Embed Size (px)
Citation preview
Roads and Urban Growth∗
P. BrandilyF. Rauch
September 12, 2019
Abstract
Roads are essential to facilitate interactions in cities. A sub-optimal road layoutin cities and towns may inhibit interactions and constrain urban population growth.Using data from over 1800 cities and towns from Sub-Saharan Africa, and instrumentsbased on the history of foundation ages of cities in Africa, we study the relationshipbetween the road density and other measures of the road layout of a city and itspopulation growth. We show that cities characterised by low road density in the citycentre experience less population growth in recent decades.
JEL-classification: O18, R4Key words: Road layout, Urban planning, Urbanization, Sub-Saharan Africa
∗Brandily: Paris School of Economics, Rauch: Oxford University. Contact email address: [email protected]. We gratefully acknowledge the generous support of a Global ResearchProgram on Spatial Development of Cities, funded by the Multi Donor Trust Fund on Sustainable Urban-ization of the World Bank and supported by the UK Department for International Development.
1
1 Introduction
Cities enhance productivity, and the literature distinguishes several channels through whichthis effect comes about. Some describe agglomeration effects related to production, oth-ers relate to consumption. Marshall (1920) suggested knowledge spillovers, labour marketexternalities and production linkages as mechanisms. More recently, Duranton and Puga(2004) highlight the importance of sharing, matching and learning. What these and otherproposed mechanisms for agglomeration effects have in common is that the essential fea-ture of an urban environment consists in facilitating short distance interactions betweeneconomic agents. In this paper we study the impact of the density of the road network,and other related measures of the road grid, on population growth in recent decades. Wegive some evidence on the properties of an effective road networks, which we can do withouttaking a stand on the underlying agglomeration effects driving it. The importance of theroad grid for urban development has been recognised before. Fuller and Romer (2014) write:“... getting the grid right will likely be more important than enforcing building codes onstructures or imposing limits on density ... ”. The last major report from the World Bankon Africa’s Cities states similar arguments in favor of “grid structure, which (. . . ) enhancedtravel efficiency” (Lall, Henderson and Venables 2017). Yet there is little empirical researchso far to assess this hypothesis.
In towns and cities that do not have an underground or extensive rail network and hencerely overwhelmingly on transportation by cars, motorbikes or busses, the distance betweenpeople can be approximated by studying the road layout. This is true in most towns andcities of Africa. The layout of roads is costly to change once they are in place. Further, theoptimal layout of roads depends on the available transport technology, which changes overtime. These observations imply that due to historic reasons some cities may be accidentallystuck with better or worse road grids, which would benefit or hurt their growth potentialin recent decades where Africa urbanised rapidly. If a city by historic accident develops agood road layout it will be rewarded by long lasting population growth, as people move into benefit from the agglomeration benefits generated in the city. If on the other hand a cityis stuck with a sub-optimal grid, it may reach a peak population level beyond which it findsit difficult to grow any further.
Africa is a useful environment to study this question for a number of reasons. First,cities and towns of Africa rely heavily on transportation on roads, and do not have manyurban train networks or underground networks. This allows us to approximate connectiv-ity with the road network in cities with less noise than elsewhere. Second, African citieshave experienced much population growth in recent years, which provides variation in thepopulation growth data to study this question. Finally, policy conclusions from this studyare particularly relevant for the fast urbanising continent, and hence evidence from Africa isof greater current importance to policy makers than evidence from other continents wouldbe. Various studies stress the importance of road investments within African cities: hiddenexternal costs of congestion were estimated at up to 5% of cities’ GDPs in Dakar or Abidjan,more than twice as much as the estimates for European cities (Cervero, 2013) while “in asample of 30 cities around the world, the 8 African cities rank in bottom 12 spots for roaddensity” (Lall, Henderson and Venables 2017).
Our main measure of the quality of the road network is a simple measure of road densityin the centre of towns and cities. We also compute a second and secondary simple measureof the distance to the nearest road for a random point in the city centre, as well as a numberof measures of regularity and orientation of the road grid. All measures we use in this paperhave the advantage to be intuitive, straight forward to compute and simple to interpret.Our conclusion focuses on the linear effect of a simple measure of road density, which is themost robust and quantitatively most important of all the measures we consider. To address
2
measurement error and potential endogeneity, we rely on an instrument for the age of a city,derived from the colonial history of Africa. We find that especially the simple measure ofroad density correlates with population growth robustly.
Our findings have important implications for policy. The results point to the importanceof carefully planned road grids when a new city emerges, as it will influence urban populationgrowth for many decades to come. Grid plans should take the growth trajectory of a cityinto account. In particular, insufficient density of roads in city centres may slow populationgrowth and harm urban development. Given that sub-optimal road grids remain in placefor many decades, it may prove more efficient to focus on the construction of new citiesand neighbourhoods with better grids rather than improving existing cities. Our resultsalso highlight that other urban policies to improve the flow of traffic (such as car sharing ortaxation approaches1) could lead to population growth.
While there is a substantial literature on the effect of roads or other forms or trans-portation on urban shape (recent examples include Baum-Snow 2007, Duranton and Turner2012, Faber 2014, Jedwab and Moradi 2016, Jedwab and Storeygard 2016, Michaels 2008and Storeygard 2016), this literature typically has concentrated on roads connecting citieswith one another. Less is known on roads and road layout within cities (exceptions includeBaum-Snow et al 2017, who consider the effect of urban highways in China and Coutureet al 2018 who suggest that restricting traffic would bring welfare gains). Our approach ismore focused on within-city variation, and includes smaller roads than these existing stud-ies. The main contributions of this paper are threefold. First we build a rich panel datasetof roads and population for more than 1800 cities and towns in Sub Saharan Africa thatmay be useful beyond this paper. Second, we provide evidence for a hypothesis advancedby Collier and Venables (2016) that the road network is an important predictor of recentpopulation growth, and that many cities and towns in Sub Saharan Africa are constraintin their relative growth due to lack of road density. Third, this leads to important policyconclusions on the importance of planning road networks long-term, and the difficulty infixing a city locked in a state of sub-optimal infrastructure provision.
The paper is organised as follows: Section 2 develops our hypothesis in greater detail,and discusses its underlying assumptions. Section 3 describes the datasets used and presentsdescriptive statistics. Section 4 develops and discusses the instrument we use for the age ofa city. Section 5 presents the main results. Section 6 concludes.
2 Hypothesis and main measures
Our thinking about this problem, our hypothesis and our choice of instrumental variable areshaped by on six findings that are commonly reported in the urban economics literature.Given that they are central to our argument, we discuss each of these assumptions one byone in this section.
First, we assume that the returns from agglomeration are largely generated in the citycentre. This is one of the standard assumptions used in monocentric city models. Thissuggests that people live around a city centre, to which they travel to work and enjoy leisuretime. The agglomeration returns are overwhelmingly generated in that centre. We takefrom this assumption that we pay special attention to the road grid in the centre of citiesrather than the whole road grid structure over the entire urban agglomeration. Althoughwe also study the city as a whole for comparison for some of our measures, our preferredmeasure captures only the centre. Empirical evidence for the monocentric model has beenfound for a variety of cities, by demonstrating gradients of wages and house prices, as wellas from commuter flows (for recent evidence for Africa see Antos et al 2016 or Larcom et
1For recent examples in developing countries see Hanna et al 2017 and Kreindler 2016 respectively.
3
al 2017). If anything, monocentricity has been found to be more pronounced in Africa thanin the developed world (Cervero 2013, Grover Groswami and Lall 2016). Following thisassumption, we define a city centre by the brightest spot of night light of a city, and drawa circle around it. In our main specification, this circle has a radius of 1km. This approachhas the additional advantage that we do not need to define the edge of a city and how itchanges over time, and that having similar sized areas makes the measure more comparableacross cities. We do, however adjust the central area of a city in the case where we findrivers, coasts, or a boarder (see section 3.1 for more details).
Second, we make the assumption that the city centre is unique for each city we study, andwe define only one centre for each town or city. This again is a standard assumption made inmonocentric city models. In the context of our study this assumption is strengthened giventhat we include not only big cities, but by large majority small and medium sized towns,for which this assumption is more plausible. We also exclude capital cities in our preferredspecification, which excludes some of the largest cities, which are more likely to have morethan one centre, from our analysis. Antos et al (2006) provide evidence that is consistentwith monocentric centres for a few large African cities, it is even more justified for smalland medium sized cities, which are not large enough to support two centres. We can verifyin our dataset that this assumption does not seem to be contradicted by the data.
Third, we assume that cities grow in a circular way around a given centre. Or to rephrase,we assume that the location of the city centre does not change much over time. This is trueif the foundation of a city followed locational fundamentals that do not change much overtime. Evidence can be found from all the many cities in the world that are still structuredaround Medieval or Ancient sites. A recent empirical paper making this assumption isHarari (2016). By this assumption, a city with a better working city centre will developa larger commuter zone around it over time. This assumption is important to inform ouridentification strategy, since it allows us to assume that the historic centre and the moderncentre coincide. Again, this assumption is easier justified for smaller towns and cities, suchas we have in our dataset.
Fourth, we assume that road variables that we compute are important to facilitate theagglomeration returns, whatever these returns may be. This could be violated if most of theagglomeration returns happen overwhelmingly within buildings, and are less dependent onconnection in the centre. Even in this case people would have to get to these buildings, butquantitatively the true agglomeration effects may be poorly approximated by our measures.This would also be violated if true connection in a city depends on factors other than thosethat we measure. It could be violated by some strict version of the ‘fundamental law of roadcongestion’ (Duranton and Turner 2011) applies, which could imply that road constructiondoes not relieve congestion.
Fifth, we assume that there is path dependence of the road layout. Once a layout hasbeen established, it is costly to change it, and the grid from one year overlaps strongly withthe grid from another year. While there are examples of radical changes in the layout ofcity maps (Haussmann’s renovation of Paris in the 19th century is a famous example), thegreat majority of cases that come to mind suggest indeed a tendency of lock-in. This is evenmore plausible in the case of small and medium sized towns and cities, which constitute themajority of our sample.2 Baruah et al (2017b) compare colonial and modern street mapsfor a few African cities. They find that “the spatial structures of cities in Sub-SaharanAfrica are strongly influenced by the type of colonial rule experienced” and they point to“high costs of acquiring new rights of way in an already built-up city”. Their examples thatcontrast colonial and modern maps show strong persistence of road layout. Shertzer et al.
