 2 Log Model Income (residential, demeaned) 4 (B) Colombo Log VIIRS Nighttime Lights 2.5 0.0 −2.5 −5.0 −2.5 0.0 2.5 Log Model Income (residential, demeaned) Notes. The two scatter plots correspond to the first column, panels A and B, in Table 3. Each circle represents residential cell phone tower, with radius proportional to residential population obtained from the cell phone data. -.4 Prop change relative to work days -.3 -.2 -.1 0 .1 Figure B.6: Model-Predicted Income by Calendar Date (Hartals, Holidays and Weekends) 9Aug 1Nov 8Nov 16Aug 23Aug 30Aug 6Sep 15Nov 22Nov 29Nov 6Dec 13Sep 20Sep 27Sep Prop change relative to work days -.4 -.3 -.2 -.1 0 .1 2Aug Weekdays Public Holiday 13Dec 20Dec 27Dec Hartal Friday, Saturday Notes. This figure shows average predicted income by calendar date. The Y axis plots the proportional change relative to the mean on workdays. The sample and outcome are as in Panel A, Column 1 in Table 4. The figure plots calendar date fixed effects from a regression of normalized trip predicted income (equal to zero when the person does not travel) on commuter and calendar date fixed effects. Hartal dates are from Ahsan and Iqbal (2015) and public holidays from https://www.timeanddate.com/holidays/bangladesh/. Friday is the main free day in Bangladesh, and Saturday is the other weekend day. August 2 is Jumatul Bidah, August 6 is Shab-e-qadr, August 9-12 is the Eid ul-Fitr (end of Ramadan), August 15 is the National Mourning Day, August 28 is Janmashtami, November 14 is Ashura, December 16 is Victory Day, and December 25th is Christmas Day. The last week in December preceded the General Election of January 5, 2014. C Additional Tables Table C.1: Cell Phone Data Coverage at User-Day Level Dhaka, Bangladesh Colombo, Sri Lanka (1) Users in sample (2) Days in sample (3) All user-days possible = (1)×(2) 5.3 · 106 122 6.5 · 108 3.0 · 106 395 1.2 · 109 (4) (5) (6) (7) User-days User-days User-days User-days 2.9 · 108 1.5 · 108 2.4 · 108 1.0 · 108 3.4 · 108 (8) Coverage rate =(7)/(3) 16.1% 28.8% with with with with data data (5-10am) data (10am-3pm) data (5-10am and 10am-3pm) Notes: This table describes data coverage in the two countries. The first row indicates the number of unique users (who appear at least once in the data set). The second row shows the total number of calendar dates with data. The third row is the product of the previous two, which is the theoretical upper bound of user-day combinations that could appear in the data. (Note that in practice some users only start using a cell phone partway through the period, so this is an overestimate.) Rows 4-6 describe the actual number of user-days in the Bangladesh data under different restrictions. The seventh row shows the number of user-days for which we have at least one location between 5 am and 10 am, and at least one location between 10 am and 3 pm – this corresponds to the data necessary to define commuting behavior for that user and that day. Table C.2: Comparison of Commuting Flows from Survey Data and Cell Phone Data Flow survey data (DHUTS) (1) Log flow cell phone data ∗∗∗ 0.69 (0.024) (2) ∗∗∗ 0.80 (0.028) Log duration Origin and destination fixed effects Observations 7915 Yes 7915 (3) ∗∗∗ (4) 0.32 (0.068) 0.75∗∗∗ (0.059) -1.13∗∗∗ (0.18) -0.14 (0.14) 7915 Yes 7915 Notes: This table shows the relationship between commuting flows from two different data sets in Dhaka: the DHUTS transportation survey (outcome) and from cell phone data (explanatory variable). An observation is a pair of survey areas from the DHUTS survey. The coefficients show the estimates from the Poisson pseudo-maximum-likelihood (PPML) estimation of DHUTS commuting flow on log flows from cell