2Examples of large and successful cities whose central road layouts still reflect their Ancient or Medievalpast abound in Europe. For evidence of path dependence for cities see Michaels and Rauch (2017). Forevidence of path dependence of infrastructure in Africa see Michaels et al (2018).
4
(2016) argue that Chicago zoning ordinances from 1923 have a larger effect on the spatialdistribution of economic activity in the city today than geography or transport networks.Without this assumption cities would be able to self correct, our first stages would not bestatistically significant, and our measures of the road grid would be poor predictors of futurepopulation growth.
Sixth, we assume that the optimal grid depends on the available transport technology.The ideal grid in the age of walking and the age of the horse differs from the ideal grid in theage of the car. For example, the latter might require broader roads and might avoid crossingsof roads harder. If so, the age of a city influences its initial grid, which in turn influences itsgrid later on. This assumption helps us to develop our instrumental variable, which uses acorrelate of the age of a city as an instrument for the initial road grid specification.
Taken together, these six results imply that initial differences of road layout, even iffrom a very distant past, can lead to different long run population developments. Due tosmall differences in geography and local history, some towns may accidentally stumble upona more successful layout than others. These initial differences depend on the age of a city,which is potentially observable.
These assumptions do not tell us what a successful layout should look like. Our mainmeasure is a simple measure of the density of roads in the city centre, which may be animportant component of a successful grid. The correct relationship between the observablegrid and the average travel speed is not obvious. Having roads with multiple lanes may behelpful in bigger cities, but wasteful in smaller ones. Crossings may slow traffic down, butenable useful travel combinations. In some cities, part of the road may serve as parkingspace, which would facilitate travel speeds, in others this effect may be limited by parkingrestrictions. A simple measure that does not take a strong stand on these and other trade-offsis simply to consider the density of roads, measured by road kilometres over area considered.This strikes us as a simple and straight forward measure of road layout in cities. Whencomputing this density variable we count a roads recorded in the open street map data withn lanes n times.
In addition, an ideal measure to characterise a useful grid should capture how wellconnected people and firms in the city centre are. This could be approximated by estimatingthe travel time between two randomly chosen points in the centre.3 We limit the means oftransportation to travel on the road and travel off the road. With an assumption on therelative travel speed on and off the road, we could compute the travel speed between tworandom points. The second measure of interest is a measure of how easy it is to access roadnetwork from a random point in the city centre. Facing a similar problem, Donaldson andHornbeck (2016) create 200 random points within US counties, calculate the distance fromeach point to the nearest railroad, and take the average of these nearest distances. Followingthis algorithm, we create 12 points in each circle, measure the distance to the nearest roadfor each of these points, and compute the average nearest distance to the road from these 12points.4 We call this measure evenness, as it tells us how evenly existing roads are spreadover the centre. To illustrate what we have in mind, consider Figure 1. This figure presentstwo cities with stylised road layouts. Both cities have the same road density in the centreby construction. The city on the right has a shorter distance from a random point to thenearest road. An example of a city with the 12 points is shown in Figure 5.
3In a strict interpretation of the monocentric city model, we could specify alternatively the travel timebetween a random point and the centre of the city, defined as a point. These two formulations of the problemlead to fairly similar answers.
4Using 12 points is a low number in comparison to 200. We do so for computational convenience. Asa justification, we also record the mean distance from the centre of the circle to the nearest road, which iseffectively the same exercise with that parameter set to one. The correlation between the exercise of 1 and12 points is around 0.9. Given this large correlation, we believe that all such measures would give fairlysimilar numeric estimates.
5
We consider a few transformation and additions to these main variables in the speci-fications below. We decompose the road density into different types of roads as classifiedby open street maps, such as motorway, primary, secondary and small roads. We considermeasures of how regular the grid is, which is explained in more detail below. We also pro-vide separate counts of nodes and edges in the road grid, and we consider measures of howaligned the direction of roads is. As mentioned before, we mainly measure these variablesfor the city centre, but also repeat the exercise considering the grid of the city as a whole.
3 Data
We build a complete and consistent sample of cities, combining satellite imagery and geospa-tial population data in addition to other data sources. For each of these cities, we observecentre’s road layout and measure a set of descriptive statistics. This section briefly de-scribes datasets and measurement, all sources and processes are detailed further in the dataappendix.
3.1 Defining Cities: Boundaries and Centres
Our main source for the measurement of cities location and boundaries is satellite imagery ofluminosity at night. This source has the double advantage of being complete (i.e. it coversthe entire continent) and consistent (i.e. all the cities are defined in the same way). First,we identify lit areas from NOAA’s DMSP-OLS satellite record (cf. data sources) by keepingpixels that emit light at least twice over the five yearly observations from 2008 to 2012.Contiguous lit areas are then aggregated using a GIS software to create unions. Each unionrepresent a city, and we consider its footprint as the city boundaries. Figure 2 illustratesour sample. As the Figure shows, we consider cities in almost all countries of Sub-SaharanAfrica, with healthy spatial variation.
Our focus is on city centres, where we assume agglomeration effects to take place. Be-cause, to the best of our knowledge, there is no dataset providing precise coordinates ofcities’ centre we rather follow a systematic data driven definition. For each city, we take as“centre” the centroid of the brightest area. This area could be a single lights pixel, but itcould be a larger area of contiguous pixels, which also allow us to compute a centroid. Incase of multiple centres of equal brightness in a city we choose the larger one. In theorythere could be a problem of multiple, separate centres of equal brightness and size for asingle city, but in practise we notice that this does not appear to occur in our dataset.
Two reasons at least support this definition of a city centre: first, because nightlightsglow, it is often the case that brightest area is close to the geographic centroid of the city.Second, at the same time, nightlights are also known to be linked with economic activity,even at the local level. In the hope of capturing historical centre, and in order to avoidtop-coding issues, we use early images of nightlights (1994-1996). Figure 3 illustrates thederived city centres for two small towns. The figure also shows the city extent, computedfrom lights data, and the road network of these towns. Note that the use of early images ofnightlights reduces our sample to cities that did emit lights in these early years.
3.2 City population
For this exercise, we need data on the population of African cities and towns, comparableacross space and time, ideally from a unified source. Several datasets fulfil these crite-ria. First, we could use night-lights data, that has been used to identify economic activityat spatial level.5 Second, we could use data from the Gridded Population of the World
5See Kocornik-Mina et al. (2019) for a recent survey of these data in the within-country context.
6
(GPW) in its 4th version, which provides population mapped into a regular grid. Third, theAfricapolis (2019) dataset by the OECD collects official population sources into a unifieddataset. Finally, the European Commission provide a GHS dataset, which is a spatial gridthat provides a population estimate for each cell (of 250 by 250m resolution) in four targetyears: 1975, 1990, 2000 and 2015. The population estimates are obtained from the spatialdistribution of census data, guided by GHS modelling of built-up presence (derived fromLandsat day-time satellite images). Of these datasets, we take the GHS dataset as our mainone, as it seems to us the most highly processed one of these. For each city and datasetin our sample, we sum all the cells falling within its boundaries and extract the populationcounts for the four target years.
We note that all of these datasets give population information that correlates fairlypositively. The coefficient of correlation for log population in the year 2000 from GHS withlog population from Africapolis is 0.72, with log population from GPW version 3 it is 0.53,with log population from GPW version 4 it is 0.56 and with light it is 0.54. We see that thecorrelation is particularly high for larger cities, while the disagreement comes mainly fromsmaller cities and towns. While most of our results in this paper use the GHS population asoutcome variable, we show below that the main result is robust to using these populationestimates from alternative sources.
3.3 Centre and Road layout
We construct city-centres’ roads layout using the publicly available data from Open StreetMap (OSM), a volunteered geographic information project consisting of information fromover 2 millions users. As with any user generated data, there is some concern that thedataset may be biased. Seidel (2019) studies OSM bias in Africa by comparing OSM mapswith information from other sources, primarily for hospitals. His findings imply that ourstandard specification with country fixed effects and additionally controls for the initial sizeof a city (which we always include) would address this bias to some degree. A specificationwith region fixed effects would absorb most of the bias. We provide evidence in this paperthat suggests that our main coefficients are robust to both these versions. Additionally, weshow that our results are robust to using large cities only, for which we expect less of a bias.
For each city we gather data on the road network at two different levels. At the citylevel, we take into account any road that falls within a city boundaries; at the centre level,we draw a 1km-radius circle around the centre point and keep only the (segments of) roadsthat fall within. Figure 4 gives an example of a small city and highlights the derived circleof the defined central part. At both city and centre level, we measure a series of descriptivestatistics about the road network. In particular we compute total length of roads, numberof nodes and intersections, as well as the density of each of these items (i.e. dividing themby the considered area)6.
As previously mentioned, we also compute two indices of the network spatial organisation.First, we create 12 points within each city centre circle and compute the average nearestdistance to the road from these 12 points. Figure 5 illustrates the construction of thisregularity measure, plotting the 12 points over a city centre (also showing the city andcentre road networks). Second, we follow G. Boeing (2018) in constructing orientationmeasure: we compute the bearing of each edge (i.e. segment of road between two nodes).We can then observe the distribution of all edges within a city centre across equal-sized binsof compass orientation (each bin covers 10 degrees). We then summarise the concentrationof roads among potential orientations by taking the Herfindahl index of this edge bearingdistribution. Figure 6 illustrates the construction of city centre orientation distributions.
6Intersections are nodes with strictly more than one street emanating from; thus excluding cul-de-sacs,Boeing, 2017.
7
The left two panels show the street grid in two city centres (Umm Ruwaba from Sudan atthe top, and Mohale’s Hoek from Lesotho at the bottom), defined as one mile radius circlesaround the centre point. The middle panels take the road segments from the first, and sortthem by their orientation, weighting segments by their length. The final two panels takethis information and display them as standard histograms. It is over these shares that wecompute the Herfindahl indices of street orientation. In this context, the Herfindahl indexis large if many roads run in parallel, and small if the grid is more chaotic.
Ideally we would want to measure the road layout at the beginning of the period overwhich we measure population growth, and not at the end. For a sample as large as oursthis is however data we were unable to obtain. Instead we use road maps that are closerto the end of the period as a proxy. We make the assumption that the road grid does notchange fast over time, particularly in the city centre, as discussed in our assumption 5. Wealso note that our instrument corrects potential bias arising from this measurement error.
3.4 Sample selection
The construction of our data set implies various steps where choices had to be done regardingsample selection. We summarise these steps here, making clear when and what decision weretaken. Further discussion is provided in the Data appendix.
When using nightlights over 2008-2012, we find 6,382 unions in Sub-Saharan Africa asdescribed above. We don’t use lights here to infer information on the amount of localeconomic activity, but simply to see the location and extent of cities, and to define theircentre. Of these, only 2,779 did emit enough light in 1994-1996 for us to define a city centre.We restrict the sample to those units. From these, we decided to keep only places that mettwo size criteria. First, we require the total city area to be greater than 2 square kilometres;and second, we require that cities have more than 2,000 inhabitants as measured in 2015.We add such steps in order to discard very small places for which the grid would not playany importance. This leaves us with mainly cities and towns. Some of those may still besmall, but we think are large enough for the grid in the centre to affect interaction of people.With these two criteria, we reduce our sample to 2,204 observations.7
We expect capital cities to follow unique development paths that are influenced by factorsof politics and international aid. Since we are trying to identify broad economic trends freefrom such intervention, we drop capital cities throughout, and focus on the other towns andcities. This represents a further reduction to 2,164 observations. We provide robustnessanalysis of this decision in the results below.
We also exclude several countries, for different reasons. First, we excluded South Africafrom the start, since we consider it too different from the rest of the sample. Second, Burundiand Somalia data were not available and we thus drop their cities. At this stage we observe2,141 cities. Third, we exclude Nigeria and Madagascar. Nigeria is a large country thatwould alone represent about 20% of the cities in our dataset. Yet, it is also a place wheremeasurement from imagery is noisy, notably because nightlight data is biased by flares morethan in the rest of the continent. We fear that including Nigeria would have our resultdepends to a large degree on one noisy set of observations. We also think this study ismore representative for Sub-Saharan Africa as a whole by excluding it, since then it doesnot depend to such a large degree on a single country. We do however verify that resultsare broadly similar when we include Nigeria.8 We also exclude Madagascar throughout thispaper, since the reasoning of our instrument does not apply there given its geography andhistory.
7We use the population estimates of 2015 in order to increase the extent of our definition of a city. Wealso verify that our main results are robust to the inclusion of these small places.
8In fact, typically statistical significance increases with the larger sample that includes Nigeria.
8
Finally, we exclude cities for which we did not find roads when extracting OSM data.These represent around 13.4% of the 2,141 cities sample. In doing so, we obtain our totalsample of 1,850. Data appendix discuss further this last step, providing a basic descriptionof the missing-data cities.
Table 1 provides a break-down of our samples by country. In 2000, the median city size isaround 33 thousand inhabitants, varying greatly across countries (from 2,500 in Djibouti toabove 100,000 in Chad and South Sudan, where we identify only 12 and 3 cities, respectively).Average population growth rate over the 2000-2015 period is estimated to be 2.04%. Andthere again, countries experienced different paths with city population growth spanningbetween 5.66% a year in Equatorial Guinea to negative growth in Swaziland (-1.04%).
As for the road network of these cities, Table 2 provides simple descriptive statistics.Here again, we find a high variation in the total amount of road length within cities (fromvirtually 0 to 28,621km) and even within centres (0 to 87km). Although the maximumvalues seem high, they reflect that cities defined by nightlights can reach very large size,magnified by lights glowing. In our sample, the largest area encompasses Lagos as well asperipheral cities like Abeokuta in the north or Ijebu Ode in the east. Table 2 also reports abasic descriptive of the distances from cities’ centre to the sea, and to the four cities (Dakar,Dijbouti, Cairo and Cape Town) we use four our IV specification.
4 Instrument and First Stage Regression
As discussed in Section 2, the age of a city is likely to vary with the initial road layout of acity, and because of path dependence also with the current layout. A city founded before theinvention and adoption of cars would have a higher density suitable for older technologies.We expect more modern cities, built in the era of increasing adoption of cars, having biggerroads and also smaller road densities.
One instrument we use is to use a measure of the foundation age of a city. Here we rely onthe Africapolis (2019) dataset, that records historic population for a large number of citiesin Sub-Saharan Africa. The oldest year in that dataset is 1950, and one instrument consistsof a simple dummy variable, indicating those cities that existed in that year according tothis dataset.
One downside of the age dataset is that it is only available for a subset of our cities,and reduces our sample. Potential endogeneity concerns could remain. Therefor we alsoconstruct another instrument, which is the distance from the cities of Cape Town, Cairo,Dakar and Djibouti. These four cities represent connection nodes of the two main colonialpowers of the continent. If a significant part of colonial expansion was conducted via theseports, distance to these four cities should influence the year of creation of modern cities onthe continent. If the age of a city correlates with the initial road layout, then the distanceto these four cities will also correlate with the initial road layout.
Urbanisation of cities in Africa was facilitated to a large degree by the colonial powers,especially the British and the French, the two main parties in the scramble for Africa. As lateas 1880, about 80 percent of Africa were ruled by Africa’s own kings and queens. Colonialpowers were found near the coast, with a British stronghold around the port of Cape Town,and French holdings around Dakar (Boahen 1985). The British invasion of Cairo of 1882established a second stronghold on the continent for the British that lasted well into the 20thcentury (Hourani et al 2004). Soon it became the vision of the British to connect these twocentres, and build the “Cape to Cairo” Railway line. Historically, these two cities played animportant part as connection points of British colonial Africa, and many expeditions, travelsand trade took off in one or the other. France’s influence on the eastern coast of Africa wasinitiated by treaties with the rulers of what is today Djibouti from 1883, and the creationof outposts in modern day Dakar. Soon the French ambition became to establish an East-
9
West link between its posts in Dakar and Djibouti. French and British expansions clashednear the town of Fashoda, in the Fashoda incident of 1898 (Bates 1984). In schoolbooks,this episode in history is sometimes illustrated with maps that show the cities of Cairo,Cape Town, Dakar and Djibouti, with four arrows meeting near Fashoda. This simplifiedschoolbook view of colonial expansion in Africa is the model we follow in the constructionof our instruments, taking the distance to these colonial origin ports as instrument for theage of a town, and consequently its initial road grid.
One concern could be that these ports still play a large role as transportation hubs. Thendistance to them might influence population growth by channels other than the historic onesuggested. To address this concern, we include the distance to the sea as an additionalcontrol variable in these instrumental variable specifications. This variable would also pickup other influences the coast may have on the local economy. We also note that the busiestports of Africa are (in order descending in size according to total cargo volume): Richard’sBay, Saldanha Bay, Alexandria, Damietta, East Port, Apapa, Casablanca, Skikda, PointeNoire, Mombasa, Abidjan, Tanger, Bejaia, Jorf Lasfar and Tin Can Island (AAPA 2015).This list suggests that there are large ports available along the African coast, apart fromthese four cities. In addition to seaports, airports now play a substantial road in connectingAfrica, which further reduces the importance of these four historic ports. In addition weprovide a robustness specification below in which we control for the market access of citiesin our dataset, and show that such measures barely change the magnitude and significanceof our coefficients.
Table 3 shows the correlation between our main right hand side variable, which is ameasure of road density, and the distance to the four instrument cities. This is effectivelya first stage regression. As expected, we observe a significant negative or no relationshipbetween the distance to the ports and road density in the city centre. It is negative in thecase of three of our four cities and insignificantly different from zero at five percent level ofstatistical significance in the case of Cape Town.9 The last four columns include a set ofcontrol variables measuring ruggedness, elevation, minimum, maximum and average annualtemperature, precipitation, distance to the big lakes, the coast and malaria probability.When we joinlty include all four cities in one regression and add various control variablesas in Column (5) we continue to observe negative coefficients, typically at a high level ofstatistical significance. Coefficients become weaker when we include country fixed effects, asin Column (6). This seems plausible to us, since we expect the distance to these colonial portsto matters less within countries than for the longer distances across the whole continent.Because of this weak instrument problem, we show the main results later with and withoutcountry fixed effects. In our IV strategy we prefer to use the four distances as separatevariables to simply using the minimum distance given that British and French coloniallegacies could have influenced cities in different ways (Baruah et al. 2017b). In Columns (7)and (8) we repeat the estimation from Columns (5) and (6) but using road density for thecity as a whole on the left hand side, rather than density in just the city centre. We find astrong relationship without country fixed effects and a much weaker one when we includecountry fixed effects. We find it plausible that our first stage strategy is stronger for thecity centre, simply because centres are older and thus shaped to a greater degree by history.
In Columns (9) and (10) we provide more direct evidence that distance to these historicports is indeed correlated with the age of a city. For this analysis we use the data from the“Africapolis” project, that includes information about population in as early as 1950. Wekeep all cities in our dataset that are also in the Africapolis, which reduces the sample to1,070 observations. We then compute an indicator for cities that, according to Africapolis,did not have any population in 1950. As Columns (9) and (10) show, there is a significant
9On average, the coefficient in columns (1)-(4) is -0.87, perhaps close to the distance coefficient of -1sometimes estimated for distance coefficients (Rauch 2016).
10
relationship between these distances and the age of a city. Once we include our controlvariables and country fixed effects, the relationship is positive, which confirms that citiesthat are further away from these ports are more likely to be new. Columns (11) and (12)are then first stage regressions when using the age indicator for cities as an instrument. Asthe columns show, a simple dummy variable for cities that are older than 1950 correlatesstrongly significantly with the density of roads in the city centre, and also the city as awhole.
Another concern for instrumental variable strategy would be if other factors that in-fluence population growth also correlate with the distance to these four cities, apart fromtheir influence on the road layout of cities. This would include all colonial legacy that de-pends on age. For example, there may be physical infrastructure other than roads whoseage depends on these distances in a similar way, and that have a similar effect on futuredevelopment. We can’t dismiss this concern entirely, yet we want to note a few things inresponse to it. First, we believe that the road grid is particularly prone to lock-in, comparedto other types of infrastructure. Additionally, road construction and city planning were ofcentral importance of all potential colonial investments. Second, building reconstructionrates in Africa are very high. Baruah et al. (2017a) estimate an annual replacement rateof five percent for housing in Tanzania, Henderson et al. (2017) find a replacement rateof 3.6 percent in Kenya. Given these high replacement rates, most houses and most otherphysical infrastructure would have been rebuilt several times since the early 20th century,and have much time to converge from an initial steady state to a new one. Path dependentbehaviour of infrastructure quality is far less likely in such a rapidly changing environment.Of course also the physical characteristics of the roads we study were changed and updatedmultiple times in the great majority of our towns and cities. It is the however the layoutof roads, the plan, that we think is harder to change, and stickier than the quality of roadsor other aspects of a city or town. Third, most public infrastructure in Sub-Saharan Africais concentrated in capital cities (Bekker and Therborn 2012), which we exclude from theanalysis. Public policy programs are observed less outside, and so there are fewer channelsthrough which administrative legacies could manifest.