phone. We use PPML to deal with the presence of zeros in DHUTS commuting flows (Silva and Tenreyro, 2006). If cell phone commuting flow data is a perfect measure of commuting flows, one would expect coefficients equal to one; the results indicate that cell phone commuting flows contain some measurement error. Standard errors are clustered at the origin survey area level. ∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01. Table C.3: Comparison of Residential Population from Cell Phone Data and Population Census log Residential Density (cell phone) log Residential Density (census) (1) (2) 1.16∗∗∗ (0.03) 1.16∗∗∗ (0.14) log Residential Population (census) City Observations Adjusted R2 Dhaka 1,866 0.61 Colombo 1,201 0.49 log Residential Population (cell phone) (3) (4) 0.57∗∗∗ (0.07) 0.40∗∗∗ (0.04) Dhaka 1,866 0.25 Colombo 1,201 0.24 Notes: This table shows the relationship between residential population and population density from cell phone data and from the population census. The unit of analysis is a Voronoi cell around each cell phone tower in the greater metropolitan area of each city (Dhaka, Gazipur, and Narayanganj districts in Bangladesh, and Western Province in Sri Lanka). In cell phone data, residential population is defined as out-commuting flow, namely the total number of commuting trips from a given origin excluding stationary trips (including them yields virtually identical results). Census residential population in a Voronoi cell is computed as the average census population in census geographic units (Mauza for Bangladesh, Grama Niladhari for Sri Lanka), weighted by their overlap with the Voronoi cell. The high adjusted R-squared in columns (1) and (2) indicates a strong association between the geographic density from the two data sources. The slope above one indicates that the cell phone data slightly over-represents residential population in denser areas. The comparatively lower adjusted R-squared in columns (3) and (4) is likely because cell phone operators tend to assign cell phone towers to equalize the subscriber coverage per tower. Conley standard errors with 5 km distance cutoff shown in parentheses. ∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01. Table C.4: Gravity Equation Robustness: Destination Fixed Effects Destination Fixed Effects (Benchmark) (1) Dest FE (Gravity w/ Model Origin FE) (2) (3) 0.94∗∗∗ (0.002) (5) (6) (7) 2.48∗∗∗ (0.04) 1.41∗∗∗ (0.01) Dest FE (Volume + 1) 1.08∗∗∗ (0.002) 0.98∗∗∗ (0.001) Dest FE (Full Sample) Dhaka 1,868 0.99 (8) 0.97∗∗∗ (0.004) 1.92∗∗∗ (0.02) Dest FE (Home-Work Flows) City Observations Adjusted R2 (4) Dhaka 1,860 0.81 Dhaka 1,868 0.94 Dhaka 1,868 1.00 1.00∗∗∗ (0.0004) Colombo 1,201 0.98 Colombo 1,201 0.75 Colombo 1,201 1.00 Colombo 1,201 1.00 Notes. This table compares destination fixed effects computed under different assumptions. The outcome in the first (last) four columns is the destination fixed effects from the first (third) column in Table 1. The first row uses fixed effects (FE) from the gravity equation estimated using the iterative procedure (columns 2 and 4 in Table 1). The second row uses FE from the gravity equation estimated using home-work flows instead of daily commuting trips (see Section 2.1 for the definition). The third row uses FE from the gravity equation estimated using log (Vij + 1) instead of log (Vij ) as outcome variable. The last row uses FE from the gravity equation estimated on all tower pairs below the 99th percentile of the travel time, with travel time censored from below at 180 seconds. Most coefficients are close to 1 and the R2 is above 0.9, except for home-work flows. This is due to slightly lower coverage for home-work flows with