Apart from physical infrastructure, there may be institutions that also depend on thesedistances in a similar way to roads. If these institutions vary at the country level, they wouldbe absorbed by the country fixed effects that we employ in some specifications. Alternatively,these institutions may vary primarily on the level of capital cities. This is part of the reasonswhy we drop capital cities in the main specifications of this analysis.
5 Results
This section presents first OLS results, followed by results from an instrumental variablespecification. The estimation equation we use for both sets of results is of the form
∆pop = f(Road density, Evenness) + µc + ln(popt=1) +X + ε. (1)
We compute a measure of annualised population growth defined as ∆pop = (ln(popt=T )−ln(popt=1))/(T − 1). In the main regression the two years for which we observe populationlevels are 2015 and 2000, and so T − 1 is equal to 14 here. This is a period of high overallpopulation growth especially in towns and cities of Africa, and so we think it is a usefulperiod to study variation on cities that facilitate growth. For the function f we experimentwith various transformations of our main measures of the road network in a city centre. Wesometimes use country fixed effects, denoted as µc above. The matrix X denotes occasionaladditional control variables, measuring ruggedness, elevation, minimum temperature, maxi-mum temperature, average temperature, precipitation, distance to the big lakes, distance to
11
the coast and climatic conditions for malaria.. In our main specifications we control for theinitial population density of a city, but we also show results including a rich set of controlvariables. We use robust standard errors throughout.
We begin by presenting OLS results. In Table 4 we report OLS results on the relationshipbetween our simple variables of road network and population growth. Column 1 showsthat a greater road density in the city centre correlates positively and strongly statisticallysignificantly with population growth in the period 2000-2015. The measure of road densityin the city centre “d centre” is measured in km/km2, and has a mean of around 6. Increasingthe road density in the city centre by one km per km2, is associated with more than onetenth of a percent of higher population growth annually. In Column (2) we show that asimilar relationship holds at the level of city also, rather than just for the city centre. Thefirst two columns do not include country fixed effects to be comparable to the IV Columnslater that partly also do not include country fixed effects. In Columns (3) and (4) we showthat coefficients remain fairly similar when we include country fixed effects, both in termsof magnitude and statistical significance. In Columns (5) and (6) we include province fixedeffects, which are the next geographic administrative unit below state. These coefficients canaddress concerns over open street map bias. We note that coefficients remain very similar interms of magnitude than in the specifications in the previous columns, which is reassuring.In Columns (7) and (8) we add our standard set of control variables, and note that they donot influence magnitudes or significance levels of coefficients by much. Finally, in Columns(9) and (10) we repeat the regression for large cities only, defined as cities with an areagreater than 10km. Again, we find coefficients that are very similar in terms of magnitudeand significance. This addresses the concern that many of the units included in our datasetmay be too small for road grids to play any significant role. This also addresses concernsover bias in the OSM data, which we expect to be less of a concern for large cities. In thiscontext, it is worth noting that the coefficient does not change meaningfully in magnitudewhen we reduce the data to this sample.
Table 5 is what we consider the main table of this paper. Here we present resultscorresponding to the results presented in Table 4, but this time estimated as 2SLS. We findsignificant and positive coefficients in all six specifications. Magnitudes are larger than inthe OLS estimate, which does not surprise us given an important reverse causality problemin the OLS regression, and some scope for measurement error. At the level of city, inColumn (2) the instrument is weak, as already observed in the First Stage Table. Thisreflects that our empirical approach makes more sense for city centres rather than cities asa whole, as city centres tend to be older and so reflect history to a larger degree. We alsofind weak instruments in the case in which we use country fixed effects in Columns (5) and(6). This problem does not appear when we omit country fixed effects, and control variablesstrengthen the first stage, as in Column (4). It is reassuring that all three specifications,without controls, with controls and with controls and country fixed effects, despite theirvarious shortcomings, give roughly similar magnitudes to the main coefficient of interest,both for city centres and whole cities. Our main specifications for both, city centre and thecity as a whole are Columns (3) and (4), which include the full set of our control variables,and have strong first stages. These magnitudes are our preferred estimates. For city centresthey are about three times the effect found using OLS.
We also use indicator for new cities as a second, alternative instrument. This variableindicates cities that were founded after 1950 according to the Africapolis dataset (cf. section4 for more details). Given that not all our towns and cities feature in the Africapolis dataset,this reduces the size of our sample to 1,070. This instrument follows the same logic as theport distances, which is that older cities are more likely to be stuck with sub-optimal grids.But it measures the age of a city more directly. The advantage of this may be that we get astronger instrument (which indeed we do). This alternative instrument also avoids potential
12
violations of the exclusion restrictions that may arise if colonial expansion left persistentmarks on towns and cities other than their foundation year. Table 6 shows that the firststages give F statistics that are well above conventionally used critical values. Coefficientsare typically less precise than with the other instrument, and show coefficients that aretypically larger. All coefficients are positive, which confirms the relationship of road densityand population growth also using this instrument.
Next we repeat the analysis using population levels on the left hand side rather thangrowth rates. We use population levels from 2015. This is equivalent to a long run popula-tion growth rate on the left hand side, starting at population levels close to zero. Table 7shows our preferred specifications only, without country fixed effects. The first two columnsare estimated in OLS. The association between road density and population levels in 2015is positive and strongly statistically significant, both for the city centre as well as the cityas a whole.10 This is hardly surprising, as larger cities will have more roads. We repeat theexercise using our instrument in Columns (3) and (4). To the degree that the instrument isvalid, it should solve any such reverse causality concern. Indeed, we find that the 2SLS spec-ification gives strongly positive and significant coefficients. We again have weak instrumentsfor cities in Column (4), strengthened by the inclusion of a set of control variables in Column(6). In Columns (7) and (8) we add annualised population growth rates for longer periodon the left hand side, the difference between the years 2015 and 1975, which sits betweenour preferred estimates, and the levels specification (the latter could be seen as a growthrate starting from zero). We find that coefficients remain positive and significant, with mag-nitudes that are lower than for the years 2000-2015. This likely reflects that urbanisationin Africa picked up after 2000, and these overall faster urban population growth manifestsalso in larger coefficients. Levels are arguably more directly affected by history than growthrates between the years 2000 and 2015, and so the justification for the instrument may beweaker here. Thus the growth rate specifications remain our preferred estimates.
We now turn to alternative measures derived from the road grid. Here we computesimple statistics on the grid that we find in the literature, or find intuitive. Our OSM roadmap allows us to distinguish between motorways primary roads, secondary roads and small,residential roads. The definition of these differences is not always specified in the roadmap, and sometimes ambiguity over road type classification remains. This suggests thatthis decomposition is noisy. We nevertheless consider it because we think that road typedifferences might reveal interesting differences. These additional specifications are in Table8. In Column (1) we replace the road density measure by these four separate measures forthe natural logarithm of the length of motorway, primary, secondary and small roads. Herewe do not include the road density measure itself, which is an aggregate of the decomposedvariables. We continue to control for initial population of cities and towns throughout thistable. All types of road have positive coefficients. Small roads seem more important thanthe other types, both in terms of magnitude and statistical significance. Primary roadscome next, with motorways and secondary roads playing a less important role. Statisticalsignificance of these coefficients at a five percent level disappears when we run these samespecifications in 2SlS (these results are not reported but available upon request), and so wedo not want to emphasise these differences too much.
We next turn to Column (2), where we test a measure of regularity. The concern hereis that two cities with identical road density as measured by our d centre measure maynevertheless have quite different flows in the centre. Consider Figure 1. Two stylised citieswith the same road density have quite different regularity of the grid. In the city on theright, the average distance from a random point in the circle to a road is shorter. To measurethese differences we plot 12 regular points throughout each circle, arranged as in Figure 5.
10When we replace the left hand side with population in 2000 we obtain similar results, both in magnitudesand statistical significance.
13
We measure the distance to the nearest road for each of these points. The average distancewe obtain this way is what we call ‘regularity’. As Column (2) shows, such a measure ofregularity indeed correlates significantly with population growth when controlling for roaddensity. The negative sign indicates that larger average distances to the nearest road harmpopulation growth, holding road density constant, as expected. Statistical significance ofthis result is however not found at 1 percent level.
The next measure we consider in Column (3) is a direction Herfindahl, which is a measureof the directional alignment of roads. Figure 6 explains this measure. The left two panelsshow two street grids in two city centres. We decompose the grid into nodes and edges.A node is the intersection of roads, or the dead end of a road. An edge is the connectionbetween nodes. We measure the direction of each edge, and plot the distribution by directionin the middle two panels, weighting road segments by their length. We also show thedistribution in the third panel. The Herfindahl index we use is then a Herfindahl indexcomputed over the bins of directions. It is large if many roads run in parallel, and small ifthe grid is more chaotic, characterised by an even distribution of directions. When we usethis measure in Column (3), we find that it is not statistically significantly correlated withpopulation growth. It is also not significant when we control for road density, as in Column(4). In Column (5) we provide separate counts of nodes and edges of roads. These countsappear not to be statistically significant. In Columns (6) we include all these measuressimultaneously, first at the level of city centre. When we use all these measures together,it seems that the regularity measure is the only one that remains statistically significantlydifferent from zero at five percent level of statistical significance at the centre. This is partof the justification why our conclusion stresses the importance of road density, and does notstress the importance of these other measures to the same degree. In Column (7) we repeatthe estimation of Column (6) but estimate all variables considered at the level of the wholecity, and not just the city centre. Magnitudes and statistical significance look fairly similaras for the centre city measures. The main conclusion from this table is that the simpledensity measure stands out amongst all the measures we use. This motivates the focus ofthis paper on it.
We explore the robustness of our results by repeating estimates of Columns (5) and (6)of Table 5 with modifications. These robustness results are displayed in Table 9. We focuson the main coefficient of interest, the coefficient corresponding to Column (5) is in theupper panel, and the coefficient corresponding to Column (6) in the lower panel. Here, inColumn (1) we take nightlights growth (from 2000 to 2012) as alternative outcome variable,a variable that is sometimes seen as a reliable measure of economic activity even at city level(Bluhm and Krause 2018). Results for the city centre remain fairly similar both qualitativelyand quantitatively. The city wide result is not statistically significant. In Column (2) wemeasure population growth using the GPW data, version 4. Results both for the city centreand for the city as a whole are positive and statistically significant. In Column (3) we use theAfricapolis data to measure population growth. While coefficients are positive, they are notstatistically significant in this sample. In Column (4) we return to the GHS data, but thistime include Nigeria. This does not seem to change our results by much. Finally, in Column(5) we include capital cities. The main effect is that now the city wide results are stronglystatistically significant, while they were marginal when we use our preferred dataset. InColumn (6) we add controls for the market access of cities and town. We measure marketaccess by distance weighted populations of all other cities in our sample as in Maurer andRauch (2019). The weights we use are -1, -2 and -4. We include all these three market accessmeasures in logs to the specification. As the table shows, this barely changes the magnitudeor significance of coefficients.
14
6 Conclusion
Providing a useful layout of the road grid in cities is an important policy concern for man-aging ongoing urbanisation in Africa. In this paper we show that the road grid, and inparticular the density of roads in the centre of cities and towns of Sub-Saharan Africa, issystematically related to population growth in recent years. As an instrument we use acorrelate of the age of a city, which we argue influences the starting grid, which then getsstuck in many cases. The main conclusion from this paper is that a lack of road density incity centres seems to slow urban population growth in cities and towns in this region.
Given the magnitudes of the problem, it is likely that local policy makers are aware ofthe problem, yet unable to retrofit better road layouts. Correcting infrastructure ex-postseems be prohibitively costly. Thus policy makers in the continent that are in the process ofdeveloping new cities and neighbourhoods should anticipate population growth trajectoriesand potential from the start, and plan and protect the road grid accordingly.
15
References
• AAPA - American Association of Port Authorities, “World Port Rankings 2015”,downloaded from http://www.aapa-ports.org on March 22nd, 2017.
• Africapolis, Aggomeration data, downloaded from http://africapolis.org/data in March,2019.
• Antos, Sarah Elizabeth, Somik V. Lall, and Nancy Lozano-Gracia. “The morphologyof African cities.” (2016).
• Ahlfeldt, Gabriel M., et al. “The economics of density: Evidence from the BerlinWall.” Econometrica 83.6 (2015): 2127-2189.
• Baruah, Neeraj, Guy Michaels, Dzhamilya Nigmatulina, Ferdinand Rauch and TannerRegan, ‘Upgrade slums or invest before they form? Evidence from Tanzania’, Mimeo,2017a.
• Baruah, Neeraj G., J. Vernon Henderson, and Cong Peng. “Colonial legacies: shapingAfrican cities.” (2017b).
• Bates, Darrell. The Fashoda incident of 1898: encounter on the Nile. Oxford Univ Pr,1984.
• Baum-Snow, Nathaniel. “Did highways cause suburbanization?.” The Quarterly Jour-nal of Economics 122.2 (2007): 775-805.
• Baum-Snow, N., Brandt, L., Henderson, J. V., Turner, M. A., & Zhang, Q. (2017).Roads, railroads, and decentralization of Chinese cities. Review of Economics andStatistics, 99(3), 435-448.2
• Bekker, Simon B., and Goran Therborn. Capital cities in Africa: power and power-lessness. Hsrc Press, 2012.
• Bluhm, Richard and Melanie Krause, ‘Top Lights - Bright cities and their contributionto economic development’, Draft paper, 2018.
• Boahen, Adu, “Africa and the colonial challenge”, in Boahen (edt.) ‘General Historyof Africa VII’, United Nations Educational, Heinemann Books, London, 1985.
• Boeing, Geoff. “OSMnx: New Methods for Acquiring, Constructing, Analyzing, andVisualizing Complex Street Networks.” Computers Environment and Urban Systems,vol. 65, pp.126-139, 2017.
• Boeing, Geoff, “Urban Spatial Order: Street Network Orientation, Configuration, andEntropy”. 2018.
• Cervero, Robert, (2013) ‘Linking Urban Transport and Land Use in Developing coun-tries’, Journal of Transport and Land Use, 6(1), 7-24.
• Collier, Paul, and Anthony J. Venables. “Urban infrastructure for development.”Oxford Review of Economic Policy 32.3 (2016): 391-409.
• Couture, Victor, Gilles Duranton, and Matthew A. Turner. “Speed.” Review ofEconomics and Statistics 100.4 (2018): 725-739.
• Donaldson, Dave, and Richard Hornbeck. ”Railroads and American economic growth:A “market access” approach.” The Quarterly Journal of Economics (2016).
16
• Duranton G, Puga D. 2004. Micro-Foundations of Urban Agglomeration Economies.In Handbook of Regional and Urban Economics, ed. J. V. Henderson and J.-F. Thisse,Amsterdam: Elsevier. Vol. 4: 2063-2117.
• Duranton, Gilles, and Matthew A. Turner. “The fundamental law of road congestion:Evidence from US cities.” American Economic Review 101.6 (2011): 2616-52.
• Duranton, Gilles, and Matthew A. Turner. “Urban growth and transportation.” Re-view of Economic Studies 79.4 (2012): 1407-1440.
• Faber, B. (2014). Trade integration, market size, and industrialization: evidence fromChina’s National Trunk Highway System. Review of Economic Studies, 81(3), 1046-1070.
• Fetzer T., Henderson V., Nigmatulina D. and Shanghavi A. “What Happens to CitiesWhen Countries Become Democratic?” (2016).
• Fuller, Brandon, and Paul Romer. “Urbanization as opportunity”. The World Bank,2014.
• Grover Goswami, Arti G., and Lall, Somik V., (2016), ‘Jobs in the City: ExplainingUrban Spatial Structure in Kampala’, World Bank Policy Research Working Paper7655, World Bank, Washington D.C.
• Hanna, Rema, Gabriel Kreindler, and Benjamin A. Olken. ”Citywide effects of high-occupancy vehicle restrictions: Evidence from “three-in-one” in Jakarta.” Science357.6346 (2017): 89-93.
• Harari, Mariaflavia. ”Cities in bad shape: Urban geometry in india.” Job marketpaper, Massachusetts Institute of Technology (2015).
• J. V. Henderson, T. Regan, A. Venables, “Building the city: Sunk capital and se-quencing” (Discussion paper 11211, Centre for Economic Policy Research, London,2016).
• Hourani, Albert Habib; Khoury, Philip Shukry; Wilson, Mary Christina (2004). TheModern Middle East: A Reader (2nd ed.). London: I.B. Tauris. ISBN 1-86064-963-7.
• Jedwab, R., & Moradi, A. (2016). The permanent effects of transportation revolutionsin poor countries: evidence from Africa. Review of economics and statistics, 98(2),268-284.
• Jedwab, R., and A. Storeygard. “The Heterogeneous Effects of Transportation In-vestments: Evidence from sub-Saharan Africa 1960-2010.” Unpublished manuscript,Tufts Univ., Medford, (2016).
• Kocornik-Mina, Adriana, Thomas McDermott, Guy Michaels and Ferdinand Rauch(2019), ‘Flooded Cities’, American Economic Journal: Applied Economics, forthcom-ing.
• Kreindler, Gabriel. ”Driving Delhi? Behavioural Responses to Driving Restrictions.”(2016).
• Lall, Somik V. and Henderson, J. Vernon and Venables, Anthony J, (2017), Africa’sCities: Opening Doors to the World, World Bank, Washington DC.
17
• Larcom, Shaun, Ferdinand Rauch, and Tim Willems. “The benefits of forced experi-mentation: striking evidence from the London underground network.” The QuarterlyJournal of Economics 132.4 (2017): 2019-2055.
• Marshall A. 1920. Principles of Economics. London: Macmillan.
• Maurer, Stephan, and Ferdinand Rauch. “Economic Geography Aspects of the PanamaCanal.” (2019), Oxford Economics Working Paper.
• Michaels, G. (2008). The effect of trade on the demand for skill: Evidence from theinterstate highway system. The Review of Economics and Statistics, 90(4), 683-701.
• Michaels, G., Nigmatulina, D., Rauch, F., Regan, T., Baruah, N., and Dahlstrand-Rudin, A. (2017). Planning Ahead for Better Neighborhoods: Long Run Evidencefrom Tanzania.
• Michaels, Guy, and Ferdinand Rauch. “Resetting the urban network: 117–2012.” TheEconomic Journal 128.608 (2018): 378-412.
• Rauch, Ferdinand. ”The geometry of the distance coefficient in gravity equations ininternational trade.” Review of International Economics 24.5 (2016): 1167-1177.
• Seidel, Andre, (2019) ”Regional Public Goods Provision, Decentralization and SocialHeterogeneity”, Draft paper.
• Shertzer, A., T. Twinam, and R. P. Walsh (2016, July). “Race, Ethnicity, and Dis-criminatory” Zoning. American Economic Journal: Applied Economics 8 (3), 217-246.
• Storeygard, Adam. “Farther on down the road: transport costs, trade and urbangrowth in sub-Saharan Africa.” The Review of economic studies 83.3 (2016): 1263-1295.
18
7 Tables and Figures
Cities Sample Share (%) Median Pop. Av. Pop. Av. Pop. Growth
Angola 33 1.78 54130 163514 4.78Benin 27 1.46 43352 58856 1.47Botswana 29 1.57 18382 32626 0.67Burkina Faso 27 1.46 44998 62813 1.94Cameroon 69 3.73 34104 93522 1.01Central African Republic 10 0.54 25628 24958 0.74Chad 12 0.65 101813 109497 4.10Congo, Dem Republic of 52 2.81 88960 177904 2.30Congo, Republic of 15 0.81 18752 65410 3.75Cote d’Ivoire 131 7.08 17728 68461 1.03Djibouti 5 0.27 2479 2581 3.86Equatorial Guinea 2 0.11 47880 47880 2.69Eritrea 5 0.27 4317 21045 5.66Ethiopia 113 6.11 50527 70516 3.41Gabon 25 1.35 10507 18436 1.93Gambia 5 0.27 22254 26840 1.75Ghana 88 4.76 41778 81420 0.77Guinea 23 1.24 40946 52615 2.54Kenya 80 4.32 29856 71457 2.92Lesotho 10 0.54 19638 38970 0.15Liberia 2 0.11 37678 37678 0.59Madagascar 33 1.78 15292 19509 2.88Malawi 57 3.08 13495 30477 2.79Mali 30 1.62 30466 46283 2.94Mauritania 15 0.81 17271 20955 2.50Mozambique 62 3.35 33707 68073 1.64Namibia 42 2.27 5246 13361 1.71Niger 34 1.84 27652 69991 3.21Nigeria 405 21.89 50889 150988 1.60Rwanda 9 0.49 63907 57128 2.40Senegal 54 2.92 29195 74327 2.20Sierra Leone 6 0.32 92277 80577 1.93South Sudan 3 0.16 101410 142367 1.64Sudan 52 2.81 91025 141978 3.63Swaziland 7 0.38 13838 30454 -1.04Tanzania, United Rep of 100 5.41 26104 56960 2.59Togo 22 1.19 51132 62932 0.50Uganda 38 2.05 33671 54929 2.97Zambia 56 3.03 45171 72605 2.12Zimbabwe 62 3.35 22623 50864 0.95
Total 1850 100.00 32972 87565 2.04
Table 1: In total, we consider 1850 cities from 40 countries. In our main specification we neverthe-less exclude Nigeria and Madagascar, reducing our sample to 1,412 cities. The first two columnsdetail each country total number of cities and contribution to the sample, respectively. The fol-lowing three column gives indication of poulation size in 2000 (median and average) as well as theaverage population growth rate in cities over the 2000-2015 period.