positive values. Some commuting pairs have zero home-work flows, and are hence omitted from the gravity equation. This sample selection tends to overestimate destination fixed effects for locations with low wages. This leads to a flatter profile of destination fixed effects. Standard errors in parentheses. ∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01. Table C.5: Robustness: Average Workplace Income and Survey Income Comparison (A) Log Survey Income log Survey Income (workplace) (1) Gravity w/ Model Origin FE log Model Income (workplace) Geographic Controls Adjusted R2 Observations (2) Home-Work Flows (3) Excluding Neighboring Towers (4) Without Area Adjustment (5) Include All Origins 0.14∗∗∗ (0.03) 0.26∗∗∗ (0.06) 0.15∗∗∗ (0.04) 0.44∗∗∗ (0.11) 0.14∗∗∗ (0.03) 0.27∗∗∗ (0.06) 0.31∗∗∗ (0.05) 0.21∗∗∗ (0.05) 0.13∗∗∗ (0.03) 0.25∗∗∗ (0.06) 0.26 88 X 0.46 88 0.19 88 X 0.44 88 0.26 88 X 0.46 88 0.3 88 X 0.45 88 0.23 89 X 0.49 89 (B) Log Survey Income Residual on Demographics log Survey Income (workplace, residual) (1) Gravity w/ Model Origin FE log Model Income (workplace) Geographic Controls Adjusted R2 Observations (2) Home-Work Flows (3) Excluding Neighboring Towers (4) Without Area Adjustment (5) Include All Origins 0.07∗∗∗ (0.02) 0.15∗∗∗ (0.03) 0.08∗∗∗ (0.02) 0.25∗∗∗ (0.07) 0.07∗∗∗ (0.02) 0.15∗∗∗ (0.03) 0.16∗∗∗ (0.03) 0.12∗∗∗ (0.03) 0.07∗∗∗ (0.02) 0.13∗∗∗ (0.03) 0.23 88 X 0.3 88 0.17 88 X 0.28 88 0.23 88 X 0.3 88 0.25 88 X 0.29 88 0.22 89 X 0.31 89 Notes. Robustness for columns 1 and 5 from Table 2 comparing model predicted workplace income and survey income. Odd and even columns correspond to the specifications in columns 1 and 5 in Table 2. The first two columns use destination fixed effects from the gravity estimated using the iterative procedure (column 2 in Table 1). The next two columns use home-work commuting flows instead of daily commuting flows (see Section 2.1 for the definition). The next two columns define workplace income at the survey-area level excluding commuters whose origin towers are within 180 seconds of the destination cell tower, when we aggregate up from cell tower level. The next two columns use destination fixed effects not adjusted for Voronoi cell tower. The last two columns include commuters from DHUTS survey whose origin locations are outside the DCC area (see Section 2.1). Table C.6: Individual Income: Model Predictions and Survey Data log Survey Income (1) (2) (3) 0.14∗∗∗ (0.02) 0.05∗∗∗ (0.02) 0.04∗∗∗ (0.01) 0.12∗∗∗ (0.02) 0.09∗∗∗ (0.01) log Dest. Dist. to CBD −0.09∗∗∗ (0.02) −0.03 (0.02) log Dest. Commuting Zone Area −0.07∗∗ (0.03) −0.02 (0.02) Model log Income (workplace) log Travel Time Male 0.48∗∗∗ (0.03) Age 0.02∗∗∗ (0.001) Level of education 0.21∗∗∗ (0.01) Origin FE Occupation and Sector FE Observations Adjusted R2 X 11,006 0.02 11,006 0.09 X X 11,006 0.50 Notes: This table regresses log income from the DHUTS survey on model-predicted income and controls. The unit of observation is a survey respondent in the sample described in Table 2. Model-predicted income for a pair of origin and destination survey areas is the weighted average of tower-pair model income, with weights given by tower-to-tower commuting flows. Formally, for P survey areas a and b, yab ≡ i∈a,j∈b Vij /Vab · yj , where i ∈ a and j ∈ b index towers, yj = ψ̂jR is P the area-adjusted destination fixed effect at j, and Vab ≡ i∈a,j∈b Vij is the total flow between a and b. We assign to each survey respondent the predicted income between his or her home and work survey areas. Columns 2 and 3 include origin survey area fixed effects, and column 3 includes occupation and job sector fixed effects. Two-way clustered standard errors clustered by