N Mean sd Min Median Max
Whole CityTotal Road Length (km) 1850 243.41 878.49 0 73 28621Meters of Road per Sq. Km 1850 2490.88 2176.68 0 1850 17569Area (Sq.Km.) 1850 94.64 278.85 2 40 7408
City Centre (1km. radius)Total Road Length (km) 1850 19.55 16.06 0 15 87Meters of Road per Sq. Km 1850 6402.65 5160.03 2 5205 27703Area (Sq.Km.) 1850 3.10 0.21 1 3 3Dist. Centre to closest Road (m) 1850 133.29 166.94 0 62 975Road Orientation Herfindahl 1850 0.19 0.19 0 0 1Average Dist. Points to Road (m) 1850 189.15 155.37 10 150 1078
Distances from City CentreDistance to Dakar (km) 1850 4065.62 2070.82 78 3487 8272Distance to Djibouti (km) 1850 3583.33 1450.88 34 3773 6507Distance to Cairo (km) 1850 4007.16 1046.31 918 3875 6738Distance to Cape Town (km) 1850 4337.90 1311.26 628 4838 7168Distance to the closest (km) 1850 2064.75 830.10 34 2135 3678Distance to the sea (km) 1850 451.14 350.46 0 389 1711
Table 2: Cities: Descriptive statistics
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
dce
ntr
ed
centr
ed
centr
ed
centr
ed
centr
ed
centr
ed
city
dci
tynew
1950
new
1950
dce
ntr
ed
city
lnd
ist
dak
-0.3
86**
-2.6
94***
-0.5
13
-1.2
66***
-0.5
69
0.1
095
0.6
75***
(0.1
64)
(0.4
95)
(1.0
97)
(0.2
02)
(0.3
92)
(0.0
998)
(0.2
39)
lnd
ist
dji
-0.5
95**
*0.0
579
3.3
67***
-0.0
752
0.6
09
0.1
28***
0.1
51
(0.2
00)
(0.4
22)
(1.2
22)
(0.1
63)
(0.4
73)
(0.0
45)
(0.1
34)
lnd
ist
cai
-2.8
76***
-9.6
33***
-8.6
81***
-2.2
06***
-1.6
47*
-0.1
03***
0.0
94
(0.5
45)
(1.1
41)
(2.6
60)
(0.4
56)
(0.9
35)
(0.0
33)
(0.0
93)
lnd
ist
cap
0.3
85
-5.5
72***
-1.3
80
-1.9
81***
-0.6
14
0.2
56
0.7
11**
(0.3
10)
(0.9
71)
(1.8
36)
(0.4
33)
(0.6
91)
(0.1
11)
(0.2
75)
lnd
ist
sea
0.01
91-0
.061
7-0
.145
**
-0.0
232
-0.1
49
-0.1
41
0.1
67***
0.1
04*
0.2
75
0.0
150
0.1
27
0.1
80***
(0.0
696)
(0.0
701)
(0.0
733)
(0.0
694)
(0.1
29)
(0.1
44)
(0.0
621)
(0.0
629)
(0.0
09)
(0.0
188)
(0.0
97)
(0.0
37)
new
-2.3
74***
-0.5
68***
(0.3
55)
(0.1
39)
lnp
op20
001.
795*
**1.
782*
**1.
702*
**
1.7
80***
1.8
50***
1.8
16***
0.6
20***
0.5
41***
-0.1
28***
-0.1
39***
1.3
16***
0.4
20***
(0.0
983)
(0.0
980)
(0.0
966)
(0.0
992)
(0.1
01)
(0.1
05)
(0.0
475)
(0.0
430)
(0.0
116)
(0.0
121)
(0.1
28)
(0.0
51)
N1,
412
1,41
21,
412
1,4
12
1,4
12
1,4
12
1,4
12
1,4
12
1,0
70
1,0
70
1,0
70
1,0
70
Con
trol
sY
esY
esY
esY
esY
esY
esY
esY
esC
ountr
yf.
e.Y
esY
esY
es
Tab
le3:
Fir
stst
age.
Lndist
refe
rsto
log
dis
tan
ceto
eith
erC
air
o,
Cap
eT
own
,D
aka
r,D
jib
ou
tior
the
sea.dcentre
mea
sure
sro
ad
den
sity
inth
eci
tyce
ntr
e,w
hil
edcity
mea
sure
sro
add
ensi
tyfo
rth
ew
hole
city
.T
he
set
ofcontrols
con
sist
sof
vari
ab
les
mea
suri
ng
rugged
nes
s,el
evat
ion
,m
inim
um
tem
per
atu
re,
max
imu
mte
mp
eratu
re,
aver
age
tem
per
atu
re,
pre
cip
itat
ion
,d
ista
nce
toth
eb
igla
kes
,d
ista
nce
toth
eco
ast
and
clim
atic
con
dit
ion
sfo
rm
alar
ia.
Var
iab
lenew
1950
isan
ind
icato
rfo
rci
ties
fou
nd
edaft
er1950.
Sta
rsd
enote
sign
ifica
nce
at
10
(*),
5(*
*)
and
1(*
**)
per
cent.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
∆p
op∆
pop
∆p
op
∆p
op
∆p
op
∆p
op
∆p
op
∆p
op
∆p
op
∆p
op
dce
ntr
e0.
0013
***
0.00
11***
0.0
013***
0.0
011***
0.0
011***
(0.0
0029
)(0
.00027)
(0.0
003))
(0.0
0026)
(0.0
0028)
dci
ty0.
0020
***
0.0
018***
0.0
021
***
0.0
020***
0.0
020***
(0.0
0037
)(0
.00044)
(0.0
066)
(0.0
0045)
(0.0
0049)
lnp
op20
00-0
.007
1***
-0.0
059*
**-0
.0083***
-0.0
072***
-0.0
078***
-0.0
061***
-0.0
099***
-0.0
089***
-0.0
084***
-0.0
075***
(0.0
016)
(0.0
013)
(0.0
018)
(0.0
016)
(0.0
018)
(0.0
014)
(0.0
018)
(0.0
016)
(0.0
019)
(0.0
017)
N1,
412
1,41
21,
412
1,4
12
1,4
12
1,4
12
1,4
12
1,4
12
1,3
39
1,3
39
Cou
ntr
yf.
e.Y
esY
esY
esY
esY
esY
esP
rovin
cef.
e.Y
esY
esC
ontr
ols
Yes
Yes
Cit
yY
esY
es
Tab
le4:
OL
S.dcentre
mea
sure
sro
add
ensi
tyin
the
city
centr
e,w
hil
edcity
mea
sure
sro
ad
den
sity
for
the
wh
ole
city
.T
he
set
ofcontrols
con
sist
sof
vari
able
sm
easu
rin
gru
gged
nes
s,el
evati
on,
min
imu
mte
mp
eratu
re,
maxim
um
tem
per
atu
re,
aver
age
tem
per
atu
re,
pre
cip
itati
on
,d
ista
nce
toth
eb
igla
kes
,d
ista
nce
toth
eco
ast
an
dcl
imati
cco
nd
itio
ns
for
mala
ria.
The
setcity
incl
ud
esonly
ob
serv
ati
on
sw
ith
an
are
ala
rger
than
10km
2.S
tars
den
ote
sign
ifica
nce
at10
(*),
5(*
*)
an
d1
(***)
per
cent.
(1) (2) (3) (4) (5) (6)∆pop ∆pop ∆pop ∆pop ∆pop ∆pop
d centre 0.0023** 0.0036*** 0.0036*(0.00089) (0.0012) (0.0021)
d city 0.0272*** 0.010*** 0.021*(0.00943) (0.0028) (0.012)
ln dist sea 0.0011** -0.0018 0.0011 -0.000864 0.0000 -0.0020(0.00049) (0.0013) (0.0013) (0.00124) (0.00069) (0.0016)
ln pop 2000 -0.0089*** -0.0195*** -0.0130*** -0.0125*** -0.0127*** -0.0161**(0.0025) (0.0059) (0.0031) (0.00259) (0.00455) (0.0063)
First stage F 46.76 4.27 24.55 12.88 4.34 1.81N 1,412 1,412 1,412 1,412 1,412 1,412Controls Yes YesCountry f. e. Yes Yes
Table 5: 2SLS using distances as instrument. d centre measures road density in the citycentre, while d city measures road density for the whole city. The Multivariate F test statis-tic reports the Sanderson-Windmeijer multivariate F test of the excluded instruments. Theset of controls consists of variables measuring ruggedness, elevation, minimum temperature,maximum temperature, average temperature, precipitation, distance to the big lakes, dis-tance to the coast and climatic conditions for malaria. Stars denote significance at 10 (*),5 (**) and 1 (***) percent.
(1) (2) (3) (4) (5) (6)∆pop ∆pop ∆pop ∆pop ∆pop ∆pop
d centre 0.0056** 0.0095** 0.0064*(0.0023) (0.0043) (0.0017)
d city 0.403 0.043 0.226***(3.074) (0.033) (0.0083)
ln dist sea 0.0015** -0.059 0.0021 -0.0095 0.0002 -0.0038*(0.0007) (0.462) (0.0018) (0.0089) (0.001) (0.0019)
ln pop 2000 -0.016*** -0.217 -0.0252*** -0.0371* -0.018*** -0.0208***(0.0048) (1.598) (0.0086) (0.0218) (0.004) (0.0052)
First stage F 64.9 31.6 24.1 14.05 33.1 24.7N 1,070 1,070 1,070 1,070 1,070 1,070Controls Yes YesCountry f. e. Yes Yes
Table 6: 2SLS using age as instrument. d centre measures road density in the city centre,while d city measures road density for the whole city. The F test statistic reports the Ftest of the excluded instruments, which is an indicator for cities that were founded after1950. The set of controls consists of variables measuring ruggedness, elevation, minimumtemperature, maximum temperature, average temperature, precipitation, distance to the biglakes, distance to the coast and climatic conditions for malaria. Stars denote significance at10 (*), 5 (**) and 1 (***) percent.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
OL
SO
LS
2S
LS
2S
LS
2S
LS
2S
LS
2S
LS
2S
LS
lnp
op20
15ln
pop
2015
lnp
op
2015
lnp
op
2015
lnp
op
2015
lnp
op
2015
∆pop
7515
∆pop
7515
dce
ntr
e0.