origin and destination survey area reported in parentheses. ∗ p ≤ 0.10, ∗∗ p ≤ 0.05, ∗∗∗ p ≤ 0.01 Table C.7: Robustness: Average Residential Income and Nightlights Comparison (A) Dhaka log VIIRS Nighttime Lights (1) Gravity w/ Model Origin FE log Model Income (residential) Geographic Controls Sub-district FE Observations Adjusted R2 0.59∗∗∗ (0.04) 0.54∗∗∗ (0.07) 1,868 0.70 X X 1,868 0.84 (2) Home-Work Flows 0.54∗∗∗ (0.04) 0.31∗∗∗ (0.08) 1,854 0.71 X X 1,854 0.84 (3) Excluding Neighboring Towers (4) log of Mean Income  = 6.84 0.65∗∗∗ (0.04) 0.30∗∗∗ (0.04) 3.99∗∗∗ (0.31) 3.17∗∗∗ (0.53) 1,867 0.65 X X 1,867 0.83 1,868 0.70 X X 1,868 0.83 (5) Without Area Adjustment −1.03∗∗∗ (0.11) −0.32∗∗∗ (0.09) 1,868 0.19 X X 1,868 0.82 (B) Colombo log VIIRS Nighttime Lights (1) Gravity w/ Model Origin FE log Model Income (residential) Geographic Controls Sub-district FE Observations Adjusted R2 0.83∗∗∗ (0.03) 0.50∗∗∗ (0.05) 1,199 0.85 X X 1,197 0.93 (2) Home-Work Flows 0.88∗∗∗ (0.03) 0.31∗∗∗ (0.07) 1,195 0.86 X X 1,193 0.93 (3) Excluding Neighboring Towers (4) log of Mean Income  = 6.84 0.90∗∗∗ (0.06) 0.39∗∗∗ (0.04) 5.50∗∗∗ (0.17) 2.93∗∗∗ (0.29) 1,199 0.82 X X 1,197 0.94 1,199 0.86 X X 1,197 0.93 (5) Without Area Adjustment 0.60 (0.41) −0.32∗∗∗ (0.04) 1,199 0.03 X X 1,197 0.93 Notes. Robustness for columns 1 and 4 in Table 3 comparing model predicted residential income and nighttime lights. Odd and even columns correspond to the specifications in columns 1 and 4 in Table 3. The first two columns use destination fixed effects from the gravity estimated using the iterative procedure (column 2 in Table 1). The next two columns use home-work commuting flows instead of daily commuting flows. The next two columns define workplace income at the survey-area level excluding commuters whose origin towers are within 180 seconds of the destination cell  tower.  The next two columns compute model-predicted log average income (rather than  P 1 R average log-income), namely log N j exp ψ̂j / , using  = 6.84 from Ahlfeldt et al. (2015) (results are not sensitive to ). The next two columns use destination fixed effects not adjusted for Voronoi cell tower. Table C.8: Robustness: Average Residential Income and Nightlights (central Dhaka only) log VIIRS Nighttime Lights (1) log Model Income (residential) (2) (3) 0.24∗∗∗ (0.03) −0.02∗ (0.01) log Residential Density −0.16∗∗∗ (0.04) log Dist. to CBD log Model Income (workplace) Sub-district FE (count) Observations Adjusted R2 (4) (5) (6) 0.15∗∗ (0.07) 0.12∗∗∗ (0.04) 0.09 (0.06) −0.06∗∗∗ (0.02) −0.02∗∗∗ (0.004) −0.02∗∗ (0.01) −0.08∗∗ (0.04) −0.16∗∗∗ (0.04) −0.16∗∗∗ (0.04) 0.05 (0.05) 1,202 0.23 1,202 0.01 1,202 0.15 1,202 0.29 0.03 (0.03) X (39) 1,202 0.66 X (39) 1,202 0.66 Notes. Version of Table 3 restricting to cell phone towers within the Dhaka City Corporation (DCC) area, which is the area covered by survey data in Table 2. (For the spatial extent of DCC, see panel B in Appendix Figure B.1). The lower slope is consistent with the curvature visible in Appendix Figure B.5, panel A. We lose significance in the last column, which includes sub-district fixed effects, geographic controls, and the model workplace income measure. Table C.9: Average Residential Income and Survey Income (A) log Survey Income log Survey Income (residential) (1) log Model Income (residential) (2) (3) 0.06 (0.06) 0.11∗ (0.06) log Residential Density −0.09∗ (0.05) log Dist. to CBD (4) (5) −0.10 (0.09) −0.12 (0.25) 0.14∗∗ (0.06) 0.14∗∗ (0.07) −0.10∗ (0.05) −0.10 (0.06) log Model Income (workplace) Adjusted R2 Observations 0.01 (0.14) 0 73 0.02 73 0.05 73 0.06 73 0.04 73 (B) Nighttime Lights at Survey Area Level (central Dhaka only) log VIIRS Nighttime Lights (1) log Model Income (residential) (2) (3) 0.25∗∗∗ (0.09) 0.15∗ (0.08) log Residential Density −0.18∗∗∗ (0.03) log Dist. to CBD (4) (5) 0.22 (0.14) 0.03 (0.24) −0.09 (0.10) −0.10 (0.09) −0.13∗∗ (0.06) −0.15∗∗ (0.06) log Model Income (workplace) Adjusted R2 Observations 0.12 (0.09) 0.2 73 0.06 73 0.24 73 0.29 73 0.3 73 Notes. Versions of Table 3 at the level of the 73 survey areas with at least one observation with reported income and residence in that survey area in the DCC. Panel A uses average residential income from survey data, panel B uses nighttime lights. The low explanatory power in all columns in panel A is consistent with low signal-to-noise ratio for average residential income measured using survey income (i.e., in Dhaka, at this level of aggregation, residential income is more equally distributed compared to workplace income). The fact that results in Panel B are more statistically significant suggests that nighttime lights may be a more precise measure of average residential income. Weaker results in panel B relative to Table 3 indicate that our model weakly performs better at finer spatial units. (This is not because of the exclusion of samples outside urban core of Dhaka; see Appendix Table C.8.) Table C.10: Impact of Hartal: Robustness with Very Frequent Callers (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∗∗∗ (0.010) −0.048∗∗∗ (0.016) Friday (free day) −0.236∗∗∗ (0.008) −0.486∗∗∗ (0.015) 7.3 · 106 7.3 · 106 Hartal −0.038∗∗∗ (0.008) −0.042∗∗∗ (0.015) −0.043∗∗∗ (0.008) −0.070∗∗∗ (0.019) −0.046∗∗∗ (0.009) −0.077∗∗∗ (0.020) Friday (free day) −0.176∗∗∗ (0.007) −0.462∗∗∗ (0.014) −0.207∗∗∗ (0.008) −0.576∗∗∗ (0.018) −0.226∗∗∗ (0.009) −0.632∗∗∗ (0.019) Hartal x Low Income 0.012∗∗∗ (0.003) 0.069∗∗∗ (0.012) 0.010∗∗∗ (0.003) 0.066∗∗∗ (0.011) Friday x Low Income 0.074∗∗∗ (0.004) 0.269∗∗∗ (0.012) 0.068∗∗∗ (0.004) 0.251∗∗∗ (0.012) Hartal x Short Commute 0.012∗∗∗ (0.002) 0.025∗∗∗ (0.006) Friday x Short Commute 0.061∗∗∗ (0.003) 0.188∗∗∗ (0.005) 7.3 · 106 0.82 7.3 · 106 0.40 Panel A. Predicted Income Observations Panel B. Make a Trip Observations Workday Mean 7.3 · 106 0.82 7.3 · 106 0.40 7.3 · 106 0.82 7.3 · 106 0.40 Notes. Version of Table 4 restricted to the sample of commuters with commuting data on at least 75% of all days (90 out of 122 days), who are 9% of all commuters. Table C.11: How Preference Shocks and Travel Time Affect Income: Estimated Structural Parameters (1) (2) Full model (3) (4) Constrained model (αd = 0) Shock productive αz 0.47 (0.06) 0.32 [3.37] 0.57 (0.07) 0.57 [0.09] Shock distance αd -0.19 (0.51) -0.47 [5.99] 0 Shape parameter  7.33 (4.22) 9.27 [38.07] 6.32 (0.78) 6.40 [0.73] Observations SE Method Bootstrap clusters 11,006 11,006 Delta Bootstrap 70 11,006 Delta 11,006 Bootstrap 70 Notes. This table reports estimates of the structural parameters that control the degree to which idiosyncratic shocks affect income (αz ), travel time affects income (αd ), and the Fréchet shape parameter , using the procedure described in Appendix A.3. We estimated equation (9) by regressing individual log survey income from the DHUTS survey on the three model-predicted terms. In columns 3 and 4, we restrict the third coefficient that corresponds to travel time to be zero (ρ3 = 0). The estimates for αz , αd and  in this table are transformations of the estimated OLS coefficients as detailed in equation (10). Columns 1 and 3 report standard errors computed using the Delta method. Columns 2 and 4 report results from 100 bootstrap runs where we cluster at the origin survey area level (70 survey areas with at least one out-commuter in DHUTS survey): the median estimate in the first row and standard errors in square parentheses.