127*
**0.0
794***
0.1
45***
0.0
016**
(0.0
0669
)(0
.0199)
(0.0
219)
(0.0
01)
dci
ty0.
207*
**1.0
47***
0.5
15***
0.0
155***
(0.0
172)
(0.2
46)
(0.0
809)
(0.0
059)
lnd
ist
sea
0.0
640***
-0.0
734
0.0
535
-0.0
484
0.0
013**
-0.0
01
(0.0
215)
(0.0
451)
(0.0
336)
(0.0
412)
(0.0
01)
(0.0
01)
N1,
412
1,41
21,4
12
1,4
12
1,4
12
1,4
12
1,3
39
1,3
39
Fir
stst
age
F36.5
64.8
528.0
221.3
735.6
5.5
Con
trol
sY
esY
es
Tab
le7:
OL
San
dIV
inle
vel
sra
ther
than
grow
thra
tes,
an
dfo
rth
elo
nger
tim
ed
iffer
ence
1975-2
015
(Colu
mn
s7
an
d8).
dcentre
mea
sure
sro
add
ensi
tyin
the
city
centr
e,w
hil
edcity
mea
sure
sro
ad
den
sity
for
the
wh
ole
city
.T
he
set
ofcontrols
con
sist
sof
vari
ab
les
mea
suri
ng
rugg
edn
ess,
elev
atio
n,
min
imu
mte
mp
erat
ure
,m
axim
um
tem
per
atu
re,
aver
age
tem
per
atu
re,
pre
cip
itati
on
,d
ista
nce
toth
eb
igla
kes,
dis
tan
ceto
the
coas
tan
dcl
imat
icco
nd
itio
ns
for
mal
aria
.S
tars
den
ote
sign
ifica
nce
at
10
(*),
5(*
*)
an
d1
(***)
per
cent.
(1)
(2)
(3)
(4)
(5)
(6)
(7)
∆p
op∆
pop
∆p
op
∆p
op
∆p
op
∆p
op
∆p
op
lnm
otor
way
0.00
046
0.0
0033
0.0
0029
(0.0
0031
)(0
.00030)
(0.0
003)
lnp
rim
ary
road
0.00
07**
*0.0
0050**
0.0
0036
(0.0
0025
)(0
.00024)
(0.0
0026)
lnse
con
dar
yro
ad0.
0004
3*0.0
0016
0.0
0022
(0.0
0024
)(0
.00024)
(0.0
0025)
lnsm
all
road
0.00
13**
*0.0
0062
0.0
004
(0.0
0042
)(0
.00039)
(0.0
0038)
regu
lari
ty-0
.000
0**
-0.0
000**
-0.0
000**
(0.0
000)
(0.0
000)
(0.0
000)
dir
ecti
onH
erf
-0.0
074
0.0
0078
-0.0
043
0.0
065
(0.0
049)
(0.0
049)
(0.0
092)
(0.0
054)
nod
eco
unt
0.0
000
0.0
000
0.0
0011**
(0.0
000)
(0.0
000)
(0.0
000)
edge
cou
nt
-0.0
000
0.0
000
-0.0
000*
(0.0
000)
(0.0
000)
(0.0
000)
dce
ntr
e0.
0011
0***
0.0
0134***
0.0
0108**
*(0
.000
297)
(0.0
00297)
(0.0
00288
)ln
pop
2000
-0.0
0622
***
-0.0
0731
***
-0.0
0492***
-0.0
0711***
-0.0
0855***
-0.0
0853***
-0.0
0727***
(0.0
0145
)(0
.001
63)
(0.0
0129)
(0.0
0160)
(0.0
0173)
(0.0
0169)
(0.0
0162)
N1,
412
1,41
21,4
12
1,4
12
1,4
12
1,4
12
1,4
12
R2
0.05
30.
068
0.0
65
0.0
65
0.0
81
0.0
80
0.0
69
Dat
aset
Cen
tre
Cen
tre
Cen
tre
Cen
tre
Cen
tre
Cit
yC
entr
e
Tab
le8:
Alt
ern
ativ
em
easu
res
ofth
ero
adgr
id.
Th
efi
rst
set
of
vari
ab
les
den
ote
sth
elo
gle
ngth
of
moto
rway
s,p
rim
ary
,se
cond
ary
an
dsm
all
road
sre
spec
tive
ly.regularity
give
sth
ere
gula
rity
mea
sure
.T
he
dir
ecti
on
Her
fin
dah
lis
am
easu
reof
the
deg
ree
tow
hic
hro
ads
run
inp
ara
llel
.T
he
nex
tse
tof
vari
able
sp
rovid
esd
ensi
tym
easu
res
of
the
nod
esan
ded
ges
of
road
sth
egr
id.dcentre
mea
sure
sro
ad
den
sity
inth
eci
tyce
ntr
e(a
nd
ofth
ew
hol
eci
tyin
Col
um
n(6
)).
Sta
rsd
enote
sign
ifica
nce
at
10
(*),
5(*
*)
an
d1
(***)
per
cent.
(1)
(2)
(3)
(4)
(5)
(6)
Lig
hts
GW
Pv4
Afr
icap
oli
sIn
clu
de
Nig
eria
Incl
ud
eC
ap
itals
Mark
etA
cces
sd
centr
e0.
0036
8***
0.00
068*
*0.
000224
0.0
0346***
0.0
033***
0.0
039***
(0.0
0097
)(0
.000
33)
(0.0
0058)
(0.0
0121)
(0.0
011)
(0.0
012)
Ob
serv
atio
ns
1,29
01,
412
926
1,8
17
1,4
12
1,4
12
dci
ty0.
0052
0.00
511*
**0.
00299
0.0
0588**
0.0
099***
0.0
11***
(0.0
045)
(0.0
012)
(0.0
0288)
(0.0
0234)
(0.0
246)
(0.0
026)
Ob
serv
atio
ns
1,29
01,
412
926
1,8
17
1,4
52
1,4
12
Tab
le9:
Rob
ust
nes
sch
ecks.
Reg
ress
ion
sin
this
tab
lere
pea
tC
olu
mn
s(5
)an
d(6
)of
Tab
le5,
wh
ere
the
coeffi
cien
tco
rres
pon
din
gto
Colu
mn
(5)
isin
the
up
per
pan
el,
and
the
coeffi
cien
tco
rres
pon
din
gto
Colu
mn
(6)
inth
elo
wer
pan
el.
Her
e,in
Colu
mn
(1)
we
mea
sure
the
ou
tcom
eby
ligh
tsgr
owth
(fro
m20
00to
2012
),in
Col
um
n(2
)w
em
easu
rep
op
ula
tion
gro
wth
usi
ng
the
GP
Wd
ata
(2000
to2015).
InC
olu
mn
(3)
we
use
the
Afr
acp
olis
dat
ato
mea
sure
pop
ula
tion
gro
wth
,(2
000
to2015).
InC
olu
mn
(4)
we
retu
rnto
the
GH
Sd
ata
,b
ut
incl
ud
eN
iger
ia.
InC
olu
mn
(5)
we
incl
ud
eca
pit
alci
ties
.C
olu
mn
(6)
incl
ud
esco
ntr
ols
for
the
mark
etacc
ess
of
citi
esin
add
itio
nto
ou
ru
sual
contr
ol
vari
ab
les.
Figure 1: This figure presents two cities with stylised road layouts. Both cities have the same roaddensity in the centre by construction. The city on the right has a shorter distance from a randompoint to the nearest road, a property we call ‘evenness’.
Cities
±
Figure 2: We identify 2,779 cities from 40 countries.
CentreOSM RoadsCity Extent 1:100,000
±
Figure 3: City extent and centre are defined using nightlight luminosity (in late - 2008-2012 - andearly - 1994-1996 - period, respectively).
1:45,000
±
CentreOSM RoadsCity Extent
Figure 4: Our measure of road layout is based on OSM roads in a 1km-radius circle around thecity centre.
E E
E E E E
E E E E
E E
1:15,000
±
CentreE Regularity
OSM Roads
Figure 5: Evenness is computed from 12 within-centre points.
Figure 6: The edge bearings histograms (right) summarise the orientation of city centre edges(left). The central panel is a polar representation of the histogram, illustrating the constructionof our measure. The first city (top) is Umm Ruwaba (Sudan) and has a typically high Herfindahlindex (0.48). Below is Mohale’s Hoek city (Lesotho), which is representative of the lowest values inour sample (0.05)
Data Appendix
A Data Sources
A.1 Cities, location and boundaries
• Lights: NOAA’s DMSP-OLS sattelites record images of earth at night (every day, at20:00 to 21:30 local). Composite images are built combining daily series in order toremove noises, cloud effects and temporary lights. The data is a 30x30 arc-second(around 1sq.km.) grid with pixels taking a value between 0 (no lights) and 63 (topcoded value). Data is available from 1992 to 2013. We use the 2008-2012 period tobuild city boundaries and the 1994-1996 to find city centres.
A.2 Road Layout
• Road Layout: Open Street Map is a volunteered geographic information project (whichinformation comes from over 2 millions users) mapping, among others, cities’ streets.Our analysis is based on the shapefiles from GeoFabrik.
• Basic network analysis: we use G. Boeing’s OSMNX python package to perform simplenetwork analysis of road layouts (counting “nodes”, i.e. intersections, measuring streetnetwork orientation, etc.)
A.3 Population
• European Commission, Global Human Settlement:
GHS-Population Grid (LDS) (2016) is a 250m x 250m grid depicting the distri-bution of population for years 1975, 1990, 2000 and 2015. The data providedthe CIESIN, Gridded Population v4, was disaggregated based on the distributionand density of built-up mapped by the GHS-Built-up Grid (that detects built-up presence from Landsat image collections: GLS1975, GLS1990, GLS2000, andad-hoc Landsat 8 collection 2013/2014)
• Oak Ridge National Laboratory, LandScan:
LandScan High Resolution global Population Data Set provides another popula-tion distribution at a 30 x 30 degree scale grid. We accessed the dataset for 2012only.
A.4 Others
• Coasts and Lakes: Natural Earth is a public domain map dataset that we use to clipcoasts and lakes when necessary (scale: 1:10m).
• Borders: we define countries using the borders from the Global Administrative Areasdataset GADM (v2)
• Nightlights growth: we test the robustness of our results using nighttime luminosityas the dependant variable; when this is the case, we also rely on NOAA’s DMSP-OLSimagery.
B Construction of the Variables of interest
B.1 Defining cities
In order to define a consistent sample of cities for many countries in Sub-Saharan Africa,we rely entirely on nightlight activity. We follow a two-steps procedure in which we firstdefine city boundaries and then find the city centre. This procedure identifies 2,779 cities inSub-Saharan Africa (excluding South Africa). The remainder of this sub-section details theGIS procedure.
B.1.1 Cities boundaries
In a first step, we use nightlights picture to define city boundaries. Specifically:
1. Contiguous lit areas are aggregated to form union of pixels;
Step 1 is based on nightlights activities during the 2008-2012 period captured as bysatellite imagery from DMSP-OLS (cf. A.1). Any pixel being lit (i.e. value being greaterthan or equal to 1) at least twice over the period is kept at this stage. Unions so defined arethen being “clipped” to coasts (sea and lakes sometimes reflect lights) and borders so as toadd natural and cultural borders that divide them.
This procedure follows Fetzer et al. (2016). When working with countries from the Gulfof Guinea, we directly rely on their extended sample. This is because Fetzer et al. manuallycorrected for gas flares issues that are numerous in this area of the world. Note that, for thesame reason, we exclude Nigeria from our main specifications.
At this stage, we observe 6,382 union in Sub-Saharan Africa (excluding South Africa).
B.1.2 City centre
In order to determine the location of each city centre, and in the absence of complete datasetof city-centres’ coordinates, we rely on nightlights. Our measure of city centres follow threesteps:
1. First, we compute city-specific maximum value of light and keep only pixels of thisvalue;
2. Second, we aggregate contiguous pixels into unions and keep only the largest union(s);
3. Third, we take the geographic (within-)centroid of these unions.
This definition ensures that only one centre is defined in each city (referring to the mono-centric assumption we make); and that this centre falls within one of the brightest pixels ofthe city.
Because we are interested in historical centres, we rely on early lights information ratherthan most recent images. We average pixel value over the three years 1994, 1995 and 1996(earliest years for which we can build a 3-year-long panel relying on only one satellite).Another advantage of taking early observations is that top-coding of the data is less likelyto be an issue (few pixels actually reach the value of 63 in mid-90’s Africa).
From the 6,382 unions defined in the previous step, only 2,779 displayed sufficient night-lights over the 1994-1996 period. These 2,779 constitutes our sample of cities.
B.2 Road Layout
We construct road layout information by extracting data from Open Street Map (OSM).OSM (cf. A.2) is a volunteered geographic information project which information comesfrom over 2 millions users. Anyone can contribute to the construction of this database.Although not the most complete, OSM has the great advantage of being publicly available.
Technically, we used two alternative ways to gather road network information for eachcity. First, OSM dataset is digitised by GeoFabrik from which we take roads informationfor 40 African countries. In particular, at the time we impored these data, Burundi andSomalia OSM maps were not available. The OSMNN roads are often (although not always)classified between primary, secondary and residential roads. Data potential extent is thetotally of roads within each country; but actually is what people contributed to at the dateof our download (last: December 2016).
In a second step, we use OSMNX python package to extract basic network descriptivestatistics such as number of nodes (i.e. crossroads), network orientation, etc. This packagewas developed by G. Boeing (cf. A.2) for urban planning studies and is open-access. (Lastdownload: July 2018.)
The two data extraction sometimes conflict with each other (some road may be missing inone but not the other, etc.) This may come from the updates that are made to OSM datasetbetween our two extractions. In any case, we restrict our analysis to cities for which bothdata exist. Our results are robust to using one or the other dataset (e.g. the correlationof total road length in centres is 0.85) as well as to the restriction on cities having bothinformation.
For any city, we gather two types of information regarding the road network: centre roadlayout and whole city layout. We call city centre a circle of 1,000m around the centre point(as defined in B.1.2). Note that we “clip out” any part of the circle that would cross thecoastline or a border (so it can happen that a centre is not a full circle) or that would falloutside the city boundary (ensuring that the area of the city centre is always less than orequal to the city total area).
B.2.1 About Open Street Map Precision
Because of its volunteered aspect, quality of OSM has long been a concern for users. Thereexists a few studies comparing OSM data to other (authoritative) datasets and generallypointing to the high and increasing quality of OSM. Most of these studies nevertheless focuson developed countries. For a quick survey, cf. G. Boeing, 2017. The instrumental variablesapproach we use in the paper would correct measurement bias arising from errors in OSM.
Another important aspect of OSM, is that roads can sometime be classified by types,such as “motorways”, “primary roads”, “path”, etc. The definition of these differences isnevertheless not always specified in the road map. Figure 7 displays the distribution of citiesacross share of classified roads. 21.18% of the cities have classification for almost all (≥90%)of their roads, and more than half of our sample has classification for more than 70% oftheir roads (unweighted by their length).
B.3 Population
We extract population information from the European Commission GHS-Population Grid(GHSP, cf. A.3). GHS is a geo-coded 250m by 250m grid providing estimates of the numberof inhabitants, based dis-aggregation and interpolations from national censuses. The GHSpopulation uses information from the presence of built-up cover as derived by the GHS-Built-up Grid, in order to spatially allocate population data. The population estimates are
provided for four target years: 1975, 1990, 2000 and 2015. Technically, we extract the totalnumber of inhabitants whose coordinates fall within the boundaries of our cities sample.
C Sample Selection
C.1 Minimum size restriction
In order to ensure our results are not driven by the large number of very small places,we exclude from our analysis any observation that does not meet one of the two followingcriteria:
1. Minimum area: cities must be at least greater than 2sq.km.
2. Minimum population: cities must have at least 2,000 inhabitants in 2015, as observedin GHS population11.
This double condition dismisses 575 out of 2,779 cities, i.e 20.69% of our sample. All ourresults are robust to including or not these 575 cities in the analysis. These restrictionsset boundaries so low that only fairly rural places get excluded. There are still quite smalltowns left in the main dataset after this exclusion.
C.2 Excluded countries and Capital cities
We focus on Sub-Saharan Africa, from which we excluded South Africa (because of itsspecific history) as well as non-continental areas.
In particular, the exclusion of Madagascar follows from the incompatibility of its geog-raphy with our historical IV.
We also exclude Nigeria from our main specification because we want to ensure it doesnot drive the results alone (Nigeria does represent a big share of the total potential sam-ple). Moreover, the presence of gas flares in the Gulf of Guinea makes the delimitation ofcities from nightlights imagery particularly noisy in this country; although our data-set wasmanually corrected for this.
In the case of these last two countries, our choice was guided by specifications issues; westill include them in the following tables as one could be interested in the robustness of ourresults. In fact, our main results are also robust to the inclusion of one or both of thesecountries in the sample.
As mentionned, when using OSM data has exported by GeoFabrik (cf. B.2) we mustalso exclude Burundi and Somalia, for which there is not data available. For consistency wealways exclude these two countries from our specifications.
Finally, we also exclude the capital city of each country. The reason is that capital citiesin Africa often have unique economic developments and histories, and may deviate fromgeneral trends observed elsewhere. But here again, the inclusion of these few cities does notchange the qualitative results.
C.3 Missing observations
We then extract OSM road networks for our 2,141 cities (i.e. the sample including Mada-gascar and Nigeria). We are nevertheless unable to observe roads (in at least one of ourextraction) in 291 of these cities.
These missing information could be caused either by an error in our definition of cities(i.e. declaring an area to be a city while it is not) or by the incompleteness of OSM data. Aquick comparison allows us to say that the missing observations represent 13.59% of the totalsample and affect some countries in particular (like Ghana, Sierra Leone, Sudan, Swaziland,etc.), cf. Table 10. Table 11 also shows that the cities with no road observation are, onaverage, smaller and less populated than the average city in the rest of the sample.
11We also exclude 5 cities that had 0 inhabitants in 2000, according to GHS data.
References
• Boeing, Geoff. (2017). OSMnx: New Methods for Acquiring, Constructing, Analyz-ing, and Visualizing Complex Street Networks. Computers Environment and UrbanSystems. 65. 126-139. 10.1016/j.compenvurbsys.2017.05.004.
Figure 7: Most cities contain roads with no precision about their type (e.g. motorway, path, etc.).In our sample, only 21.18% of the cities have classification for almost all (≥90%) of their roads.
Table 10: Missing OSM Data
OSM Missing
Angola 80.49 19.51Benin 84.38 15.63Botswana 87.88 12.12Burkina Faso 96.43 3.57Cameroon 95.83 4.17Central African Republic 90.91 9.09Chad 85.71 14.29Congo, Dem Republic of 98.11 1.89Congo, Republic of 88.24 11.76Cote d’Ivoire 64.53 35.47Djibouti 100.00 0.00Equatorial Guinea 100.00 0.00Eritrea 100.00 0.00Ethiopia 99.12 0.88Gabon 89.29 10.71Gambia 100.00 0.00Ghana 79.28 20.72Guinea 95.83 4.17Kenya 89.89 10.11Lesotho 100.00 0.00Liberia 100.00 0.00Madagascar 97.06 2.94Malawi 95.00 5.00Mali 90.91 9.09Mauritania 100.00 0.00Mozambique 93.94 6.06Namibia 91.30 8.70Niger 85.00 15.00Nigeria 88.96 11.04Rwanda 100.00 0.00Senegal 98.18 1.82Sierra Leone 85.71 14.29South Sudan 100.00 0.00Sudan 76.47 23.53Swaziland 77.78 22.22Tanzania, United Rep of 94.34 5.66Togo 95.65 4.35Uganda 92.68 7.32Zambia 88.89 11.11Zimbabwe 80.52 19.48
Total 86.41 13.59
Table 11: Samples Comparison
OSM Missing DifferenceMean Sd Mean Sd Diff. t
Ln.Area (sq.km) 3.77 (1.10) 3.39 (0.92) 0.38***(5.90)GHS Population DataLn.Pop. 2000 10.38 (1.37) 9.05 (1.30) 1.32***(16.03)Ln.Pop. 2015 10.68 (1.37) 9.40 (1.06) 1.28***(18.39)Av. Growth Rate (2000-15) 2.04 (2.99) 2.30 (4.92) -0.26 (-0.87)
Observations 1850 291 2141
